why kl loss is not used?

[Doragd]

hello, thanks for you code about this paper. it helps me a lot. but from your code at linguistic_style_transfer_pytorch/model.py line 157: vae_and_classifier_loss

 vae_and_classifier_loss = mconfig.content_adversary_loss_weight * content_entropy_loss + \
            mconfig.style_adversary_loss_weight * style_entropy_loss + \
            mconfig.style_multitask_loss_weight * style_mul_loss + \
            mconfig.content_multitask_loss_weight * content_mul_loss + \
            reconstruction_loss

why kl loss is not used?

by the way, at train.py line 19

model = AdversarialVAE(inference=False, weight=weights)

This parameter inference does not seem to define?

[h3lio5]

@Doragd I'm glad that my code could be of help to you 😊.

I have unknowingly omitted KL loss' contribution to the final loss. I also forgot to remove the inference parameter before pushing the final train.py script. I will fix the errors at the earliest. Thank you very much for reporting the errors.

[Doragd]

thanks for your quick reply. 👍 By the way, i find some other errors. it maybe typos. linguistic_style_transfer_pytorch/utils/train_w2v.py line 11:

model = Word2Vec(sentences=LineSentence(text_file_path),
                          min_count=1, size=config.embedding_size)

the config is not found and should be added into the function parameters or defines the global variable.

and in linguistic_style_transfer_pytorch/utils/vocab.py line 77: self.vocab_save_path+'word2index.json')) should be self.vocab_save_path+'/word2index.json')) Missing backslash~ the same errors are also seen in: line 81 and line 112

😄 by by by by the way, the repo is missing requirements.txt ?

[h3lio5]

@Doragd I am really sorry that I couldn't give you a quick reply as I was busy this whole week with exams at my college. All the errors that you have mentioned above seem to arise because I replaced all the hard-coded file path names in the config.py file using os.path.join(... , ...) which doesn't append \, in my final push. I'll fix all the bugs you have mentioned above right away. I'm sorry that I forgot to add the requirements.txt file, I'll add it right away.

I'm really thankful to you for pointing out all the bugs 😃. Feel free to ask any other queries!

[Doragd]

@h3lio5 thanks for your help~ by the way, there is another bug when running on the GPU at model.py line 249.

epsilon = torch.randn(mu.size(1))

should be

epsilon = torch.randn(mu.size(1)).cuda()

[h3lio5]

@Doragd Fixed it. And please feel free to report any other bugs 😃.

[Doragd]

@h3lio5 I have run the code on my server successfully! It's so wonderful！But I cannot find the paper‘s experiment code: Experiment I: Disentangling Latent Space. Could you have time to reproduce this part?

[h3lio5]

@Doragd I don't have the trained model weights as the whole project folder has been deleted from the computer in the CS Dept. where I trained the model on. I'll have to retrain the model from scratch but I'm not sure if I will be able to get enough gpu compute time at the lab right now.

[michaeldu1]

@Doragd @h3lio5 I've opened another issue, but it seems like generate.py is working well for you guys. Would it be at all possible to check out the issue I've opened/offer any advice on how to avoid shape errors when running target_tokenids = model.transfer_style(token_ids, target_style_id) in generate.py? Thanks so much!! :)

[arijit1410]

@Doragd @h3lio5 I've opened another issue, but it seems like generate.py is working well for you guys. Would it be at all possible to check out the issue I've opened/offer any advice on how to avoid shape errors when running target_tokenids = model.transfer_style(token_ids, target_style_id) in generate.py? Thanks so much!! :)

I'm facing the same problem

[Doragd]

@michaeldu1 @h3lio5 @arijit1410 about the shape error, you should change the final hidden state shape and concat with the word_emb at every time step. the author seems to forget it. at transfer_style function, model.py line 504: final_hidden_state = final_hidden_state.transpose(1,0).contiguous().view(embedded_seq.size(0), -1) at generate sentence function, model.py line 466:

for idx in range(mconfig.max_seq_len):
                    hidden_state = self.decoder(torch.cat([word_emb,latent_emb], axis=1), hidden_state)

[sharan21]

Hey everyone, I also faced many many trivial device/shape mismatch mistakes and have corrected almost all of them. I am trying to run generate.py but again there is a shape mismatch in generate_sentences() in model.py. Can you please report of any of you finally managed to get good results from this code? I just hope all my efforts arent in vain.

[Doragd]

Hey everyone, I also faced many many trivial device/shape mismatch mistakes and have corrected almost all of them. I am trying to run generate.py but again there is a shape mismatch in generate_sentences() in model.py. Can you please report of any of you finally managed to get good results from this code? I just hope all my efforts arent in vain.

Sorry, I've given up this project and forgotten about any details.

[arijit1410]

Yeah same. This guy is just padding repos to this GitHub, and this codebase was an absolute train wreck and horribly written.

On Sat, 17 Apr, 2021, 8:53 pm Gordon Lee, @.***> wrote:

Hey everyone, I also faced many many trivial device/shape mismatch mistakes and have corrected almost all of them. I am trying to run generate.py but again there is a shape mismatch in generate_sentences() in model.py. Can you please report of any of you finally managed to get good results from this code? I just hope all my efforts arent in vain.

Sorry, I've given up this project and forgotten about any details.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/h3lio5/linguistic-style-transfer-pytorch/issues/2#issuecomment-821839892, or unsubscribe https://github.com/notifications/unsubscribe-auth/AE5ZI6NC76NFPXR47YLITYLTJGRYHANCNFSM4K6ZZHCQ .