mratsim / Amazon-Forest-Computer-Vision

Amazon Forest Computer Vision: Satellite Image tagging code using PyTorch / Keras with lots of PyTorch tricks

Amazon Forest Computer Vision

Satellite Image tagging code using PyTorch / Keras

Here is a sample of images we had to work with

Source: https://www.kaggle.com/c/planet-understanding-the-amazon-from-space/data

Note: the repo was developed in May 2017 on PyTorch 0.1. PyTorch was publicly announced in January 2017 and has seen tremendous changes since then.

You will find:

• A script that output the mean and stddev of your image if you want to train from scratch

• Using weighted loss function

• Logging your experiment

• Composing data augmentations, also here. Note use Pillow-SIMD instead of PIL/Pillow. It is even faster than OpenCV

• Loading from a CSV that contains image path - 61 lines yeah

• Equivalent in Keras - 216 lines ugh. Note: so much lines were needed because by default in Keras you either have the data augmentation with ImageDataGenerator or lazy loading of images with "flow_from_directory" and there is no flow_from_csv

• Model finetuning with custom PyCaffe weights

• Train_test_split, PyTorch version and Keras version

• Weighted sampling training so that the model view rare cases more often

• Custom Sampler creation, example for the balanced sampler

• Saving snapshots each epoch

• Loading the best snapshot for prediction

• Failed word embeddings experiments to combine image and text data

• Combined weighted loss function (softmax for unique weather tags, BCE for multilabel tags)

• Selecting the best F2-threshold via stochastic search at the end of each epoch to maximize validation score. This is then saved along model parameter.

• CNN-RNN combination (work in progress)