월간 데이콘 이미지 기반 질의 응답 AI 경진대회

결과

• ★ PUBLIC 3위 / PRIVATE 2위 ★
• LLaVA 모델 from pretrained train data

1. Introduction

[배경]

• 멀티모달 AI는 서로 다른 유형의 데이터를 결합하여 사용하는 기술로, 텍스트와 이미지 등 다양한 데이터를 종합적으로 다루는 기술입니다.
• 서비스적으로 활용 가치가 높은 멀티모달 AI 모델 개발 및 고도화에 도전해 보세요!

[주제] 이미지 기반 질의 응답 AI 모델 개발

[기간] 2023.07.10. ~ 2023.08.07.

[링크] https://dacon.io/competitions/official/236118/overview/description

2. Data

data
├─  image
│   ├─  train : 107,231개
│   │   ├─  train_000000.png
│   │   ├─  train_000001.png
│   │   └─  ...
│   └─  test : 11,915개
│       ├─  test_00000.png
│       ├─  test_00001.png
│       └─  ...
├─  train.csv
|    ├─  ID : 질문 ID
|    ├─  image_id : 이미지 ID
|    ├─  question : 이미지 관련 질문
|    └─  answer : 질문에 대한 답변
├─  test.csv
|    ├─  ID : 질문 ID
|    ├─  image_id : 이미지 ID
|    └─  question : 이미지 관련 질문
└─  sample_submission.csv
     ├─  ID : 질문 ID
     └─  *answer : 질문에 대한 답변

3. Setup

• In Colab-PRO or PRO+ Users only
• Set up for sure GPU A100

Clone LLaVA

!git clone https://github.com/haotian-liu/LLaVA.git
%cd /content/LLaVA

Install

!pip install --upgrade pip
!pip install -e .
!pip install ninja
!pip install flash-attn --no-build-isolation

Clone Vicuna

!git clone https://huggingface.co/lmsys/vicuna-7b-v1.3

Download Data

# Download directly
!gdown https://drive.google.com/u/0/uc?id=1a9XB3r83ZCFWLOHBp8ooz3zQFl9rEIei&export=download

Preprocessing

• You could get 'output.json' and 'test.json' file
• If else, download our file and run it in your '/content' directory

%cd /content
!git clone https://github.com/pimang62/dacon-multimodal-vqa.git

%cd /content/dacon-multimodal-vqa
!python preprocessing.py

4. Run

• For recording wandb
  • put your API

%cd /content/LLaVA
!pip install wandb
!wandb login

• Train
  • /content/LLaVA/llava/train/train.py in 928 line, resume_from_checkpoint=False

!python /content/LLaVA/llava/train/train_mem.py \
    --model_name_or_path /content/LLaVA/vicuna-7b-v1.3 \
    --version v1 \
    --data_path /content/dacon-multimodal-vqa/output.json \
    --image_folder /content/dacon-multimodal-vqa/image/train \
    --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 /content/drive/MyDrive/llava \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 16 \
    --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 128 \
    --gradient_checkpointing True \
    --lazy_preprocess True \
    --report_to wandb

5. Re-training

• You should put 'vicuna' to your model-name
• output_dir folder should be contained 'checkpoint-*'
• num_train_epochs must have started from 2 or more

!python /content/LLaVA/llava/train/train_mem.py \
    --model_name_or_path /content/LLaVA/vicuna-7b-v1.3\
    --version v1 \
    --data_path /content/dacon-multimodal-vqa/output.json \
    --image_folder /content/dacon-multimodal-vqa/train \
    --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 /content/drive/MyDrive/llava/checkpoint-2400 \
    --num_train_epochs 2 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 16 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 2400 \
    --save_total_limit 1 \
    --learning_rate 2e-3 \
    --weight_decay 0. \
    --warmup_ratio 0.00 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 128 \
    --gradient_checkpointing True \
    --lazy_preprocess True \
    --report_to wandb

6. Inference

%cd /content

# go to your output directory
from google.colab import drive
drive.mount('/content/drive')

• You should change output_dir name 'checkpoint-' to 'llava-"
  • May be you might get a difference whether the name contains 'llava' or not

%cd /content/LLaVA
!python /content/dacon-multimodal-vqa/eval/model_vqa.py \
    --model-path /content/drive/MyDrive/llava/checkpoint/llava-2400 \
    --model-base lmsys/vicuna-7b-v1.3 \
    --question-file \
    /content/dacon-multimodal-vqa/test.jsonl \
    --image-folder \
   /content/image/test \
    --answers-file \
    /content/result.jsonl \

7. Submission

%cd /content/dacon-multimodal-vqa
!python submission.py