Skip to content

Latest commit

聽

History

History
393 lines (345 loc) 路 12.8 KB

TRAIN_AND_VALIDATE.md

File metadata and controls

393 lines (345 loc) 路 12.8 KB
Raw

1. Conda environment

Install packages.

conda create -n chatunivi python=3.10 -y
conda activate chatunivi
pip install --upgrade pip
pip install -e .
pip install ninja
pip install flash-attn --no-build-isolation

2. Pre-trained model

Download the pre-trained model.

Name Weight
Chat-UniVi-7B Download
Chat-UniVi-13B Download
Chat-UniVi-ScienceQA-7B Download

3. Train the model

Prepare the base model.

Base model Weight
Vicuna-7B Download
Vicuna-13B Download

Stage1: Multimodal Pre-training

deepspeed \
--include localhost:0,1,2,3,4,5,6,7 \
--master_port=29602 \
ChatUniVi/train/train_mem.py \
--deepspeed scripts/zero3.json \
--model_name_or_path ${LLM model path} \
--version v1 \
--model_use PRETUNE \
--dataset_use Pretrain \
--vision_tower openai/clip-vit-large-patch14 \
--tune_mm_mlp_adapter True \
--mm_vision_select_layer -2 \
--mm_use_im_start_end False \
--mm_use_im_patch_token False \
--bf16 True \
--output_dir ${stage1 save path} \
--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 24000 \
--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 \
--dataloader_num_workers 4 \
--lazy_preprocess True \
--report_to wandb

Stage2: Joint Instruction Tuning

deepspeed \
--include localhost:0,1,2,3,4,5,6,7 \
--master_port=29601 \
ChatUniVi/train/train_mem.py \
--deepspeed scripts/zero2.json \
--model_name_or_path ${LLM model path} \
--version v1 \
--model_use FINETUNE \
--dataset_use FINETUNE \
--vision_tower openai/clip-vit-large-patch14 \
--pretrain_mm_mlp_adapter ${stage1 save path}/mm_projector.bin \
--mm_vision_select_layer -2 \
--mm_use_im_start_end False \
--mm_use_im_patch_token False \
--bf16 True \
--output_dir ${stage2 save path} \
--num_train_epochs 2 \
--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-5 \
--weight_decay 0. \
--warmup_ratio 0.03 \
--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

4. Evaluate the model

GPT-based evaluation for image understanding

Our quantitative evaluation protocol follows that of LLaVA. Following LLaVA, we employ 90 questions based on 30 COCO validation images, covering various aspects, including conversation, detail description, and complex reasoning. For more details, please refer to LLaVA.

Step 1: Load the model to generate results

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python ChatUniVi/eval/model_vqa.py \
--model-path ${model path} \
--question-file ChatUniVi/eval/questions/coco2014_val_qa_eval/qa90_questions.jsonl \
--image-folder ${image folder} \
--answers-file results/answer-file-vqa.jsonl

Step 2: GPT evaluation

OPENAI_API_KEY=[openai api key] \
python ChatUniVi/eval/evaluate/evaluate_gpt_review_visual.py \
--question ChatUniVi/eval/questions/coco2014_val_qa_eval/qa90_questions.jsonl \
--context ChatUniVi/eval/table/caps_boxes_coco2014_val_80.jsonl \
--answer-list ChatUniVi/eval/questions/coco2014_val_qa_eval/qa90_gpt4_answer.jsonl  results/answer-file-vqa.jsonl \
--rule ChatUniVi/eval/table/rule.json \
--output results/review-file-vqa.json

Step 3: Calculate score

python ChatUniVi/eval/evaluate/summarize_gpt_review.py \
-d  results/review-file-vqa.json

GPT-based evaluation for video understanding

The quantitative evaluation protocol for video understanding follows the methodology introduced by Video-ChatGPT. Specifically, Video-ChatGPT curates a test set based on the ActivityNet-200 dataset, which includes videos with rich, dense descriptive captions and associated question-answer pairs from human annotations. For more details, please refer to Video-ChatGPT.

It is worth noting that the results span a range from 0 to 5. To standardize the metrics, we normalized all scores to a scale of 0 to 100 in the paper.

