Meissonic/train/train_video.sh

#!/bin/bash
# 8-GPU training script for video diffusion model
# Usage: bash train/train_video.sh

accelerate launch --multi_gpu --gpu_ids '0,1,2,3,4,5,6,7' --main_process_port 25011 --num_processes 8 \
    train/train_mei_video.py \
    --use_precomputed_features \
    --features_dir /mnt/VideoGen/dataset/OpenVid1M/extracted_features_17_128_128_with_mask_attn_debug \
    --text_encoder_architecture umt5-xxl \
    --wan_pretrained_path /mnt/Wan2.1-T2V-1.3B \
    --training_from_scratch True \
    --pretrained_model_name_or_path "dummy" \
    --wan_backbone_lr_ratio 0.2 \
    --num_frames 17 \
    --video_height 128 \
    --video_width 128 \
    --dataloader_num_workers 8 \
    --video_tokenizer_model_id "Cosmos-0.1-Tokenizer-DV4x8x8" \
    --instance_dataset OpenVid1MDataset \
    --instance_data_dir "/mnt/VideoGen/dataset/OpenVid1M/video_reorg/OpenVid1M_reorganized.csv" \
    --train_batch_size 8 \
    --gradient_accumulation_steps 4 \
    --learning_rate 3e-4 \
    --max_train_steps 10000 \
    --checkpointing_steps 500 \
    --validation_steps 500 \
    --logging_steps 10 \
    --validation_prompts "a cat playing" "a girl walking" \
    --output_dir "./output_128x128_17f_2*4bs_4*8*8vqvae_0_2_ratio" \
    --mixed_precision bf16 \
    --lr_scheduler constant \
    --lr_warmup_steps 0 \
    --use_8bit_adam \
    --gradient_checkpointing \
    --min_masking_rate 0.0 \
    --cond_dropout_prob 0.1 \
    --split_vae_encode 1 \
    --allow_tf32 \
    --seed 42 \
    --report_to wandb

    # --use_precomputed_features \
    # --features_dir /mnt/VideoGen/dataset/OpenVid1M/extracted_features \

# accelerate launch --multi_gpu --gpu_ids '0,1,2,3,4,5,6,7' --main_process_port 25011 --num_processes 8 \
#     train/train_mei_video.py \
#     --use_precomputed_features \
#     --features_dir /mnt/VideoGen/dataset/OpenVid1M/extracted_features \
#     --text_encoder_architecture umt5-xxl \
#     --wan_pretrained_path Wan-AI/Wan2.1-T2V-1.3B \
#     --training_from_scratch True \
#     --pretrained_model_name_or_path "dummy" \
#     --wan_backbone_lr_ratio 1 \
#     --num_frames 4 \
#     --video_height 256 \
#     --video_width 448 \
#     --dataloader_num_workers 8 \
#     --video_tokenizer_model_id "Cosmos-0.1-Tokenizer-DV4x8x8" \
#     --instance_dataset OpenVid1MDataset \
#     --instance_data_dir "/mnt/VideoGen/dataset/OpenVid1M/video_reorg/OpenVid1M_reorganized.csv" \
#     --train_batch_size 1 \
#     --gradient_accumulation_steps 1 \
#     --learning_rate 3e-4 \
#     --max_train_steps 10000 \
#     --checkpointing_steps 500 \
#     --validation_steps 500 \
#     --logging_steps 10 \
#     --validation_prompts "a cat playing" "a girl walking" \
#     --output_dir "./output_256x448_4f_2bs_4*8*8vqvae_0_00_ratio_continue_tmp" \
#     --mixed_precision bf16 \
#     --lr_scheduler constant \
#     --lr_warmup_steps 0 \
#     --use_8bit_adam \
#     --gradient_checkpointing \
#     --min_masking_rate 0.0 \
#     --cond_dropout_prob 0.1 \
#     --split_vae_encode 1 \
#     --allow_tf32 \
#     --seed 42 \
#     --report_to wandb

# --pretrained_model_name_or_path "/mnt/Meissonic/output_256x448_4f_2bs_4*8*8vqvae_0_00_ratio/checkpoint-4000" \