# Description

Trained on top of llava-13b-v1: https://huggingface.co/wza/llava-13b-v1 (github: https://github.com/haotian-liu/LLaVA)

# Dataset

Constructed dataset on stock k lines, both pre-train and instrcution-tune

# Training scripts

pre-train:
```
torchrun --nnodes=1 --nproc_per_node=8 --master_port=25001 \
    LLaVA/llava/train/train_mem.py \
    --model_name_or_path llava-13b-v1 \
    --data_path JsonFormatDataset/PretrainData/data.json \
    --image_folder JsonFormatDataset/PretrainData/images \
    --vision_tower openai/clip-vit-large-patch14 \
    --tune_mm_mlp_adapter True \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end \
    --bf16 True \
    --output_dir ./checkpoints/llava-13b-pretrain \
    --num_train_epochs 1 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 2 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 2400 \
    --save_total_limit 1 \
    --learning_rate 2e-3 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --lazy_preprocess True \
    --report_to wandb
```

instruction:
```
torchrun --nnodes=1 --nproc_per_node=8 --master_port=25001 \
    LLaVA/llava/train/train_mem.py \
    --model_name_or_path ./checkpoints/llava-13b-pretrain \
    --data_path JsonFormatDataset/InstructionTuneData/data.json \
    --image_folder JsonFormatDataset/InstructionTuneData/images/ \
    --vision_tower openai/clip-vit-large-patch14 \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end \
    --bf16 True \
    --output_dir ./checkpoints/llava-13b-instruction \
    --num_train_epochs 3 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 5000 \
    --save_total_limit 3 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --fsdp "full_shard auto_wrap" \
    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --lazy_preprocess True \
    --report_to wandb

```

# Training settings

8xA100-80G-sxm4

Pre-train: https://wandb.ai/wzaa/huggingface/runs/cd5ou876/overview?workspace=user-wangziao1993

Fine-tune: https://wandb.ai/wzaa/huggingface/runs/y5bsz8dw/overview?workspace=user-wangziao1993