limingwu8 / UNet-pytorch

UNet in pytorch for Kaggle 2018 data science bowl nuclei segmentation

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

UNet-pytorch

Overview

This is the code for kaggle 2018 data science bowl nuclei segmentation (https://www.kaggle.com/c/data-science-bowl-2018). We will use UNet to perform the segmentation task.

Dependencies

  • numpy
  • scipy
  • tqdm
  • pillow
  • scikit-image
  • pytorch
  • pandas

Usage

  1. Download the dataset from Kaggle (https://www.kaggle.com/c/data-science-bowl-2018/data).

  2. Create two folders called combined and testing_data. Run script utils.py to prepare training image and testing image, the prepared image will be inside combined and testing_data folder.

  3. In class Option under script utils.py, set is_train = True and adjust three dirs and other parameters.

  4. Run script train.py. The model will be saved under folder checkpoint.

  5. When making prediction using testing data, set train=False in utils.py, and run script train.py again. The prediction masks will be saved to the folder specified in Option class under utils.py, and the run-length-encoding csv file will be saved in current folder.

Training results

U-Net Architecture

image1

Some examples of prediction masks

image2

Evaluation

image3

About

UNet in pytorch for Kaggle 2018 data science bowl nuclei segmentation

License:MIT License


Languages

Language:Python 100.0%