This is some unit test for ML algorithms.

This is my origin implementation of SimCLR.

The experiment is made on mini-imagenet dataset.(If the task is too easy, there will be no different between using SimCLR and not)

My experiment shows there's difference on 3 layers convolution network, but no difference for restnet20.

This is result of 3 layers conv version, showing improvment with SimCLR:

This is resultof of resnet20 version, showing no difference betweeen using SimCLR or not:

make the nerual network to know coordinate by channel of up-down and left-right.

The easiest way to implement discrete latent variable. I know it by the paper of model-based learning paper.

For updating the codebook, there're two different approach: one is normal gradient update; another one is Exponential Moving Average. The implementation of EMA is pretty lousy. But the result is astonishing: it convergs much faster since it wouldn't affect by bad gradient of decoder and encoder.

I adapt the decomposed trick for ease the index collapse problem. Decomposed trick also accelerate converge.

This is result of my experiment: