Home

Awesome

GridMixup

A GridMixup augmentation, inspired by GridMask and CutMix

Easy install

pip install git+https://github.com/IlyaDobrynin/GridMixup.git

Overview

This simple augmentation is inspired by the GridMask and CutMix augmentations. The combination of this two augmentations forms proposed method.

Example

To run simple examples notebooks, you should install requirements:

pip install -r requirements.txt

Simple examples are here: demo and pipeline demo

TlDr:

from gridmix import GridMixupLoss

gridmix_cls = GridMixupLoss(
    alpha=(0.4, 0.7),
    hole_aspect_ratio=1.,
    crop_area_ratio=(0.5, 1),
    crop_aspect_ratio=(0.5, 2),
    n_holes_x=(2, 6)
)

images, targets = batch['images'], batch['targets']
images_mixed, targets_mixed = gridmix_cls.get_sample(images=images, targets=targets)
preds = model(images_mixed)
loss = criterion(preds, targets_mixed) 

Before<br> <br>

After<br>

GridMixup loss defined as:

lam * CrossEntropyLoss(preds, trues1) + (1 - lam) * CrossEntropyLoss(preds, trues2)

where:

Parameters

GridMixupLoss takes follow arguments: