Introduction

A repository contains WIP code for 2020 NAIC competitions. Competition details can be found at https://naic.pcl.ac.cn/frame/2.

Install

Requirements

Testing environment:

• CentOS Linux 7 (Core)
• Python 3.6
• PyTorch 1.6
• CUDA 10.1
• segmentation-model-pytorch 0.1.2

Installation

a. Create a conda virtual environment and activate it.

conda create -n naic python=3.6 -y
conda activate naic

b. Install PyTorch and torchvision following the official instructions. Here we use PyTorch and CUDA 10.1. You may also switch to other vision by specifying the version number.

conda install pytorch=1.6.0 torchvision cudatoolkit=10.1 -c pytorch

c. Install project requirements by.

pip install -r requirements.txt

Getting Started

Prepare datasets

Project has following structure. Dataset is recommended to be placed at $PROJECTROOT/data although it can be configured via cfg.ini.

├── cfg.ini
├── inference.py
├── plot_latest_log.sh
├── requirements.txt
├── src
│   ├── data
│   │   ├── aug.py
│   │   ├── dataset.py
│   │   ├── histequal.py
│   │   └── __init__.py
│   ├── __init__.py
│   ├── logger.py
│   ├── loss.py
│   ├── metrics.py
│   ├── optimizer.py
│   ├── scheduler.py
│   └── utils.py
├── tools
│   ├── check_mask.py
│   ├── correct_small_area.py
│   ├── count_connected_component.py
│   ├── __init__.py
│   ├── mixup.py
│   ├── plot_logs.py
│   └── voting.py
├── data
│   ├── image_A
│   ├── results
│   └── train
│       ├── image
│       ├── label
│       └── train_aug
│           ├── image
│           └── label
└── train.py

Training

We created label.csv file to split data into train and validation set and used this label file in our dataset class src/data/dataset.py.

image,mask,label
48366.tif,48366.png,train
60305.tif,60305.png,train
267981.tif,267981.png,train
230520.tif,230520.png,val
432758.tif,432758.png,val
49156.tif,49156.png,val

You can also write your own dataset class and ignore this file. Please change line 162 and line 172 correspondingly in train.py if you used your own dataset.

Currently, we are using segmentation-model-pytorch API for our model's architectures and encoders. All supportive models can be found at https://github.com/qubvel/segmentation_models.pytorch.

Train a model

You can set your training parameters via CML and training a model like:

python train.py --encoder resnext50_32x4d -w imagenet --arch unet -b 64 -lr 5e-5 -wd 5e-6 --num_workers 12 --num_epoch 100 --parallel

By default, model will be saved at $PROJECTROOT/checkpoints and log will be saved at $PROJECTROOT/logs. You can resume your model via --reusme.

Inference

We provide inference script to output the prediction to desired directory. Simple using the following command to set checkpoint, input_dir and output_dir.

python inference.py --checkpoint checkpoints/deeplabv3plus_efficientnet-b3/202010041423/epoch_209_0.8191.pth --input_dir data/image_A --output_dir ./results

Acknowledgement

@misc{Yakubovskiy:2019,
  Author = {Pavel Yakubovskiy},
  Title = {Segmentation Models Pytorch},
  Year = {2020},
  Publisher = {GitHub},
  Journal = {GitHub repository},
  Howpublished = {\url{https://github.com/qubvel/segmentation_models.pytorch}}
}