Add files using upload-large-folder tool

versatile_diffusion/lib/model_zoo/clip_justin/model.py

@@ -0,0 +1,436 @@
+from collections import OrderedDict
+from typing import Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1):
+        super().__init__()
+
+        # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
+        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.relu1 = nn.ReLU(inplace=True)
+
+        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.relu2 = nn.ReLU(inplace=True)
+
+        self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()
+
+        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+        self.relu3 = nn.ReLU(inplace=True)
+
+        self.downsample = None
+        self.stride = stride
+
+        if stride > 1 or inplanes != planes * Bottleneck.expansion:
+            # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
+            self.downsample = nn.Sequential(OrderedDict([
+                ("-1", nn.AvgPool2d(stride)),
+                ("0", nn.Conv2d(inplanes, planes * self.expansion, 1, stride=1, bias=False)),
+                ("1", nn.BatchNorm2d(planes * self.expansion))
+            ]))
+
+    def forward(self, x: torch.Tensor):
+        identity = x
+
+        out = self.relu1(self.bn1(self.conv1(x)))
+        out = self.relu2(self.bn2(self.conv2(out)))
+        out = self.avgpool(out)
+        out = self.bn3(self.conv3(out))
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu3(out)
+        return out
+
+
+class AttentionPool2d(nn.Module):
+    def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(torch.randn(spacial_dim ** 2 + 1, embed_dim) / embed_dim ** 0.5)
+        self.k_proj = nn.Linear(embed_dim, embed_dim)
+        self.q_proj = nn.Linear(embed_dim, embed_dim)
+        self.v_proj = nn.Linear(embed_dim, embed_dim)
+        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
+        self.num_heads = num_heads
+
+    def forward(self, x):
+        x = x.flatten(start_dim=2).permute(2, 0, 1)  # NCHW -> (HW)NC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (HW+1)NC
+        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (HW+1)NC
+        x, _ = F.multi_head_attention_forward(
+            query=x[:1], key=x, value=x,
+            embed_dim_to_check=x.shape[-1],
+            num_heads=self.num_heads,
+            q_proj_weight=self.q_proj.weight,
+            k_proj_weight=self.k_proj.weight,
+            v_proj_weight=self.v_proj.weight,
+            in_proj_weight=None,
+            in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]),
+            bias_k=None,
+            bias_v=None,
+            add_zero_attn=False,
+            dropout_p=0,
+            out_proj_weight=self.c_proj.weight,
+            out_proj_bias=self.c_proj.bias,
+            use_separate_proj_weight=True,
+            training=self.training,
+            need_weights=False
+        )
+        return x.squeeze(0)
+
+
+class ModifiedResNet(nn.Module):
+    """
+    A ResNet class that is similar to torchvision's but contains the following changes:
+    - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool.
+    - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1
+    - The final pooling layer is a QKV attention instead of an average pool
+    """
+
+    def __init__(self, layers, output_dim, heads, input_resolution=224, width=64):
+        super().__init__()
+        self.output_dim = output_dim
+        self.input_resolution = input_resolution
+
+        # the 3-layer stem
+        self.conv1 = nn.Conv2d(3, width // 2, kernel_size=3, stride=2, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(width // 2)
+        self.relu1 = nn.ReLU(inplace=True)
+        self.conv2 = nn.Conv2d(width // 2, width // 2, kernel_size=3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(width // 2)
+        self.relu2 = nn.ReLU(inplace=True)
+        self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(width)
+        self.relu3 = nn.ReLU(inplace=True)
+        self.avgpool = nn.AvgPool2d(2)
+
+        # residual layers
+        self._inplanes = width  # this is a *mutable* variable used during construction
+        self.layer1 = self._make_layer(width, layers[0])
+        self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
+        self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
+        self.layer4 = self._make_layer(width * 8, layers[3], stride=2)
+
+        embed_dim = width * 32  # the ResNet feature dimension
+        self.attnpool = AttentionPool2d(input_resolution // 32, embed_dim, heads, output_dim)
+
+    def _make_layer(self, planes, blocks, stride=1):
+        layers = [Bottleneck(self._inplanes, planes, stride)]
+
+        self._inplanes = planes * Bottleneck.expansion
+        for _ in range(1, blocks):
+            layers.append(Bottleneck(self._inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        def stem(x):
+            x = self.relu1(self.bn1(self.conv1(x)))
+            x = self.relu2(self.bn2(self.conv2(x)))
+            x = self.relu3(self.bn3(self.conv3(x)))
+            x = self.avgpool(x)
+            return x
+
+        x = x.type(self.conv1.weight.dtype)
+        x = stem(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.attnpool(x)
+
+        return x
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        ret = super().forward(x.type(torch.float32))
+        return ret.type(orig_type)
+
+
+class QuickGELU(nn.Module):
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
+        super().__init__()
+
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ln_1 = LayerNorm(d_model)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, d_model * 4)),
+            ("gelu", QuickGELU()),
+            ("c_proj", nn.Linear(d_model * 4, d_model))
+        ]))
+        self.ln_2 = LayerNorm(d_model)
+        self.attn_mask = attn_mask
+
+    def attention(self, x: torch.Tensor):
+        self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
+        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
+
+    def forward(self, x: torch.Tensor):
+        x = x + self.attention(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])
+
+    def forward(self, x: torch.Tensor):
+        return self.resblocks(x)
+
+
+class VisionTransformer(nn.Module):
+    def __init__(self, input_resolution: int, patch_size: int, width: int, layers: int, heads: int, output_dim: int):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.output_dim = output_dim
+        self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False)
+
+        scale = width ** -0.5
+        self.class_embedding = nn.Parameter(scale * torch.randn(width))
+        self.positional_embedding = nn.Parameter(scale * torch.randn((input_resolution // patch_size) ** 2 + 1, width))
+        self.ln_pre = LayerNorm(width)
+
+        self.transformer = Transformer(width, layers, heads)
+
+        self.ln_post = LayerNorm(width)
+        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
+
+    def forward(self, x: torch.Tensor):
+        x = self.conv1(x)  # shape = [*, width, grid, grid]
+        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+        x = torch.cat([self.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1)  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.positional_embedding.to(x.dtype)
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        x = self.ln_post(x[:, 0, :])
+
+        if self.proj is not None:
+            x = x @ self.proj
+
+        return x
+
+
+class CLIP(nn.Module):
+    def __init__(self,
+                 embed_dim: int,
+                 # vision
+                 image_resolution: int,
+                 vision_layers: Union[Tuple[int, int, int, int], int],
+                 vision_width: int,
+                 vision_patch_size: int,
+                 # text
+                 context_length: int,
+                 vocab_size: int,
+                 transformer_width: int,
+                 transformer_heads: int,
+                 transformer_layers: int
+                 ):
+        super().__init__()
+
+        self.context_length = context_length
+
+        if isinstance(vision_layers, (tuple, list)):
+            vision_heads = vision_width * 32 // 64
+            self.visual = ModifiedResNet(
+                layers=vision_layers,
+                output_dim=embed_dim,
+                heads=vision_heads,
+                input_resolution=image_resolution,
+                width=vision_width
+            )
+        else:
+            vision_heads = vision_width // 64
+            self.visual = VisionTransformer(
+                input_resolution=image_resolution,
+                patch_size=vision_patch_size,
+                width=vision_width,
+                layers=vision_layers,
+                heads=vision_heads,
+                output_dim=embed_dim
+            )
+
+        self.transformer = Transformer(
+            width=transformer_width,
+            layers=transformer_layers,
+            heads=transformer_heads,
+            attn_mask=self.build_attention_mask()
+        )
+
+        self.vocab_size = vocab_size
+        self.token_embedding = nn.Embedding(vocab_size, transformer_width)
+        self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width))
+        self.ln_final = LayerNorm(transformer_width)
+
+        self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim))
+        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+
+        self.initialize_parameters()
+
+    def initialize_parameters(self):
+        nn.init.normal_(self.token_embedding.weight, std=0.02)
+        nn.init.normal_(self.positional_embedding, std=0.01)
+
+        if isinstance(self.visual, ModifiedResNet):
+            if self.visual.attnpool is not None:
+                std = self.visual.attnpool.c_proj.in_features ** -0.5
+                nn.init.normal_(self.visual.attnpool.q_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.k_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.v_proj.weight, std=std)
+                nn.init.normal_(self.visual.attnpool.c_proj.weight, std=std)
+
+            for resnet_block in [self.visual.layer1, self.visual.layer2, self.visual.layer3, self.visual.layer4]:
+                for name, param in resnet_block.named_parameters():
+                    if name.endswith("bn3.weight"):
+                        nn.init.zeros_(param)
+
+        proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
+        attn_std = self.transformer.width ** -0.5
+        fc_std = (2 * self.transformer.width) ** -0.5
+        for block in self.transformer.resblocks:
+            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+
+        if self.text_projection is not None:
+            nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)
+
+    def build_attention_mask(self):
+        # lazily create causal attention mask, with full attention between the vision tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.context_length, self.context_length)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    @property
+    def dtype(self):
+        return self.visual.conv1.weight.dtype
+
+    def encode_image(self, image):
+        return self.visual(image.type(self.dtype))
+
+    def encode_text(self, text):
+        x = self.token_embedding(text).type(self.dtype)  # [batch_size, n_ctx, d_model]
+
+        x = x + self.positional_embedding.type(self.dtype)
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.ln_final(x).type(self.dtype)
+
+        # x.shape = [batch_size, n_ctx, transformer.width]
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection
+
+        return x
+
+    def forward(self, image, text):
+        image_features = self.encode_image(image)
+        text_features = self.encode_text(text)
+
+        # normalized features
+        image_features = image_features / image_features.norm(dim=1, keepdim=True)
+        text_features = text_features / text_features.norm(dim=1, keepdim=True)
+
+        # cosine similarity as logits
+        logit_scale = self.logit_scale.exp()
+        logits_per_image = logit_scale * image_features @ text_features.t()
+        logits_per_text = logits_per_image.t()
+
+        # shape = [global_batch_size, global_batch_size]
+        return logits_per_image, logits_per_text
+
+
+def convert_weights(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+
+    def _convert_weights_to_fp16(l):
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.half()
+            if l.bias is not None:
+                l.bias.data = l.bias.data.half()
+
+        if isinstance(l, nn.MultiheadAttention):
+            for attr in [*[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]], "in_proj_bias", "bias_k", "bias_v"]:
+                tensor = getattr(l, attr)
+                if tensor is not None:
+                    tensor.data = tensor.data.half()
+
+        for name in ["text_projection", "proj"]:
+            if hasattr(l, name):
+                attr = getattr(l, name)
+                if attr is not None:
+                    attr.data = attr.data.half()
+
+    model.apply(_convert_weights_to_fp16)
+
+
+def build_model(state_dict: dict):
+    vit = "visual.proj" in state_dict
+
+    if vit:
+        vision_width = state_dict["visual.conv1.weight"].shape[0]
+        vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")])
+        vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
+        grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5)
+        image_resolution = vision_patch_size * grid_size
+    else:
+        counts: list = [len(set(k.split(".")[2] for k in state_dict if k.startswith(f"visual.layer{b}"))) for b in [1, 2, 3, 4]]
+        vision_layers = tuple(counts)
+        vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0]
+        output_width = round((state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5)
+        vision_patch_size = None
+        assert output_width ** 2 + 1 == state_dict["visual.attnpool.positional_embedding"].shape[0]
+        image_resolution = output_width * 32
+
+    embed_dim = state_dict["text_projection"].shape[1]
+    context_length = state_dict["positional_embedding"].shape[0]
+    vocab_size = state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_heads = transformer_width // 64
+    transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith("transformer.resblocks")))
+
+    model = CLIP(
+        embed_dim,
+        image_resolution, vision_layers, vision_width, vision_patch_size,
+        context_length, vocab_size, transformer_width, transformer_heads, transformer_layers
+    )
+
+    for key in ["input_resolution", "context_length", "vocab_size"]:
+        if key in state_dict:
+            del state_dict[key]
+
+    convert_weights(model)
+    model.load_state_dict(state_dict)
+    return model.eval()

versatile_diffusion/lib/model_zoo/common/utils.py

@@ -0,0 +1,292 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+import copy
+import functools
+import itertools
+
+import matplotlib.pyplot as plt
+
+########
+# unit #
+########
+
+def singleton(class_):
+    instances = {}
+    def getinstance(*args, **kwargs):
+        if class_ not in instances:
+            instances[class_] = class_(*args, **kwargs)
+        return instances[class_]
+    return getinstance
+
+def str2value(v):
+    v = v.strip()
+    try:
+        return int(v)
+    except:
+        pass
+    try:
+        return float(v)
+    except:
+        pass
+    if v in ('True', 'true'):
+        return True
+    elif v in ('False', 'false'):
+        return False
+    else:
+        return v
+
+@singleton
+class get_unit(object):
+    def __init__(self):
+        self.unit = {}
+        self.register('none', None)
+
+        # general convolution
+        self.register('conv'  , nn.Conv2d)
+        self.register('bn'    , nn.BatchNorm2d)
+        self.register('relu'  , nn.ReLU)
+        self.register('relu6' , nn.ReLU6)
+        self.register('lrelu' , nn.LeakyReLU)
+        self.register('dropout'  , nn.Dropout)
+        self.register('dropout2d', nn.Dropout2d)
+        self.register('sine',     Sine)
+        self.register('relusine', ReLUSine)
+
+    def register(self, 
+                 name, 
+                 unitf,):
+
+        self.unit[name] = unitf
+
+    def __call__(self, name):
+        if name is None:
+            return None
+        i = name.find('(')
+        i = len(name) if i==-1 else i
+        t = name[:i]
+        f = self.unit[t]
+        args = name[i:].strip('()')
+        if len(args) == 0:
+            args = {}
+            return f
+        else:
+            args = args.split('=')
+            args = [[','.join(i.split(',')[:-1]), i.split(',')[-1]] for i in args]
+            args = list(itertools.chain.from_iterable(args))
+            args = [i.strip() for i in args if len(i)>0]
+            kwargs = {}
+            for k, v in zip(args[::2], args[1::2]):
+                if v[0]=='(' and v[-1]==')':
+                    kwargs[k] = tuple([str2value(i) for i in v.strip('()').split(',')])
+                elif v[0]=='[' and v[-1]==']':
+                    kwargs[k] = [str2value(i) for i in v.strip('[]').split(',')]
+                else:
+                    kwargs[k] = str2value(v)
+            return functools.partial(f, **kwargs)
+
+def register(name):
+    def wrapper(class_):
+        get_unit().register(name, class_)
+        return class_
+    return wrapper
+
+class Sine(object):
+    def __init__(self, freq, gain=1):
+        self.freq = freq
+        self.gain = gain
+        self.repr = 'sine(freq={}, gain={})'.format(freq, gain)
+
+    def __call__(self, x, gain=1):
+        act_gain = self.gain * gain
+        return torch.sin(self.freq * x) * act_gain
+
+    def __repr__(self,):
+        return self.repr
+
+class ReLUSine(nn.Module):
+    def __init(self):
+        super().__init__()
+
+    def forward(self, input):
+        a = torch.sin(30 * input)
+        b = nn.ReLU(inplace=False)(input)
+        return a+b
+
+@register('lrelu_agc')
+# class lrelu_agc(nn.Module):
+class lrelu_agc(object):
+    """
+    The lrelu layer with alpha, gain and clamp
+    """
+    def __init__(self, alpha=0.1, gain=1, clamp=None):
+        # super().__init__()
+        self.alpha = alpha
+        if gain == 'sqrt_2':
+            self.gain = np.sqrt(2)
+        else:
+            self.gain = gain
+        self.clamp = clamp
+        self.repr = 'lrelu_agc(alpha={}, gain={}, clamp={})'.format(
+            alpha, gain, clamp)
+
+    # def forward(self, x, gain=1):
+    def __call__(self, x, gain=1):
+        x = F.leaky_relu(x, negative_slope=self.alpha, inplace=True)
+        act_gain = self.gain * gain
+        act_clamp = self.clamp * gain if self.clamp is not None else None
+        if act_gain != 1:
+            x = x * act_gain
+        if act_clamp is not None:
+            x = x.clamp(-act_clamp, act_clamp)
+        return x
+
+    def __repr__(self,):
+        return self.repr
+
+####################
+# spatial encoding #
+####################
+
+@register('se')
+class SpatialEncoding(nn.Module):
+    def __init__(self, 
+                 in_dim, 
+                 out_dim, 
+                 sigma = 6,
+                 cat_input=True,
+                 require_grad=False,):
+
+        super().__init__()
+        assert out_dim % (2*in_dim) == 0, "dimension must be dividable"
+
+        n = out_dim // 2 // in_dim
+        m = 2**np.linspace(0, sigma, n)
+        m = np.stack([m] + [np.zeros_like(m)]*(in_dim-1), axis=-1)
+        m = np.concatenate([np.roll(m, i, axis=-1) for i in range(in_dim)], axis=0)
+        self.emb = torch.FloatTensor(m)
+        if require_grad:
+            self.emb = nn.Parameter(self.emb, requires_grad=True)    
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+        self.sigma = sigma
+        self.cat_input = cat_input
+        self.require_grad = require_grad
+
+    def forward(self, x, format='[n x c]'):
+        """
+        Args:
+            x: [n x m1],
+                m1 usually is 2
+        Outputs:
+            y: [n x m2]         
+                m2 dimention number
+        """
+        if format == '[bs x c x 2D]':
+            xshape = x.shape
+            x = x.permute(0, 2, 3, 1).contiguous()
+            x = x.view(-1, x.size(-1))
+        elif format == '[n x c]':
+            pass
+        else:
+            raise ValueError
+
+        if not self.require_grad:
+            self.emb = self.emb.to(x.device)
+        y = torch.mm(x, self.emb.T)
+        if self.cat_input:
+            z = torch.cat([x, torch.sin(y), torch.cos(y)], dim=-1)
+        else:
+            z = torch.cat([torch.sin(y), torch.cos(y)], dim=-1)
+
+        if format == '[bs x c x 2D]':
+            z = z.view(xshape[0], xshape[2], xshape[3], -1)
+            z = z.permute(0, 3, 1, 2).contiguous()
+        return z
+
+    def extra_repr(self):
+        outstr = 'SpatialEncoding (in={}, out={}, sigma={}, cat_input={}, require_grad={})'.format(
+            self.in_dim, self.out_dim, self.sigma, self.cat_input, self.require_grad)
+        return outstr
+
+@register('rffe')
+class RFFEncoding(SpatialEncoding):
+    """
+    Random Fourier Features
+    """
+    def __init__(self, 
+                 in_dim, 
+                 out_dim, 
+                 sigma = 6,
+                 cat_input=True,
+                 require_grad=False,):
+
+        super().__init__(in_dim, out_dim, sigma, cat_input, require_grad)
+        n = out_dim // 2
+        m = np.random.normal(0, sigma, size=(n, in_dim))
+        self.emb = torch.FloatTensor(m)
+        if require_grad:
+            self.emb = nn.Parameter(self.emb, requires_grad=True)    
+
+    def extra_repr(self):
+        outstr = 'RFFEncoding (in={}, out={}, sigma={}, cat_input={}, require_grad={})'.format(
+            self.in_dim, self.out_dim, self.sigma, self.cat_input, self.require_grad)
+        return outstr
+
+##########
+# helper #
+##########
+
+def freeze(net):
+    for m in net.modules():
+        if isinstance(m, (
+                nn.BatchNorm2d, 
+                nn.SyncBatchNorm,)):
+            # inplace_abn not supported
+            m.eval()
+    for pi in net.parameters():
+        pi.requires_grad = False
+    return net
+
+def common_init(m):
+    if isinstance(m, (
+            nn.Conv2d, 
+            nn.ConvTranspose2d,)):
+        nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+        if m.bias is not None:
+            nn.init.constant_(m.bias, 0)
+    elif isinstance(m, (
+            nn.BatchNorm2d, 
+            nn.SyncBatchNorm,)):
+        nn.init.constant_(m.weight, 1)
+        nn.init.constant_(m.bias, 0)
+    else:
+        pass
+
+def init_module(module):
+    """
+    Args:
+        module: [nn.module] list or nn.module
+            a list of module to be initialized.
+    """
+    if isinstance(module, (list, tuple)):
+        module = list(module)
+    else:
+        module = [module]
+
+    for mi in module:
+        for mii in mi.modules():
+            common_init(mii)
+
+def get_total_param(net):
+    if getattr(net, 'parameters', None) is None:
+        return 0
+    return sum(p.numel() for p in net.parameters())
+
+def get_total_param_sum(net):
+    if getattr(net, 'parameters', None) is None:
+        return 0
+    with torch.no_grad():
+        s = sum(p.cpu().detach().numpy().sum().item() for p in net.parameters())
+    return s 

