junwenbai / c-dsvae Goto Github PK

Hello, I have a question about the generation of time variant latent z_t

I think in the paper z_t depends on z_0,...,z_t-1.

while in your code latents are sampled independently, although with a rnn hidden state.

features, _ = self.z_rnn(lstm_out)
z_mean = self.z_mean(features)
z_logvar = self.z_logvar(features)
z_post = self.reparameterize(z_mean, z_logvar, random_sampling=True)

I think the randomness of previous latent are not introduced to the cureent latent, I'm wondering whether this is the same meaning with the paper?

Thank you for sharing the code!

I have a question about the grid search. In Section 4.3 from the paper, it is written

Models are trained until convergence and the one with the best validation accuracy would be chosen for testing.

I wonder if the accuracy refers to the metric ACC. For example, were the TIMIT hyperparameters chosen based on the best speaker identification accuracy on the validation set?

Hi,
If you don't mind, could you share '../dataset/Sprite/data.pkl' with us?

Thank you for sharing your programs.
I attempted to reproduce C-DSVAE but couldn't find the edited dataset(data.pkl).
If you might see #1, Could you share the dataset or the program to create it?