Skip to content

Latest commit

 

History

History
170 lines (124 loc) · 6.11 KB

README.md

File metadata and controls

170 lines (124 loc) · 6.11 KB
Raw

LLaVA-1.5

Model Training

1. model preparation

The LLaVA version we use in HA-DPO is LLaVA-v1.5-7b. Before training, prepare following:

# download command
from huggingface_hub import snapshot_download
snapshot_download(repo_id="liuhaotian/llava-v1.5-7b")

If you download language model weights to a user-specified path using git lfs rather than huggingface provided download API (such as from_pretrained or snapshot_download), replace all liuhaotian/llava-v1.5-7b to path of your downloaded model in training and evaluation.

2. data Preparation

follow instructions in data preparation for LLaVA-1.5 data preparation.

3. model training

We use LoRA adapters to fine-tune the language model of LLaVA-1.5 to train the model. Following training settings in LLaVA, all linear layers in the model are set to trainable. 8 A100 GPUs are used during fine-tuning.

Training command 
deepspeed ha_dpo/models/llava-v1_5/train_dpo.py \
    --lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 0 \
    --deepspeed ha_dpo/models/llava-v1_5/scripts/zero3.json \
    --model_name_or_path liuhaotian/llava-v1.5-7b \
    --version v1 \
    --vg_path ha_dpo/data/VG \
    --desc_data_path ha_dpo/data/hadpo/llava-v1.5/desc_data.json \
    --pope_data_path ha_dpo/data/hadpo/llava-v1.5/pope_data.json \
    --vision_tower openai/clip-vit-large-patch14-336 \
    --mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio pad \
    --group_by_modality_length True \
    --bf16 True \
    --output_dir ha_dpo/models/llava-v1_5/checkpoints/{model_name} \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 50000 \
    --save_total_limit 1 \
    --learning_rate 2e-6 \
    --weight_decay 0. \
    --warmup_steps 0 \
    --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 \
    --run_name "llava-v1.5" \
    --beta 0.1

default parameters are as follows:

epochs learning rate lora_r lora_alpha beta
1 2e-6 128 256 0.1

Model Evaluation

SHR Evaluation

Run the following command to evaluate on SHR:

python ha_dpo/models/llava-v1_5/shr_eval.py \
--api-key {openai_apikey} \
--vg-path ha_dpo/data/VG \
--shr-path ha_dpo/data/shr \
--model-base liuhaotian/llava-v1.5-7b \
--model-path ha_dpo/models/llava-v1_5/checkpoints/{model_name}
  1. --api-key: SHR evaluation relies on GPT-4. Provide the openai key, begin with sk.
  2. --model-path: path to the the trained adapter weights.

After evaluation is finished, results are saved in ha_dpo/models/llava-v1_5/shr_eval_results/{localtime}/metrics.json.

  1. judgement.json: detailed judgements in SHR evaluation.
  2. metrics.json: detailed metrics in SHR evaluation. mean_hal_ratio indicates the ration of hallucinated sentences, which is the main SHR result.

to reproduce results, use trained adapter weights and set --model-path juliozhao/hadpo-llava-1.5.

SHR results
Model HA-DPO SHR
LLaVA-1.5 ✖️ 36.7
LLaVA-1.5 ✔️ 34.0

POPE Evaluation

step 1. Firstly, inference answers using:

torchrun --nproc_per_node {NGPUS} --master_port $RANDOM ha_dpo/models/llava-v1_5/pope_eval.py \
    --coco_path ha_dpo/data/coco2014 \
    --pope_path ha_dpo/data/POPE \
    --model-path ha_dpo/models/llava-v1_5/checkpoints/{model_name} \
    --model-base liuhaotian/llava-v1.5-7b \
    --set {random/popular/adv}
  1. --set: validation sets in POPE, choose random/popular/adv. After inference, the answer file will be generated under the folder of LLaVA.
  2. --model-path: path to the the trained adapter weights.

to reproduce results, use trained adapter weights and set --model-path juliozhao/hadpo-llava-1.5.

step 2. Set the path of answer file and the label file in ha_dpo/data/POPE/evaluate.py.

Set ans_file to the path of the generated answer file in step 1, and set label_file to the path of label files under ha_dpo/data/POPE/output/coco.

step 3. Evaluate.

run python ha_dpo/data/POPE/evaluate.py to get results.

POPE results

POPE Random

Model HA-DPO Accuracy Precision Recall F1 Score Yes Ratio (%)
LLaVA-1.5 ✖️ 89.60 88.77 90.66 89.70 51.06
LLaVA-1.5 ✔️ 90.53 92.99 87.66 90.25 47.13

POPE Popular

Model HA-DPO Accuracy Precision Recall F1 Score Yes Ratio (%)
LLaVA-1.5 ✖️ 86.20 83.23 90.66 86.79 54.46
LLaVA-1.5 ✔️ 87.90 88.07 87.66 87.81 49.76

POPE Adversarial

Model HA-DPO Accuracy Precision Recall F1 Score Yes Ratio (%)
LLaVA-1.5 ✖️ 79.76 74.43 90.66 81.75 60.90
LLaVA-1.5 ✔️ 81.46 77.99 87.66 82.54 56.20

⚠️ NOTICE:

  1. The optimal parameters can differ according to the environment of your machine, you can adjust these parameters according to the behavior of LVLM.
  2. For baseline LLaVA-1.5 results, don't set --model-base and set --model-path to liuhaotian/llava-v1.5-7b in the evaluation command.