kaggle_severstal_2019

Parts of my python/ pytorch code from Kaggle Severstal Steel Defect Detection competition 2019.

Description of competition

The task of this segmentation competition was to localize and classify surface defects on a steel sheet images. There were 12k training images (~6k with defects and ~6k without), 4 classes of defects. Labeled ground truth masks of defects were very noisy.

Files

• 05_1fold_experiments.ipynb - pipeline for segmentation 1 fold experiments;
• 05_5folds_experiments.ipynb - pipeline for segmentation 5 fold experiments;
• trainers.py - Trainer Class for segmentation 1fold and cross-validation;
• trainers_classification.py - Trainer class for multilabel classificaiton 1fold and cross-validation;
• samplers.py - Pytorch and my custom Samplers to sample images from dataset, including: SubsetSequentSampler, SubsetRandomSampler, ClassProbSampler;
• losses.py - several losses types for training, including: BCE, Dice, Focal, Tversky. Also they class weighted variants and combinations (ex. BCE-Dice);
• datasets.py - Pytorch datasets for segmentation and multilable classification;
• meter.py - Class for computing and monitoring of metrics, should be refactored;
• utils.py - visualization, seeds, rle coding methods, etc.;
• configs.py - Some constansts and lists of images to exclude from training.