update feat train

configs/opensora-v1-2/misc/extract.py

@@ -58,5 +58,5 @@ text_encoder = dict(
 # feature extraction settings
 save_text_features = True
 save_compressed_text_features = True
-bin_size = 10
+bin_size = 250  # 1GB, 4195 bins
 log_time = False

configs/opensora-v1-2/train/stage1_feat.py

@@ -25,7 +25,7 @@ scheduler = dict(
 
 vae_out_channels = 4
 model_max_length = 300
-text_encoder_output_dim = 1152
+text_encoder_output_dim = 4096
 load_video_features = True
 load_text_features = True
 

configs/opensora-v1-2/train/train_load_batch.py

@@ -1,68 +0,0 @@
-# Dataset settings
-dataset = dict(
-    type="BatchFeatureDataset",
-)
-
-grad_checkpoint = True
-
-# Acceleration settings
-num_workers = 8
-dtype = "bf16"
-plugin = "zero2"
-
-# Model settings
-model = dict(
-    type="STDiT3-XL/2",
-    from_pretrained=None,
-    qk_norm=True,
-    enable_flash_attn=True,
-    enable_layernorm_kernel=True,
-    freeze_y_embedder=True,
-)
-vae = dict(
-    type="OpenSoraVAE_V1_2",
-    from_pretrained="pretrained_models/vae-pipeline",
-    micro_frame_size=17,
-    micro_batch_size=4,
-)
-text_encoder = dict(
-    type="t5",
-    from_pretrained="DeepFloyd/t5-v1_1-xxl",
-    model_max_length=300,
-    shardformer=True,
-    local_files_only=True,
-)
-scheduler = dict(
-    type="rflow",
-    use_timestep_transform=True,
-    sample_method="logit-normal",
-)
-
-# Mask settings
-mask_ratios = {
-    "random": 0.2,
-    "intepolate": 0.01,
-    "quarter_random": 0.01,
-    "quarter_head": 0.01,
-    "quarter_tail": 0.01,
-    "quarter_head_tail": 0.01,
-    "image_random": 0.05,
-    "image_head": 0.1,
-    "image_tail": 0.05,
-    "image_head_tail": 0.05,
-}
-
-# Log settings
-seed = 42
-outputs = "outputs"
-wandb = False
-epochs = 1
-log_every = 10
-ckpt_every = 500
-
-# optimization settings
-load = None
-grad_clip = 1.0
-lr = 2e-4
-ema_decay = 0.99
-adam_eps = 1e-15

scripts/misc/extract_feat.py

@@ -61,7 +61,7 @@ def main():
         pin_memory=True,
         process_group=get_data_parallel_group(),
     )
-    dataloader, sampler = prepare_dataloader(
+    dataloader, _ = prepare_dataloader(
         bucket_config=cfg.get("bucket_config", None),
         num_bucket_build_workers=cfg.get("num_bucket_build_workers", 1),
         **dataloader_args,

scripts/train.py

@@ -113,7 +113,7 @@ def main():
         text_encoder_output_dim = text_encoder.output_dim
         text_encoder_model_max_length = text_encoder.model_max_length
     else:
-        text_encoder_output_dim = cfg.get("text_encoder_output_dim", 1152)
+        text_encoder_output_dim = cfg.get("text_encoder_output_dim", 4096)
         text_encoder_model_max_length = cfg.get("text_encoder_model_max_length", 300)
 
     # == build vae ==