Binary Cross Entropy instead of MSE

[JanoschMenke]

Hi,

maybe a quite simple question but in your Regression example you pass the

'nll_func': <function neuralfingerprint.util.rmse as the nll_func via build_conv_deep_net to the build_standard_net function. However, I am not sure how the utils.rmse relates to the mean_squared_error function that is being used in the build_standard_net.

My goal is to adapt the example code so that I can do a binary classification.
I tried to replace the default loss from build_standard_net with the binary_cross_entropy. But I think I am missing something because the results do not make sense:

[duvenaud]

There might be a bug in that binary_cross_entropy method, try using the code from here:

https://github.com/HIPS/autograd/blob/master/examples/variational_autoencoder.py#L26