VAIPE: Medicine Pill Image Recognition Challenge - Team: UIT4meow solution

Table of contents

• Environment Setup:
• Inference on testing data
• Train

---

Environment Setup:

We recommend to use Docker for building environment.

1. Build Docker Image with Dockerfile:

docker build -t uit4meow - < Dockerfile

2. Run Docker container:

docker run -it -v DATA_DIR:/data/:ro -v "$PWD":/workspace  --gpus all --ipc=host --name uit4meow uit4meow bash

Where DATA_DIR=<path to public_test folder>. The directory public_test should look like this:

data
|__ public_train
|     |__ pill 
|     |     |__ image
|     |           |__ VAIPE_P_0_0.jpg
|     |           |__ VAIPE_P_0_1.jpg
|     |           |__ VAIPE_P_0_2.jpg
|     |           |__ ...
|     |__ prescription
|           |__ image
|                 |__ VAIPE_P_TEST_0.png
|                 |__ VAIPE_P_TEST_1.png
|                 |__ VAIPE_P_TEST_2.png
|                 |__ ...
|__ public_valid
|     |__ pill 
|          |__ image
|                |__ VAIPE_P_0_0.jpg
|                |__ VAIPE_P_0_1.jpg
|                |__ VAIPE_P_0_2.jpg
|                |__ ...
|__ public_test
      |__ pill 
      |     |__ image
      |           |__ VAIPE_P_0_0.jpg
      |           |__ VAIPE_P_0_1.jpg
      |           |__ VAIPE_P_0_2.jpg
      |           |__ ...
      |__ prescription
            |__ image
                  |__ VAIPE_P_TEST_0.png
                  |__ VAIPE_P_TEST_1.png
                  |__ VAIPE_P_TEST_2.png
                  |__ ...

3. Install

pip install -r requirements.txt

# Compiling CUDA operators
cd DINO/models/dino/ops
python setup.py build install
cd ../../../..

---

Inference on testing data

For inference to submit, we provide a file script to run the complete pipeline. The file results.zip to submit will be in foler output.

Download weight

Download the pretrained models, you can download these files here and move them to folder weights , backbone to DINO/pretrained or run script below.

bash scripts/download_infer.sh

Inference

bash run.sh /data/public_test 0     # 0 is gpu-id

---

Train

Train on 2 GPUs Tesla P40 24GB

Yolov7

  python tools/generate_train_detect_pres.py
  cd yolov7
  CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node 2 train.py    \
        --epoch 50 --single-cls --workers 8 --device 0,1 --sync-bn     \
        --batch-size 8 --data data/coco.yaml --img 640 640      \
        --cfg cfg/training/yolov7x.yaml --weights ''     \
        --name yolov7x --hyp data/hyp.scratch.p5.yaml

VietOCR

  python tools/crop_pres.py
  CUDA_VISIBLE_DEVICES=0 python train_vietocr.py

DINO

  prepare data
  ln -s /data/public_train ./dataset/dino/train2017
  ln -s /data/public_val ./dataset/dino/val2017
  cp -r ./DINO/annotations ./dataset/dino

  prepare pretrained
  bash scripts/download_train.sh
  cd DINO
  CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 main.py \
        --pretrain_model_path './checkpoint0011_4scale_swin'	\
        --finetune_ignore label_enc.weight class_embed	\
        --output_dir logs_swin/4scale -c config/DINO/DINO_4scale_swin.py --coco_path ../dataset/dino_data \
        --options dn_scalar=100 embed_init_tgt=TRUE \
        dn_label_coef=1.0 dn_bbox_coef=1.0 use_ema=False \
        dn_box_noise_scale=1.0 backbone_dir='./pretrained'

FGVC-PIM

  python tools/crop.py
  cd classify
  CUDA_VISIBLE_DEVICES=0,1 python main.py --c ./configs/cfg.yaml