Step 1: Load the model to generate results

python ChatUniVi/eval/model_video_general.py \
--model-path ${model path} \
--question-file ChatUniVi/eval/questions/video_qa/generic_qa.json \
--video-folder ${video folder} \
--answers-file results/answer-video-generic.jsonl


CUDA_VISIBLE_DEVICES=0,1,2,3 \
python ChatUniVi/eval/model_video_general.py \
--model-path ${model path} \
--question-file ChatUniVi/eval/questions/video_qa/temporal_qa.json \
--video-folder ${video folder} \
--answers-file results/answer-video-temporal.jsonl


CUDA_VISIBLE_DEVICES=0,1,2,3 \
python ChatUniVi/eval/model_video_consistency.py \
--model-path ${model path} \
--question-file ChatUniVi/eval/questions/video_qa/consistency_qa.json \
--video-folder ${video folder} \
--answers-file results/answer-video-consistency.jsonl

Step 2: GPT evaluation

# Correctness of Information
python ChatUniVi/eval/evaluate/evaluate_benchmark_1_correctness.py \
--pred_path results/answer-video-generic.jsonl \
--output_dir results/correctness \
--output_json results/review-video-correctness.jsonl \
--api_key [openai api key] \
--num_tasks 1

# Detail Orientation
python ChatUniVi/eval/evaluate/evaluate_benchmark_2_detailed_orientation.py \
--pred_path results/answer-video-generic.jsonl \
--output_dir results/detailed_orientation \
--output_json results/review-video-detailed_orientation.jsonl \
--api_key [openai api key] \
--num_tasks 1

# Contextual Understanding
python ChatUniVi/eval/evaluate/evaluate_benchmark_3_context.py \
--pred_path results/answer-video-generic.jsonl \
--output_dir results/context \
--output_json results/review-video-context.jsonl \
--api_key [openai api key] \
--num_tasks 1

# Temporal Understanding
python ChatUniVi/eval/evaluate/evaluate_benchmark_4_temporal.py \
--pred_path results/answer-video-temporal.jsonl \
--output_dir results/temporal \
--output_json results/review-video-temporal.jsonl \
--api_key [openai api key] \
--num_tasks 1

# Consistency
python ChatUniVi/eval/evaluate/evaluate_benchmark_5_consistency.py \
--pred_path results/answer-video-consistency.jsonl \
--output_dir results/consistency \
--output_json results/review-video-consistency.jsonl \
--api_key [openai api key] \
--num_tasks 1

ScienceQA

ScienceQA is a comprehensive multimodal science question-answering dataset comprising 21k multiple-choice questions. It covers a wide range of domains, spanning 3 subjects, 26 topics, 127 categories, and 379 skills. Each example in ScienceQA contains a visual context, a textual context, a question, multiple options, and the correct answer. For the input of Chat-UniVi, we concatenate the question, textual context, and options sequentially into a single sentence.

ScienceQA Fine-tuning

deepspeed \
--include localhost:0,1,2,3,4,5,6,7 \
--master_port=29603 ChatUniVi/train/train.py \
--deepspeed scripts/zero.json \
--model_name_or_path ${LLM model path} \
--version v1 \
--model_use FINETUNE \
--dataset_use SQA \
--vision_tower openai/clip-vit-large-patch14  \
--pretrain_mm_mlp_adapter ${stage1 save path}/mm_projector.bin \
--mm_vision_select_layer -2 \
--mm_use_im_start_end False \
--mm_use_im_patch_token False \
--bf16 True \
--output_dir ${save path} \
--num_train_epochs 9 \
--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 1 \
--learning_rate 2e-5 \
--weight_decay 0. \
--warmup_ratio 0.03 \
--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

Step 1: Load the model to generate results

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python ChatUniVi/eval/model_vqa_scienceqa.py \
--model-path ${model path} \
--question-file ChatUniVi/eval/questions/scienceqa/test_QCM-LEA.json \
--image-folder ${image folder} \
--answers-file results/answer-scienceqa.jsonl

Step 2: Calculate score

python ChatUniVi/eval/evaluate/evaluate_science_qa.py \
--base-dir ChatUniVi/eval/questions/scienceqa \
--result-file results/answer-scienceqa.jsonl \
--output-file results/output-scienceqa.json \
--output-result results/output-result-scienceqa.json

Zero-shot Video Question Evaluation

Our evaluation protocol follows that of Video-ChatGPT, utilizing GPT-assisted evaluation to assess the capabilities of models. For more details, please refer to Video-ChatGPT.

