【Hackathon 8th No.9】在 PaddleSpeech 中复现 DAC 训练需要用到的 loss #3954
+332
−2
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Thanks for your contribution! |
|
In the original DAC repository, the training data is generated randomly. To assess accuracy, I sampled ten loss values and saved them in PyTorch tensor (.pt) format. The original repository showed no numerical errors, but in the Paddle implementation, bias is observed and I still tracing it . Here is the test result
|
|
With these patch, the new test result is below diff --git a/paddlespeech/t2s/modules/losses.py b/paddlespeech/t2s/modules/losses.py
index 029ad1be..ce5f441d 100644
--- a/paddlespeech/t2s/modules/losses.py
+++ b/paddlespeech/t2s/modules/losses.py
@@ -501,7 +502,7 @@ def stft(x,
real = x_stft.real()
imag = x_stft.imag()
- return paddle.sqrt(paddle.clip(real**2 + imag**2, min=1e-7)).transpose(
+ return paddle.clip(paddle.sqrt(real**2 + imag**2), min=clamp_eps).transpose(
[0, 2, 1])
@@ -930,7 +930,7 @@ class MelSpectrogram(nn.Layer):
real = real.transpose([0, 2, 1])
imag = imag.transpose([0, 2, 1])
x_power = real**2 + imag**2
- x_amp = paddle.sqrt(paddle.clip(x_power, min=self.eps))
+ x_amp = paddle.clip(paddle.sqrt(x_power), min=self.eps)
|
|
开发者你好，感谢你的参与！由于你的黑客松赛题完成度较高，其PR已被锁定，请尽快完善锁定的PR，并确保在2025年1月3日前完成合入。逾期未合入PR将无法获得奖金发放。 |
…ust the clipping threshold
…n calculation methods - Change precision threshold to ’1e-5‘ - Use relative error instead of absolute error
|
📢：请尽快完善锁定的PR，并确保在2025年1月10日（不再延期）前完成合入。逾期未合入PR将无法获得奖金发放。 |
zxcd
reviewed
Jan 13, 2025
|
|
||
| import librosa | ||
| import numpy as np | ||
| import paddle | ||
| from paddle import nn | ||
| from paddle.nn import functional as F | ||
| from paddleaudio.audiotools import AudioSignal | ||
| from paddleaudio.audiotools import STFTParams |
| loss_1.backward() | ||
| loss_1_grad = signal.audio_data.grad.sum() | ||
|
|
||
| assert abs( |
There was a problem hiding this comment.
suggest use np.testing.assert_allclose, these losses can pass 1e-6?
There was a problem hiding this comment.
Currently not. After debugging, I find out that the loss is generated by 'paddle.signal.stft' (without cuda), so I have to compare the implement with '_VF' and paddle. I'm sure that the loss can decrease to 0 if fixing this
luotao1
changed the title
luotao1
changed the title
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
PR types
New features
PR changes
APIs
Describe