Error When I run scripts/training.py about mad-twinnet HOT 17 CLOSED

Hello !

Thank you so much ! I think I found the solution for that.
In ' settings.py ' I changed the files_per_pass value from 4 to 2 and it starts working normally.
maybe because I'm using the small DSD100 dataset '120MB' for test (even I tested before with full DSD100 dataset)?
Do you have any explanation for that?

# Process constants
training_constants = {
    'epochs': 2 if debug else 100,
    'batch_size': 16,
    'files_per_pass': 2
}

Thank you so much

from mad-twinnet.

Hi!

It seems that there are no values produced for the loss of the Masker, during training.

Let me check it on Tuesday and I will post here again.

from mad-twinnet.

Also, could you please tell me which version of PyTorch are you using?

from mad-twinnet.

from mad-twinnet.

Version of PyTorch? Not version of Python.

from mad-twinnet.

Oh ! sorry my bad .. the version 0.3.0.post4

from mad-twinnet.

Hi

I could not reproduce your error. Seeing from your error message that you posted, the variable epoch_l_m (the epoch loss for the masker) is empty, that is why your get "empty Tensor" error.

I would suggest to see if your python path is setup properly (according to the guidelines given in the README file).

Please check your python path for the project and make sure that you are using the export PYTHONPATH... part, given in the README file.

from mad-twinnet.

from mad-twinnet.

I would suggest you to do debug. I cannot figure out why you have this problem.

If you could get a breakpoint at the line where the values are appended at the epoch_l_m list, and see what values are appended.

from mad-twinnet.

Thank you so much for
I think that the problem can be Dataset set-up
My Dataset DSD100 folder looks like default structure
my problem is that I don't know which files should I rename them to numbers and I don't understand this line "each file name should have an identifier whether the file is about the mixture, the voice, the bass, and other." it means like 001(the bass).wav ?
so could you please help me.

Best regards,

from mad-twinnet.

Hi,

Sorry for the not clear enough description on how the naming conventions of the dataset are.

• The mixture files should be named as "mixture.wav"
• The voice files should be names as "vocals.wav"
• The bass files as "bass.wav"
• The drums as "drums.wav"
• The other as "other.wav".

Do the above help a bit?

Cheers!

from mad-twinnet.

from mad-twinnet.

I have not understood what exactly you are trying to do. The line that you removed was calculating the mean loss of the masker for one epoch.

I would guess that now you do not get any indication for the losses in one epoch?

from mad-twinnet.

from mad-twinnet.

Hi !

Is there any update about this issue? I'm having the same problem even I'm trying my best to resolve it.
by the way I'm using the CPU version of Torch, do you think it can be the reason?

-- Starting training process. Debug mode: False
-- Setting up modules... done.
-- Setting up optimizes and losses... done.
-- Training starts

Traceback (most recent call last):
  File "scripts/training.py", line 199, in <module>
    main()
  File "scripts/training.py", line 195, in main
    training_process()
  File "scripts/training.py", line 176, in training_process
    l_m=torch.mean(torch.FloatTensor(epoch_l_m)),
RuntimeError: invalid argument 1: empty Tensor at /pytorch/torch/lib/TH/generic/THTensorMath.c:3311

Thank you in advance.

from mad-twinnet.

Hey,

I tried to reproduce your error but I cannot.

If you want, we can try together and solve your problem. What I would do, is to put a break point at line 166 of training.py

epoch_l_m.append(l_m.data[0])

and see what the value of l_m.data.

Could you please tell me the above?

Cheers and thank you for trying to solve this with me :)

from mad-twinnet.

Hi!

I'm very glad that finally worked for you.

The only explanation that I can think of now is due to some rounding in divisions. That is, with the number of files that you had, the division with the files_per_pass gave some invalid numbers and this resulted in not getting files, thus not giving output for the l_m, and this resulting to the empty list of epoch_l_m.

The above is just a speculation though. I would have to dig a bit more to this particular report. Since I'm planning to update the code for Python 3.X and PyTorch 0.4.X, I think that I will do it but for the new version :)

In any case, thank you so much for reporting this and staying here with me to fix it.

Enjoy using the code!!!

from mad-twinnet.