MNIST datasets are used for training to get their respective embeddings.
I have here used 2-dimensional embeddings which is not the best choice for other complicated datasets but works fine for easy datasets like "MNIST".
The architecture details of the embedding network used is:
(32 conv 5x5 --> PReLU --> MaxPool 2x2 --> 64 Conv 5x5 --> PReLU --> MaxPool 2x2 --> Dense 256 --> PReLU --> Dense 256 --> PReLU --> Dense 2)
The architecture details of the classification network used is:
The loss function used is the "cross entropy loss". The embeddings obtained for the datasets are in the "outputs" folders respectively.