Introduction

This is an open-source research library for polyp segmentation. Model configuration follows OpenMMLab style while other is implemented by us to conduct experiments faster.

It helps you have better understanding of your model by providing some useful debug tools:

• CAM-based visualization
• Per channel visualization
• Model params and flops count

This project belongs to Sun-Asterisk Inc.

Installation

Install pytorch

• Version: 1.10.1 (recommended)
  conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio==0.10.1 cudatoolkit=11.3 -c pytorch -c conda-forge

Install MMSegmentation

• Install openmim: pip install openmim
• Install mmcv: mim install mmcv-full==1.6.0
• Install mmseg:

cd sun-polyp
pip install -v -e .

Install dependencies

• Install wandb (for logging): pip install wandb
• Install pytorch-lightning: pip install pytorch-lightning
• Install segmentation model pytorch: pip install segmentation-models-pytorch

Config

Config everything in mcode/config.py
What to config?

• Model:
  • Follow mmseg config
  • pretrained: path to ImageNet pretrained MiT backbone
  • Please change pretrained in backbone to pretrained=pretrained
  • Config model head to head of your choice
• Wandb:
  • use_wandb: True, False if debug
  • wandb_key: Please use your wandb authorize key
  • wandb_name: Name of your experiments, please make it as specific as possible
  • wandb_group: We need 5 runs/experiments, grouping makes it easier to see on wandb UI
• Dataset:
  • train_images: path to image in Train Dataset
  • train_masks: path to mask in Train Dataset
  • test_folder: path to Test Dataset
  • test_images and test_masks: leave it
  • num_workers: torch workers
  • save_path: path to save checkpoints and logs
  • bs: this should be 16 if possible
  • grad_accumulate_rate: num iters to backward, if bs=16, this should be 1