• ColonFormer: An Efficient Transformer based Method for Colon Polyp Segmentation
• • • Environment
    • Dataset
    • Training
    • Evaluation

ColonFormer: An Efficient Transformer based Method for Colon Polyp Segmentation

This repository contains the official Pytorch implementation of training & evaluation code for ColonFormer.

Environment

• Creating a virtual environment in terminal: conda create -n ColonFormer
• Install CUDA 11.1 and pytorch 1.7.1
• Install other requirements: pip install -r requirements.txt

Dataset

Downloading necessary data:

1. For Experiment 1 in our paper:
   • Download testing dataset and move it into ./data/TestDataset/, which can be found in this download link (Google Drive).
   • Download training dataset and move it into ./data/TrainDataset/, which can be found in this download link (Google Drive).
2. For Experiment 2 and Experiment 3:
   • All datasets we use in this experiments can be found in this download link (Google Drive)

Training

Download MiT's pretrained weights on ImageNet-1K, and put them in a folder pretrained/.
Config hyper-parameters and run train.py for training. For example:

python train.py --backbone b3 --train_path ./data/TrainDataset --train_save ColonFormerB3

Here is an example in Google Colab

Evaluation

For evaluation, specific your backbone version, weight's path and dataset and run test.py. For example:

python test.py --backbone b3 --weight ./snapshots/ColonFormerB3/last.pth --test_path ./data/TestDataset

We provide some pretrained weights in case you need.