versatile_diffusion/lib/model_zoo/optimus_models/configuration_bert.py

@@ -0,0 +1,113 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" BERT model configuration """
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import sys
+from io import open
+
+from .configuration_utils import PretrainedConfig
+
+logger = logging.getLogger(__name__)
+
+BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    'bert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-config.json",
+    'bert-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-config.json",
+    'bert-base-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-config.json",
+    'bert-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-config.json",
+    'bert-base-multilingual-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased-config.json",
+    'bert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased-config.json",
+    'bert-base-chinese': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese-config.json",
+    'bert-base-german-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-cased-config.json",
+    'bert-large-uncased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-config.json",
+    'bert-large-cased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-config.json",
+    'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-config.json",
+    'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-config.json",
+    'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-config.json",
+}
+
+
+class BertConfig(PretrainedConfig):
+    r"""
+        :class:`~pytorch_transformers.BertConfig` is the configuration class to store the configuration of a
+        `BertModel`.
+
+
+        Arguments:
+            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `BertModel`.
+            hidden_size: Size of the encoder layers and the pooler layer.
+            num_hidden_layers: Number of hidden layers in the Transformer encoder.
+            num_attention_heads: Number of attention heads for each attention layer in
+                the Transformer encoder.
+            intermediate_size: The size of the "intermediate" (i.e., feed-forward)
+                layer in the Transformer encoder.
+            hidden_act: The non-linear activation function (function or string) in the
+                encoder and pooler. If string, "gelu", "relu" and "swish" are supported.
+            hidden_dropout_prob: The dropout probabilitiy for all fully connected
+                layers in the embeddings, encoder, and pooler.
+            attention_probs_dropout_prob: The dropout ratio for the attention
+                probabilities.
+            max_position_embeddings: The maximum sequence length that this model might
+                ever be used with. Typically set this to something large just in case
+                (e.g., 512 or 1024 or 2048).
+            type_vocab_size: The vocabulary size of the `token_type_ids` passed into
+                `BertModel`.
+            initializer_range: The sttdev of the truncated_normal_initializer for
+                initializing all weight matrices.
+            layer_norm_eps: The epsilon used by LayerNorm.
+    """
+    pretrained_config_archive_map = BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+
+    def __init__(self,
+                 vocab_size_or_config_json_file=30522,
+                 hidden_size=768,
+                 num_hidden_layers=12,
+                 num_attention_heads=12,
+                 intermediate_size=3072,
+                 hidden_act="gelu",
+                 hidden_dropout_prob=0.1,
+                 attention_probs_dropout_prob=0.1,
+                 max_position_embeddings=512,
+                 type_vocab_size=2,
+                 initializer_range=0.02,
+                 layer_norm_eps=1e-12,
+                 **kwargs):
+        super(BertConfig, self).__init__(**kwargs)
+        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
+                        and isinstance(vocab_size_or_config_json_file, unicode)):
+            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
+                json_config = json.loads(reader.read())
+            for key, value in json_config.items():
+                self.__dict__[key] = value
+        elif isinstance(vocab_size_or_config_json_file, int):
+            self.vocab_size = vocab_size_or_config_json_file
+            self.hidden_size = hidden_size
+            self.num_hidden_layers = num_hidden_layers
+            self.num_attention_heads = num_attention_heads
+            self.hidden_act = hidden_act
+            self.intermediate_size = intermediate_size
+            self.hidden_dropout_prob = hidden_dropout_prob
+            self.attention_probs_dropout_prob = attention_probs_dropout_prob
+            self.max_position_embeddings = max_position_embeddings
+            self.type_vocab_size = type_vocab_size
+            self.initializer_range = initializer_range
+            self.layer_norm_eps = layer_norm_eps
+        else:
+            raise ValueError("First argument must be either a vocabulary size (int)"
+                             " or the path to a pretrained model config file (str)")

versatile_diffusion/lib/model_zoo/optimus_models/configuration_gpt2.py

@@ -0,0 +1,143 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" OpenAI GPT-2 configuration """
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import sys
+from io import open
+
+from .configuration_utils import PretrainedConfig
+
+logger = logging.getLogger(__name__)
+
+GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP = {"gpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-config.json",
+                                      "gpt2-medium": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-medium-config.json",
+                                      "gpt2-large": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-large-config.json"}
+
+class GPT2Config(PretrainedConfig):
+    """Configuration class to store the configuration of a `GPT2Model`.
+
+    Args:
+        vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
+        n_positions: Number of positional embeddings.
+        n_ctx: Size of the causal mask (usually same as n_positions).
+        n_embd: Dimensionality of the embeddings and hidden states.
+        n_layer: Number of hidden layers in the Transformer encoder.
+        n_head: Number of attention heads for each attention layer in
+            the Transformer encoder.
+        layer_norm_epsilon: epsilon to use in the layer norm layers
+        resid_pdrop: The dropout probabilitiy for all fully connected
+            layers in the embeddings, encoder, and pooler.
+        attn_pdrop: The dropout ratio for the attention
+            probabilities.
+        embd_pdrop: The dropout ratio for the embeddings.
+        initializer_range: The sttdev of the truncated_normal_initializer for
+            initializing all weight matrices.
+    """
+    pretrained_config_archive_map = GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP
+
+    def __init__(
+        self,
+        vocab_size_or_config_json_file=50257,
+        n_positions=1024,
+        n_ctx=1024,
+        n_embd=768,
+        n_layer=12,
+        n_head=12,
+        resid_pdrop=0.1,
+        embd_pdrop=0.1,
+        attn_pdrop=0.1,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+
+        num_labels=1,
+        summary_type='cls_index',
+        summary_use_proj=True,
+        summary_activation=None,
+        summary_proj_to_labels=True,
+        summary_first_dropout=0.1,
+        **kwargs
+    ):
+        """Constructs GPT2Config.
+
+        Args:
+            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
+            n_positions: Number of positional embeddings.
+            n_ctx: Size of the causal mask (usually same as n_positions).
+            n_embd: Dimensionality of the embeddings and hidden states.
+            n_layer: Number of hidden layers in the Transformer encoder.
+            n_head: Number of attention heads for each attention layer in
+                the Transformer encoder.
+            layer_norm_epsilon: epsilon to use in the layer norm layers
+            resid_pdrop: The dropout probabilitiy for all fully connected
+                layers in the embeddings, encoder, and pooler.
+            attn_pdrop: The dropout ratio for the attention
+                probabilities.
+            embd_pdrop: The dropout ratio for the embeddings.
+            initializer_range: The sttdev of the truncated_normal_initializer for
+                initializing all weight matrices.
+        """
+        super(GPT2Config, self).__init__(**kwargs)
+
+        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
+                        and isinstance(vocab_size_or_config_json_file, unicode)):
+            with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
+                json_config = json.loads(reader.read())
+            for key, value in json_config.items():
+                self.__dict__[key] = value
+        elif isinstance(vocab_size_or_config_json_file, int):
+            self.vocab_size = vocab_size_or_config_json_file
+            self.n_ctx = n_ctx
+            self.n_positions = n_positions
+            self.n_embd = n_embd
+            self.n_layer = n_layer
+            self.n_head = n_head
+            self.resid_pdrop = resid_pdrop
+            self.embd_pdrop = embd_pdrop
+            self.attn_pdrop = attn_pdrop
+            self.layer_norm_epsilon = layer_norm_epsilon
+            self.initializer_range = initializer_range
+
+            self.num_labels = num_labels
+            self.summary_type = summary_type
+            self.summary_use_proj = summary_use_proj
+            self.summary_activation = summary_activation
+            self.summary_first_dropout = summary_first_dropout
+            self.summary_proj_to_labels = summary_proj_to_labels
+        else:
+            raise ValueError(
+                "First argument must be either a vocabulary size (int)"
+                "or the path to a pretrained model config file (str)"
+            )
+
+    @property
+    def max_position_embeddings(self):
+        return self.n_positions
+
+    @property
+    def hidden_size(self):
+        return self.n_embd
+
+    @property
+    def num_attention_heads(self):
+        return self.n_head
+
+    @property
+    def num_hidden_layers(self):
+        return self.n_layer

versatile_diffusion/lib/model_zoo/optimus_models/configuration_utils.py

@@ -0,0 +1,205 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Configuration base class and utilities."""
+
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+import copy
+import json
+import logging
+import os
+from io import open
+
+from .file_utils import cached_path, CONFIG_NAME
+
+logger = logging.getLogger(__name__)
+
+class PretrainedConfig(object):
+    r""" Base class for all configuration classes.
+        Handles a few parameters common to all models' configurations as well as methods for loading/downloading/saving configurations.
+
+        Note:
+            A configuration file can be loaded and saved to disk. Loading the configuration file and using this file to initialize a model does **not** load the model weights.
+            It only affects the model's configuration.
+
+        Class attributes (overridden by derived classes):
+            - ``pretrained_config_archive_map``: a python ``dict`` of with `short-cut-names` (string) as keys and `url` (string) of associated pretrained model configurations as values.
+
+        Parameters:
+            ``finetuning_task``: string, default `None`. Name of the task used to fine-tune the model. This can be used when converting from an original (TensorFlow or PyTorch) checkpoint.
+            ``num_labels``: integer, default `2`. Number of classes to use when the model is a classification model (sequences/tokens)
+            ``output_attentions``: boolean, default `False`. Should the model returns attentions weights.
+            ``output_hidden_states``: string, default `False`. Should the model returns all hidden-states.
+            ``torchscript``: string, default `False`. Is the model used with Torchscript.
+    """
+    pretrained_config_archive_map = {}
+
+    def __init__(self, **kwargs):
+        self.finetuning_task = kwargs.pop('finetuning_task', None)
+        self.num_labels = kwargs.pop('num_labels', 2)
+        self.output_attentions = kwargs.pop('output_attentions', False)
+        self.output_hidden_states = kwargs.pop('output_hidden_states', False)
+        self.torchscript = kwargs.pop('torchscript', False)
+        self.pruned_heads = kwargs.pop('pruned_heads', {})
+
+    def save_pretrained(self, save_directory):
+        """ Save a configuration object to the directory `save_directory`, so that it
+            can be re-loaded using the :func:`~pytorch_transformers.PretrainedConfig.from_pretrained` class method.
+        """
+        assert os.path.isdir(save_directory), "Saving path should be a directory where the model and configuration can be saved"
+
+        # If we save using the predefined names, we can load using `from_pretrained`
+        output_config_file = os.path.join(save_directory, CONFIG_NAME)
+
+        self.to_json_file(output_config_file)
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        r""" Instantiate a :class:`~pytorch_transformers.PretrainedConfig` (or a derived class) from a pre-trained model configuration.
+
+        Parameters:
+            pretrained_model_name_or_path: either:
+
+                - a string with the `shortcut name` of a pre-trained model configuration to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a path to a `directory` containing a configuration file saved using the :func:`~pytorch_transformers.PretrainedConfig.save_pretrained` method, e.g.: ``./my_model_directory/``.
+                - a path or url to a saved configuration JSON `file`, e.g.: ``./my_model_directory/configuration.json``.
+
+            cache_dir: (`optional`) string:
+                Path to a directory in which a downloaded pre-trained model
+                configuration should be cached if the standard cache should not be used.
+
+            kwargs: (`optional`) dict: key/value pairs with which to update the configuration object after loading.
+
+                - The values in kwargs of any keys which are configuration attributes will be used to override the loaded values.
+                - Behavior concerning key/value pairs whose keys are *not* configuration attributes is controlled by the `return_unused_kwargs` keyword parameter.
+
+            force_download: (`optional`) boolean, default False:
+                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
+
+            proxies: (`optional`) dict, default None:
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
+                The proxies are used on each request.
+
+            return_unused_kwargs: (`optional`) bool:
+
+                - If False, then this function returns just the final configuration object.
+                - If True, then this functions returns a tuple `(config, unused_kwargs)` where `unused_kwargs` is a dictionary consisting of the key/value pairs whose keys are not configuration attributes: ie the part of kwargs which has not been used to update `config` and is otherwise ignored.
+
+        Examples::
+
+            # We can't instantiate directly the base class `PretrainedConfig` so let's show the examples on a
+            # derived class: BertConfig
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            config = BertConfig.from_pretrained('./test/saved_model/')  # E.g. config (or model) was saved using `save_pretrained('./test/saved_model/')`
+            config = BertConfig.from_pretrained('./test/saved_model/my_configuration.json')
+            config = BertConfig.from_pretrained('bert-base-uncased', output_attention=True, foo=False)
+            assert config.output_attention == True
+            config, unused_kwargs = BertConfig.from_pretrained('bert-base-uncased', output_attention=True,
+                                                               foo=False, return_unused_kwargs=True)
+            assert config.output_attention == True
+            assert unused_kwargs == {'foo': False}
+
+        """
+        cache_dir = kwargs.pop('cache_dir', None)
+        force_download = kwargs.pop('force_download', False)
+        proxies = kwargs.pop('proxies', None)
+        return_unused_kwargs = kwargs.pop('return_unused_kwargs', False)
+
+        if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
+            config_file = cls.pretrained_config_archive_map[pretrained_model_name_or_path]
+        elif os.path.isdir(pretrained_model_name_or_path):
+            config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
+        else:
+            config_file = pretrained_model_name_or_path
+        # redirect to the cache, if necessary
+        try:
+            resolved_config_file = cached_path(config_file, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
+        except EnvironmentError as e:
+            if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
+                logger.error(
+                    "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
+                        config_file))
+            else:
+                logger.error(
+                    "Model name '{}' was not found in model name list ({}). "
+                    "We assumed '{}' was a path or url but couldn't find any file "
+                    "associated to this path or url.".format(
+                        pretrained_model_name_or_path,
+                        ', '.join(cls.pretrained_config_archive_map.keys()),
+                        config_file))
+            raise e
+        if resolved_config_file == config_file:
+            logger.info("loading configuration file {}".format(config_file))
+        else:
+            logger.info("loading configuration file {} from cache at {}".format(
+                config_file, resolved_config_file))
+
+        # Load config
+        config = cls.from_json_file(resolved_config_file)
+
+        if hasattr(config, 'pruned_heads'):
+            config.pruned_heads = dict((int(key), set(value)) for key, value in config.pruned_heads.items())
+
+        # Update config with kwargs if needed
+        to_remove = []
+        for key, value in kwargs.items():
+            if hasattr(config, key):
+                setattr(config, key, value)
+                to_remove.append(key)
+        for key in to_remove:
+            kwargs.pop(key, None)
+
+        logger.info("Model config %s", config)
+        if return_unused_kwargs:
+            return config, kwargs
+        else:
+            return config
+
+    @classmethod
+    def from_dict(cls, json_object):
+        """Constructs a `Config` from a Python dictionary of parameters."""
+        config = cls(vocab_size_or_config_json_file=-1)
+        for key, value in json_object.items():
+            config.__dict__[key] = value
+        return config
+
+    @classmethod
+    def from_json_file(cls, json_file):
+        """Constructs a `BertConfig` from a json file of parameters."""
+        with open(json_file, "r", encoding='utf-8') as reader:
+            text = reader.read()
+        return cls.from_dict(json.loads(text))
+
+    def __eq__(self, other):
+        return self.__dict__ == other.__dict__
+
+    def __repr__(self):
+        return str(self.to_json_string())
+
+    def to_dict(self):
+        """Serializes this instance to a Python dictionary."""
+        output = copy.deepcopy(self.__dict__)
+        return output
+
+    def to_json_string(self):
+        """Serializes this instance to a JSON string."""
+        return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"
+
+    def to_json_file(self, json_file_path):
+        """ Save this instance to a json file."""
+        with open(json_file_path, "w", encoding='utf-8') as writer:
+            writer.write(self.to_json_string())

