Meissonic/train/extract.sh

# mv /mnt/VideoGen/dataset/OpenVid1M/extracted_features_17_128_128_full_set /opt/dlami/nvme
    

# accelerate launch --multi_gpu --gpu_ids '0,1,2,3,4,5,6,7' --main_process_port 25011 --num_processes 8 \
#     train/extract_features.py \
#     --csv_path /mnt/VideoGen/dataset/OpenVid1M/video_reorg/OpenVid1M_reorganized.csv \
#     --output_dir /mnt/VideoGen/dataset/OpenVid1M/extracted_features_17_256_256_full_set \
#     --text_encoder_architecture umt5-xxl \
#     --video_tokenizer_model_id Cosmos-0.1-Tokenizer-DV4x8x8 \
#     --num_frames 17 \
#     --video_height 256 \
#     --video_width 256 \
#     --batch_size 64 \
#     --num_workers 8 \
#     --extract_text \
#     --extract_video \
#     --save_attention_mask \
#     --skip_existing

accelerate launch --multi_gpu --gpu_ids '0,1,2,3,4,5,6,7' --main_process_port 25012 --num_processes 8 \
  train/extract_features.py \
  --csv_path /mnt/VideoGen/dataset/OpenVid1M/video_reorg/OpenVid1M_reorganized.csv \
  --output_dir /mnt/VideoGen/dataset/OpenVid1M/extracted_features_17_256_256_full_set \
  --text_encoder_architecture umt5-xxl \
  --video_tokenizer_model_id Cosmos-0.1-Tokenizer-DV4x8x8 \
  --num_frames 17 \
  --video_height 256 \
  --video_width 256 \
  --batch_size 64 \
  --num_workers 8 \
  --extract_text --extract_video --save_attention_mask \
  --skip_existing \
  --index_file /mnt/VideoGen/dataset/OpenVid1M/extracted_features_17_256_256_full_set/missing_process_{rank}.txt \
  --run_tag resume \
  --merge_world_size 8


    
# python train/extract_empty_embeds.py \
#     --text_encoder_architecture umt5-base \
#     --output_path /path/to/empty_embeds.pt \
#     --dtype float16


# python train/train_mei_video.py \
#     --use_precomputed_features \
#     --features_dir /path/to/extracted_features \
#     --text_encoder_architecture umt5-base \
#     --video_tokenizer_model_id Cosmos-1.0-Tokenizer-DV8x16x16 \
#     --num_frames 16 \
#     --video_height 480 \
#     --video_width 848 \
#     --train_batch_size 8 \
#     --learning_rate 3e-4 \
#     --max_train_steps 10000 \
#     --output_dir ./output \
#     --mixed_precision bf16


# python train/check_codebook_range.py \
#     --csv_path /mnt/VideoGen/dataset/OpenVid1M/video_reorg/OpenVid1M_reorganized.csv \
#     --video_tokenizer_model_id Cosmos-0.1-Tokenizer-DV4x8x8 \
#     --num_frames 16 \
#     --video_height 480 \
#     --video_width 848 \
#     --check_interval 10 \
#     --max_samples 1000  # 可选：限制检查的样本数