Step 1: Load the model to generate results

# MSRVTT QA
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python ChatUniVi/eval/model_video_qa.py \
--model-path ${model path} \
--question-file ChatUniVi/eval/questions/video_qa/msrvtt_qa.json \
--video-folder ${video folder} \
--answers-list ChatUniVi/eval/questions/video_qa/msrvtt_a_list.json \
--answers-file results/answer-msrvtt-qa.jsonl

# MSVD QA
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python ChatUniVi/eval/model_video_qa.py \
--model-path ${model path} \
--question-file ChatUniVi/eval/questions/video_qa/msvd_qa.json \
--video-folder ${video folder} \
--answers-list ChatUniVi/eval/questions/video_qa/msvd_a_list.json \
--answers-file results/answer-msvd-qa.jsonl

# TGIF QA
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python ChatUniVi/eval/model_video_qa.py \
--model-path ${model path} \
--question-file ChatUniVi/eval/questions/video_qa/tgif_qa.json \
--video-folder ${video folder} \
--answers-list ChatUniVi/eval/questions/video_qa/tgif_a_list.json \
--answers-file results/answer-tgif-qa.jsonl

# ActivityNet QA
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python ChatUniVi/eval/model_video_qa.py \
--model-path ${model path} \
--question-file ChatUniVi/eval/questions/video_qa/activitynet_qa.json \
--video-folder ${video folder} \
--answers-list ChatUniVi/eval/questions/video_qa/activitynet_a_list.json \
--answers-file results/answer-activitynet-qa.jsonl

Step 2: Calculate score

# MSRVTT QA
python ChatUniVi/eval/evaluate/evaluate_video_qa.py \
--pred_path results/answer-msrvtt-qa.jsonl \
--output_dir results/msrvtt-qa \
--output_json results/review-msrvtt-qa.jsonl \
--api_key [openai api key] \
--num_tasks 1

# MSVD QA
python ChatUniVi/eval/evaluate/evaluate_video_qa.py \
--pred_path results/answer-msvd-qa.jsonl \
--output_dir results/msvd-qa \
--output_json results/review-msvd-qa.jsonl \
--api_key [openai api key] \
--num_tasks 1

# TGIF QA
python ChatUniVi/eval/evaluate/evaluate_video_qa.py \
--pred_path results/answer-tgif-qa.jsonl \
--output_dir results/tgif-qa \
--output_json results/review-tgif-qa.jsonl \
--api_key [openai api key] \
--num_tasks 1

# ActivityNet QA
python ChatUniVi/eval/evaluate/evaluate_video_qa.py \
--pred_path results/answer-activitynet-qa.jsonl \
--output_dir results/activitynet-qa \
--output_json results/review-activitynet-qa.jsonl \
--api_key [openai api key] \
--num_tasks 1

Zero-shot Object Hallucination Evaluation

To quantitatively evaluate the hallucination problem of the model, we adopt the polling-based object probing evaluation (POPE) process.

# Random
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python ChatUniVi/eval/model_coco_vqa.py \
--model-path ${model path} \
--question-file ChatUniVi/eval/questions/coco_pope/coco_pope_random.jsonl \
--image-folder ${image folder} \
--answers-file results/pope-random.jsonl

# Popular
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python ChatUniVi/eval/model_coco_vqa.py \
--model-path ${model path} \
--question-file ChatUniVi/eval/questions/coco_pope/coco_pope_popular.jsonl \
--image-folder ${image folder} \
--answers-file results/pope-popular.jsonl

# Adversarial
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python ChatUniVi/eval/model_coco_vqa.py \
--model-path ${model path} \
--question-file ChatUniVi/eval/questions/coco_pope/coco_pope_adversarial.jsonl \
--image-folder ${image folder} \
--answers-file results/pope-adversarial.jsonl