versatile_diffusion/lib/model_zoo/optimus_models/file_utils.py

@@ -0,0 +1,294 @@
+"""
+Utilities for working with the local dataset cache.
+This file is adapted from the AllenNLP library at https://github.com/allenai/allennlp
+Copyright by the AllenNLP authors.
+"""
+from __future__ import (absolute_import, division, print_function, unicode_literals)
+
+import sys
+import json
+import logging
+import os
+import six
+import shutil
+import tempfile
+import fnmatch
+from functools import wraps
+from hashlib import sha256
+from io import open
+
+# import boto3
+# from botocore.config import Config
+# from botocore.exceptions import ClientError
+import requests
+from tqdm import tqdm
+
+try:
+    from torch.hub import _get_torch_home
+    torch_cache_home = _get_torch_home()
+except ImportError:
+    torch_cache_home = os.path.expanduser(
+        os.getenv('TORCH_HOME', os.path.join(
+            os.getenv('XDG_CACHE_HOME', '~/.cache'), 'torch')))
+default_cache_path = os.path.join(torch_cache_home, 'pytorch_transformers')
+
+try:
+    from urllib.parse import urlparse
+except ImportError:
+    from urlparse import urlparse
+
+try:
+    from pathlib import Path
+    PYTORCH_PRETRAINED_BERT_CACHE = Path(
+        os.getenv('PYTORCH_TRANSFORMERS_CACHE', os.getenv('PYTORCH_PRETRAINED_BERT_CACHE', default_cache_path)))
+except (AttributeError, ImportError):
+    PYTORCH_PRETRAINED_BERT_CACHE = os.getenv('PYTORCH_TRANSFORMERS_CACHE',
+                                              os.getenv('PYTORCH_PRETRAINED_BERT_CACHE',
+                                                        default_cache_path))
+
+PYTORCH_TRANSFORMERS_CACHE = PYTORCH_PRETRAINED_BERT_CACHE  # Kept for backward compatibility
+
+WEIGHTS_NAME = "pytorch_model.bin"
+TF_WEIGHTS_NAME = 'model.ckpt'
+CONFIG_NAME = "config.json"
+
+logger = logging.getLogger(__name__)  # pylint: disable=invalid-name
+
+if not six.PY2:
+    def add_start_docstrings(*docstr):
+        def docstring_decorator(fn):
+            fn.__doc__ = ''.join(docstr) + fn.__doc__
+            return fn
+        return docstring_decorator
+
+    def add_end_docstrings(*docstr):
+        def docstring_decorator(fn):
+            fn.__doc__ = fn.__doc__ + ''.join(docstr)
+            return fn
+        return docstring_decorator
+else:
+    # Not possible to update class docstrings on python2
+    def add_start_docstrings(*docstr):
+        def docstring_decorator(fn):
+            return fn
+        return docstring_decorator
+
+    def add_end_docstrings(*docstr):
+        def docstring_decorator(fn):
+            return fn
+        return docstring_decorator
+
+def url_to_filename(url, etag=None):
+    """
+    Convert `url` into a hashed filename in a repeatable way.
+    If `etag` is specified, append its hash to the url's, delimited
+    by a period.
+    """
+    url_bytes = url.encode('utf-8')
+    url_hash = sha256(url_bytes)
+    filename = url_hash.hexdigest()
+
+    if etag:
+        etag_bytes = etag.encode('utf-8')
+        etag_hash = sha256(etag_bytes)
+        filename += '.' + etag_hash.hexdigest()
+
+    return filename
+
+
+def filename_to_url(filename, cache_dir=None):
+    """
+    Return the url and etag (which may be ``None``) stored for `filename`.
+    Raise ``EnvironmentError`` if `filename` or its stored metadata do not exist.
+    """
+    if cache_dir is None:
+        cache_dir = PYTORCH_TRANSFORMERS_CACHE
+    if sys.version_info[0] == 3 and isinstance(cache_dir, Path):
+        cache_dir = str(cache_dir)
+
+    cache_path = os.path.join(cache_dir, filename)
+    if not os.path.exists(cache_path):
+        raise EnvironmentError("file {} not found".format(cache_path))
+
+    meta_path = cache_path + '.json'
+    if not os.path.exists(meta_path):
+        raise EnvironmentError("file {} not found".format(meta_path))
+
+    with open(meta_path, encoding="utf-8") as meta_file:
+        metadata = json.load(meta_file)
+    url = metadata['url']
+    etag = metadata['etag']
+
+    return url, etag
+
+
+def cached_path(url_or_filename, cache_dir=None, force_download=False, proxies=None):
+    """
+    Given something that might be a URL (or might be a local path),
+    determine which. If it's a URL, download the file and cache it, and
+    return the path to the cached file. If it's already a local path,
+    make sure the file exists and then return the path.
+    Args:
+        cache_dir: specify a cache directory to save the file to (overwrite the default cache dir).
+        force_download: if True, re-dowload the file even if it's already cached in the cache dir.
+    """
+    if cache_dir is None:
+        cache_dir = PYTORCH_TRANSFORMERS_CACHE
+    if sys.version_info[0] == 3 and isinstance(url_or_filename, Path):
+        url_or_filename = str(url_or_filename)
+    if sys.version_info[0] == 3 and isinstance(cache_dir, Path):
+        cache_dir = str(cache_dir)
+
+    parsed = urlparse(url_or_filename)
+
+    if parsed.scheme in ('http', 'https', 's3'):
+        # URL, so get it from the cache (downloading if necessary)
+        return get_from_cache(url_or_filename, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
+    elif os.path.exists(url_or_filename):
+        # File, and it exists.
+        return url_or_filename
+    elif parsed.scheme == '':
+        # File, but it doesn't exist.
+        raise EnvironmentError("file {} not found".format(url_or_filename))
+    else:
+        # Something unknown
+        raise ValueError("unable to parse {} as a URL or as a local path".format(url_or_filename))
+
+
+def split_s3_path(url):
+    """Split a full s3 path into the bucket name and path."""
+    parsed = urlparse(url)
+    if not parsed.netloc or not parsed.path:
+        raise ValueError("bad s3 path {}".format(url))
+    bucket_name = parsed.netloc
+    s3_path = parsed.path
+    # Remove '/' at beginning of path.
+    if s3_path.startswith("/"):
+        s3_path = s3_path[1:]
+    return bucket_name, s3_path
+
+
+def s3_request(func):
+    """
+    Wrapper function for s3 requests in order to create more helpful error
+    messages.
+    """
+
+    @wraps(func)
+    def wrapper(url, *args, **kwargs):
+        try:
+            return func(url, *args, **kwargs)
+        except ClientError as exc:
+            if int(exc.response["Error"]["Code"]) == 404:
+                raise EnvironmentError("file {} not found".format(url))
+            else:
+                raise
+
+    return wrapper
+
+
+@s3_request
+def s3_etag(url, proxies=None):
+    """Check ETag on S3 object."""
+    s3_resource = boto3.resource("s3", config=Config(proxies=proxies))
+    bucket_name, s3_path = split_s3_path(url)
+    s3_object = s3_resource.Object(bucket_name, s3_path)
+    return s3_object.e_tag
+
+
+@s3_request
+def s3_get(url, temp_file, proxies=None):
+    """Pull a file directly from S3."""
+    s3_resource = boto3.resource("s3", config=Config(proxies=proxies))
+    bucket_name, s3_path = split_s3_path(url)
+    s3_resource.Bucket(bucket_name).download_fileobj(s3_path, temp_file)
+
+
+def http_get(url, temp_file, proxies=None):
+    req = requests.get(url, stream=True, proxies=proxies)
+    content_length = req.headers.get('Content-Length')
+    total = int(content_length) if content_length is not None else None
+    progress = tqdm(unit="B", total=total)
+    for chunk in req.iter_content(chunk_size=1024):
+        if chunk: # filter out keep-alive new chunks
+            progress.update(len(chunk))
+            temp_file.write(chunk)
+    progress.close()
+
+
+def get_from_cache(url, cache_dir=None, force_download=False, proxies=None):
+    """
+    Given a URL, look for the corresponding dataset in the local cache.
+    If it's not there, download it. Then return the path to the cached file.
+    """
+    if cache_dir is None:
+        cache_dir = PYTORCH_TRANSFORMERS_CACHE
+    if sys.version_info[0] == 3 and isinstance(cache_dir, Path):
+        cache_dir = str(cache_dir)
+    if sys.version_info[0] == 2 and not isinstance(cache_dir, str):
+        cache_dir = str(cache_dir)
+
+    if not os.path.exists(cache_dir):
+        os.makedirs(cache_dir)
+
+    # Get eTag to add to filename, if it exists.
+    if url.startswith("s3://"):
+        etag = s3_etag(url, proxies=proxies)
+    else:
+        try:
+            response = requests.head(url, allow_redirects=True, proxies=proxies)
+            if response.status_code != 200:
+                etag = None
+            else:
+                etag = response.headers.get("ETag")
+        except EnvironmentError:
+            etag = None
+
+    if sys.version_info[0] == 2 and etag is not None:
+        etag = etag.decode('utf-8')
+    filename = url_to_filename(url, etag)
+
+    # get cache path to put the file
+    cache_path = os.path.join(cache_dir, filename)
+
+    # If we don't have a connection (etag is None) and can't identify the file
+    # try to get the last downloaded one
+    if not os.path.exists(cache_path) and etag is None:
+        matching_files = fnmatch.filter(os.listdir(cache_dir), filename + '.*')
+        matching_files = list(filter(lambda s: not s.endswith('.json'), matching_files))
+        if matching_files:
+            cache_path = os.path.join(cache_dir, matching_files[-1])
+
+    if not os.path.exists(cache_path) or force_download:
+        # Download to temporary file, then copy to cache dir once finished.
+        # Otherwise you get corrupt cache entries if the download gets interrupted.
+        with tempfile.NamedTemporaryFile() as temp_file:
+            logger.info("%s not found in cache or force_download set to True, downloading to %s", url, temp_file.name)
+
+            # GET file object
+            if url.startswith("s3://"):
+                s3_get(url, temp_file, proxies=proxies)
+            else:
+                http_get(url, temp_file, proxies=proxies)
+
+            # we are copying the file before closing it, so flush to avoid truncation
+            temp_file.flush()
+            # shutil.copyfileobj() starts at the current position, so go to the start
+            temp_file.seek(0)
+
+            logger.info("copying %s to cache at %s", temp_file.name, cache_path)
+            with open(cache_path, 'wb') as cache_file:
+                shutil.copyfileobj(temp_file, cache_file)
+
+            logger.info("creating metadata file for %s", cache_path)
+            meta = {'url': url, 'etag': etag}
+            meta_path = cache_path + '.json'
+            with open(meta_path, 'w') as meta_file:
+                output_string = json.dumps(meta)
+                if sys.version_info[0] == 2 and isinstance(output_string, str):
+                    output_string = unicode(output_string, 'utf-8')  # The beauty of python 2
+                meta_file.write(output_string)
+
+            logger.info("removing temp file %s", temp_file.name)
+
+    return cache_path

