Simple Self-Supervised Learning Experiments
ImageNet
Preprocess
Copy the training set all into one folder.
python preprocess/imagenet/move_to_one_folder_train.py
And then manually copy over the validation folder since there's no subfolders.
Preprocess to generate the trainval folder
python preprocess/imagenet/2_preprocess_rot.py
Vanilla training
python code/imagenet_rotation/vanilla_train.py
Training
Old VOC stuff
Get all the images into a training and validation folders.
python preprocess/voc/1_organize.py
Rotate images and put them into training and validation folders
python preprocess/voc/2_preprocess_jigsaw.py
Train the model and print out the validation loss each time.
python code/voc_rotation/vanilla_train.py
Visualize the training and validation accuracy of the vanilla model.
python code/visualization/plot_training_graphs.py
Get the predictions and confidences for the model.
python code/voc_rotation/get_val_confidences.py
Visualize the predicted distributions of the vanilla model (now old code).
python code/visualization/plot_predicted_distributions.py
Get data subsets.
python code/voc_rotation/create_data_subsets.py
Retrain from a certain epoch using a subset of the data.
code/voc_rotation/train_continue_correct.py