pretrain.sh

#!/bin/bash
# deepspeed --include=localhost:1 pretrain_image_caption.py
# --image_folder /home/cv/sxp/LLaVA_1_0/dataset/cc3m_595k_images \
deepspeed --include=localhost:0,1 --master_port 9913 llava/train/train_mem.py \
    --deepspeed ./scripts/zero2.json \
    --model_name_or_path ./ocr_mllm_toy/pretrain_weight/qwen_pretrain \
    --version plain \
    --data_path ./datasets/caption_train.json \
    --image_folder "" \
    --vision_tower ./ocr_mllm_toy/pretrain_weight/qwen_vit_448 \
    --mm_projector_type linear \
    --tune_mm_mlp_adapter True \
    --mm_vision_select_layer -1 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio, pad \
    --bf16 True \
    --output_dir ./checkpoints/llava-v1.5-qwen14b-pretrain-5m-448 \
    --num_train_epochs 1 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 500 \
    --save_total_limit 4 \
    --learning_rate 1e-3 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 1024 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --report_to wandb