write domain generalization on color mnist into documentation and show an algorithm which works better than deepall

[smilesun]

make a markdown file which document

1. an image of comparing color mnist from different domains and how to get this done from domainlab via --sanitycheck
2. how to do classification using deepall from domainlab
3. how to do classification using diva, write the commandline (find a hyper-parameter that diva works better than deepall)
4. show the conterfactorial image generation of different domains of diva.

[Car-la-F]

@smilesun

I started a markdown file on the branch cf_colorMNIST_doc in docs/doc_MNIST_classification.md. I got until the point where I start to write about the image generation of diva. I also uploaded all images which are generated by dive using the --gen keyword (they can be found in docs/figs/colored_MNIST/gen_...).

I realised I have no idea how the images are generated exactly: I know that the generation is done by modifying the latent representation, for example by changing the latent variable for the domain, but I don't know how I can interpret the filenames to know how the latent space was modified.

Maybe you can have a look and point some thinks out.