edersantana / gumbel Goto Github PK

Thanks for uploading this. However, there is an error.

In the loss computation, you have elbo = data_dim * bce(x, x_hat) - KL.

It should be elbo = data_dim * bce(x, x_hat) + KL instead.

I can't find this module in here:
https://github.com/AgnezIO/agnez.git@55657e5148ce58493f1847a
would u uploAD it again pls?

Hi， thanks for the good work. However I am confused how you sample the Gumbel noise.
In the sampling function, you just sample the 'Gumbel noise' from a 'uniform' distribution:
U = K.random_uniform(K.shape(logits_y), 0, 1)
y = logits_y - K.log(-K.log(U + 1e-20) + 1e-20) # logits + gumbel noise

I wander if it is right??? Looking forward for the answer~ Thanks

Hi Eder,

Thanks for the Keras implementation! I'm excited about this idea as well. (Hopefully) quick question: I'm interested in learning an autoencoder for the output features of some other CNN (VGG16), so I'd like to maximize the cosine similarity of x and x_hat. Presumably this means I have to change

return data_dim * binary_crossentropy(x, x_hat) - KL 

to something else, since it doesn't seem like binary_crossentropy makes sense in this context. Any thoughts on what it should be changed to? (I'm new to variational methods, so I'm trying to derive the answer myself right now, but thought I'd post the question here as well..)

Thanks!
Ben