search

galib360 / FaceXHuBERT

Other
95 stars 7 forks source link

About the training #1

Open TCK1999 opened 1 year ago

TCK1999 commented 1 year ago

Hi, I trained the "BIWI" dataset according to your solution, but I found that the results are not as expected. the current loss does not decrease and the test set does not predict well,so I wonder if you have encountered a similar problem or know some possible solutions. Thank you!

galib360 commented 1 year ago

Hi, Did you prepare the dataset and run the preprocessing script according to the documentation in the repository README? In my experience, using the dataset without preprocessing did not learn anything like you mentioned.

TCK1999 commented 1 year ago

Thanks for your reply! Are you referring to the vertex_process.py script? Yes, I have run that script and pre-processed the dataset according to the README instructions. I also tried to visualize the preprocessed vertices_npy file by rendering it with render_result.py, but I found that the results are not good, the difference between the before and after frames is very far. So I think there is indeed some error in preprocessing the dataset, but I don't know what the problem is. I am confused about this, so do you have any possible suggestions for this? Again, thank you!