记得把.gitigore补回来
playground/data/ /checkpoints/ pycache /hf_models/ /hf_datas/
下载仓库
Install Package
cd VoCo-LLaMA
conda create -n voco python=3.10 -y
conda activate voco
pip install --upgrade pip # enable PEP 660 support
pip install -e .
- Install additional packages for training cases
pip install -e ".[train]"
- 找到conda环境里的hf代码:
miniconda3/envs/voco/lib/python3.10/site-packages/transformers/modeling_attn_mask_utils.py把VoCo-LLaMA/llava/model/language_model/cache_py/modeling_attn_mask_utils.py文件复制过去(直接覆盖)
cp VoCo-LLaMA/llava/model/language_model/cache_py/modeling_attn_mask_utils.py /data/miniconda3/envs/voco/lib/python3.10/site-packages/transformers/modeling_attn_mask_utils.py
- 重新安装deepspeed
pip install deepspeed==0.15.4
- 训练
bash scripts/finetune_voco_llama.sh
- 评估
pip install openpyxl
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/eval/vqav2.sh
CUDA_VISIBLE_DEVICES=1 bash scripts/eval/mmbench.sh
CUDA_VISIBLE_DEVICES=2 bash scripts/eval/sqa.sh
- 提交结果(只有sqa可以直接出结果,其他两个应该是闭源评测)
把VoCo-LLaMA/playground/data/eval/vqav2/answers_upload/llava_vqav2_mscoco_test-dev2015/voco_llava.json提交到https://eval.ai/web/challenges/challenge-page/830/my-submission
把VoCo-LLaMA/playground/data/eval/mmbench/answers_upload/mmbench_dev_20230712/voco_llama.xlsx提交到https://mmbench.opencompass.org.cn/mmbench-submission