Before training, the scripts will first download datasets to a dataset folder. For image classification, train with

bash train_mnist.sh

And evaluate the model with best checkpoint with this script:

bash visualize_mnist.sh

This script will generate visualization results in images folder and corresponding information on test set. mnist_feature_map%d_%s.png represents the feature map on %d layer with %s visualization method. For example, mnist_feature_map3_tsne.png is the result of feature map of the third layer with t-SNE visualization method. mnist-train-loss-acc.png represents the change of loss and accuracy on training set.

Train with the following script:

bash train_fashion_mnish.sh

And evaluate the model with best checkpoint with this script:

bash visualize_vae.sh

This script will generate visualization results in images folder. vae-image1/2.png are randomly generated images with two arbitrary Gaussian noise. vae-%.2f-merge.png is the image generated with interpolation.

Train with the following script:

bash train_sst2.sh

And evaluate the model with best checkpoint with this script:

bash visualize_sst2.sh

This script will generate visualization results in images folder. sst2_feature_map%d_%s.png represents the feature map on %d layer with %s visualization method. sst2-%s-loss-acc.png represents the loss and accuracy change figure either in training set or in validation set.