This git is an unofficial implementation of the following paper UNSUPERVISED REPRESENTATION LEARNING PREDICTING IMAGE ROTATIONS - 2018

The idea is to first augment the dataset by generating random rotations of the images by 0, 90, 180, or 270 degrees. The second step is to train the model on this dataset to learn to predict the image rotation. Finally, use the obtained weights as initialization to train on a much smaller labeled dataset.

Results with 30 epochs and default training parameters, trained on GPU from Colab:

Training only the FC layer on whole dataset CASE1:: test_accuracy: 66.05% ; test_XELoss: 1.0 -- time 00:05:36

Training only the FC layer on 10% of dataset CASE1:: test_accuracy: 57.43% ; test_XELoss: 1.33 -- time 00:00:54

Training only the CNN layers on whole rotation dataset ROTATION MODEL:: test_accuracy: 98.0% ; test_XELoss: 0.07 -- time 00:01:26

Training FC only on 10% of dataset while using the CNN layers trained on rotation model CASE2:: test_accuracy: 73.08% ; test_XELoss: 0.77 -- total time 00:02:22