versatile_diffusion/lib/model_zoo/optimus_models/modeling_utils.py

@@ -0,0 +1,780 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch BERT model."""
+
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+
+import pdb
+import copy
+import json
+import logging
+import os
+from io import open
+
+import six
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from torch.nn import functional as F
+
+from .configuration_utils import PretrainedConfig
+from .file_utils import cached_path, WEIGHTS_NAME, TF_WEIGHTS_NAME
+
+logger = logging.getLogger(__name__)
+
+
+try:
+    from torch.nn import Identity
+except ImportError:
+    # Older PyTorch compatibility
+    class Identity(nn.Module):
+        r"""A placeholder identity operator that is argument-insensitive.
+        """
+        def __init__(self, *args, **kwargs):
+            super(Identity, self).__init__()
+
+        def forward(self, input):
+            return input
+
+class PreTrainedModel(nn.Module):
+    r""" Base class for all models.
+
+        :class:`~pytorch_transformers.PreTrainedModel` takes care of storing the configuration of the models and handles methods for loading/downloading/saving models
+        as well as a few methods commons to all models to (i) resize the input embeddings and (ii) prune heads in the self-attention heads.
+
+        Class attributes (overridden by derived classes):
+            - ``config_class``: a class derived from :class:`~pytorch_transformers.PretrainedConfig` to use as configuration class for this model architecture.
+            - ``pretrained_model_archive_map``: a python ``dict`` of with `short-cut-names` (string) as keys and `url` (string) of associated pretrained weights as values.
+            - ``load_tf_weights``: a python ``method`` for loading a TensorFlow checkpoint in a PyTorch model, taking as arguments:
+
+                - ``model``: an instance of the relevant subclass of :class:`~pytorch_transformers.PreTrainedModel`,
+                - ``config``: an instance of the relevant subclass of :class:`~pytorch_transformers.PretrainedConfig`,
+                - ``path``: a path (string) to the TensorFlow checkpoint.
+
+            - ``base_model_prefix``: a string indicating the attribute associated to the base model in derived classes of the same architecture adding modules on top of the base model.
+    """
+    config_class = None
+    pretrained_model_archive_map = {}
+    load_tf_weights = lambda model, config, path: None
+    base_model_prefix = ""
+
+    def __init__(self, config, *inputs, **kwargs):
+        super(PreTrainedModel, self).__init__()
+        if not isinstance(config, PretrainedConfig):
+            raise ValueError(
+                "Parameter config in `{}(config)` should be an instance of class `PretrainedConfig`. "
+                "To create a model from a pretrained model use "
+                "`model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`".format(
+                    self.__class__.__name__, self.__class__.__name__
+                ))
+        # Save config in model
+        self.config = config
+
+    def _get_resized_embeddings(self, old_embeddings, new_num_tokens=None):
+        """ Build a resized Embedding Module from a provided token Embedding Module.
+            Increasing the size will add newly initialized vectors at the end
+            Reducing the size will remove vectors from the end
+
+        Args:
+            new_num_tokens: (`optional`) int
+                New number of tokens in the embedding matrix.
+                Increasing the size will add newly initialized vectors at the end
+                Reducing the size will remove vectors from the end
+                If not provided or None: return the provided token Embedding Module.
+        Return: ``torch.nn.Embeddings``
+            Pointer to the resized Embedding Module or the old Embedding Module if new_num_tokens is None
+        """
+        if new_num_tokens is None:
+            return old_embeddings
+
+        old_num_tokens, old_embedding_dim = old_embeddings.weight.size()
+        if old_num_tokens == new_num_tokens:
+            return old_embeddings
+
+        # Build new embeddings
+        new_embeddings = nn.Embedding(new_num_tokens, old_embedding_dim)
+        new_embeddings.to(old_embeddings.weight.device)
+
+        # initialize all new embeddings (in particular added tokens)
+        self._init_weights(new_embeddings)
+
+        # Copy word embeddings from the previous weights
+        num_tokens_to_copy = min(old_num_tokens, new_num_tokens)
+        new_embeddings.weight.data[:num_tokens_to_copy, :] = old_embeddings.weight.data[:num_tokens_to_copy, :]
+
+        return new_embeddings
+
+    def _tie_or_clone_weights(self, first_module, second_module):
+        """ Tie or clone module weights depending of weither we are using TorchScript or not
+        """
+        if self.config.torchscript:
+            first_module.weight = nn.Parameter(second_module.weight.clone())
+        else:
+            first_module.weight = second_module.weight
+
+        if hasattr(first_module, 'bias') and first_module.bias is not None:
+            first_module.bias.data = torch.nn.functional.pad(
+                first_module.bias.data,
+                (0, first_module.weight.shape[0] - first_module.bias.shape[0]),
+                'constant',
+                0
+            )
+
+    def resize_token_embeddings(self, new_num_tokens=None):
+        """ Resize input token embeddings matrix of the model if new_num_tokens != config.vocab_size.
+        Take care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
+
+        Arguments:
+
+            new_num_tokens: (`optional`) int:
+                New number of tokens in the embedding matrix. Increasing the size will add newly initialized vectors at the end. Reducing the size will remove vectors from the end.
+                If not provided or None: does nothing and just returns a pointer to the input tokens ``torch.nn.Embeddings`` Module of the model.
+
+        Return: ``torch.nn.Embeddings``
+            Pointer to the input tokens Embeddings Module of the model
+        """
+
+
+        base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
+        
+        model_embeds = base_model._resize_token_embeddings(new_num_tokens)
+        if new_num_tokens is None:
+            return model_embeds
+
+        # Update base model and current model config
+        self.config.vocab_size = new_num_tokens
+        base_model.vocab_size = new_num_tokens
+
+        # Tie weights again if needed
+        if hasattr(self, 'tie_weights'):
+            self.tie_weights()
+
+        return model_embeds
+
+    def init_weights(self):
+        """ Initialize and prunes weights if needed. """
+        # Initialize weights
+        self.apply(self._init_weights)
+
+        # Prune heads if needed
+        if self.config.pruned_heads:
+            self.prune_heads(self.config.pruned_heads)
+
+    def prune_heads(self, heads_to_prune):
+        """ Prunes heads of the base model.
+
+            Arguments:
+
+                heads_to_prune: dict with keys being selected layer indices (`int`) and associated values being the list of heads to prune in said layer (list of `int`).
+                E.g. {1: [0, 2], 2: [2, 3]} will prune heads 0 and 2 on layer 1 and heads 2 and 3 on layer 2.
+        """
+        base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
+
+        # save new sets of pruned heads as union of previously stored pruned heads and newly pruned heads
+        for layer, heads in heads_to_prune.items():
+            union_heads = set(self.config.pruned_heads.get(layer, [])) | set(heads)
+            self.config.pruned_heads[layer] = list(union_heads)  # Unfortunately we have to store it as list for JSON
+
+        base_model._prune_heads(heads_to_prune)
+
+    def save_pretrained(self, save_directory):
+        """ Save a model and its configuration file to a directory, so that it
+            can be re-loaded using the `:func:`~pytorch_transformers.PreTrainedModel.from_pretrained`` class method.
+        """
+        assert os.path.isdir(save_directory), "Saving path should be a directory where the model and configuration can be saved"
+
+        # Only save the model it-self if we are using distributed training
+        model_to_save = self.module if hasattr(self, 'module') else self
+
+        # Save configuration file
+        model_to_save.config.save_pretrained(save_directory)
+
+        # If we save using the predefined names, we can load using `from_pretrained`
+        output_model_file = os.path.join(save_directory, WEIGHTS_NAME)
+
+        torch.save(model_to_save.state_dict(), output_model_file)
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        r"""Instantiate a pretrained pytorch model from a pre-trained model configuration.
+
+        The model is set in evaluation mode by default using ``model.eval()`` (Dropout modules are deactivated)
+        To train the model, you should first set it back in training mode with ``model.train()``
+
+        The warning ``Weights from XXX not initialized from pretrained model`` means that the weights of XXX do not come pre-trained with the rest of the model.
+        It is up to you to train those weights with a downstream fine-tuning task.
+
+        The warning ``Weights from XXX not used in YYY`` means that the layer XXX is not used by YYY, therefore those weights are discarded.
+
+        Parameters:
+            pretrained_model_name_or_path: either:
+
+                - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a path to a `directory` containing model weights saved using :func:`~pytorch_transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
+                - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
+
+            model_args: (`optional`) Sequence of positional arguments:
+                All remaning positional arguments will be passed to the underlying model's ``__init__`` method
+
+            config: (`optional`) instance of a class derived from :class:`~pytorch_transformers.PretrainedConfig`:
+                Configuration for the model to use instead of an automatically loaded configuation. Configuration can be automatically loaded when:
+
+                - the model is a model provided by the library (loaded with the ``shortcut-name`` string of a pretrained model), or
+                - the model was saved using :func:`~pytorch_transformers.PreTrainedModel.save_pretrained` and is reloaded by suppling the save directory.
+                - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
+
+            state_dict: (`optional`) dict:
+                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                This option can be used if you want to create a model from a pretrained configuration but load your own weights.
+                In this case though, you should check if using :func:`~pytorch_transformers.PreTrainedModel.save_pretrained` and :func:`~pytorch_transformers.PreTrainedModel.from_pretrained` is not a simpler option.
+
+            cache_dir: (`optional`) string:
+                Path to a directory in which a downloaded pre-trained model
+                configuration should be cached if the standard cache should not be used.
+
+            force_download: (`optional`) boolean, default False:
+                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
+
+            proxies: (`optional`) dict, default None:
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
+                The proxies are used on each request.
+
+            output_loading_info: (`optional`) boolean:
+                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
+
+            kwargs: (`optional`) Remaining dictionary of keyword arguments:
+                Can be used to update the configuration object (after it being loaded) and initiate the model. (e.g. ``output_attention=True``). Behave differently depending on whether a `config` is provided or automatically loaded:
+
+                - If a configuration is provided with ``config``, ``**kwargs`` will be directly passed to the underlying model's ``__init__`` method (we assume all relevant updates to the configuration have already been done)
+                - If a configuration is not provided, ``kwargs`` will be first passed to the configuration class initialization function (:func:`~pytorch_transformers.PretrainedConfig.from_pretrained`). Each key of ``kwargs`` that corresponds to a configuration attribute will be used to override said attribute with the supplied ``kwargs`` value. Remaining keys that do not correspond to any configuration attribute will be passed to the underlying model's ``__init__`` function.
+
+        Examples::
+
+            model = BertModel.from_pretrained('bert-base-uncased')    # Download model and configuration from S3 and cache.
+            model = BertModel.from_pretrained('./test/saved_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+            model = BertModel.from_pretrained('bert-base-uncased', output_attention=True)  # Update configuration during loading
+            assert model.config.output_attention == True
+            # Loading from a TF checkpoint file instead of a PyTorch model (slower)
+            config = BertConfig.from_json_file('./tf_model/my_tf_model_config.json')
+            model = BertModel.from_pretrained('./tf_model/my_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+
+        """
+        config = kwargs.pop('config', None)
+        state_dict = kwargs.pop('state_dict', None)
+        cache_dir = kwargs.pop('cache_dir', None)
+        from_tf = kwargs.pop('from_tf', False)
+        force_download = kwargs.pop('force_download', False)
+        proxies = kwargs.pop('proxies', None)
+        output_loading_info = kwargs.pop('output_loading_info', False)
+
+        # Load config
+        if config is None:
+            config, model_kwargs = cls.config_class.from_pretrained(
+                pretrained_model_name_or_path, *model_args,
+                cache_dir=cache_dir, return_unused_kwargs=True,
+                force_download=force_download,
+                **kwargs
+            )
+        else:
+            model_kwargs = kwargs
+
+        # Load model
+        if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
+            archive_file = cls.pretrained_model_archive_map[pretrained_model_name_or_path]
+        elif os.path.isdir(pretrained_model_name_or_path):
+            if from_tf:
+                # Directly load from a TensorFlow checkpoint
+                archive_file = os.path.join(pretrained_model_name_or_path, TF_WEIGHTS_NAME + ".index")
+            else:
+                archive_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)
+        else:
+            if from_tf:
+                # Directly load from a TensorFlow checkpoint
+                archive_file = pretrained_model_name_or_path + ".index"
+            else:
+                archive_file = pretrained_model_name_or_path
+        # redirect to the cache, if necessary
+        try:
+            resolved_archive_file = cached_path(archive_file, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
+        except EnvironmentError as e:
+            if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
+                logger.error(
+                    "Couldn't reach server at '{}' to download pretrained weights.".format(
+                        archive_file))
+            else:
+                logger.error(
+                    "Model name '{}' was not found in model name list ({}). "
+                    "We assumed '{}' was a path or url but couldn't find any file "
+                    "associated to this path or url.".format(
+                        pretrained_model_name_or_path,
+                        ', '.join(cls.pretrained_model_archive_map.keys()),
+                        archive_file))
+            raise e
+        if resolved_archive_file == archive_file:
+            logger.info("loading weights file {}".format(archive_file))
+        else:
+            logger.info("loading weights file {} from cache at {}".format(
+                archive_file, resolved_archive_file))
+
+        # Instantiate model.
+        model = cls(config, *model_args, **model_kwargs)
+
+        if state_dict is None and not from_tf:
+            state_dict = torch.load(resolved_archive_file, map_location='cpu')
+        if from_tf:
+            # Directly load from a TensorFlow checkpoint
+            return cls.load_tf_weights(model, config, resolved_archive_file[:-6])  # Remove the '.index'
+
+        # Convert old format to new format if needed from a PyTorch state_dict
+        old_keys = []
+        new_keys = []
+        for key in state_dict.keys():
+            new_key = None
+            if 'gamma' in key:
+                new_key = key.replace('gamma', 'weight')
+            if 'beta' in key:
+                new_key = key.replace('beta', 'bias')
+            if new_key:
+                old_keys.append(key)
+                new_keys.append(new_key)
+        for old_key, new_key in zip(old_keys, new_keys):
+            state_dict[new_key] = state_dict.pop(old_key)
+
+        # Load from a PyTorch state_dict
+        missing_keys = []
+        unexpected_keys = []
+        error_msgs = []
+        # copy state_dict so _load_from_state_dict can modify it
+        metadata = getattr(state_dict, '_metadata', None)
+        state_dict = state_dict.copy()
+        if metadata is not None:
+            state_dict._metadata = metadata
+
+        def load(module, prefix=''):
+            local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
+            module._load_from_state_dict(
+                state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs)
+            for name, child in module._modules.items():
+                if child is not None:
+                    load(child, prefix + name + '.')
+
+        # Make sure we are able to load base models as well as derived models (with heads)
+        start_prefix = ''
+        model_to_load = model
+        if not hasattr(model, cls.base_model_prefix) and any(s.startswith(cls.base_model_prefix) for s in state_dict.keys()):
+            start_prefix = cls.base_model_prefix + '.'
+        if hasattr(model, cls.base_model_prefix) and not any(s.startswith(cls.base_model_prefix) for s in state_dict.keys()):
+            model_to_load = getattr(model, cls.base_model_prefix)
+
+        load(model_to_load, prefix=start_prefix)
+        if len(missing_keys) > 0:
+            logger.info("Weights of {} not initialized from pretrained model: {}".format(
+                model.__class__.__name__, missing_keys))
+        if len(unexpected_keys) > 0:
+            logger.info("Weights from pretrained model not used in {}: {}".format(
+                model.__class__.__name__, unexpected_keys))
+        if len(error_msgs) > 0:
+            raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
+                               model.__class__.__name__, "\n\t".join(error_msgs)))
+
+        if hasattr(model, 'tie_weights'):
+            model.tie_weights()  # make sure word embedding weights are still tied
+
+        # Set model in evaluation mode to desactivate DropOut modules by default
+        model.eval()
+
+        if output_loading_info:
+            loading_info = {"missing_keys": missing_keys, "unexpected_keys": unexpected_keys, "error_msgs": error_msgs}
+            return model, loading_info
+
+        return model
+
+
+class Conv1D(nn.Module):
+    def __init__(self, nf, nx):
+        """ Conv1D layer as defined by Radford et al. for OpenAI GPT (and also used in GPT-2)
+            Basically works like a Linear layer but the weights are transposed
+        """
+        super(Conv1D, self).__init__()
+        self.nf = nf
+        w = torch.empty(nx, nf)
+        nn.init.normal_(w, std=0.02)
+        self.weight = nn.Parameter(w)
+        self.bias = nn.Parameter(torch.zeros(nf))
+
+    def forward(self, x):
+        size_out = x.size()[:-1] + (self.nf,)
+        x = torch.addmm(self.bias, x.view(-1, x.size(-1)), self.weight)
+        x = x.view(*size_out)
+        return x
+
+
+class PoolerStartLogits(nn.Module):
+    """ Compute SQuAD start_logits from sequence hidden states. """
+    def __init__(self, config):
+        super(PoolerStartLogits, self).__init__()
+        self.dense = nn.Linear(config.hidden_size, 1)
+
+    def forward(self, hidden_states, p_mask=None):
+        """ Args:
+            **p_mask**: (`optional`) ``torch.FloatTensor`` of shape `(batch_size, seq_len)`
+                invalid position mask such as query and special symbols (PAD, SEP, CLS)
+                1.0 means token should be masked.
+        """
+        x = self.dense(hidden_states).squeeze(-1)
+
+        if p_mask is not None:
+            if next(self.parameters()).dtype == torch.float16:
+                x = x * (1 - p_mask) - 65500 * p_mask
+            else:
+                x = x * (1 - p_mask) - 1e30 * p_mask
+
+        return x
+
+
+class PoolerEndLogits(nn.Module):
+    """ Compute SQuAD end_logits from sequence hidden states and start token hidden state.
+    """
+    def __init__(self, config):
+        super(PoolerEndLogits, self).__init__()
+        self.dense_0 = nn.Linear(config.hidden_size * 2, config.hidden_size)
+        self.activation = nn.Tanh()
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dense_1 = nn.Linear(config.hidden_size, 1)
+
+    def forward(self, hidden_states, start_states=None, start_positions=None, p_mask=None):
+        """ Args:
+            One of ``start_states``, ``start_positions`` should be not None.
+            If both are set, ``start_positions`` overrides ``start_states``.
+
+            **start_states**: ``torch.LongTensor`` of shape identical to hidden_states
+                hidden states of the first tokens for the labeled span.
+            **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
+                position of the first token for the labeled span:
+            **p_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, seq_len)``
+                Mask of invalid position such as query and special symbols (PAD, SEP, CLS)
+                1.0 means token should be masked.
+        """
+        assert start_states is not None or start_positions is not None, "One of start_states, start_positions should be not None"
+        if start_positions is not None:
+            slen, hsz = hidden_states.shape[-2:]
+            start_positions = start_positions[:, None, None].expand(-1, -1, hsz) # shape (bsz, 1, hsz)
+            start_states = hidden_states.gather(-2, start_positions) # shape (bsz, 1, hsz)
+            start_states = start_states.expand(-1, slen, -1) # shape (bsz, slen, hsz)
+
+        x = self.dense_0(torch.cat([hidden_states, start_states], dim=-1))
+        x = self.activation(x)
+        x = self.LayerNorm(x)
+        x = self.dense_1(x).squeeze(-1)
+
+        if p_mask is not None:
+            x = x * (1 - p_mask) - 1e30 * p_mask
+
+        return x
+
+
+class PoolerAnswerClass(nn.Module):
+    """ Compute SQuAD 2.0 answer class from classification and start tokens hidden states. """
+    def __init__(self, config):
+        super(PoolerAnswerClass, self).__init__()
+        self.dense_0 = nn.Linear(config.hidden_size * 2, config.hidden_size)
+        self.activation = nn.Tanh()
+        self.dense_1 = nn.Linear(config.hidden_size, 1, bias=False)
+
+    def forward(self, hidden_states, start_states=None, start_positions=None, cls_index=None):
+        """
+        Args:
+            One of ``start_states``, ``start_positions`` should be not None.
+            If both are set, ``start_positions`` overrides ``start_states``.
+
+            **start_states**: ``torch.LongTensor`` of shape identical to ``hidden_states``.
+                hidden states of the first tokens for the labeled span.
+            **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
+                position of the first token for the labeled span.
+            **cls_index**: torch.LongTensor of shape ``(batch_size,)``
+                position of the CLS token. If None, take the last token.
+
+            note(Original repo):
+                no dependency on end_feature so that we can obtain one single `cls_logits`
+                for each sample
+        """
+        hsz = hidden_states.shape[-1]
+        assert start_states is not None or start_positions is not None, "One of start_states, start_positions should be not None"
+        if start_positions is not None:
+            start_positions = start_positions[:, None, None].expand(-1, -1, hsz) # shape (bsz, 1, hsz)
+            start_states = hidden_states.gather(-2, start_positions).squeeze(-2) # shape (bsz, hsz)
+
+        if cls_index is not None:
+            cls_index = cls_index[:, None, None].expand(-1, -1, hsz) # shape (bsz, 1, hsz)
+            cls_token_state = hidden_states.gather(-2, cls_index).squeeze(-2) # shape (bsz, hsz)
+        else:
+            cls_token_state = hidden_states[:, -1, :] # shape (bsz, hsz)
+
+        x = self.dense_0(torch.cat([start_states, cls_token_state], dim=-1))
+        x = self.activation(x)
+        x = self.dense_1(x).squeeze(-1)
+
+        return x
+
+
+class SQuADHead(nn.Module):
+    r""" A SQuAD head inspired by XLNet.
+
+    Parameters:
+        config (:class:`~pytorch_transformers.XLNetConfig`): Model configuration class with all the parameters of the model.
+
+    Inputs:
+        **hidden_states**: ``torch.FloatTensor`` of shape ``(batch_size, seq_len, hidden_size)``
+            hidden states of sequence tokens
+        **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
+            position of the first token for the labeled span.
+        **end_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
+            position of the last token for the labeled span.
+        **cls_index**: torch.LongTensor of shape ``(batch_size,)``
+            position of the CLS token. If None, take the last token.
+        **is_impossible**: ``torch.LongTensor`` of shape ``(batch_size,)``
+            Whether the question has a possible answer in the paragraph or not.
+        **p_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, seq_len)``
+            Mask of invalid position such as query and special symbols (PAD, SEP, CLS)
+            1.0 means token should be masked.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned if both ``start_positions`` and ``end_positions`` are provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification loss as the sum of start token, end token (and is_impossible if provided) classification losses.
+        **start_top_log_probs**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
+            ``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top)``
+            Log probabilities for the top config.start_n_top start token possibilities (beam-search).
+        **start_top_index**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
+            ``torch.LongTensor`` of shape ``(batch_size, config.start_n_top)``
+            Indices for the top config.start_n_top start token possibilities (beam-search).
+        **end_top_log_probs**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
+            ``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``
+            Log probabilities for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
+        **end_top_index**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
+            ``torch.LongTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``
+            Indices for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
+        **cls_logits**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
+            ``torch.FloatTensor`` of shape ``(batch_size,)``
+            Log probabilities for the ``is_impossible`` label of the answers.
+    """
+    def __init__(self, config):
+        super(SQuADHead, self).__init__()
+        self.start_n_top = config.start_n_top
+        self.end_n_top = config.end_n_top
+
+        self.start_logits = PoolerStartLogits(config)
+        self.end_logits = PoolerEndLogits(config)
+        self.answer_class = PoolerAnswerClass(config)
+
+    def forward(self, hidden_states, start_positions=None, end_positions=None,
+                cls_index=None, is_impossible=None, p_mask=None):
+        outputs = ()
+
+        start_logits = self.start_logits(hidden_states, p_mask=p_mask)
+
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, let's remove the dimension added by batch splitting
+            for x in (start_positions, end_positions, cls_index, is_impossible):
+                if x is not None and x.dim() > 1:
+                    x.squeeze_(-1)
+
+            # during training, compute the end logits based on the ground truth of the start position
+            end_logits = self.end_logits(hidden_states, start_positions=start_positions, p_mask=p_mask)
+
+            loss_fct = CrossEntropyLoss()
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+            if cls_index is not None and is_impossible is not None:
+                # Predict answerability from the representation of CLS and START
+                cls_logits = self.answer_class(hidden_states, start_positions=start_positions, cls_index=cls_index)
+                loss_fct_cls = nn.BCEWithLogitsLoss()
+                cls_loss = loss_fct_cls(cls_logits, is_impossible)
+
+                # note(zhiliny): by default multiply the loss by 0.5 so that the scale is comparable to start_loss and end_loss
+                total_loss += cls_loss * 0.5
+
+            outputs = (total_loss,) + outputs
+
+        else:
+            # during inference, compute the end logits based on beam search
+            bsz, slen, hsz = hidden_states.size()
+            start_log_probs = F.softmax(start_logits, dim=-1) # shape (bsz, slen)
+
+            start_top_log_probs, start_top_index = torch.topk(start_log_probs, self.start_n_top, dim=-1) # shape (bsz, start_n_top)
+            start_top_index_exp = start_top_index.unsqueeze(-1).expand(-1, -1, hsz) # shape (bsz, start_n_top, hsz)
+            start_states = torch.gather(hidden_states, -2, start_top_index_exp) # shape (bsz, start_n_top, hsz)
+            start_states = start_states.unsqueeze(1).expand(-1, slen, -1, -1) # shape (bsz, slen, start_n_top, hsz)
+
+            hidden_states_expanded = hidden_states.unsqueeze(2).expand_as(start_states) # shape (bsz, slen, start_n_top, hsz)
+            p_mask = p_mask.unsqueeze(-1) if p_mask is not None else None
+            end_logits = self.end_logits(hidden_states_expanded, start_states=start_states, p_mask=p_mask)
+            end_log_probs = F.softmax(end_logits, dim=1) # shape (bsz, slen, start_n_top)
+
+            end_top_log_probs, end_top_index = torch.topk(end_log_probs, self.end_n_top, dim=1) # shape (bsz, end_n_top, start_n_top)
+            end_top_log_probs = end_top_log_probs.view(-1, self.start_n_top * self.end_n_top)
+            end_top_index = end_top_index.view(-1, self.start_n_top * self.end_n_top)
+
+            start_states = torch.einsum("blh,bl->bh", hidden_states, start_log_probs)
+            cls_logits = self.answer_class(hidden_states, start_states=start_states, cls_index=cls_index)
+
+            outputs = (start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits) + outputs
+
+        # return start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits
+        # or (if labels are provided) (total_loss,)
+        return outputs
+
+
+class SequenceSummary(nn.Module):
+    r""" Compute a single vector summary of a sequence hidden states according to various possibilities:
+        Args of the config class:
+            summary_type:
+                - 'last' => [default] take the last token hidden state (like XLNet)
+                - 'first' => take the first token hidden state (like Bert)
+                - 'mean' => take the mean of all tokens hidden states
+                - 'cls_index' => supply a Tensor of classification token position (GPT/GPT-2)
+                - 'attn' => Not implemented now, use multi-head attention
+            summary_use_proj: Add a projection after the vector extraction
+            summary_proj_to_labels: If True, the projection outputs to config.num_labels classes (otherwise to hidden_size). Default: False.
+            summary_activation: 'tanh' => add a tanh activation to the output, Other => no activation. Default
+            summary_first_dropout: Add a dropout before the projection and activation
+            summary_last_dropout: Add a dropout after the projection and activation
+    """
+    def __init__(self, config):
+        super(SequenceSummary, self).__init__()
+
+        self.summary_type = config.summary_type if hasattr(config, 'summary_use_proj') else 'last'
+        if self.summary_type == 'attn':
+            # We should use a standard multi-head attention module with absolute positional embedding for that.
+            # Cf. https://github.com/zihangdai/xlnet/blob/master/modeling.py#L253-L276
+            # We can probably just use the multi-head attention module of PyTorch >=1.1.0
+            raise NotImplementedError
+
+        self.summary = Identity()
+        if hasattr(config, 'summary_use_proj') and config.summary_use_proj:
+            if hasattr(config, 'summary_proj_to_labels') and config.summary_proj_to_labels and config.num_labels > 0:
+                num_classes = config.num_labels
+            else:
+                num_classes = config.hidden_size
+            self.summary = nn.Linear(config.hidden_size, num_classes)
+
+        self.activation = Identity()
+        if hasattr(config, 'summary_activation') and config.summary_activation == 'tanh':
+            self.activation = nn.Tanh()
+
+        self.first_dropout = Identity()
+        if hasattr(config, 'summary_first_dropout') and config.summary_first_dropout > 0:
+            self.first_dropout = nn.Dropout(config.summary_first_dropout)
+
+        self.last_dropout = Identity()
+        if hasattr(config, 'summary_last_dropout') and config.summary_last_dropout > 0:
+            self.last_dropout = nn.Dropout(config.summary_last_dropout)
+
+    def forward(self, hidden_states, cls_index=None):
+        """ hidden_states: float Tensor in shape [bsz, seq_len, hidden_size], the hidden-states of the last layer.
+            cls_index: [optional] position of the classification token if summary_type == 'cls_index',
+                shape (bsz,) or more generally (bsz, ...) where ... are optional leading dimensions of hidden_states.
+                if summary_type == 'cls_index' and cls_index is None:
+                    we take the last token of the sequence as classification token
+        """
+        if self.summary_type == 'last':
+            output = hidden_states[:, -1]
+        elif self.summary_type == 'first':
+            output = hidden_states[:, 0]
+        elif self.summary_type == 'mean':
+            output = hidden_states.mean(dim=1)
+        elif self.summary_type == 'cls_index':
+            if cls_index is None:
+                cls_index = torch.full_like(hidden_states[..., :1, :], hidden_states.shape[-2]-1, dtype=torch.long)
+            else:
+                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
+                cls_index = cls_index.expand((-1,) * (cls_index.dim()-1) + (hidden_states.size(-1),))
+            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
+            output = hidden_states.gather(-2, cls_index).squeeze(-2) # shape (bsz, XX, hidden_size)
+        elif self.summary_type == 'attn':
+            raise NotImplementedError
+
+        output = self.first_dropout(output)
+        output = self.summary(output)
+        output = self.activation(output)
+        output = self.last_dropout(output)
+
+        return output
+
+
+def prune_linear_layer(layer, index, dim=0):
+    """ Prune a linear layer (a model parameters) to keep only entries in index.
+        Return the pruned layer as a new layer with requires_grad=True.
+        Used to remove heads.
+    """
+    index = index.to(layer.weight.device)
+    W = layer.weight.index_select(dim, index).clone().detach()
+    if layer.bias is not None:
+        if dim == 1:
+            b = layer.bias.clone().detach()
+        else:
+            b = layer.bias[index].clone().detach()
+    new_size = list(layer.weight.size())
+    new_size[dim] = len(index)
+    new_layer = nn.Linear(new_size[1], new_size[0], bias=layer.bias is not None).to(layer.weight.device)
+    new_layer.weight.requires_grad = False
+    new_layer.weight.copy_(W.contiguous())
+    new_layer.weight.requires_grad = True
+    if layer.bias is not None:
+        new_layer.bias.requires_grad = False
+        new_layer.bias.copy_(b.contiguous())
+        new_layer.bias.requires_grad = True
+    return new_layer
+
+
+def prune_conv1d_layer(layer, index, dim=1):
+    """ Prune a Conv1D layer (a model parameters) to keep only entries in index.
+        A Conv1D work as a Linear layer (see e.g. BERT) but the weights are transposed.
+        Return the pruned layer as a new layer with requires_grad=True.
+        Used to remove heads.
+    """
+    index = index.to(layer.weight.device)
+    W = layer.weight.index_select(dim, index).clone().detach()
+    if dim == 0:
+        b = layer.bias.clone().detach()
+    else:
+        b = layer.bias[index].clone().detach()
+    new_size = list(layer.weight.size())
+    new_size[dim] = len(index)
+    new_layer = Conv1D(new_size[1], new_size[0]).to(layer.weight.device)
+    new_layer.weight.requires_grad = False
+    new_layer.weight.copy_(W.contiguous())
+    new_layer.weight.requires_grad = True
+    new_layer.bias.requires_grad = False
+    new_layer.bias.copy_(b.contiguous())
+    new_layer.bias.requires_grad = True
+    return new_layer
+
+
+def prune_layer(layer, index, dim=None):
+    """ Prune a Conv1D or nn.Linear layer (a model parameters) to keep only entries in index.
+        Return the pruned layer as a new layer with requires_grad=True.
+        Used to remove heads.
+    """
+    if isinstance(layer, nn.Linear):
+        return prune_linear_layer(layer, index, dim=0 if dim is None else dim)
+    elif isinstance(layer, Conv1D):
+        return prune_conv1d_layer(layer, index, dim=1 if dim is None else dim)
+    else:
+        raise ValueError("Can't prune layer of class {}".format(layer.__class__))

