File size: 12,577 Bytes
523bea4 faf4cb2 523bea4 faf4cb2 523bea4 faf4cb2 523bea4 faf4cb2 523bea4 faf4cb2 523bea4 faf4cb2 523bea4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 | """
mLSTM Cell and Block for Vision-LSTM (ViL) backbone.
Architecture follows the official NX-AI ViL-S implementation:
- LinearHeadwiseExpand for Q/K/V projections (block-diagonal, ~3K params each)
- Depthwise causal Conv1d on the mLSTM branch
- Gates (igate, fgate) take concatenated [q, k, v] as input
- Output gate from second half of proj_up output
- Parallel mLSTM scan with matrix memory
Reference: Beck et al., "xLSTM: Extended Long Short-Term Memory" (arXiv:2405.04517)
Alkin et al., "Vision-LSTM: xLSTM as Generic Vision Backbone" (arXiv:2406.04303)
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange, einsum
class LinearHeadwiseExpand(nn.Module):
"""Block-diagonal linear projection: each head has its own small weight matrix.
Instead of a full Linear(inner_dim, inner_dim) with inner_dim^2 params,
this uses num_heads independent (head_dim, head_dim) matrices.
For inner_dim=768, num_heads=192, head_dim=4:
Full linear: 768*768 = 589,824 params
Headwise: 192*4*4 = 3,072 params (192x fewer!)
"""
def __init__(self, in_features: int, num_heads: int, bias: bool = False):
super().__init__()
assert in_features % num_heads == 0, f"{in_features} not divisible by {num_heads}"
self.num_heads = num_heads
self.head_dim = in_features // num_heads
self.in_features = in_features
# Weight: (num_heads, head_dim_out, head_dim_in)
self.weight = nn.Parameter(torch.empty(num_heads, self.head_dim, self.head_dim))
self.bias = nn.Parameter(torch.zeros(in_features)) if bias else None
self._reset_parameters()
def _reset_parameters(self):
nn.init.normal_(self.weight, std=math.sqrt(2.0 / (5.0 * self.head_dim)))
if self.bias is not None:
nn.init.zeros_(self.bias)
def forward(self, x: torch.Tensor) -> torch.Tensor:
# x: (..., in_features)
x = rearrange(x, '... (nh d) -> ... nh d', nh=self.num_heads)
x = einsum(x, self.weight, '... nh d, nh od d -> ... nh od')
x = rearrange(x, '... nh od -> ... (nh od)')
if self.bias is not None:
x = x + self.bias
return x
class StochasticDepth(nn.Module):
"""Drop entire residual path with probability `drop_prob` during training."""
def __init__(self, drop_prob: float = 0.0):
super().__init__()
self.drop_prob = drop_prob
def forward(self, x: torch.Tensor) -> torch.Tensor:
if not self.training or self.drop_prob == 0.0:
return x
keep_prob = 1 - self.drop_prob
shape = (x.shape[0],) + (1,) * (x.ndim - 1)
mask = torch.bernoulli(torch.full(shape, keep_prob, device=x.device, dtype=x.dtype))
return x * mask / keep_prob
class mLSTMCell(nn.Module):
"""Parallel mLSTM cell with matrix memory.
Official architecture from xLSTM/ViL:
- proj_up: Linear(D, 2*inner_dim) → split into mLSTM branch + output gate branch
- CausalConv1d on mLSTM branch (depthwise, k=4)
- LinearHeadwiseExpand for Q, K, V projections
- igate, fgate: Linear(3*inner_dim, num_heads) from concat(q,k,v)
- Parallel scan: C_t = f_t*C_{t-1} + i_t*(v_t ⊗ k_t), h_t = C_t*q_t
- Output: (h + skip*conv_act) * SiLU(z), then proj_down
ViL-S config: D=384, proj_factor=2.0, inner_dim=768,
qkv_proj_blocksize=4, num_heads=4 (memory heads)
Note: GroupNorm uses num_proj_heads (192) groups, matching official
MultiHeadLayerNorm — one group per projection head, NOT per memory head.
