Link Search Menu Expand Document
UnMICST

UnMICST-M - (Mask R-CNN)

Requirements

  • Linux
  • Python 3.6
  • PyTorch 1.5.1
  • CUDA 10.1

Installation Instructions 

conda create --name unmicst python=3.6
conda activate unmicst
pip install -r requirements.txt

Operation Instructions

Dataset

Training data can be downloaded from /training data from https://www.dropbox.com/sh/3aqp83f5w1pxk0y/AABFgNRMJD2EvfSLFgCrXrBba?dl=0
The dataset is supposed to be arranged below

RootFolder
├── test  
│   ├── *_Img.tif
├── train
│   ├── *_Img.tif
├── valid
│   ├── *_Img.tif
├── coco_cellsegm_test.json
├── coco_cellsegm_train.json
├── coco_cellsegm_valid.json

Train

  • Set nproc_per_node and world-size as the number of GPUs to use
  • root-path is a path to a folder that contains train / val / test data
  • output-dir is a path to save trained models

DNA Channel / Real Augmentation

./DNA_Aug.sh

DNA Channel / Gaussian Augmentation

./DNA_GaussianAug.sh

DNA Channel / No Augmentation

./DNA_NoAug.sh

DNA + NES Channels / Real Augmentation

./DNA_NES_Aug.sh

DNA + NES Channels / No Augmentation

./DNA_NES_NoAug.sh

Test

  • Set nproc_per_node and world-size as the number of GPUs to use
  • use-channel is either dapi/both
  • testdomain is one of clean/topblur/bottomblur
  • resume is a path to a saved model to test
  • root-path is a path to a folder that contains train / val / test data
  • output-dir is a path to save trained models

Command for Testing

./UnMICST_M_Test.sh