versatile_diffusion/lib/model_zoo/optimus_models/optimus_bert.py

@@ -0,0 +1,1439 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch BERT model. """
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import math
+import os
+import sys
+from io import open
+
+import pdb
+
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from .modeling_utils import PreTrainedModel, prune_linear_layer
+from .configuration_bert import BertConfig
+from .file_utils import add_start_docstrings
+
+logger = logging.getLogger(__name__)
+
+BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'bert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-pytorch_model.bin",
+    'bert-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-pytorch_model.bin",
+    'bert-base-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-pytorch_model.bin",
+    'bert-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-pytorch_model.bin",
+    'bert-base-multilingual-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased-pytorch_model.bin",
+    'bert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased-pytorch_model.bin",
+    'bert-base-chinese': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese-pytorch_model.bin",
+    'bert-base-german-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-cased-pytorch_model.bin",
+    'bert-large-uncased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-pytorch_model.bin",
+    'bert-large-cased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-pytorch_model.bin",
+    'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-pytorch_model.bin",
+    'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-pytorch_model.bin",
+    'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-pytorch_model.bin",
+}
+
+def load_tf_weights_in_bert(model, config, tf_checkpoint_path):
+    """ Load tf checkpoints in a pytorch model.
+    """
+    try:
+        import re
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error("Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions.")
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info("Loading TF weight {} with shape {}".format(name, shape))
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        name = name.split('/')
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(n in ["adam_v", "adam_m", "global_step"] for n in name):
+            logger.info("Skipping {}".format("/".join(name)))
+            continue
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
+                l = re.split(r'_(\d+)', m_name)
+            else:
+                l = [m_name]
+            if l[0] == 'kernel' or l[0] == 'gamma':
+                pointer = getattr(pointer, 'weight')
+            elif l[0] == 'output_bias' or l[0] == 'beta':
+                pointer = getattr(pointer, 'bias')
+            elif l[0] == 'output_weights':
+                pointer = getattr(pointer, 'weight')
+            elif l[0] == 'squad':
+                pointer = getattr(pointer, 'classifier')
+            else:
+                try:
+                    pointer = getattr(pointer, l[0])
+                except AttributeError:
+                    logger.info("Skipping {}".format("/".join(name)))
+                    continue
+            if len(l) >= 2:
+                num = int(l[1])
+                pointer = pointer[num]
+        if m_name[-11:] == '_embeddings':
+            pointer = getattr(pointer, 'weight')
+        elif m_name == 'kernel':
+            array = np.transpose(array)
+        try:
+            assert pointer.shape == array.shape
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info("Initialize PyTorch weight {}".format(name))
+        pointer.data = torch.from_numpy(array)
+    return model
+
+
+def gelu(x):
+    """Implementation of the gelu activation function.
+        For information: OpenAI GPT's gelu is slightly different (and gives slightly different results):
+        0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
+        Also see https://arxiv.org/abs/1606.08415
+    """
+    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
+
+
+def swish(x):
+    return x * torch.sigmoid(x)
+
+
+ACT2FN = {"gelu": gelu, "relu": torch.nn.functional.relu, "swish": swish}
+
+
+try:
+    from apex.normalization.fused_layer_norm import FusedLayerNorm as BertLayerNorm
+except (ImportError, AttributeError) as e:
+    logger.info("Better speed can be achieved with apex installed from https://www.github.com/nvidia/apex .")
+    BertLayerNorm = torch.nn.LayerNorm
+
+class BertEmbeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings.
+    """
+    def __init__(self, config):
+        super(BertEmbeddings, self).__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=0)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, input_ids, token_type_ids=None, position_ids=None):
+        seq_length = input_ids.size(1)
+        if position_ids is None:
+            position_ids = torch.arange(seq_length, dtype=torch.long, device=input_ids.device)
+            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros_like(input_ids)
+
+        words_embeddings = self.word_embeddings(input_ids)
+        position_embeddings = self.position_embeddings(position_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = words_embeddings + position_embeddings + token_type_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class BertSelfAttention(nn.Module):
+    def __init__(self, config):
+        super(BertSelfAttention, self).__init__()
+        if config.hidden_size % config.num_attention_heads != 0:
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, config.num_attention_heads))
+        self.output_attentions = config.output_attentions
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(self, hidden_states, attention_mask, head_mask=None):
+        mixed_query_layer = self.query(hidden_states)
+        mixed_key_layer = self.key(hidden_states)
+        mixed_value_layer = self.value(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+        key_layer = self.transpose_for_scores(mixed_key_layer)
+        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+        attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if self.output_attentions else (context_layer,)
+        return outputs
+
+
+class BertSelfOutput(nn.Module):
+    def __init__(self, config):
+        super(BertSelfOutput, self).__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertAttention(nn.Module):
+    def __init__(self, config):
+        super(BertAttention, self).__init__()
+        self.self = BertSelfAttention(config)
+        self.output = BertSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        mask = torch.ones(self.self.num_attention_heads, self.self.attention_head_size)
+        heads = set(heads) - self.pruned_heads  # Convert to set and emove already pruned heads
+        for head in heads:
+            # Compute how many pruned heads are before the head and move the index accordingly
+            head = head - sum(1 if h < head else 0 for h in self.pruned_heads)
+            mask[head] = 0
+        mask = mask.view(-1).contiguous().eq(1)
+        index = torch.arange(len(mask))[mask].long()
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(self, input_tensor, attention_mask, head_mask=None):
+        self_outputs = self.self(input_tensor, attention_mask, head_mask)
+        attention_output = self.output(self_outputs[0], input_tensor)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+class BertIntermediate(nn.Module):
+    def __init__(self, config):
+        super(BertIntermediate, self).__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str) or (sys.version_info[0] == 2 and isinstance(config.hidden_act, unicode)):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class BertOutput(nn.Module):
+    def __init__(self, config):
+        super(BertOutput, self).__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertLayer(nn.Module):
+    def __init__(self, config):
+        super(BertLayer, self).__init__()
+        self.attention = BertAttention(config)
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config)
+
+    def forward(self, hidden_states, attention_mask, head_mask=None):
+        attention_outputs = self.attention(hidden_states, attention_mask, head_mask)
+        attention_output = attention_outputs[0]
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        outputs = (layer_output,) + attention_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+class BertEncoder(nn.Module):
+    def __init__(self, config):
+        super(BertEncoder, self).__init__()
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.layer = nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
+
+    def forward(self, hidden_states, attention_mask, head_mask=None):
+        all_hidden_states = ()
+        all_attentions = ()
+        for i, layer_module in enumerate(self.layer):
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_outputs = layer_module(hidden_states, attention_mask, head_mask[i])
+            hidden_states = layer_outputs[0]
+
+            if self.output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        # Add last layer
+        if self.output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (all_attentions,)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
+
+
+class BertPooler(nn.Module):
+    def __init__(self, config):
+        super(BertPooler, self).__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class BertPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super(BertPredictionHeadTransform, self).__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str) or (sys.version_info[0] == 2 and isinstance(config.hidden_act, unicode)):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class BertLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super(BertLMPredictionHead, self).__init__()
+        self.transform = BertPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size,
+                                 config.vocab_size,
+                                 bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states) + self.bias
+        return hidden_states
+
+
+class BertOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super(BertOnlyMLMHead, self).__init__()
+        self.predictions = BertLMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+class BertOnlyNSPHead(nn.Module):
+    def __init__(self, config):
+        super(BertOnlyNSPHead, self).__init__()
+        self.seq_relationship = nn.Linear(config.hidden_size, 2)
+
+    def forward(self, pooled_output):
+        seq_relationship_score = self.seq_relationship(pooled_output)
+        return seq_relationship_score
+
+
+class BertPreTrainingHeads(nn.Module):
+    def __init__(self, config):
+        super(BertPreTrainingHeads, self).__init__()
+        self.predictions = BertLMPredictionHead(config)
+        self.seq_relationship = nn.Linear(config.hidden_size, 2)
+
+    def forward(self, sequence_output, pooled_output):
+        prediction_scores = self.predictions(sequence_output)
+        seq_relationship_score = self.seq_relationship(pooled_output)
+        return prediction_scores, seq_relationship_score
+
+
+class BertPreTrainedModel(PreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = BertConfig
+    pretrained_model_archive_map = BERT_PRETRAINED_MODEL_ARCHIVE_MAP
+    load_tf_weights = load_tf_weights_in_bert
+    base_model_prefix = "bert"
+
+    def _init_weights(self, module):
+        """ Initialize the weights """
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, BertLayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+BERT_START_DOCSTRING = r"""    The BERT model was proposed in
+    `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_
+    by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova. It's a bidirectional transformer
+    pre-trained using a combination of masked language modeling objective and next sentence prediction
+    on a large corpus comprising the Toronto Book Corpus and Wikipedia.
+
+    This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
+    refer to the PyTorch documentation for all matter related to general usage and behavior.
+
+    .. _`BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`:
+        https://arxiv.org/abs/1810.04805
+
+    .. _`torch.nn.Module`:
+        https://pytorch.org/docs/stable/nn.html#module
+
+    Parameters:
+        config (:class:`~pytorch_transformers.BertConfig`): Model configuration class with all the parameters of the model. 
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~pytorch_transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+BERT_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, BERT input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
+                
+                ``token_type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+                
+                ``token_type_ids:   0   0   0   0  0     0   0``
+
+            Bert is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
+
+            Indices can be obtained using :class:`pytorch_transformers.BertTokenizer`.
+            See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
+            :func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Segment token indices to indicate first and second portions of the inputs.
+            Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
+            corresponds to a `sentence B` token
+            (see `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
+        **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+"""
+
+@add_start_docstrings("The bare Bert Model transformer outputting raw hidden-states without any specific head on top.",
+                      BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+class BertModel(BertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            objective during Bert pretraining. This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertModel.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    def __init__(self, config):
+        super(BertModel, self).__init__(config)
+
+        self.embeddings = BertEmbeddings(config)
+        self.encoder = BertEncoder(config)
+        self.pooler = BertPooler(config)
+
+        self.init_weights()
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        old_embeddings = self.embeddings.word_embeddings
+        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
+        self.embeddings.word_embeddings = new_embeddings
+        return self.embeddings.word_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            See base class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None):
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros_like(input_ids)
+
+        # We create a 3D attention mask from a 2D tensor mask.
+        # Sizes are [batch_size, 1, 1, to_seq_length]
+        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+        # this attention mask is more simple than the triangular masking of causal attention
+        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if head_mask is not None:
+            if head_mask.dim() == 1:
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.expand(self.config.num_hidden_layers, -1, -1, -1, -1)
+            elif head_mask.dim() == 2:
+                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+        else:
+            head_mask = [None] * self.config.num_hidden_layers
+
+        embedding_output = self.embeddings(input_ids, position_ids=position_ids, token_type_ids=token_type_ids)
+        encoder_outputs = self.encoder(embedding_output,
+                                       extended_attention_mask,
+                                       head_mask=head_mask)
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output)
+
+        outputs = (sequence_output, pooled_output,) + encoder_outputs[1:]  # add hidden_states and attentions if they are here
+        return outputs  # sequence_output, pooled_output, (hidden_states), (attentions)
+
+
+
+
+
+
+@add_start_docstrings("The bare Bert Model transformer outputting raw hidden-states without any specific head on top.",
+                      BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+class BertForLatentConnector(BertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            objective during Bert pretraining. This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertModel.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    def __init__(self, config, latent_size):
+        super(BertForLatentConnector, self).__init__(config)
+
+        self.embeddings = BertEmbeddings(config)
+        self.encoder = BertEncoder(config)
+        self.pooler = BertPooler(config)
+
+        self.linear = nn.Linear(config.hidden_size, 2 * latent_size, bias=False)
+
+        self.init_weights()
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        old_embeddings = self.embeddings.word_embeddings
+        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
+        self.embeddings.word_embeddings = new_embeddings
+        return self.embeddings.word_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            See base class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None):
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros_like(input_ids)
+
+        # We create a 3D attention mask from a 2D tensor mask.
+        # Sizes are [batch_size, 1, 1, to_seq_length]
+        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+        # this attention mask is more simple than the triangular masking of causal attention
+        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if head_mask is not None:
+            if head_mask.dim() == 1:
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.expand(self.config.num_hidden_layers, -1, -1, -1, -1)
+            elif head_mask.dim() == 2:
+                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+        else:
+            head_mask = [None] * self.config.num_hidden_layers
+
+        embedding_output = self.embeddings(input_ids, position_ids=position_ids, token_type_ids=token_type_ids)
+        encoder_outputs = self.encoder(embedding_output,
+                                       extended_attention_mask,
+                                       head_mask=head_mask)
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output)
+
+        outputs = (sequence_output, pooled_output,) + encoder_outputs[1:]  # add hidden_states and attentions if they are here
+        return outputs  # sequence_output, pooled_output, (hidden_states), (attentions)
+
+
+
+@add_start_docstrings("""Bert Model with two heads on top as done during the pre-training:
+    a `masked language modeling` head and a `next sentence prediction (classification)` head. """,
+    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+class BertForPreTraining(BertPreTrainedModel):
+    r"""
+        **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the masked language modeling loss.
+            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
+        **next_sentence_label**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair (see ``input_ids`` docstring)
+            Indices should be in ``[0, 1]``.
+            ``0`` indicates sequence B is a continuation of sequence A,
+            ``1`` indicates sequence B is a random sequence.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when both ``masked_lm_labels`` and ``next_sentence_label`` are provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Total loss as the sum of the masked language modeling loss and the next sequence prediction (classification) loss.
+        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **seq_relationship_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, 2)``
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForPreTraining.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        prediction_scores, seq_relationship_scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(BertForPreTraining, self).__init__(config)
+
+        self.bert = BertModel(config)
+        self.cls = BertPreTrainingHeads(config)
+
+        self.init_weights()
+        self.tie_weights()
+
+    def tie_weights(self):
+        """ Make sure we are sharing the input and output embeddings.
+            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
+        """
+        self._tie_or_clone_weights(self.cls.predictions.decoder,
+                                   self.bert.embeddings.word_embeddings)
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                masked_lm_labels=None, next_sentence_label=None):
+
+        outputs = self.bert(input_ids,
+                            attention_mask=attention_mask,
+                            token_type_ids=token_type_ids,
+                            position_ids=position_ids, 
+                            head_mask=head_mask)
+
+        sequence_output, pooled_output = outputs[:2]
+        prediction_scores, seq_relationship_score = self.cls(sequence_output, pooled_output)
+
+        outputs = (prediction_scores, seq_relationship_score,) + outputs[2:]  # add hidden states and attention if they are here
+
+        if masked_lm_labels is not None and next_sentence_label is not None:
+            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
+            next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
+            total_loss = masked_lm_loss + next_sentence_loss
+            outputs = (total_loss,) + outputs
+
+        return outputs  # (loss), prediction_scores, seq_relationship_score, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Bert Model with a `language modeling` head on top. """,
+    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+class BertForMaskedLM(BertPreTrainedModel):
+    r"""
+        **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the masked language modeling loss.
+            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``masked_lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Masked language modeling loss.
+        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForMaskedLM.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, masked_lm_labels=input_ids)
+        loss, prediction_scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(BertForMaskedLM, self).__init__(config)
+
+        self.bert = BertModel(config)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+        self.tie_weights()
+
+    def tie_weights(self):
+        """ Make sure we are sharing the input and output embeddings.
+            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
+        """
+        self._tie_or_clone_weights(self.cls.predictions.decoder,
+                                   self.bert.embeddings.word_embeddings)
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                masked_lm_labels=None):
+
+        outputs = self.bert(input_ids,
+                            attention_mask=attention_mask,
+                            token_type_ids=token_type_ids,
+                            position_ids=position_ids, 
+                            head_mask=head_mask)
+
+        sequence_output = outputs[0]
+        prediction_scores = self.cls(sequence_output)
+
+        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+        if masked_lm_labels is not None:
+            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
+            outputs = (masked_lm_loss,) + outputs
+
+        return outputs  # (masked_lm_loss), prediction_scores, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Bert Model with a `next sentence prediction (classification)` head on top. """,
+    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+class BertForNextSentencePrediction(BertPreTrainedModel):
+    r"""
+        **next_sentence_label**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair (see ``input_ids`` docstring)
+            Indices should be in ``[0, 1]``.
+            ``0`` indicates sequence B is a continuation of sequence A,
+            ``1`` indicates sequence B is a random sequence.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``next_sentence_label`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Next sequence prediction (classification) loss.
+        **seq_relationship_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, 2)``
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForNextSentencePrediction.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        seq_relationship_scores = outputs[0]
+
+    """
+    def __init__(self, config):
+        super(BertForNextSentencePrediction, self).__init__(config)
+
+        self.bert = BertModel(config)
+        self.cls = BertOnlyNSPHead(config)
+
+        self.init_weights()
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                next_sentence_label=None):
+
+        outputs = self.bert(input_ids,
+                            attention_mask=attention_mask,
+                            token_type_ids=token_type_ids,
+                            position_ids=position_ids, 
+                            head_mask=head_mask)
+
+        pooled_output = outputs[1]
+
+        seq_relationship_score = self.cls(pooled_output)
+
+        outputs = (seq_relationship_score,) + outputs[2:]  # add hidden states and attention if they are here
+        if next_sentence_label is not None:
+            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
+            outputs = (next_sentence_loss,) + outputs
+
+        return outputs  # (next_sentence_loss), seq_relationship_score, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Bert Model transformer with a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks. """,
+    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+class BertForSequenceClassification(BertPreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
+            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification (or regression if config.num_labels==1) loss.
+        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(BertForSequenceClassification, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
+        self.use_freeze = False
+
+        self.init_weights()
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, labels=None):
+
+        outputs = self.bert(input_ids,
+                            attention_mask=attention_mask,
+                            token_type_ids=token_type_ids,
+                            position_ids=position_ids, 
+                            head_mask=head_mask)
+
+        pooled_output = outputs[1]
+
+        if self.use_freeze:
+            pooled_output = pooled_output.detach()
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+        
+        # pdb.set_trace()
+        return outputs, pooled_output  # (loss), logits, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Bert Model transformer with a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks. """,
+    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+class BertForSequenceClassificationLatentConnector(BertPreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
+            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification (or regression if config.num_labels==1) loss.
+        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForSequenceClassificationLatentConnector.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
+
+    """
+    def __init__(self, config, latent_size):
+        super(BertForSequenceClassificationLatentConnector, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
+        self.linear = nn.Linear(config.hidden_size, 2 * latent_size, bias=False)
+        self.use_freeze = False
+
+        self.init_weights()
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, labels=None):
+
+        outputs = self.bert(input_ids,
+                            attention_mask=attention_mask,
+                            token_type_ids=token_type_ids,
+                            position_ids=position_ids, 
+                            head_mask=head_mask)
+
+
+        pooled_output = outputs[1]
+        # mean, logvar = self.linear(pooled_output).chunk(2, -1)
+
+        if self.use_freeze:
+            pooled_output = pooled_output.detach()
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs, pooled_output   # (loss), logits, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Bert Model with a multiple choice classification head on top (a linear layer on top of
+    the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
+    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+class BertForMultipleChoice(BertPreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for computing the multiple choice classification loss.
+            Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
+            of the input tensors. (see `input_ids` above)
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification loss.
+        **classification_scores**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices)`` where `num_choices` is the size of the second dimension
+            of the input tensors. (see `input_ids` above).
+            Classification scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForMultipleChoice.from_pretrained('bert-base-uncased')
+        choices = ["Hello, my dog is cute", "Hello, my cat is amazing"]
+        input_ids = torch.tensor([tokenizer.encode(s) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
+        labels = torch.tensor(1).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, classification_scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(BertForMultipleChoice, self).__init__(config)
+
+        self.bert = BertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        self.init_weights()
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, labels=None):
+        num_choices = input_ids.shape[1]
+
+        input_ids = input_ids.view(-1, input_ids.size(-1))
+        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+
+        outputs = self.bert(input_ids,
+                            attention_mask=attention_mask,
+                            token_type_ids=token_type_ids,
+                            position_ids=position_ids,
+                            head_mask=head_mask)
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        outputs = (reshaped_logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Bert Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+class BertForTokenClassification(BertPreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the token classification loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification loss.
+        **scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
+            Classification scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForTokenClassification.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1] * input_ids.size(1)).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(BertForTokenClassification, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None,
+                position_ids=None, head_mask=None, labels=None):
+
+        outputs = self.bert(input_ids,
+                            attention_mask=attention_mask,
+                            token_type_ids=token_type_ids,
+                            position_ids=position_ids, 
+                            head_mask=head_mask)
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)[active_loss]
+                active_labels = labels.view(-1)[active_loss]
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), scores, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Bert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
+    the hidden-states output to compute `span start logits` and `span end logits`). """,
+    BERT_START_DOCSTRING, BERT_INPUTS_DOCSTRING)
+class BertForQuestionAnswering(BertPreTrainedModel):
+    r"""
+        **start_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+        **end_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        **start_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
+            Span-start scores (before SoftMax).
+        **end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
+            Span-end scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForQuestionAnswering.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        start_positions = torch.tensor([1])
+        end_positions = torch.tensor([3])
+        outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
+        loss, start_scores, end_scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(BertForQuestionAnswering, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                start_positions=None, end_positions=None):
+
+        outputs = self.bert(input_ids,
+                            attention_mask=attention_mask,
+                            token_type_ids=token_type_ids,
+                            position_ids=position_ids, 
+                            head_mask=head_mask)
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = end_logits.squeeze(-1)
+
+        outputs = (start_logits, end_logits,) + outputs[2:]
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions.clamp_(0, ignored_index)
+            end_positions.clamp_(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+            outputs = (total_loss,) + outputs
+
+        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+
+
+############
+# XX Added #
+############
+
+class BertForLatentConnector_XX(nn.Module):
+    def __init__(self, config, latent_size):
+        super().__init__()
+        self.config = config
+        self.embeddings = BertEmbeddings(config)
+        self.encoder = BertEncoder(config)
+        self.pooler = BertPooler(config)
+        self.linear = nn.Linear(config.hidden_size, 2 * latent_size, bias=False)
+        self.init_weights()
+
+    def init_weights(self):
+        """ Initialize and prunes weights if needed. """
+        # Initialize weights
+        self.apply(self._init_weights)
+
+        # Prune heads if needed
+        if self.config.pruned_heads:
+            self.prune_heads(self.config.pruned_heads)
+
+    def _init_weights(self, module):
+        """ Initialize the weights """
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, BertLayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        old_embeddings = self.embeddings.word_embeddings
+        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
+        self.embeddings.word_embeddings = new_embeddings
+        return self.embeddings.word_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            See base class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None):
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros_like(input_ids)
+
+        # We create a 3D attention mask from a 2D tensor mask.
+        # Sizes are [batch_size, 1, 1, to_seq_length]
+        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+        # this attention mask is more simple than the triangular masking of causal attention
+        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if head_mask is not None:
+            if head_mask.dim() == 1:
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.expand(self.config.num_hidden_layers, -1, -1, -1, -1)
+            elif head_mask.dim() == 2:
+                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+        else:
+            head_mask = [None] * self.config.num_hidden_layers
+
+        embedding_output = self.embeddings(input_ids, position_ids=position_ids, token_type_ids=token_type_ids)
+        encoder_outputs = self.encoder(embedding_output,
+                                       extended_attention_mask,
+                                       head_mask=head_mask)
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output)
+
+        outputs = (sequence_output, pooled_output,) + encoder_outputs[1:]  # add hidden_states and attentions if they are here
+        return outputs  # sequence_output, pooled_output, (hidden_states), (attentions)
+
+

