SIIM-ISIC Melanoma Classification

This repo contains my code for training classification model for kaggle recently held competition SIIM-ISIC Melanoma Classification.This code reproduce one of my best single model.

Table of contents

• SIIM-ISIC Melanoma Classification
• Table of contents
• Project Structure
• Installation
• Training

Project Structure

(top)

./
├── config.yaml
├── data
├── logs
├── README.md
├── requirements.txt
├── src
│   ├── classifier.py
│   ├── create_folds.py
│   ├── dataset.py
│   ├── early_stopping.py
│   ├── logger.py
│   ├── models.py
│   ├── resize_data.py
│   ├── train.py
│   └── utils.py
├── subs
└── weights

Installation

(top)

1. Setup virtual environment(using conda or virtualenv).

   conda create -n kaggle python=3.7
   conda activate kaggle
   pip install -r requirements.txt
   

   NOTE : I am using pytorch 1.6 in this code, earlier version of pytorch wont work without making changes to the code.

2. Download the dataset(jpeg folder) and zip it in data folder from kaggle. After downloading data folder should look like this

   data/
   ├── jpeg
   │   ├── test
   │   └── train
   ├── sample_submission.csv
   ├── test.csv
   └── train.csv
   

Training

(top)

1. Create folds.

   python create_folds.py

2. Resize images. Original images are quite large so loading large images while training will slow training. So resize dataset to smaller size (384,512,1024 etc.) by changing shape in resize_data.py and then run

   python resize_data.py

   This will create two folders train_image and test_images in the data folder containing training image and testing images respectively.

3. Start training

   python train.py

   weights, logs and submission file will be saved in weights, logs, subs folder respectively.

Reference

• wtfml