cvqluu / factorized-tdnn Goto Github PK

Hello!

Thank you so much for the great work you've shared, would it possible for you to share the pre-processing methodology and the training data you worked with as a demo.

It would be great help.

Thank you!

Hey man, and thank you for this great repository, it saved me from implementing and debugging TDNN/FTDNN and instead I can focus on experiments. I have one question regarding the correct way to step the optimizer, because it seems that I have been doing it wrong until now. After reading the "usage" part of the README it seems to me that the correct way to step during the training is to do something like that: assuming that the most basic way to train network in pytorch is to:

for input, target in dataset:
    optimizer.zero_grad()
    output = model(input)
    loss = loss_fn(output, target)
    loss.backward()
    optimizer.step()

and you wrote

tdnn_f.step_semi_orth() # The key method to constrain the first two convolutions, perform after every SGD step

Than I should add this call tdnn_f.step_semi_orth() after the optimizer.step() call, making it like this, right?

for input, target in dataset:
    optimizer.zero_grad()
    output = model(input)
    loss = loss_fn(output, target)
    loss.backward()
    optimizer.step()
    tdnn_f.step_semi_orth()

Thank you for your time if you happen to answer this question!

Hi!

How are you?
I was wondering if there is a way to load thes weight of a pretrained kaldi network into this architechture you are implementing. Say, going from http://kaldi-asr.org/models/m13 to your architecture. Did you encounter this problem before and per chance solved it?

Thanks anyways :D

Hello. I use the FTDNN model in your model.py to train the VoxCeleb1 dataset but the model works terribly compared to the original TDNN system, it can only reach 40% acc after 100 epochs however TDNN can reach nearly 100%. And the loss is still relatively high after 100 epochs. It clearly failed to reach convergence. Have you tried this on Voxceleb1 dataset before? I hope to know the configuration. Thank you .

Hello,
I am trying to include the FTDNN model that you wrote in the framework pytorch-kaldi. I managed to implement the model ut I have some issues when training which I suspect are related to the input and output I am using.
Could you possibly provide the code that you used for the demonstration of FTDNN or what king of Loss function should be used at the end ?
Thank you.