versatile_diffusion/lib/model_zoo/optimus_models/optimus_gpt2.py

@@ -0,0 +1,1122 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch OpenAI GPT-2 model."""
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import pdb
+
+import collections
+import json
+import logging
+import math
+import os
+import sys
+from io import open
+
+import torch
+import torch.nn as nn
+from torch.nn import CrossEntropyLoss
+from torch.nn.parameter import Parameter
+
+from .modeling_utils import PreTrainedModel, Conv1D, prune_conv1d_layer, SequenceSummary
+from .configuration_gpt2 import GPT2Config
+from .file_utils import add_start_docstrings
+
+logger = logging.getLogger(__name__)
+
+GPT2_PRETRAINED_MODEL_ARCHIVE_MAP = {"gpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-pytorch_model.bin",
+                                     "gpt2-medium": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-medium-pytorch_model.bin",
+                                     "gpt2-large": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-large-pytorch_model.bin"}
+
+def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
+    """ Load tf checkpoints in a pytorch model
+    """
+    try:
+        import re
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error("Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions.")
+        raise
+    tf_path = os.path.abspath(gpt2_checkpoint_path)
+    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info("Loading TF weight {} with shape {}".format(name, shape))
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array.squeeze())
+
+    for name, array in zip(names, arrays):
+        name = name[6:]  # skip "model/"
+        name = name.split('/')
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r'[A-Za-z]+\d+', m_name):
+                l = re.split(r'(\d+)', m_name)
+            else:
+                l = [m_name]
+            if l[0] == 'w' or l[0] == 'g':
+                pointer = getattr(pointer, 'weight')
+            elif l[0] == 'b':
+                pointer = getattr(pointer, 'bias')
+            elif l[0] == 'wpe' or l[0] == 'wte':
+                pointer = getattr(pointer, l[0])
+                pointer = getattr(pointer, 'weight')
+            else:
+                pointer = getattr(pointer, l[0])
+            if len(l) >= 2:
+                num = int(l[1])
+                pointer = pointer[num]
+        try:
+            assert pointer.shape == array.shape
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info("Initialize PyTorch weight {}".format(name))
+        pointer.data = torch.from_numpy(array)
+    return model
+
+
+def gelu(x):
+    return 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
+
+
+class Attention(nn.Module):
+    def __init__(self, nx, n_ctx, config, scale=False):
+        super(Attention, self).__init__()
+        self.output_attentions = config.output_attentions
+
+        n_state = nx  # in Attention: n_state=768 (nx=n_embd)
+        # [switch nx => n_state from Block to Attention to keep identical to TF implem]
+        assert n_state % config.n_head == 0
+        self.register_buffer("bias", torch.tril(torch.ones(n_ctx, n_ctx)).view(1, 1, n_ctx, n_ctx))
+        self.n_head = config.n_head
+        self.split_size = n_state
+        self.scale = scale
+
+        self.c_attn = Conv1D(n_state * 3, nx)
+        self.c_proj = Conv1D(n_state, nx)
+        self.attn_dropout = nn.Dropout(config.attn_pdrop)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        mask = torch.ones(self.n_head, self.split_size // self.n_head)
+        heads = set(heads) - self.pruned_heads  # Convert to set and emove already pruned heads
+        for head in heads:
+            # Compute how many pruned heads are before the head and move the index accordingly
+            head = head - sum(1 if h < head else 0 for h in self.pruned_heads)
+            mask[head] = 0
+        mask = mask.view(-1).contiguous().eq(1)
+        index = torch.arange(len(mask))[mask].long()
+        index_attn = torch.cat([index, index + self.split_size, index + (2*self.split_size)])
+
+        # Prune conv1d layers
+        self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
+        self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
+
+        # Update hyper params
+        self.split_size = (self.split_size // self.n_head) * (self.n_head - len(heads))
+        self.n_head = self.n_head - len(heads)
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def _attn(self, q, k, v, attention_mask=None, head_mask=None):
+        w = torch.matmul(q, k)
+        if self.scale:
+            w = w / math.sqrt(v.size(-1))
+        nd, ns = w.size(-2), w.size(-1)
+        b = self.bias[:, :, ns-nd:ns, :ns]
+        w = w * b - 1e4 * (1 - b)
+
+        if attention_mask is not None:
+            # Apply the attention mask
+            w = w + attention_mask
+
+        w = nn.Softmax(dim=-1)(w)
+        w = self.attn_dropout(w)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            w = w * head_mask
+
+        outputs = [torch.matmul(w, v)]
+        if self.output_attentions:
+            outputs.append(w)
+        return outputs
+
+    def merge_heads(self, x):
+        x = x.permute(0, 2, 1, 3).contiguous()
+        new_x_shape = x.size()[:-2] + (x.size(-2) * x.size(-1),)
+        return x.view(*new_x_shape)  # in Tensorflow implem: fct merge_states
+
+    def split_heads(self, x, k=False):
+        new_x_shape = x.size()[:-1] + (self.n_head, x.size(-1) // self.n_head)
+        x = x.view(*new_x_shape)  # in Tensorflow implem: fct split_states
+        if k:
+            return x.permute(0, 2, 3, 1)  # (batch, head, head_features, seq_length)
+        else:
+            return x.permute(0, 2, 1, 3)  # (batch, head, seq_length, head_features)
+
+    def forward(self, x, layer_past=None, attention_mask=None, head_mask=None):
+        x = self.c_attn(x)
+        query, key, value = x.split(self.split_size, dim=2)
+        query = self.split_heads(query)
+        key = self.split_heads(key, k=True)
+        value = self.split_heads(value)
+
+        
+        if layer_past is not None:
+            past_key, past_value = layer_past[0], layer_past[1]  # transpose back cf below
+            
+            past_key = self.split_heads(past_key, k=True)
+            past_value = self.split_heads(past_value)
+            # pdb.set_trace()
+            key = torch.cat((past_key, key), dim=-1)
+            value = torch.cat((past_value, value), dim=-2)
+        present = torch.stack((key.transpose(-2, -1), value))  # transpose to have same shapes for stacking
+
+        attn_outputs = self._attn(query, key, value, attention_mask, head_mask)
+        a = attn_outputs[0]
+
+        a = self.merge_heads(a)
+        a = self.c_proj(a)
+        a = self.resid_dropout(a)
+
+        outputs = [a, present] + attn_outputs[1:]
+        return outputs  # a, present, (attentions)
+
+
+class MLP(nn.Module):
+    def __init__(self, n_state, config):  # in MLP: n_state=3072 (4 * n_embd)
+        super(MLP, self).__init__()
+        nx = config.n_embd
+        self.c_fc = Conv1D(n_state, nx)
+        self.c_proj = Conv1D(nx, n_state)
+        self.act = gelu
+        self.dropout = nn.Dropout(config.resid_pdrop)
+
+    def forward(self, x):
+        h = self.act(self.c_fc(x))
+        h2 = self.c_proj(h)
+        return self.dropout(h2)
+
+
+class Block(nn.Module):
+    def __init__(self, n_ctx, config, scale=False):
+        super(Block, self).__init__()
+        nx = config.n_embd
+        self.ln_1 = nn.LayerNorm(nx, eps=config.layer_norm_epsilon)
+        self.attn = Attention(nx, n_ctx, config, scale)
+        self.ln_2 = nn.LayerNorm(nx, eps=config.layer_norm_epsilon)
+        self.mlp = MLP(4 * nx, config)
+
+    def forward(self, x, layer_past=None, attention_mask=None, head_mask=None):
+        output_attn = self.attn(self.ln_1(x),
+                                layer_past=layer_past,
+                                attention_mask=attention_mask,
+                                head_mask=head_mask)
+        a = output_attn[0]  # output_attn: a, present, (attentions)
+
+        x = x + a
+        m = self.mlp(self.ln_2(x))
+        x = x + m
+
+        outputs = [x] + output_attn[1:]
+        return outputs  # x, present, (attentions)
+
+
+class GPT2PreTrainedModel(PreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = GPT2Config
+    pretrained_model_archive_map = GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
+    load_tf_weights = load_tf_weights_in_gpt2
+    base_model_prefix = "transformer"
+
+    def __init__(self, *inputs, **kwargs):
+        super(GPT2PreTrainedModel, self).__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module):
+        """ Initialize the weights.
+        """
+        if isinstance(module, (nn.Linear, nn.Embedding, Conv1D)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if isinstance(module, (nn.Linear, Conv1D)) and module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+GPT2_START_DOCSTRING = r"""    OpenAI GPT-2 model was proposed in
+    `Language Models are Unsupervised Multitask Learners`_
+    by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
+    It's a causal (unidirectional) transformer pre-trained using  language modeling on a very large
+    corpus of ~40 GB of text data.
+
+    This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
+    refer to the PyTorch documentation for all matter related to general usage and behavior.
+
+    .. _`Language Models are Unsupervised Multitask Learners`:
+        https://openai.com/blog/better-language-models/
+
+    .. _`torch.nn.Module`:
+        https://pytorch.org/docs/stable/nn.html#module
+
+    Parameters:
+        config (:class:`~pytorch_transformers.GPT2Config`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~pytorch_transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+GPT2_INPUTS_DOCSTRING = r"""    Inputs:
+        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            GPT-2 is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
+            Indices can be obtained using :class:`pytorch_transformers.GPT2Tokenizer`.
+            See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
+            :func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **past**:
+            list of ``torch.FloatTensor`` (one for each layer):
+            that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
+            (see `past` output below). Can be used to speed up sequential decoding.
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            A parallel sequence of tokens (can be used to indicate various portions of the inputs).
+            The embeddings from these tokens will be summed with the respective token embeddings.
+            Indices are selected in the vocabulary (unlike BERT which has a specific vocabulary for segment indices).
+        **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+"""
+
+@add_start_docstrings("The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
+                      GPT2_START_DOCSTRING, GPT2_INPUTS_DOCSTRING)
+class GPT2Model(GPT2PreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the last layer of the model.
+        **past**:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            that contains pre-computed hidden-states (key and values in the attention blocks).
+            Can be used (see `past` input) to speed up sequential decoding.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+        model = GPT2Model.from_pretrained('gpt2')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    def __init__(self, config):
+        super(GPT2Model, self).__init__(config)
+        self.output_hidden_states = config.output_hidden_states
+        self.output_attentions = config.output_attentions
+
+        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
+        self.wpe = nn.Embedding(config.n_positions, config.n_embd)
+        self.drop = nn.Dropout(config.embd_pdrop)
+        self.h = nn.ModuleList([Block(config.n_ctx, config, scale=True) for _ in range(config.n_layer)])
+        self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+
+        try:
+            self.latent_size = config.latent_size
+        except: 
+            self.latent_size = 32 # default size is 32
+
+        self.linear = nn.Linear(self.latent_size, config.hidden_size * config.n_layer, bias=False) # different latent vector for each layer 
+        self.linear_emb = nn.Linear(self.latent_size, config.hidden_size, bias=False) # share the same latent vector as the embeddings
+
+        self.config = config
+        self.init_weights()
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        self.wte = self._get_resized_embeddings(self.wte, new_num_tokens)
+        return self.wte
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+        """
+        for layer, heads in heads_to_prune.items():
+            self.h[layer].attn.prune_heads(heads)
+
+    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, latent_as_gpt_emb=False, latent_as_gpt_memory=True):
+
+        if past is None:
+            past_length = 0
+            past = [None] * len(self.h)
+        else:
+            
+
+            if latent_as_gpt_emb:
+                past_emb = self.linear_emb(past) # used as embeddings to add on other three embeddings
+
+            if latent_as_gpt_memory:
+                past = self.linear(past)
+                share_latent = False
+                if share_latent: 
+                    # the same latent vector shared by all layers
+                    past = [past.unsqueeze(-2), past.unsqueeze(-2)] # query, key
+                    past = [past] * len(self.h)
+                    past_length = past[0][0].size(-2)
+                else: 
+                    # different latent vectors for each layer
+                    past_split = torch.split(past.unsqueeze(1), self.config.hidden_size, dim=2)
+                    past = list(zip(past_split,past_split))
+                    
+                    # past = past.view(batch_size,len(self.h),-1)
+                    # past = [[past[:,i,:].unsqueeze(-2), past[:,i,:].unsqueeze(-2) ] for i in range(len(self.h))]
+                    past_length = 1 # past[0][0].size(-2)
+            else:
+                past_length = 0
+                past = [None] * len(self.h)
+
+
+        if position_ids is None:
+            position_ids = torch.arange(past_length, input_ids.size(-1) + past_length, dtype=torch.long, device=input_ids.device)
+            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
+
+
+        # Attention mask.
+        if attention_mask is not None:
+            # We create a 3D attention mask from a 2D tensor mask.
+            # Sizes are [batch_size, 1, 1, to_seq_length]
+            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+            # this attention mask is more simple than the triangular masking of causal attention
+            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+            attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+
+            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+            # masked positions, this operation will create a tensor which is 0.0 for
+            # positions we want to attend and -10000.0 for masked positions.
+            # Since we are adding it to the raw scores before the softmax, this is
+            # effectively the same as removing these entirely.
+            attention_mask = attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
+            attention_mask = (1.0 - attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # head_mask has shape n_layer x batch x n_heads x N x N
+        if head_mask is not None:
+            if head_mask.dim() == 1:
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.expand(self.config.n_layer, -1, -1, -1, -1)
+            elif head_mask.dim() == 2:
+                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+        else:
+            head_mask = [None] * self.config.n_layer
+
+        
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_ids.size(-1))
+        position_ids = position_ids.view(-1, position_ids.size(-1))
+
+
+        inputs_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1))
+            token_type_embeds = self.wte(token_type_ids)
+        else:
+            token_type_embeds = 0
+
+
+        hidden_states = inputs_embeds + position_embeds + token_type_embeds
+        if latent_as_gpt_emb:
+            # pdb.set_trace()
+            hidden_states = hidden_states + past_emb.unsqueeze(1)
+
+        hidden_states = self.drop(hidden_states)
+
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        presents = ()
+        all_attentions = []
+        all_hidden_states = ()
+        for i, (block, layer_past) in enumerate(zip(self.h, past)):
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states.view(*output_shape),)
+
+            
+            outputs = block(hidden_states,
+                            layer_past=layer_past,
+                            attention_mask=attention_mask,
+                            head_mask=head_mask[i])
+
+            
+            hidden_states, present = outputs[:2]
+            presents = presents + (present,)
+
+            if self.output_attentions:
+                all_attentions.append(outputs[2])
+
+        hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = hidden_states.view(*output_shape)
+        # Add last hidden state
+        if self.output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states, presents)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            # let the number of heads free (-1) so we can extract attention even after head pruning
+            attention_output_shape = input_shape[:-1] + (-1,) + all_attentions[0].shape[-2:]
+            all_attentions = tuple(t.view(*attention_output_shape) for t in all_attentions)
+            outputs = outputs + (all_attentions,)
+        return outputs  # last hidden state, presents, (all hidden_states), (attentions)
+
+
+@add_start_docstrings("""The GPT2 Model transformer with a language modeling head on top
+(linear layer with weights tied to the input embeddings). """, GPT2_START_DOCSTRING, GPT2_INPUTS_DOCSTRING)
+class GPT2LMHeadModel(GPT2PreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for language modeling.
+            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
+            Indices are selected in ``[-1, 0, ..., config.vocab_size]``
+            All labels set to ``-1`` are ignored (masked), the loss is only
+            computed for labels in ``[0, ..., config.vocab_size]``
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Language modeling loss.
+        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **past**:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            that contains pre-computed hidden-states (key and values in the attention blocks).
+            Can be used (see `past` input) to speed up sequential decoding.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import torch
+        from pytorch_transformers import GPT2Tokenizer, GPT2LMHeadModel
+
+        tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+        model = GPT2LMHeadModel.from_pretrained('gpt2')
+
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=input_ids)
+        loss, logits = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(GPT2LMHeadModel, self).__init__(config)
+        self.transformer = GPT2Model(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+        self.init_weights()
+        self.tie_weights()
+
+
+    def tie_weights(self):
+        """ Make sure we are sharing the input and output embeddings.
+            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
+        """
+        self._tie_or_clone_weights(self.lm_head,
+                                   self.transformer.wte)
+
+    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                labels=None, label_ignore=None):
+        transformer_outputs = self.transformer(input_ids,
+                                               past=past,
+                                               attention_mask=attention_mask,
+                                               token_type_ids=token_type_ids,
+                                               position_ids=position_ids,
+                                               head_mask=head_mask)
+        hidden_states = transformer_outputs[0]
+
+        lm_logits = self.lm_head(hidden_states)
+
+        outputs = (lm_logits,) + transformer_outputs[1:]
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss(ignore_index=label_ignore, reduce=False) # 50258 is the padding id, otherwise -1 is used for masked LM.
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
+                            shift_labels.view(-1))
+            loss = torch.sum(loss.view(-1, shift_labels.shape[-1]), -1)
+            outputs = (loss,) + outputs
+
+
+        return outputs  # (loss), lm_logits, presents, (all hidden_states), (attentions)
+
+
+
+@add_start_docstrings("""The GPT2 Model transformer with a language modeling head on top
