From b74f9d781225cc0cf56e3ec1b35b613eaad1f58d Mon Sep 17 00:00:00 2001 From: longrookie Date: Fri, 24 Mar 2023 19:59:58 +0800 Subject: [PATCH] Format code using pre-commit --- examples/csmsc/voc5/iSTFTNet.md | 2 +- paddlespeech/t2s/models/hifigan/hifigan.py | 24 +++++++++++++++------- 2 files changed, 18 insertions(+), 8 deletions(-) diff --git a/examples/csmsc/voc5/iSTFTNet.md b/examples/csmsc/voc5/iSTFTNet.md index 4fb0e4622..e83da6875 100644 --- a/examples/csmsc/voc5/iSTFTNet.md +++ b/examples/csmsc/voc5/iSTFTNet.md @@ -141,4 +141,4 @@ The pretained hifigan model int the comparison can be downloaded here: ## Acknowledgement -We adapted some code from https://github.com/rishikksh20/iSTFTNet-pytorch.git. \ No newline at end of file +We adapted some code from https://github.com/rishikksh20/iSTFTNet-pytorch.git. diff --git a/paddlespeech/t2s/models/hifigan/hifigan.py b/paddlespeech/t2s/models/hifigan/hifigan.py index 25873da66..a2592c140 100644 --- a/paddlespeech/t2s/models/hifigan/hifigan.py +++ b/paddlespeech/t2s/models/hifigan/hifigan.py @@ -21,7 +21,6 @@ from typing import Optional import paddle import paddle.nn.functional as F from paddle import nn -import numpy as np from paddlespeech.t2s.modules.activation import get_activation from paddlespeech.t2s.modules.nets_utils import initialize @@ -103,7 +102,8 @@ class HiFiGANGenerator(nn.Layer): assert len(upsample_scales) >= istft_layer_id if use_istft else True # define modules - self.num_upsamples = len(upsample_kernel_sizes) if not use_istft else istft_layer_id + self.num_upsamples = len( + upsample_kernel_sizes) if not use_istft else istft_layer_id self.num_blocks = len(resblock_kernel_sizes) self.input_conv = nn.Conv1D( in_channels, @@ -155,8 +155,13 @@ class HiFiGANGenerator(nn.Layer): self.istft_layer_id = istft_layer_id self.istft_n_fft = int(self.istft_hop_size * overlap_ratio) self.istft_win_size = self.istft_n_fft - self.reflection_pad = nn.Pad1D(padding=[1,0], mode='reflect') - self.conv_post = nn.Conv1D(channels// (2**(i + 1)), (self.istft_n_fft // 2 + 1)*2, kernel_size, 1, padding=(kernel_size - 1) // 2, ) + self.reflection_pad = nn.Pad1D(padding=[1, 0], mode='reflect') + self.conv_post = nn.Conv1D( + channels // (2**(i + 1)), + (self.istft_n_fft // 2 + 1) * 2, + kernel_size, + 1, + padding=(kernel_size - 1) // 2, ) else: self.istft_layer_id = len(upsample_scales) @@ -191,7 +196,7 @@ class HiFiGANGenerator(nn.Layer): for j in range(self.num_blocks): cs += self.blocks[i * self.num_blocks + j](c) c = cs / self.num_blocks - + if self.use_istft: c = F.leaky_relu(c) c = self.reflection_pad(c) @@ -204,8 +209,13 @@ class HiFiGANGenerator(nn.Layer): spec = paddle.exp(c[:, :self.istft_n_fft // 2 + 1, :]) phase = paddle.sin(c[:, self.istft_n_fft // 2 + 1:, :]) - c = paddle.complex(spec*(paddle.cos(phase)), spec*(paddle.sin(phase))) - c = paddle.signal.istft(c, n_fft=self.istft_n_fft, hop_length=self.istft_hop_size, win_length=self.istft_win_size) + c = paddle.complex(spec * (paddle.cos(phase)), + spec * (paddle.sin(phase))) + c = paddle.signal.istft( + c, + n_fft=self.istft_n_fft, + hop_length=self.istft_hop_size, + win_length=self.istft_win_size) c = c.unsqueeze(1) else: c = self.output_conv(c)