Views
No views yet
1#!/bin/bash
2
3PRETRAIN_CKPT_PATH=checkpoints/llava-qwen1.5-0.5b-pretrain-vista_description-3ep
4BASE_MODEL=Qwen/Qwen1.5-0.5B
5ROOT_DATA=data/llm_data
6
7WANDB_PROJECT=chart-vision-llm CUDA_VISIBLE_DEVICES=0,1,2,3,4 deepspeed moellava/train/train_mem.py \
8 --lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 0.00000125 \
9 --deepspeed ./scripts/zero2.json \
10 --model_name_or_path $BASE_MODEL \
11 --version qwen \
12 --data_path $ROOT_DATA/json_files/vista_llava.json \
13 --image_folder ${ROOT_DATA}/coco2017/train2017 \
14 --image_tower google/siglip-base-patch16-256-multilingual \
15 --image_projector_type mlp2x_gelu \
16 --pretrain_mm_mlp_adapter $PRETRAIN_CKPT_PATH/mm_projector.bin \
17 --tune_mm_mlp_adapter True \
18 --mm_vision_select_layer -2 \
19 --mm_use_im_start_end False \
20 --mm_use_im_patch_token False \
21 --image_aspect_ratio pad \
22 --group_by_modality_length True \
23 --bf16 True \
24 --output_dir ./checkpoints/ft-llava-qwen1.5-0.5b-vista_llava-lora-2ep \
25 --num_train_epochs 2 \
26 --per_device_train_batch_size 8 \
27 --per_device_eval_batch_size 2 \
28 --gradient_accumulation_steps 2 \
29 --evaluation_strategy "no" \
30 --save_strategy "steps" \
31 --save_steps 5000 \
32 --save_total_limit 1 \
33 --learning_rate 2e-5 \
34 --weight_decay 0. \
35 --warmup_ratio 0.03 \
36 --lr_scheduler_type "cosine" \
37 --logging_steps 10 \
38 --tf32 True \
39 --model_max_length 2048 \
40 --gradient_checkpointing True \
41 --dataloader_num_workers 4 \
42 --lazy_preprocess True \
43 --report_to wandb \
44 --push_to_hub True \
45 --cache_dir cache_dir \
46 --run_name ft-llava-qwen1.5-0.5b-vista_llava-lora-2ep