psyai-net / selftalk_release Goto Github PK

https://github.com/psyai-net/SelfTalk_release/blob/f9b5e0a0597ff9690761b58242e0d91c51040ba5/SelfTalk.py#L133C53-L133C63
Maybe the second param of the self.transformer should be the lip_gt ?

I have imported the .ply file of the template into blender. I found that the Measuring Unit is m not mm. When I use the measure tool to roughly measure the face size of BIWI and vocaset, I found that vocaset might be accurate while BIWI was so big. This seems inconsistent with the evaluation error in paper, not only this model, but alse CodeTalker, FaceFormer and other related model. I think if I have missed something? Or the measurement in blender is incompatible with the .ply data? Or have you done some process on the data or evaluation?
By the way, I wonder how to evaluate the error. I found the one hot input only refers to 8 persons as training set according to the class Dataset "self.one_hot_labels = np.eye(len(subjects_dict["train"]))" in data_loader.py. But for the test_pred.py, when input the test data of the test subject, the one hot input is still the 8 persons'. I am confused about this. Thanks for answering the issue!

May I ask if anyone can provide the metric calculation code?

Hi,
This is a nice job!
Is there any way to convert flame vertex to 52 blendshapes?

Hello author - thanks for the fantastic work!

I am able to test the pretrained models with the subjects present and I am currently looking to try with my custom video. Can you please guide me how to create vertices_npy folder for my video?

Thank you in advance!

@ZiqiaoPeng Hello. I have learn your paper and code rencently, and I have serval questions. May I ask you how to cal the FDD metric of the vocaset-test ? I know the FLAME has some masks like forehead or lips, are you use some of them to cal it ? And I also want to know that the model how to learn the 'talking sytle' of a unseen subject (the subject in voca-test dataset), or it just learn a 'mean' talking style ?

Hi ZiQiao, thank you so much for your work
I found in your paper that you have calculated LVE on VOCA-test dataset,since I found that after using FLAME_masks. pkl to get the lip vertices, the results calculated by GT are not quite the same as yours, I suspect that there is a problem with my processing could you please provide your testing process?
Thank you again for your work

wav2vec.py line.111

hidden_states = self.feature_projection(hidden_states)[0]

should be

hidden_states = self.feature_projection(hidden_states)

or you'll encounter an error.

    batch_size, sequence_length, hidden_size = hidden_states.size()
ValueError: not enough values to unpack (expected 3, got 2)