[Hackathon 7th No.56] 在 PaddleSpeech 中复现 DAC 训练需要用到的 loss

[suzakuwcx]

PR types

New features

PR changes

APIs

Describe

• 【Hackathon 7th】开源贡献个人挑战赛 Paddle#68244
• [Hackathon 7th No.56】rfc: 在 PaddleSpeech 中复现 DAC 训练需要用到的 loss community#1029

[paddle-bot]

Thanks for your contribution!

[suzakuwcx]

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

index	AudioSignal_STFT	AudioSignal_MEL	Tensor_STFT	Tensor_MEL
0	2.86102294921875e-06	9.5367431640625e-07	6.103515625e-05	6.198883056640625e-05
1	9.5367431640625e-07	4.76837158203125e-07	6.103515625e-05	0.000194549560546875
2	0.0	9.5367431640625e-07	3.0517578125e-05	0.0005588531494140625
3	9.5367431640625e-07	4.76837158203125e-07	2.09808349609375e-05	4.9114227294921875e-05
4	4.76837158203125e-07	4.76837158203125e-07	0.00012445449829101562	7.05718994140625e-05
5	4.76837158203125e-06	0.0	0.03401947021484375	0.012701988220214844
6	4.76837158203125e-07	9.5367431640625e-07	0.00014448165893554688	0.0003781318664550781
7	3.814697265625e-06	9.5367431640625e-07	0.2675790786743164	0.03282880783081055
8	9.5367431640625e-07	0.0	0.22318553924560547	0.03190279006958008
9	2.384185791015625e-06	0.0	0.22658443450927734	0.03314781188964844

[suzakuwcx]

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)

index	Tensor_STFT	Tensor_MEL
0	2.86102294921875e-06	9.5367431640625e-07
1	9.5367431640625e-07	4.76837158203125e-07
2	9.5367431640625e-07	9.5367431640625e-07
3	9.5367431640625e-07	4.76837158203125e-07
4	4.76837158203125e-07	4.76837158203125e-07
5	3.814697265625e-06	0.0
6	0.0	9.5367431640625e-07
7	2.86102294921875e-06	4.76837158203125e-06
8	0.0	1.430511474609375e-06
9	2.384185791015625e-06	4.76837158203125e-07

[zxcd]

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。