Meissonic/train/run_overfit.sh

#!/bin/bash
# Overfitting experiment script
# This script runs a small overfitting experiment to verify implementation correctness

accelerate launch --multi_gpu --gpu_ids '0,1,2,3,4,5,6,7' --main_process_port 25011 --num_processes 8 train/train_overfit.py \
    --text_encoder_architecture umt5-base \
    --video_tokenizer_model_id "Cosmos-1.0-Tokenizer-DV8x16x16" \
    --instance_data_dir "/mnt/VideoGen/dataset/OpenVid1M/video_reorg/OpenVid1M_reorganized.csv" \
    --max_samples 256 \
    --num_frames 8 \
    --video_height 64 \
    --video_width 112 \
    --dataloader_num_workers 8 \
    --train_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --learning_rate 5e-4 \
    --max_train_steps 3000 \
    --lr_warmup_steps 100 \
    --gradient_checkpointing \
    --mixed_precision bf16 \
    --seed 42 \
    --output_dir "./output_overfit" \
    --logging_steps 50 \
    --save_steps 500 \
    --inference_steps 500 \
    --num_inference_samples 4 \
    --num_inference_steps 48 \
    --wan_pretrained_path Wan-AI/Wan2.1-T2V-1.3B   # Optional: path to pretrained Wan weights