vcoder_train.sh

#!/bin/bash

export WANDB_PROJECT= "vcoder" 
deepspeed vcoder_llava/train/vcoder_train_mem.py \
    --deepspeed ./scripts/zero3.json \
    --model_name_or_path liuhaotian/llava-v1.5-7b \
    --version v1 \
    --seg_data_path ./playground/data \
    --seg_image_folder ./playground/data \
    --image_folder ./playground/data \
    --use_mm2_proj True \
    --pretrain_mm2_mlp_adapter llava-v1.5-mlp2x-336px-pretrain-vicuna-7b-v1.5/mm_projector.bin \
    --tune_mm_mlp_adapter False \
    --freeze_mm_mlp_adapter True \
    --freeze_llm True \
    --seg_tune_adapter True \
    --mm_projector_type mlp2x_gelu \
    --seg_mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --image_aspect_ratio pad \
    --group_by_modality_length True \
    --bf16 True \
    --output_dir ./outputs/vcoder_llava-v1.5-7b \
    --num_train_epochs 2 \
    --per_device_train_batch_size 32 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 1100 \
    --save_total_limit 1 \
    --learning_rate 1e-3 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --report_to wandb