Why is the gumbel_softmax not enabled for the YELP sentiment transfer?

[naveen-kinnal]

Hello. I want to know why is the gumbel_softmax not set to True in the model params for the YELP sentiment transfer task? Is it because we have a pre-trained language model and hence do want to compute the KL loss with respect to the gumbel softmax log prior? Thanks.

[jxhe]

I think we use gumbel_softmax the same way across all tasks to propagate gradients from the language model loss. For reference, gumbel_softmax is used here and is enabled in Yelp sentiment transfer:

https://github.com/cindyxinyiwang/deep-latent-sequence-model/blob/8a798582b1af5ef7f6ac4ca1f2138fd382a1cb06/src/model.py#L444

[naveen-kinnal]

Thank you. But the above line doesn't get reached at all. In this below code, the self.hparams.gumbel_softmax is not set to True and hence it will reach the if condition (get_translations), but not the get_soft_translations.

if self.hparams.bt:
      if eval or (not self.hparams.gumbel_softmax):
        with torch.no_grad():
          x_trans, x_trans_mask, x_trans_len, index = self.get_translations(x_train, x_mask, x_len, y_sampled, y_sampled_mask, y_sampled_len, temperature)
        index = torch.tensor(index.copy(), dtype=torch.long, requires_grad=False, device=self.hparams.device)
      else:
        # with torch.no_grad():
        lm_flag = True
        x_trans, x_trans_mask, x_trans_len, index, org_index, neg_entropy = self.get_soft_translations(x_train, x_mask, x_len,
          y_sampled, y_sampled_mask, y_sampled_len)
      trans_length = sum(x_trans_len)
    else:
      trans_length = 0.

https://github.com/cindyxinyiwang/deep-latent-sequence-model/blob/8a798582b1af5ef7f6ac4ca1f2138fd382a1cb06/src/model.py#L250

[jxhe]

The back-translation loss does not use Gumbel-softmax as explained in the paper, only the LM loss uses Gumbel-softmax here:

https://github.com/cindyxinyiwang/deep-latent-sequence-model/blob/8a798582b1af5ef7f6ac4ca1f2138fd382a1cb06/src/model.py#L266

[naveen-kinnal]

I think the param 'lm' is also not set to true for the YELP style transfer. Could you please let me know in which file we set the 'lm' to true?

[jxhe]

https://github.com/cindyxinyiwang/deep-latent-sequence-model/blob/8a798582b1af5ef7f6ac4ca1f2138fd382a1cb06/scripts/yelp/train_yelp.sh#L45

[shentianxiao]

But I don't see "--gs_soft" set. As a result, x_trans_lm is a hard generation, then how can the gradients of log_prior backprop?

[jxhe]

Hi,

When --gs_soft is not set, gumbel_softmax will output discretized one-hot vectors, but will be differentiated as if it is the soft sample in autograd. Maybe also note that our language model multiplies the embedding weights with those one-hot vectors to obtain word embeddings, thus the entire process is differentiable:

https://github.com/cindyxinyiwang/deep-latent-sequence-model/blob/8a798582b1af5ef7f6ac4ca1f2138fd382a1cb06/src/lm_lstm.py#L77