twairball / keras_lstm_vae Goto Github PK

It looks like the LSTM layers have a fixed number of timesteps. I think it would be even better if we could use a variable number of timesteps because the input sequences are not fixed in length.

Hello,

How can I use my own data, which contains categoric encoded sequences of length 1000?
(e.g. one sample of sequence is represented like this [1, 23, 6, 12, 4, 0, 0, 0 ...] )
And how should I decide the time step?

Thanks in advance!

should be
return z_mean + K.exp(z_log_sigma/) * epsilon

Is there a reason you compute a normal distributed tensor as your epsilon in the sampling Lambda, but actually use the epsilon value given as argument? Doesn't this go against the idea of the VAE?

Hello,

Thanks for your great implementation. I have one question: Somehow it is very hard for me to implement KL term annealing in Keras. Do you have any idea how to fix this problem?

Many thanks!

Hello,

I believe for h_decoded, we need to concat both z (latent variable) and x (input word). It cannot be using only z as input for the decoder.

Please correct me if I was wrong.

Best,

For batch_size = 2, I see ValueError: could not broadcast input array from shape (2,3,13) into shape (1,3,13).

I doubt it very much. Take care of using it. The author should be more responsible by sharing your code. Please consider to improve the code or write a warning to notice in your README.md. Thanks.

Hi,

The default config for the vae has an intermediate layer size of 32 and a latent layer size of 100. The example data are 3*13, so the input dimension is 39. Is there some reason that unlike most autoencoders, the "bottleneck" middle layer is larger than the input? There are auto-encoder architectures that do this, but usually they require weight regularization somewhere in the layer sequence.