Per-cell params: ~920K (vs 2.66M with full Linear Q/K/V)
"""
def __init__(
self,
dim: int = 384,
proj_factor: float = 2.0,
qkv_proj_blocksize: int = 4,
num_heads: int = 4,
conv_kernel: int = 4,
bias: bool = False,
):
super().__init__()
self.dim = dim
# inner_dim rounded up to multiple of 64
self.inner_dim = math.ceil(proj_factor * dim / 64) * 64
self.num_heads = num_heads
self.head_dim = self.inner_dim // num_heads # 768/4 = 192
# Number of projection heads for Q/K/V (block-diagonal)
num_proj_heads = self.inner_dim // qkv_proj_blocksize
self.num_proj_heads = num_proj_heads
# Up-projection: D -> 2*inner_dim (mLSTM branch + output gate branch)
self.proj_up = nn.Linear(dim, 2 * self.inner_dim, bias=bias)
# Depthwise causal conv1d on mLSTM branch
self.conv1d = nn.Conv1d(
self.inner_dim, self.inner_dim,
kernel_size=conv_kernel,
padding=conv_kernel - 1, # causal: pad left
groups=self.inner_dim,
bias=True,
)
self.conv_kernel = conv_kernel
# Block-diagonal Q/K/V projections
self.q_proj = LinearHeadwiseExpand(self.inner_dim, num_proj_heads, bias=bias)
self.k_proj = LinearHeadwiseExpand(self.inner_dim, num_proj_heads, bias=bias)
self.v_proj = LinearHeadwiseExpand(self.inner_dim, num_proj_heads, bias=bias)
# Gates: take concat(q, k, v) as input
self.igate = nn.Linear(3 * self.inner_dim, num_heads, bias=True)
self.fgate = nn.Linear(3 * self.inner_dim, num_heads, bias=True)
# Output normalization: per-projection-head group norm (192 groups for ViL-S)
# Matches official MultiHeadLayerNorm — one group per projection head
self.outnorm = nn.GroupNorm(num_proj_heads, self.inner_dim, affine=True)
# Down-projection: inner_dim -> D
self.proj_down = nn.Linear(self.inner_dim, dim, bias=bias)
# Learnable skip connection and layer scale
self.learnable_skip = nn.Parameter(torch.ones(self.inner_dim))
self.layerscale = nn.Parameter(torch.ones(self.inner_dim))
self._reset_gate_bias()
def _reset_gate_bias(self):
"""Initialize forget gate bias high (encourages remembering) and input gate low."""
with torch.no_grad():
nn.init.zeros_(self.igate.bias)
# Forget gate bias: initialize to encourage remembering
nn.init.constant_(self.fgate.bias, 3.0)
def forward(self, x: torch.Tensor, reverse: bool = False) -> torch.Tensor:
"""
Args:
x: (B, S, D) input sequence
reverse: if True, process sequence right-to-left (for bidirectional scanning)
Returns:
(B, S, D) output
"""
B, S, D = x.shape
if reverse:
x = x.flip(1)
# 1. Up-project to 2*inner_dim
up = self.proj_up(x) # (B, S, 2*inner)
x_mlstm = up[..., :self.inner_dim] # mLSTM branch
z = up[..., self.inner_dim:] # output gate branch
# 2. Causal conv1d on mLSTM branch
x_conv = self.conv1d(x_mlstm.transpose(1, 2)) # (B, inner, S+pad)
x_conv = x_conv[..., :S].transpose(1, 2) # causal: keep first S
x_conv_act = F.silu(x_conv)
# 3. Q/K/V projections (block-diagonal, very lightweight)
q = self.q_proj(x_conv_act) # (B, S, inner)
k = self.k_proj(x_conv_act) # (B, S, inner)
v = self.v_proj(x_mlstm) # V from pre-conv branch
# 4. Gates from concat(q, k, v)
qkv_cat = torch.cat([q, k, v], dim=-1) # (B, S, 3*inner)
i_gate = self.igate(qkv_cat) # (B, S, num_heads)
f_gate = self.fgate(qkv_cat) # (B, S, num_heads)
# Stabilized gates
i_tilde = torch.exp(i_gate) # (B, S, H)
f_tilde = torch.sigmoid(f_gate) # (B, S, H)
# Log-space stabilization
log_f = torch.log(f_tilde.clamp(min=1e-6)) # (B, S, H)
# 5. Reshape Q/K/V for multi-head matrix memory
q = rearrange(q, 'b s (h d) -> b h s d', h=self.num_heads) # (B, H, S, D_h)
k = rearrange(k, 'b s (h d) -> b h s d', h=self.num_heads)
