I ran the original code for NVDM and found that the kld loss curve decreased and then quickly increased.

According to the ELBO, B = -KL[q(z|x) || p(z)] + E[p(x|z)]. Maximizing this bound is equivalent to minimizing KL[q(z|x) || p(z)] and maximizing E[p(x|z)]. The KLD curve should approach to zero, because KL[q(z|x) || p(z)] >= 0.

The experimental result is strange.

I can't understand why setting matrix_start_zero=True when computing logsigm at line 54, nvdm.py , would you mind making a brief explanation, thanks.

I ran into minor issues when using Python 3.

1. 'xrange' changed to 'range'
2. change the lines of the function, 'create_batches'
   from ids=range(data_size) to ids = np.arange(data_size) which consequently, changing the padding line to
   batches.append(np.append(ids[-rest:], ([-1] * (batch_size - rest)))) # -1 as padding

An updated version of Tensorflow changed the matrix multiplication from 'tf.mul' to 'tf.multiply'.

Hi,

I am having trouble with running the script. Here is the error code:

could anyone take a look at the code? It seems there are errors in utility code( e.g., batch size and indexing).

Thank you!

I assume that row number of each word in vocabs.new is the index of the word, and the integer after each word is counts of that word in corpus, am I right ?

In the paper, it says that the 20news dataset has 11,314 training and 7,531 test articles.
However, why the data you give in this repository has only 11,268 training and 7,505 test articles？

Hi! So the perplexity calculation here is (per line 140 from "train" in nvdm.py):
print_ppx = np.exp(loss_sum / word_count)

However, loss_sum is based on the sum of "loss" which is the result of "model.objective" i.e. the sum of reconstruction loss (cross-entropy) and K-L Divergence.
Lines 129-132 from "train" in nvdm.py

_, (loss, kld) = sess.run((optim, 
                                    [model.objective, model.kld]),
                                    input_feed)
          loss_sum += np.sum(loss)

Line 78 the model definition in nvdm.py
self.objective = self.recons_loss + self.kld

I thought Perplexity is usually the exponentiated form of the normalized cross-entropy, so is there a technical reason for using the result of model.objective instead of model.recons_loss to calculate the perplexity or is that a bug? I bet numbers should only get better if this is corrected (as KL Divergence is non-negative)

Thank you! (for my purposes a commercially-friendly license, e.g. Apache / MIT / Berkeley would be awesome.)