+(linear layer with weights tied to the input embeddings). """, GPT2_START_DOCSTRING, GPT2_INPUTS_DOCSTRING)
+class GPT2ForLatentConnector(GPT2PreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for language modeling.
+            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
+            Indices are selected in ``[-1, 0, ..., config.vocab_size]``
+            All labels set to ``-1`` are ignored (masked), the loss is only
+            computed for labels in ``[0, ..., config.vocab_size]``
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Language modeling loss.
+        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **past**:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            that contains pre-computed hidden-states (key and values in the attention blocks).
+            Can be used (see `past` input) to speed up sequential decoding.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import torch
+        from pytorch_transformers import GPT2Tokenizer, GPT2LMHeadModel
+
+        tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+        model = GPT2LMHeadModel.from_pretrained('gpt2')
+
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=input_ids)
+        loss, logits = outputs[:2]
+
+    """
+    def __init__(self, config, latent_size=32, latent_as_gpt_emb=True, latent_as_gpt_memory=True):
+        
+        super(GPT2ForLatentConnector, self).__init__(config)
+
+        
+        self.transformer = GPT2Model(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+        self.init_weights()
+        self.tie_weights()
+
+        self.latent_as_gpt_emb = latent_as_gpt_emb
+        self.latent_as_gpt_memory = latent_as_gpt_memory
+
+
+
+    def tie_weights(self):
+        """ Make sure we are sharing the input and output embeddings.
+            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
+        """
+        self._tie_or_clone_weights(self.lm_head,
+                                   self.transformer.wte)
+
+    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                labels=None, label_ignore=None):
+
+
+        transformer_outputs = self.transformer(input_ids,
+                                               past=past,
+                                               attention_mask=attention_mask,
+                                               token_type_ids=token_type_ids,
+                                               position_ids=position_ids,
+                                               head_mask=head_mask, 
+                                               latent_as_gpt_emb=self.latent_as_gpt_emb,
+                                               latent_as_gpt_memory=self.latent_as_gpt_memory)
+        hidden_states = transformer_outputs[0]
+
+        lm_logits = self.lm_head(hidden_states)
+
+        outputs = (lm_logits,) + transformer_outputs[1:]
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss(ignore_index=label_ignore, reduce=False) # 50258 is the padding id, otherwise -1 is used for masked LM.
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
+                            shift_labels.view(-1))
+            loss = torch.sum(loss.view(-1, shift_labels.shape[-1]), -1)
+            outputs = (loss,) + outputs
+
+
+        return outputs  # (loss), lm_logits, presents, (all hidden_states), (attentions)
+
+@add_start_docstrings("""The GPT2 Model transformer with a language modeling and a multiple-choice classification
+head on top e.g. for RocStories/SWAG tasks. The two heads are two linear layers.
+The language modeling head has its weights tied to the input embeddings,
+the classification head takes as input the input of a specified classification token index in the input sequence).
+""", GPT2_START_DOCSTRING, GPT2_INPUTS_DOCSTRING)
+class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
+    r"""
+        **mc_token_ids**: (`optional`, default to index of the last token of the input) ``torch.LongTensor`` of shape ``(batch_size, num_choices)``:
+            Index of the classification token in each input sequence.
+            Selected in the range ``[0, input_ids.size(-1) - 1[``.
+        **lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for language modeling.
+            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
+            Indices are selected in ``[-1, 0, ..., config.vocab_size]``
+            All labels set to ``-1`` are ignored (masked), the loss is only
+            computed for labels in ``[0, ..., config.vocab_size]``
+        **mc_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size)``:
+            Labels for computing the multiple choice classification loss.
+            Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
+            of the input tensors. (see `input_ids` above)
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **lm_loss**: (`optional`, returned when ``lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Language modeling loss.
+        **mc_loss**: (`optional`, returned when ``multiple_choice_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Multiple choice classification loss.
+        **lm_prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **mc_prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices)``
+            Prediction scores of the multiplechoice classification head (scores for each choice before SoftMax).
+        **past**:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            that contains pre-computed hidden-states (key and values in the attention blocks).
+            Can be used (see `past` input) to speed up sequential decoding.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import torch
+        from pytorch_transformers import GPT2Tokenizer, GPT2DoubleHeadsModel
+        
+        tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+        model = GPT2DoubleHeadsModel.from_pretrained('gpt2')
+        
+        # Add a [CLS] to the vocabulary (we should train it also!)
+        tokenizer.add_special_tokens({'cls_token': '[CLS]'})
+        model.resize_token_embeddings(len(tokenizer))  # Update the model embeddings with the new vocabulary size
+        print(tokenizer.cls_token_id, len(tokenizer))  # The newly token the last token of the vocabulary
+        
+        choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]
+        encoded_choices = [tokenizer.encode(s) for s in choices]
+        cls_token_location = [tokens.index(tokenizer.cls_token_id) for tokens in encoded_choices]
+
+        input_ids = torch.tensor(encoded_choices).unsqueeze(0)  # Batch size: 1, number of choices: 2
+        mc_token_ids = torch.tensor([cls_token_location])  # Batch size: 1
+
+        outputs = model(input_ids, mc_token_ids=mc_token_ids)
+        lm_prediction_scores, mc_prediction_scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(GPT2DoubleHeadsModel, self).__init__(config)
+        self.transformer = GPT2Model(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        self.multiple_choice_head = SequenceSummary(config)
+
+        self.init_weights()
+        self.tie_weights()
+
+    def tie_weights(self):
+        """ Make sure we are sharing the input and output embeddings.
+            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
+        """
+        self._tie_or_clone_weights(self.lm_head,
+                                   self.transformer.wte)
+
+    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                mc_token_ids=None, lm_labels=None, mc_labels=None):
+        transformer_outputs = self.transformer(input_ids,
+                                               past=past,
+                                               attention_mask=attention_mask,
+                                               token_type_ids=token_type_ids,
+                                               position_ids=position_ids,
+                                               head_mask=head_mask)
+
+        hidden_states = transformer_outputs[0]
+
+        lm_logits = self.lm_head(hidden_states)
+        mc_logits = self.multiple_choice_head(hidden_states, mc_token_ids).squeeze(-1)
+
+        outputs = (lm_logits, mc_logits) + transformer_outputs[1:]
+        if mc_labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(mc_logits.view(-1, mc_logits.size(-1)),
+                            mc_labels.view(-1))
+            outputs = (loss,) + outputs
+        if lm_labels is not None:
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = lm_labels[..., 1:].contiguous()
+            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
+                            shift_labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (lm loss), (mc loss), lm logits, mc logits, presents, (all hidden_states), (attentions)
+
+############
+# XX Added #
+############
+
+class GPT2Model_XX(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.output_hidden_states = config.output_hidden_states
+        self.output_attentions = config.output_attentions
+
+        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
+        self.wpe = nn.Embedding(config.n_positions, config.n_embd)
+        self.drop = nn.Dropout(config.embd_pdrop)
+        self.h = nn.ModuleList([Block(config.n_ctx, config, scale=True) for _ in range(config.n_layer)])
+        self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+
+        try:
+            self.latent_size = config.latent_size
+        except: 
+            self.latent_size = 32 # default size is 32
+
+        self.linear = nn.Linear(self.latent_size, config.hidden_size * config.n_layer, bias=False) # different latent vector for each layer 
+        self.linear_emb = nn.Linear(self.latent_size, config.hidden_size, bias=False) # share the same latent vector as the embeddings
+
+        self.config = config
+        self.init_weights()
+
+    def init_weights(self):
+        """ Initialize and prunes weights if needed. """
+        # Initialize weights
+        self.apply(self._init_weights)
+
+        # Prune heads if needed
+        if self.config.pruned_heads:
+            self.prune_heads(self.config.pruned_heads)
+
+    def _init_weights(self, module):
+        """ Initialize the weights.
+        """
+        if isinstance(module, (nn.Linear, nn.Embedding, Conv1D)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if isinstance(module, (nn.Linear, Conv1D)) and module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        self.wte = self._get_resized_embeddings(self.wte, new_num_tokens)
+        return self.wte
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+        """
+        for layer, heads in heads_to_prune.items():
+            self.h[layer].attn.prune_heads(heads)
+
+    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, latent_as_gpt_emb=False, latent_as_gpt_memory=True):
+        if past is None:
+            past_length = 0
+            past = [None] * len(self.h)
+        else:
+            if latent_as_gpt_emb:
+                past_emb = self.linear_emb(past) # used as embeddings to add on other three embeddings
+
+            if latent_as_gpt_memory:
+                past = self.linear(past)
+                share_latent = False
+                if share_latent: 
+                    # the same latent vector shared by all layers
+                    past = [past.unsqueeze(-2), past.unsqueeze(-2)] # query, key
+                    past = [past] * len(self.h)
+                    past_length = past[0][0].size(-2)
+                else: 
+                    # different latent vectors for each layer
+                    past_split = torch.split(past.unsqueeze(1), self.config.hidden_size, dim=2)
+                    past = list(zip(past_split,past_split))
+                    
+                    # past = past.view(batch_size,len(self.h),-1)
+                    # past = [[past[:,i,:].unsqueeze(-2), past[:,i,:].unsqueeze(-2) ] for i in range(len(self.h))]
+                    past_length = 1 # past[0][0].size(-2)
+            else:
+                past_length = 0
+                past = [None] * len(self.h)
+
+
+        if position_ids is None:
+            position_ids = torch.arange(past_length, input_ids.size(-1) + past_length, dtype=torch.long, device=input_ids.device)
+            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
+
+
+        # Attention mask.
+        if attention_mask is not None:
+            # We create a 3D attention mask from a 2D tensor mask.
+            # Sizes are [batch_size, 1, 1, to_seq_length]
+            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+            # this attention mask is more simple than the triangular masking of causal attention
+            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+            attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+
+            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+            # masked positions, this operation will create a tensor which is 0.0 for
+            # positions we want to attend and -10000.0 for masked positions.
+            # Since we are adding it to the raw scores before the softmax, this is
+            # effectively the same as removing these entirely.
+            attention_mask = attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
+            attention_mask = (1.0 - attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # head_mask has shape n_layer x batch x n_heads x N x N
+        if head_mask is not None:
+            if head_mask.dim() == 1:
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.expand(self.config.n_layer, -1, -1, -1, -1)
+            elif head_mask.dim() == 2:
+                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+        else:
+            head_mask = [None] * self.config.n_layer
+
+        
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_ids.size(-1))
+        position_ids = position_ids.view(-1, position_ids.size(-1))
+
+
+        inputs_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1))
+            token_type_embeds = self.wte(token_type_ids)
+        else:
+            token_type_embeds = 0
+
+
+        hidden_states = inputs_embeds + position_embeds + token_type_embeds
+        if latent_as_gpt_emb:
+            # pdb.set_trace()
+            hidden_states = hidden_states + past_emb.unsqueeze(1)
+
+        hidden_states = self.drop(hidden_states)
+
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        presents = ()
+        all_attentions = []
+        all_hidden_states = ()
+        for i, (block, layer_past) in enumerate(zip(self.h, past)):
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states.view(*output_shape),)
+
+            
+            outputs = block(hidden_states,
+                            layer_past=layer_past,
+                            attention_mask=attention_mask,
+                            head_mask=head_mask[i])
+
+            
+            hidden_states, present = outputs[:2]
+            presents = presents + (present,)
+
+            if self.output_attentions:
+                all_attentions.append(outputs[2])
+
+        hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = hidden_states.view(*output_shape)
+        # Add last hidden state
+        if self.output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states, presents)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            # let the number of heads free (-1) so we can extract attention even after head pruning
+            attention_output_shape = input_shape[:-1] + (-1,) + all_attentions[0].shape[-2:]
+            all_attentions = tuple(t.view(*attention_output_shape) for t in all_attentions)
+            outputs = outputs + (all_attentions,)
+        return outputs  # last hidden state, presents, (all hidden_states), (attentions)
+
+    def _get_resized_embeddings(self, old_embeddings, new_num_tokens=None):
+        """ Build a resized Embedding Module from a provided token Embedding Module.
+            Increasing the size will add newly initialized vectors at the end
+            Reducing the size will remove vectors from the end
+
+        Args:
+            new_num_tokens: (`optional`) int
+                New number of tokens in the embedding matrix.
+                Increasing the size will add newly initialized vectors at the end
+                Reducing the size will remove vectors from the end
+                If not provided or None: return the provided token Embedding Module.
+        Return: ``torch.nn.Embeddings``
+            Pointer to the resized Embedding Module or the old Embedding Module if new_num_tokens is None
+        """
+        if new_num_tokens is None:
+            return old_embeddings
+        old_num_tokens, old_embedding_dim = old_embeddings.weight.size()
+        if old_num_tokens == new_num_tokens:
+            return old_embeddings
+        # Build new embeddings
+        new_embeddings = nn.Embedding(new_num_tokens, old_embedding_dim)
+        new_embeddings.to(old_embeddings.weight.device)
+        # initialize all new embeddings (in particular added tokens)
+        self._init_weights(new_embeddings)
+        # Copy word embeddings from the previous weights
+        num_tokens_to_copy = min(old_num_tokens, new_num_tokens)
+        new_embeddings.weight.data[:num_tokens_to_copy, :] = old_embeddings.weight.data[:num_tokens_to_copy, :]
+        return new_embeddings
+
+class GPT2ForLatentConnector_XX(nn.Module):
+    def __init__(self, 
+                 config, 
+                 latent_size=32, 
+                 latent_as_gpt_emb=True, 
+                 latent_as_gpt_memory=True):
+        
+        super().__init__()
+        self.config = config
+        self.transformer = GPT2Model_XX(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        self.init_weights()
+        self.tie_weights()
+        self.latent_as_gpt_emb = latent_as_gpt_emb
+        self.latent_as_gpt_memory = latent_as_gpt_memory
+
+    def init_weights(self):
+        """ Initialize and prunes weights if needed. """
+        # Initialize weights
+        self.apply(self._init_weights)
+
+        # Prune heads if needed
+        if self.config.pruned_heads:
+            self.prune_heads(self.config.pruned_heads)
+
+    def _init_weights(self, module):
+        """ Initialize the weights.
+        """
+        if isinstance(module, (nn.Linear, nn.Embedding, Conv1D)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if isinstance(module, (nn.Linear, Conv1D)) and module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def _tie_or_clone_weights(self, first_module, second_module):
+        """ Tie or clone module weights depending of weither we are using TorchScript or not
+        """
+        if self.config.torchscript:
+            first_module.weight = nn.Parameter(second_module.weight.clone())
+        else:
+            first_module.weight = second_module.weight
+
+        if hasattr(first_module, 'bias') and first_module.bias is not None:
+            first_module.bias.data = torch.nn.functional.pad(
+                first_module.bias.data,
+                (0, first_module.weight.shape[0] - first_module.bias.shape[0]),
+                'constant', 0,)
+
+    def tie_weights(self):
+        """ Make sure we are sharing the input and output embeddings.
+            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
+        """
+        self._tie_or_clone_weights(self.lm_head,
+                                   self.transformer.wte)
+
+    def forward(self, input_ids, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
+                labels=None, label_ignore=None):
+
+
+        transformer_outputs = self.transformer(input_ids,
+                                               past=past,
+                                               attention_mask=attention_mask,
+                                               token_type_ids=token_type_ids,
+                                               position_ids=position_ids,
+                                               head_mask=head_mask, 
+                                               latent_as_gpt_emb=self.latent_as_gpt_emb,
+                                               latent_as_gpt_memory=self.latent_as_gpt_memory)
+        hidden_states = transformer_outputs[0]
+
+        lm_logits = self.lm_head(hidden_states)
+
+        outputs = (lm_logits,) + transformer_outputs[1:]
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss(ignore_index=label_ignore, reduce=False) # 50258 is the padding id, otherwise -1 is used for masked LM.
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
+                            shift_labels.view(-1))
+            loss = torch.sum(loss.view(-1, shift_labels.shape[-1]), -1)
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), lm_logits, presents, (all hidden_states), (attentions)
+
+    def resize_token_embeddings(self, new_num_tokens=None):
+        model_embeds = self.transformer._resize_token_embeddings(new_num_tokens)
+        if new_num_tokens is None:
+            return model_embeds
+        self.config.vocab_size = new_num_tokens
+        self.transformer.vocab_size = new_num_tokens
+        if hasattr(self, 'tie_weights'):
+            self.tie_weights()
+        return model_embeds

versatile_diffusion/lib/model_zoo/optimus_models/tokenization_bert.py

@@ -0,0 +1,457 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes."""
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import collections
+import logging
+import os
+import unicodedata
+from io import open
+
+from .tokenization_utils import PreTrainedTokenizer
+
+logger = logging.getLogger(__name__)
+
+VOCAB_FILES_NAMES = {'vocab_file': 'vocab.txt'}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    'vocab_file':
+    {
+        'bert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-vocab.txt",
+        'bert-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-vocab.txt",
+        'bert-base-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-vocab.txt",
+        'bert-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-vocab.txt",
+        'bert-base-multilingual-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased-vocab.txt",
+        'bert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased-vocab.txt",
+        'bert-base-chinese': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese-vocab.txt",
+        'bert-base-german-cased': "https://int-deepset-models-bert.s3.eu-central-1.amazonaws.com/pytorch/bert-base-german-cased-vocab.txt",
+        'bert-large-uncased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-vocab.txt",
+        'bert-large-cased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-vocab.txt",
+        'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-vocab.txt",
+        'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-vocab.txt",
+        'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-vocab.txt",
+    }
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    'bert-base-uncased': 512,
+    'bert-large-uncased': 512,
+    'bert-base-cased': 512,
+    'bert-large-cased': 512,
+    'bert-base-multilingual-uncased': 512,
+    'bert-base-multilingual-cased': 512,
+    'bert-base-chinese': 512,
+    'bert-base-german-cased': 512,
+    'bert-large-uncased-whole-word-masking': 512,
+    'bert-large-cased-whole-word-masking': 512,
+    'bert-large-uncased-whole-word-masking-finetuned-squad': 512,
+    'bert-large-cased-whole-word-masking-finetuned-squad': 512,
+    'bert-base-cased-finetuned-mrpc': 512,
+}
+
+PRETRAINED_INIT_CONFIGURATION = {
+    'bert-base-uncased': {'do_lower_case': True},
+    'bert-large-uncased': {'do_lower_case': True},
+    'bert-base-cased': {'do_lower_case': False},
+    'bert-large-cased': {'do_lower_case': False},
+    'bert-base-multilingual-uncased': {'do_lower_case': True},
+    'bert-base-multilingual-cased': {'do_lower_case': False},
+    'bert-base-chinese': {'do_lower_case': False},
+    'bert-base-german-cased': {'do_lower_case': False},
+    'bert-large-uncased-whole-word-masking': {'do_lower_case': True},
+    'bert-large-cased-whole-word-masking': {'do_lower_case': False},
+    'bert-large-uncased-whole-word-masking-finetuned-squad': {'do_lower_case': True},
+    'bert-large-cased-whole-word-masking-finetuned-squad': {'do_lower_case': False},
+    'bert-base-cased-finetuned-mrpc': {'do_lower_case': False},
+}
+
+
+def load_vocab(vocab_file):
+    """Loads a vocabulary file into a dictionary."""
+    vocab = collections.OrderedDict()
+    with open(vocab_file, "r", encoding="utf-8") as reader:
+        tokens = reader.readlines()
+    for index, token in enumerate(tokens):
+        token = token.rstrip('\n')
+        vocab[token] = index
+    return vocab
+
+
+def whitespace_tokenize(text):
+    """Runs basic whitespace cleaning and splitting on a piece of text."""
+    text = text.strip()
+    if not text:
+        return []
+    tokens = text.split()
+    return tokens
+
+
+class BertTokenizer(PreTrainedTokenizer):
+    r"""
+    Constructs a BertTokenizer.
+    :class:`~pytorch_transformers.BertTokenizer` runs end-to-end tokenization: punctuation splitting + wordpiece
+
+    Args:
+        vocab_file: Path to a one-wordpiece-per-line vocabulary file
+        do_lower_case: Whether to lower case the input. Only has an effect when do_wordpiece_only=False
+        do_basic_tokenize: Whether to do basic tokenization before wordpiece.
+        max_len: An artificial maximum length to truncate tokenized sequences to; Effective maximum length is always the
+            minimum of this value (if specified) and the underlying BERT model's sequence length.
+        never_split: List of tokens which will never be split during tokenization. Only has an effect when
+            do_wordpiece_only=False
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+
+    def __init__(self, vocab_file, do_lower_case=True, do_basic_tokenize=True, never_split=None,
+                 unk_token="[UNK]", sep_token="[SEP]", pad_token="[PAD]", cls_token="[CLS]",
+                 mask_token="[MASK]", tokenize_chinese_chars=True, **kwargs):
+        """Constructs a BertTokenizer.
+
+        Args:
+            **vocab_file**: Path to a one-wordpiece-per-line vocabulary file
+            **do_lower_case**: (`optional`) boolean (default True)
+                Whether to lower case the input
+                Only has an effect when do_basic_tokenize=True
+            **do_basic_tokenize**: (`optional`) boolean (default True)
+                Whether to do basic tokenization before wordpiece.
+            **never_split**: (`optional`) list of string
+                List of tokens which will never be split during tokenization.
+                Only has an effect when do_basic_tokenize=True
+            **tokenize_chinese_chars**: (`optional`) boolean (default True)
+                Whether to tokenize Chinese characters.
+                This should likely be deactivated for Japanese:
+                see: https://github.com/huggingface/pytorch-pretrained-BERT/issues/328
+        """
+        super(BertTokenizer, self).__init__(unk_token=unk_token, sep_token=sep_token,
+                                            pad_token=pad_token, cls_token=cls_token,
+                                            mask_token=mask_token, **kwargs)
+        self.max_len_single_sentence = self.max_len - 2  # take into account special tokens
+        self.max_len_sentences_pair = self.max_len - 3  # take into account special tokens
+
+        if not os.path.isfile(vocab_file):
+            raise ValueError(
+                "Can't find a vocabulary file at path '{}'. To load the vocabulary from a Google pretrained "
+                "model use `tokenizer = BertTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`".format(vocab_file))
+        self.vocab = load_vocab(vocab_file)
+        self.ids_to_tokens = collections.OrderedDict(
+            [(ids, tok) for tok, ids in self.vocab.items()])
+        self.do_basic_tokenize = do_basic_tokenize
+        if do_basic_tokenize:
+            self.basic_tokenizer = BasicTokenizer(do_lower_case=do_lower_case,
+                                                  never_split=never_split,
+                                                  tokenize_chinese_chars=tokenize_chinese_chars)
+        self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab, unk_token=self.unk_token)
+
+    @property
+    def vocab_size(self):
+        return len(self.vocab)
+
+    def _tokenize(self, text):
+        split_tokens = []
+        if self.do_basic_tokenize:
+            for token in self.basic_tokenizer.tokenize(text, never_split=self.all_special_tokens):
+                for sub_token in self.wordpiece_tokenizer.tokenize(token):
+                    split_tokens.append(sub_token)
+        else:
+            split_tokens = self.wordpiece_tokenizer.tokenize(text)
+        return split_tokens
+
+    def _convert_token_to_id(self, token):
+        """ Converts a token (str/unicode) in an id using the vocab. """
+        return self.vocab.get(token, self.vocab.get(self.unk_token))
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (string/unicode) using the vocab."""
+        return self.ids_to_tokens.get(index, self.unk_token)
+
+    def convert_tokens_to_string(self, tokens):
+        """ Converts a sequence of tokens (string) in a single string. """
+        out_string = ' '.join(tokens).replace(' ##', '').strip()
+        return out_string
+
+    def add_special_tokens_single_sentence(self, token_ids):
+        """
+        Adds special tokens to the a sequence for sequence classification tasks.
+        A BERT sequence has the following format: [CLS] X [SEP]
+        """
+        return [self.cls_token_id] + token_ids + [self.sep_token_id]
+
+    def add_special_tokens_sentences_pair(self, token_ids_0, token_ids_1):
+        """
+        Adds special tokens to a sequence pair for sequence classification tasks.
+        A BERT sequence pair has the following format: [CLS] A [SEP] B [SEP]
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        return cls + token_ids_0 + sep + token_ids_1 + sep
+
+    def save_vocabulary(self, vocab_path):
+        """Save the tokenizer vocabulary to a directory or file."""
+        index = 0
+        if os.path.isdir(vocab_path):
+            vocab_file = os.path.join(vocab_path, VOCAB_FILES_NAMES['vocab_file'])
+        else:
+            vocab_file = vocab_path
+        with open(vocab_file, "w", encoding="utf-8") as writer:
+            for token, token_index in sorted(self.vocab.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning("Saving vocabulary to {}: vocabulary indices are not consecutive."
+                                   " Please check that the vocabulary is not corrupted!".format(vocab_file))
+                    index = token_index
+                writer.write(token + u'\n')
+                index += 1
+        return (vocab_file,)
+
+
+class BasicTokenizer(object):
+    """Runs basic tokenization (punctuation splitting, lower casing, etc.)."""
+
+    def __init__(self, do_lower_case=True, never_split=None, tokenize_chinese_chars=True):
+        """ Constructs a BasicTokenizer.
+
+        Args:
+            **do_lower_case**: Whether to lower case the input.
+            **never_split**: (`optional`) list of str
+                Kept for backward compatibility purposes.
+                Now implemented directly at the base class level (see :func:`PreTrainedTokenizer.tokenize`)
+                List of token not to split.
+            **tokenize_chinese_chars**: (`optional`) boolean (default True)
+                Whether to tokenize Chinese characters.
+                This should likely be deactivated for Japanese:
+                see: https://github.com/huggingface/pytorch-pretrained-BERT/issues/328
+        """
+        if never_split is None:
+            never_split = []
+        self.do_lower_case = do_lower_case
+        self.never_split = never_split
+        self.tokenize_chinese_chars = tokenize_chinese_chars
+
+    def tokenize(self, text, never_split=None):
+        """ Basic Tokenization of a piece of text.
+            Split on "white spaces" only, for sub-word tokenization, see WordPieceTokenizer.
+
+        Args:
+            **never_split**: (`optional`) list of str
+                Kept for backward compatibility purposes.
+                Now implemented directly at the base class level (see :func:`PreTrainedTokenizer.tokenize`)
+                List of token not to split.
+        """
+        never_split = self.never_split + (never_split if never_split is not None else [])
+        text = self._clean_text(text)
+        # This was added on November 1st, 2018 for the multilingual and Chinese
+        # models. This is also applied to the English models now, but it doesn't
+        # matter since the English models were not trained on any Chinese data
+        # and generally don't have any Chinese data in them (there are Chinese
+        # characters in the vocabulary because Wikipedia does have some Chinese
+        # words in the English Wikipedia.).
+        if self.tokenize_chinese_chars:
+            text = self._tokenize_chinese_chars(text)
+        orig_tokens = whitespace_tokenize(text)
+        split_tokens = []
+        for token in orig_tokens:
+            if self.do_lower_case and token not in never_split:
+                token = token.lower()
+                token = self._run_strip_accents(token)
+            split_tokens.extend(self._run_split_on_punc(token))
+
+        output_tokens = whitespace_tokenize(" ".join(split_tokens))
+        return output_tokens
+
+    def _run_strip_accents(self, text):
+        """Strips accents from a piece of text."""
+        text = unicodedata.normalize("NFD", text)
+        output = []
+        for char in text:
+            cat = unicodedata.category(char)
+            if cat == "Mn":
+                continue
+            output.append(char)
+        return "".join(output)
+
+    def _run_split_on_punc(self, text, never_split=None):
+        """Splits punctuation on a piece of text."""
+        if never_split is not None and text in never_split:
+            return [text]
+        chars = list(text)
+        i = 0
+        start_new_word = True
+        output = []
+        while i < len(chars):
+            char = chars[i]
+            if _is_punctuation(char):
+                output.append([char])
+                start_new_word = True
+            else:
+                if start_new_word:
+                    output.append([])
+                start_new_word = False
+                output[-1].append(char)
+            i += 1
+
+        return ["".join(x) for x in output]
+
+    def _tokenize_chinese_chars(self, text):
+        """Adds whitespace around any CJK character."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if self._is_chinese_char(cp):
+                output.append(" ")
+                output.append(char)
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+    def _is_chinese_char(self, cp):
+        """Checks whether CP is the codepoint of a CJK character."""
+        # This defines a "chinese character" as anything in the CJK Unicode block:
+        #   https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
+        #
+        # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
+        # despite its name. The modern Korean Hangul alphabet is a different block,
+        # as is Japanese Hiragana and Katakana. Those alphabets are used to write
+        # space-separated words, so they are not treated specially and handled
+        # like the all of the other languages.
+        if ((cp >= 0x4E00 and cp <= 0x9FFF) or  #
+                (cp >= 0x3400 and cp <= 0x4DBF) or  #
+                (cp >= 0x20000 and cp <= 0x2A6DF) or  #
+                (cp >= 0x2A700 and cp <= 0x2B73F) or  #
+                (cp >= 0x2B740 and cp <= 0x2B81F) or  #
+                (cp >= 0x2B820 and cp <= 0x2CEAF) or
+                (cp >= 0xF900 and cp <= 0xFAFF) or  #
+                (cp >= 0x2F800 and cp <= 0x2FA1F)):  #
+            return True
+
+        return False
+
+    def _clean_text(self, text):
+        """Performs invalid character removal and whitespace cleanup on text."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if cp == 0 or cp == 0xfffd or _is_control(char):
+                continue
+            if _is_whitespace(char):
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+
+class WordpieceTokenizer(object):
+    """Runs WordPiece tokenization."""
+
+    def __init__(self, vocab, unk_token, max_input_chars_per_word=100):
+        self.vocab = vocab
+        self.unk_token = unk_token
+        self.max_input_chars_per_word = max_input_chars_per_word
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text into its word pieces.
+
+        This uses a greedy longest-match-first algorithm to perform tokenization
+        using the given vocabulary.
+
+        For example:
+          input = "unaffable"
+          output = ["un", "##aff", "##able"]
+
+        Args:
+          text: A single token or whitespace separated tokens. This should have
+            already been passed through `BasicTokenizer`.
+
+        Returns:
+          A list of wordpiece tokens.
+        """
+
+        output_tokens = []
+        for token in whitespace_tokenize(text):
+            chars = list(token)
+            if len(chars) > self.max_input_chars_per_word:
+                output_tokens.append(self.unk_token)
+                continue
+
+            is_bad = False
+            start = 0
+            sub_tokens = []
+            while start < len(chars):
+                end = len(chars)
+                cur_substr = None
+                while start < end:
+                    substr = "".join(chars[start:end])
+                    if start > 0:
+                        substr = "##" + substr
+                    if substr in self.vocab:
+                        cur_substr = substr
+                        break
+                    end -= 1
+                if cur_substr is None:
+                    is_bad = True
+                    break
+                sub_tokens.append(cur_substr)
+                start = end
+
+            if is_bad:
+                output_tokens.append(self.unk_token)
+            else:
+                output_tokens.extend(sub_tokens)
+        return output_tokens
+
+
+def _is_whitespace(char):
+    """Checks whether `chars` is a whitespace character."""
+    # \t, \n, and \r are technically contorl characters but we treat them
+    # as whitespace since they are generally considered as such.
+    if char == " " or char == "\t" or char == "\n" or char == "\r":
+        return True
+    cat = unicodedata.category(char)
+    if cat == "Zs":
+        return True
+    return False
+
+
+def _is_control(char):
+    """Checks whether `chars` is a control character."""
+    # These are technically control characters but we count them as whitespace
+    # characters.
+    if char == "\t" or char == "\n" or char == "\r":
+        return False
+    cat = unicodedata.category(char)
+    if cat.startswith("C"):
+        return True
+    return False
+
+
+def _is_punctuation(char):
+    """Checks whether `chars` is a punctuation character."""
+    cp = ord(char)
+    # We treat all non-letter/number ASCII as punctuation.
+    # Characters such as "^", "$", and "`" are not in the Unicode
+    # Punctuation class but we treat them as punctuation anyways, for
+    # consistency.
+    if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or
+            (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)):
+        return True
+    cat = unicodedata.category(char)
+    if cat.startswith("P"):
+        return True
+    return False