v = rearrange(v, 'b s (h d) -> b h s d', h=self.num_heads)
# 6. Parallel mLSTM computation (log-space stabilized)
# Cumulative sum of log forget gates for parallel scan
log_f_cumsum = torch.cumsum(log_f.permute(0, 2, 1), dim=-1) # (B, H, S)
# Compute pairwise log forget gate differences for attention-like matrix
# log_f_cumsum[:,:,j] - log_f_cumsum[:,:,i] gives product of f gates from i+1 to j
log_D = log_f_cumsum.unsqueeze(-1) - log_f_cumsum.unsqueeze(-2) # (B, H, S, S)
# Causal mask: only attend to past positions
causal_mask = torch.tril(torch.ones(S, S, device=x.device, dtype=torch.bool))
log_D = log_D.masked_fill(~causal_mask, -1e9)
# Add input gate contribution
i_tilde_perm = i_tilde.permute(0, 2, 1) # (B, H, S)
log_D = log_D + torch.log(i_tilde_perm.clamp(min=1e-6)).unsqueeze(-2) # broadcast over queries
# Stabilize with max trick
max_log_D = log_D.max(dim=-1, keepdim=True).values.clamp(min=-10)
D = torch.exp(log_D - max_log_D) # (B, H, S, S)
D = D.masked_fill(~causal_mask, 0.0)
# Compute attention: h = D @ v, then normalize by D @ k·q
# Scale queries
q_scaled = q / math.sqrt(self.head_dim)
# Attention scores: (q @ k^T) * D
attn = torch.matmul(q_scaled, k.transpose(-1, -2)) * D # (B, H, S, S)
# Normalizer
normalizer = attn.sum(dim=-1, keepdim=True).clamp(min=1.0)
attn = attn / normalizer
# Output
h = torch.matmul(attn, v) # (B, H, S, D_h)
h = rearrange(h, 'b h s d -> b s (h d)')
# 7. Output norm
h = self.outnorm(h.transpose(1, 2)).transpose(1, 2) # GroupNorm on channel dim
# 8. Skip connection + output gate
h_skip = h + self.learnable_skip * x_conv_act
output = h_skip * F.silu(z) # output gate: SiLU (not sigmoid) per official ViL
# 9. Down-project + layer scale
output = self.proj_down(output)
output = output * self.layerscale[:self.dim] # Note: layerscale is inner_dim sized, we need dim
if reverse:
output = output.flip(1)
return output
class SwiGLUMLP(nn.Module):
"""SwiGLU MLP as used in ViL blocks.
SwiGLU(x) = (W1·x ⊙ Swish(V·x)) then W2·hidden → output
"""
def __init__(self, dim: int, mlp_ratio: float = 4.0, bias: bool = False, drop: float = 0.0):
super().__init__()
hidden_dim = int(dim * mlp_ratio)
# SwiGLU: two parallel projections, one gated
self.w1 = nn.Linear(dim, hidden_dim, bias=bias) # value path
self.w2 = nn.Linear(hidden_dim, dim, bias=bias) # down projection
self.v = nn.Linear(dim, hidden_dim, bias=bias) # gate path
self.drop = nn.Dropout(drop)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.drop(self.w2(F.silu(self.v(x)) * self.w1(x)))
class mLSTMBlock(nn.Module):
"""Single ViL block: LayerNorm → mLSTMCell → residual.
Following the official ViL-S architecture, there is NO separate MLP/FFN layer.
The gated output (proj_up → split → z-gate → proj_down) inside the mLSTMCell
already performs the role of dimension expansion + nonlinearity + projection.
This matches ViL-S: ~0.92M params per block, 24 blocks ≈ 22M backbone.
"""
def __init__(
self,
dim: int = 384,
proj_factor: float = 2.0,
qkv_proj_blocksize: int = 4,
num_heads: int = 4,
conv_kernel: int = 4,
mlp_ratio: float = 4.0, # kept for API compat but unused in standard blocks
drop_path: float = 0.0,
bias: bool = False,
):
super().__init__()
self.norm1 = nn.LayerNorm(dim, bias=False)
self.mlstm = mLSTMCell(
dim=dim,
proj_factor=proj_factor,
qkv_proj_blocksize=qkv_proj_blocksize,
num_heads=num_heads,
conv_kernel=conv_kernel,
bias=bias,
)
self.drop_path = StochasticDepth(drop_path)
def forward(self, x: torch.Tensor, reverse: bool = False) -> torch.Tensor:
x = x + self.drop_path(self.mlstm(self.norm1(x), reverse=reverse))
return x
|