Thanks for your implement. It quites help me. But I am confused about the adversarial loss. For the Voice conversion task, it aims to convert the voice from A to speaker B's. And the discriminator trains iteratively to measure the current distribution and the target distribution. And why in the implement the adversarial loss compute the distribution distance between A and B but not B and B' ( I denote the G(A, y_b) as B').
Thank you. Look forward to your reply.

Hi,

I can see from this line:

g_loss = g_loss_fake + self.lambda_rec * g_loss_rec + self.lambda_cls * g_loss_cls_spks

in solver.py that adv. loss, reconstruction loss and domian class loss are being computed however, am I correct in saying that this paper does not implement the id mapping loss introduced in the StarGAN-VC paper?

Thank you,
Sam

Hello, I wanted to ask a question.
When we want to compute the MCEPs of the wav files, we do it using the function coded_sp = pyworld.code_spectral_envelope(sp, fs, dim) where dim=36.
I wanted to ask you why you are using 36 dimensions o 36 MCEPs?

Hi,
I want to know what should be put into the path "./mc/train" appeared in the file "data_loader.py". Besides, could you tell me what files need to be run and the specific order of running these files?

In Requirements you say python 3.6 (or 3.5).
However Python 3.5 cannot be used here, because there are a few f-strings which requires python 3.6+, (though we can manually change them to %-formatted string and work with python 3.5)

The resample_to_16k function of preprocessing.py just prints:

[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

when finished.

The next section begins with number of workers: 10 however, here the program just halts.

I think this may be a sox issue as i have just installed it with conda forge. I am working on a VM and am not entirely sure how to install sox from binary.

Thanks

Hello can someone explain me in a bit more detail what each loss function says about the training?
Thanks in advance

I used a model that iterated 7000/8000 times and it didn't work well.
I'm continuing to train.

Can implementation of the author share 200000G&D iteration model for comparative study?

tks.

sox WARN rate: rate clipped 27 samples; decrease volume?
sox WARN dither: dither clipped 25 samples; decrease volume?
sox WARN rate: rate clipped 3 samples; decrease volume?
sox WARN dither: dither clipped 2 samples; decrease volume?
60%|██████████████████████████▍ | 6/10 [00:27<00:36, 9.02s/it]sox WARN rate: rate clipped 4 samples; decrease volume?
sox WARN dither: dither clipped 4 samples; decrease volume?
100%|███████████████████████████████████████████| 10/10 [00:34<00:00, 5.55s/it]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
number of workers: 10
0%| | 0/10 [00:00<?, ?it/s]concurrent.futures.process._RemoteTraceback:
"""
Traceback (most recent call last):
File "/root/anaconda3/envs/tensorflow/lib/python3.6/concurrent/futures/process.py", line 175, in _process_worker
r = call_item.fn(*call_item.args, **call_item.kwargs)
File "preprocess.py", line 50, in get_spk_world_feats
train_paths, test_paths = split_data(paths)
File "preprocess.py", line 42, in split_data
train_indices, test_indices = train_test_split(indices, test_size=test_size, random_state=1234)
File "/root/anaconda3/envs/tensorflow/lib/python3.6/site-packages/sklearn/model_selection/_split.py", line 2100, in train_test_split
default_test_size=0.25)
File "/root/anaconda3/envs/tensorflow/lib/python3.6/site-packages/sklearn/model_selection/_split.py", line 1782, in _validate_shuffle_split
train_size)
ValueError: With n_samples=0, test_size=0.1 and train_size=None, the resulting train set will be empty. Adjust any of the aforementioned parameters.
"""

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
File "preprocess.py", line 130, in
result_list = [future.result() for future in tqdm(futures)]
File "preprocess.py", line 130, in
result_list = [future.result() for future in tqdm(futures)]
File "/root/anaconda3/envs/tensorflow/lib/python3.6/concurrent/futures/_base.py", line 405, in result
return self.__get_result()
File "/root/anaconda3/envs/tensorflow/lib/python3.6/concurrent/futures/_base.py", line 357, in __get_result
raise self._exception
ValueError: With n_samples=0, test_size=0.1 and train_size=None, the resulting train set will be empty. Adjust any of the aforementioned parameters.
^CError in atexit._run_exitfuncs:

hi, thanks for you sharing.
There is GLU(Gated CNN) in Generator and Discriminator. But I didn't find them.
I just find RELU, in model.py line:30 Generator:

    layers.append(nn.ReLU(inplace=True))

hi,
(中文熟悉些😆）貌似模型是cyclegan-vc改的，并不是stargan-vc的模型，不知转换结果怎么样呢？

When I run the command
python preprocess.py --sample_rate 16000 --origin_wavpath data/VCTK-Corpus/wav48 --target_wavpath data/VCTK-Corpus/wav16 --mc_dir_train data/mc/train --mc_dir_test data/mc/test

