odach 0.1.5-2107010420

ODAch is a test-time-augmentation tool for pytorch 2d object detectors.

ODAch, An Object Detection TTA tool for Pytorch

ODA is a test-time-augmentation (TTA) tool for 2d object detectors.

For use in Kaggle object detection competitions.

:star: if it helps you! ;)

Install

pip install odach

Usage

See Example.ipynb.

The setup is very simple, similar to ttach.

Singlescale TTA

import odach as oda
# Declare TTA variations
tta = [oda.HorizontalFlip(), oda.VerticalFlip(), oda.Rotate90(), oda.Multiply(0.9), oda.Multiply(1.1)]

# load image
img = loadimg(impath)
# wrap model and tta
tta_model = oda.TTAWrapper(model, tta)
# Execute TTA!
boxes, scores, labels = tta_model(img)

Multiscale TTA

import odach as oda
# Declare TTA variations
tta = [oda.HorizontalFlip(), oda.VerticalFlip(), oda.Rotate90(), oda.Multiply(0.9), oda.Multiply(1.1)]
# Declare scales to tta
scale = [0.8, 0.9, 1, 1.1, 1.2]

# load image
img = loadimg(impath)
# wrap model and tta
tta_model = oda.TTAWrapper(model, tta, scale)
# Execute TTA!
boxes, scores, labels = tta_model(img)

• The boxes are also filtered by nms(wbf default).

• The image size should be square.

model output wrapping

• Wrap your detection model so that the output is similar to torchvision frcnn format: [["box":[[x,y,x2,y2], [], ..], "labels": [0,1,..], "scores": [1.0, 0.8, ..]]

• Example for EfficientDets https://www.kaggle.com/kyoshioka47/example-of-2d-single-scale-tta-with-odach/

# wrap effdet
oda_effdet = oda.wrap_effdet(effdet)
# Declare TTA variations
tta = [oda.HorizontalFlip(), oda.VerticalFlip(), oda.Rotate90()]
# Declare scales to tta
scale = [1]
# wrap model and tta
tta_model = oda.TTAWrapper(oda_effdet, tta, scale)

Example

Global Wheat Detection

Example notebook

Thanks

nms, wbf are from https://kaggle.com/zfturbo

tta is based on https://github.com/qubvel/ttach, https://github.com/andrewekhalel/edafa/tree/master/edafa and https://www.kaggle.com/shonenkov/wbf-over-tta-single-model-efficientdet