[Bug] No grad in YourTTS speaker encoder

@Edresson

Describe the bug

Hello guys (CC: @Edresson @WeberJulian), when going through YourTTS code & paper, I noticed that you are calculating the inputs for the speaker encoder with no grads:

TTS/TTS/encoder/models/resnet.py

Lines 153 to 200 in d46fbc2

    
               def forward(self, x, l2_norm=False): 
        
                   """Forward pass of the model. 
        
                   Args: 
        
                       x (Tensor): Raw waveform signal or spectrogram frames. If input is a waveform, `torch_spec` must be `True` 
        
                           to compute the spectrogram on-the-fly. 
        
                       l2_norm (bool): Whether to L2-normalize the outputs. 
        
                   Shapes: 
        
                       - x: :math:`(N, 1, T_{in})` or :math:`(N, D_{spec}, T_{in})` 
        
                   """ 
        
                   with torch.no_grad(): 
        
                       with torch.cuda.amp.autocast(enabled=False): 
        
                           x.squeeze_(1) 
        
                           # if you torch spec compute it otherwise use the mel spec computed by the AP 
        
                           if self.use_torch_spec: 
        
                               x = self.torch_spec(x) 
        
                           if self.log_input: 
        
                               x = (x + 1e-6).log() 
        
                           x = self.instancenorm(x).unsqueeze(1) 
        
                   x = self.conv1(x) 
        
                   x = self.relu(x) 
        
                   x = self.bn1(x) 
        
                   x = self.layer1(x) 
        
                   x = self.layer2(x) 
        
                   x = self.layer3(x) 
        
                   x = self.layer4(x) 
        
                   x = x.reshape(x.size()[0], -1, x.size()[-1]) 
        
                   w = self.attention(x) 
        
                   if self.encoder_type == "SAP": 
        
                       x = torch.sum(x * w, dim=2) 
        
                   elif self.encoder_type == "ASP": 
        
                       mu = torch.sum(x * w, dim=2) 
        
                       sg = torch.sqrt((torch.sum((x**2) * w, dim=2) - mu**2).clamp(min=1e-5)) 
        
                       x = torch.cat((mu, sg), 1) 
        
                   x = x.view(x.size()[0], -1) 
        
                   x = self.fc(x) 
        
                   if l2_norm: 
        
                       x = torch.nn.functional.normalize(x, p=2, dim=1) 
        
                   return x

I suspect that the speaker encoder is not producing any gradients, and the speaker consistency loss has no effect.
It looks like this happens:

forward gets x with grads
the spectrogram is extracted via torch_spec with no grads
the output of the speaker encoder has requires grad False (and would produce an exception when you called backward on it since it did not keep activations for grad calculation), but is added to the total loss (which has requires grads se to True)
so the call to loss.backward() works as usually, but the speaker encoder does not contribute to the gradients flowing to the generator at all

Could you please check on that?

To Reproduce

import torch

a = torch.tensor(1.0, requires_grad=True)
b = torch.tensor(2.0, requires_grad=True)

with torch.no_grad():
     c = a + b

d = c + 1
e = a + d
e.backward()

print(a.grad) # 1
print(b.grad) # not set

d.backward() # RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

Expected behavior

No response

Logs

No response

Environment

Just reading the code and asking :)

Additional context

No response

Nice catch, that's concerning. Thanks for reporting it. We'll look more into it but it looks like you are right.
This torch.no_grad doesn't affect the speaker encoder during training because there is no parameter before it. But in our case, using the original implementation for SCL, the TTS parameters are before this torch.no_grad. I guess in the paper the slightly better SECS score are explained by the extra training steps...

Nice catch. Indeed it is an issue. I will submit a PR to fix it.

@Edresson

Thank you @Edresson

Are you also planning to retrain the models and update the YourTTS paper at least on arxiv? 😇

@Edresson

Thank you @Edresson

Are you also planning to retrain the models and update the YourTTS paper at least on arxiv? innocent

I think it is not worth because if we do it we will need to recompute the MOS and Sim-MOS. I'm thinking about update the preprint, removing Speaker Consistency Loss from the methodology. And given that the Speaker Consistency Loss had no effect on the results, Speaker Consistency Loss experiments are equal than keep the model training per more 50k steps. In addition, I will try to retracting this issue on ICML published paper as well. Fortunately, It is a minor issue and the reported results are not effected (only the method description that is wrong).

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. You might also look our discussion channels.

@Tomiinek

@Tomiinek Thanks so much for finding the bug and reporting it. I talked with all authors and the final decision was added a Erratum on YourTTS Github repository and on the last page of the preprint. It is done :).

Navigation Menu

Repository navigation

[Bug] No grad in YourTTS speaker encoder #2348

Describe the bug

To Reproduce

Expected behavior

Logs

Environment

Additional context

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	def forward(self, x, l2_norm=False):
	"""Forward pass of the model.

	Args:
	x (Tensor): Raw waveform signal or spectrogram frames. If input is a waveform, `torch_spec` must be `True`
	to compute the spectrogram on-the-fly.
	l2_norm (bool): Whether to L2-normalize the outputs.

	Shapes:
	- x: :math:`(N, 1, T_{in})` or :math:`(N, D_{spec}, T_{in})`
	"""
	with torch.no_grad():
	with torch.cuda.amp.autocast(enabled=False):
	x.squeeze_(1)
	# if you torch spec compute it otherwise use the mel spec computed by the AP
	if self.use_torch_spec:
	x = self.torch_spec(x)

	if self.log_input:
	x = (x + 1e-6).log()
	x = self.instancenorm(x).unsqueeze(1)

	x = self.conv1(x)
	x = self.relu(x)
	x = self.bn1(x)

	x = self.layer1(x)
	x = self.layer2(x)
	x = self.layer3(x)
	x = self.layer4(x)

	x = x.reshape(x.size()[0], -1, x.size()[-1])

	w = self.attention(x)

	if self.encoder_type == "SAP":
	x = torch.sum(x * w, dim=2)
	elif self.encoder_type == "ASP":
	mu = torch.sum(x * w, dim=2)
	sg = torch.sqrt((torch.sum((x*2) w, dim=2) - mu**2).clamp(min=1e-5))
	x = torch.cat((mu, sg), 1)

	x = x.view(x.size()[0], -1)
	x = self.fc(x)

	if l2_norm:
	x = torch.nn.functional.normalize(x, p=2, dim=1)
	return x

[Bug] No grad in YourTTS speaker encoder #2348

Description

Describe the bug

To Reproduce

Expected behavior

Logs

Environment

Additional context

Activity

WeberJulian commented on Feb 16, 2023

Edresson commented on Feb 23, 2023

Tomiinek commented on Feb 23, 2023

Edresson commented on Feb 25, 2023

stale commented on Mar 28, 2023

Edresson commented on Apr 30, 2023

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions