RulinShao / VL-Instruct

Codes for vision-language instruction tuning. Currently support BLIP2-t5 and BLIP2-vicuna.

VL-Instruct

Reference codes:

1. https://github.com/salesforce/BLIP
2. https://github.com/salesforce/LAVIS

Setup

Hard-coded the vicuna7b checkpoint in the LAVIS pkg. Install from the local source

cd LAVIS
pip install -e .

Prepare the Data

Modify the data paths in the personalized dataloader in data/vqa_dataset.py

Train BLIP2-FlanT5-xl

python -m torch.distributed.run --nproc_per_node=8 train_vqa.py --model_type blip2_t5 --train_qformer

Train BLIP2-Vicuna-7b

TODO: double check if the model is loaded properly. Pay attention to qformer_text_input according to this issue.

python -m torch.distributed.run --nproc_per_node=8 train_vqa.py --model_type blip2_vicuna --train_qformer

To also finetune the LLM

python -m torch.distributed.run --nproc_per_node=8 train_vqa.py --model_type blip2_vicuna --train_qformer --train_llm

change HF cache dir

export TRANSFORMERS_CACHE=/projects/nlp_lab/zhiyang/.cache/

install flash attention for cuda older than 11.4

pip install flash-attn==2.1.1 --no-build-isolation