I get the following error
Traceback (most recent call last): File "C:\Users\user pc\Anaconda3\lib\concurrent\futures\process.py", line 232, in _process_worker r = call_item.fn(*call_item.args, **call_item.kwargs) File "C:\Users\user pc\PycharmProjects\song\preprocess.py", line 25, in resample subprocess.call(['sox', wav_from, "-r", "16000", wav_to]) File "C:\Users\user pc\Anaconda3\lib\subprocess.py", line 323, in call with Popen(*popenargs, **kwargs) as p: File "C:\Users\user pc\Anaconda3\lib\subprocess.py", line 775, in __init__ restore_signals, start_new_session) File "C:\Users\user pc\Anaconda3\lib\subprocess.py", line 1178, in _execute_child startupinfo) FileNotFoundError: [WinError 2] The system cannot find the file specified

Hi there,
Thanks for the code. I am able to work with the code with 4 speakers in folders Speaker and Speaker_test but if I increase the number of the speaker to 6, preprocess.py file runs fine but training through main.py is throwing an error saying that "target x is out of bounds" I am keeping data very small (100 samples for all 6 speakers to run successfully and then increasing the voice sample size. I try to read the code but not able to debug thoroughly yet.

UserWarning: Implicit dimension choice for log_softmax has been deprecated. Change the call to include dim=X as an argument.
input = module(input)
Traceback (most recent call last):
File "main.py", line 92, in
main(config)
File "main.py", line 34, in main
solver.train()
File "/Users/shalinisaini/pytorch-StarGAN-VCtk-oct-6small/solver.py", line 160, in train
cls_loss_real = CELoss(input=cls_real, target=speaker_idx_org)
File "/Users/shalinisaini/opt/anaconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 550, in call
result = self.forward(*input, **kwargs)
File "/Users/shalinisaini/opt/anaconda3/lib/python3.8/site-packages/torch/nn/modules/loss.py", line 931, in forward
return F.cross_entropy(input, target, weight=self.weight,
File "/Users/shalinisaini/opt/anaconda3/lib/python3.8/site-packages/torch/nn/functional.py", line 2317, in cross_entropy
return nll_loss(log_softmax(input, 1), target, weight, None, ignore_index, None, reduction)
File "/Users/shalinisaini/opt/anaconda3/lib/python3.8/site-packages/torch/nn/functional.py", line 2115, in nll_loss
ret = torch._C._nn.nll_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index)
IndexError: Target 4 is out of bounds.

---can anyone tell if this code can be expanded for more than 4 speakers. I am not sure if I am missing something stupid but please see if anyone can help in the right direction. I would like to run the training for more speakers.

Thanks for your help.

Python 3.6 is needed because of the f-strings
And also sox is needed (sudo apt-get install sox on Ubuntu)

could you please tell me which cuda and cudnn you used??

I put my own Chinese corpus into the training model, and the conversion effect is not as good as that of the original English corpus.
Are there any requirements for training datasets?

Looking forward to answer.

thks.

Firstly thanks to the author for the implementation of this work. In ./converted_samples/readme, you mentioned that 'These converted samples were obtained from the not-well-fine-tuned model. If you want to get better results, please tune the hyper-parameters carefully'. Thus please give me some advice on how to fine-tune this model? Thanks a lot.

I found that the identity loss seems like missing in the "train the G ", and in my opinion, the code is supposed to be :

            id_mc = self.G ( mc_real, spk_c_org )
            g_loss_id = torch.mean( torch.abs( id_mc - mc_real ))

I wish that it can be explained by you . Thank you very much .

The paper claims that this could allow at least real time, I would like to know the inference time and the hardware that you are using. Thanks.