add checkpoints

pretrained_models/autoenc/autoencoder_config.yaml

@@ -0,0 +1,45 @@
+model:
+  base_learning_rate: 0.0001
+  scale_lr: false
+  target: cbgen.autoencoder.VQModel
+  init:
+    checkpoint: pretrained_models/llamagen/vq_ds16_c2i.pt
+    ignore_keys: ["quantize."]
+    freeze_encoder: true
+    freeze_decoder: true
+    freeze_decoder_last_layer: true
+    freeze_quant_convs: true
+  params:
+    monitor: "val/qloss"
+    embed_dim: 8
+    n_embed: 4096
+    #n_embed: 16384
+    quantize_prob: 1.0
+    p_train_random_scale: 0.9
+    normalize: true
+    quantize_iter_start: 1000
+    requant_pyramid_levels: 1  # -1 means all levels, 1 means only the top level
+    resolution: 256
+    # uses defaults for llg_vq_model class defs
+    ddconfig: {}
+
+data:
+  target: main_train_ae.DataModuleFromConfig
+  params:
+    num_workers: 32
+    batch_size: 32
+    wrap: true
+    train:
+      target: arexper.ldm_dataset.DatasetTrain
+      params:
+        train_dir: /datasets/imagenet/ILSVRC/Data/CLS-LOC/train
+        dataset_name: imagenet
+        size: 256
+        degradation: pil_nearest
+    validation:
+      target: arexper.ldm_dataset.DatasetVal
+      params:
+        val_dir: /datasets/imagenet/ILSVRC/Data/CLS-LOC/val
+        dataset_name: imagenet
+        size: 256
+        degradation: pil_nearest

pretrained_models/autoenc/checkpoint.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7503e2afd945546a0ba77cbb863227e8e5b1e1498190725c9358c6fed5720afa
+size 287552963

pretrained_models/checkerboard-L-2x/checkpoint-last.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0b1c226a3b6bc84827855d2228bc21642774d99145d5978636f2a1e3553f297
+size 4698506735

pretrained_models/checkerboard-L-2x/config.yaml

@@ -0,0 +1,37 @@
+model:
+    img_size: 256
+    scale_ratio: 2
+    class_cond: true
+    num_classes: 1000
+    random_block_size: true
+    random_replace: 0.1
+    random_replace_sampled: false  # "sampled" gt or just upscaled for random replace
+    use_head_mlp: false
+    num_mixing_layers: 2
+    autoenc:
+        config_path: pretrained_models/autoenc/autoencoder_config.yaml
+        checkpoint_path: pretrained_models/autoenc/checkpoint.pth
+        params:
+            requant_pyramid_levels: -1  # needed for loading pretrained from config
+
+    llama_config:
+        hidden_size: 1024
+        intermediate_size: 3072
+        num_attention_heads: 16
+        num_hidden_layers: 20
+
+args:
+    batch_size: 64
+    lr: 5.e-5
+    lr_step_factor: 0.2
+    warmup_epochs: 1
+    optimizer: adamw
+    weight_decay: 0.01
+    num_workers: 8
+    data_path: data/imagenet
+    lr_schedule: step
+    epochs: 205
+    lr_step_epochs: [200]
+    eval_freq: 5
+    cfg: 2.0
+    cfg_start_step: 5

pretrained_models/checkerboard-L-4x/checkpoint-last.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e359552cb2b1d7ace912d0aaac2450e80680e2635b9cb537442aba3823692fa9
+size 4697565524

pretrained_models/checkerboard-L-4x/config.yaml

@@ -0,0 +1,37 @@
+model:
+    img_size: 256
+    scale_ratio: 4
+    class_cond: true
+    num_classes: 1000
+    random_block_size: true
+    random_replace: 0.1
+    random_replace_sampled: false  # "sampled" gt or just upscaled for random replace
+    use_head_mlp: false
+    num_mixing_layers: 2
+    autoenc:
+        config_path: pretrained_models/autoenc/autoencoder_config.yaml
+        checkpoint_path: pretrained_models/autoenc/checkpoint.pth
+        params:
+            requant_pyramid_levels: -1  # needed for loading pretrained from config
+
+    llama_config:
+        hidden_size: 1024
+        intermediate_size: 3072
+        num_attention_heads: 16
+        num_hidden_layers: 20
+
+args:
+    batch_size: 64
+    lr: 5.e-5
+    lr_step_factor: 0.2
+    warmup_epochs: 1
+    optimizer: adamw
+    weight_decay: 0.01
+    num_workers: 8
+    data_path: data/imagenet
+    lr_schedule: step
+    epochs: 205
+    lr_step_epochs: [200]
+    eval_freq: 5
+    cfg: 2.0
+    cfg_start_step: 5