Invalid compatible shapes caused by op LogicalAnd in triplet_loss.py

[bjfu-ai-institute]

Invalid compatible shapes caused by op LogicalAnd in triplet_loss.py

[vzxxbacq]

Hello @pranoot , I'm gald to that my code is helpful to you.

In fact, this error is caused by one-hot encoding. It has been solved by adding np.argmax(labels, axis=1) . contact me if you still have question.

Thanks for your report.

[pranoot]

Hello @vzxxbacq ,

Thanks for the revert and the fix both were helpful and now that script is running without any bug.

Now I am trying to train the same model on my own dataset.
I have few doubts can you please help me with the same.

1. For pre processing feature extraction in deep speaker I used ext_fbank_feature as specified in the paper. There is SLIDE_WINDOW present in it. Is it required and also if yes what should be the appropriate shape for the same?
2. The shape of the input place holder is ( None, 100, 64, 1) I want to know is 100 num of frames of individual clip, 64 the number of f-bank coefficients?
3. What should be the duration of individual clip which is used for training ?
4. Should the clip contain silence along with utterances or should clip not have any silence?

Thank you very much! :)

[vzxxbacq]

Hi @pranoot ,
Actually, if you understand slide_window function, all problems will be solved. In the paper, author use fixed length audio ( 10ms ), but we have many dataset with changable length. So, I write the slide_window to apply our model to these dataset. slide_window parameters is a list [l, r] and the function will do feature[i] = feature[i-l : i+r], so obviously 100 means feature[i] = feature[ i-49 : i+50 ] that slide_window=[49, 50]. And 64 is feature dims in the paper.
And I didn't write VAD method, I used other toolkit to do this part job.
If you still have question feel free to contact me.

[pranoot]

Oh cool got it !
Thanks a lot !! :D