[tts] add adversarial loss (#2588)
liangym
2 years ago
committed by
GitHub
parent
9aab706cba
commit
e18170228c
@ -0,0 +1,13 @@
|
||||
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
@ -0,0 +1,58 @@
|
||||
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
import paddle
|
||||
import paddle.nn as nn
|
||||
from paddle.autograd import PyLayer
|
||||
|
||||
|
||||
class GradientReversalFunction(PyLayer):
|
||||
"""Gradient Reversal Layer from:
|
||||
Unsupervised Domain Adaptation by Backpropagation (Ganin & Lempitsky, 2015)
|
||||
|
||||
Forward pass is the identity function. In the backward pass,
|
||||
the upstream gradients are multiplied by -lambda (i.e. gradient is reversed)
|
||||
"""
|
||||
|
||||
@staticmethod
|
||||
def forward(ctx, x, lambda_=1):
|
||||
"""Forward in networks
|
||||
"""
|
||||
ctx.save_for_backward(lambda_)
|
||||
return x.clone()
|
||||
|
||||
@staticmethod
|
||||
def backward(ctx, grads):
|
||||
"""Backward in networks
|
||||
"""
|
||||
lambda_, = ctx.saved_tensor()
|
||||
dx = -lambda_ * grads
|
||||
return paddle.clip(dx, min=-0.5, max=0.5)
|
||||
|
||||
|
||||
class GradientReversalLayer(nn.Layer):
|
||||
"""Gradient Reversal Layer from:
|
||||
Unsupervised Domain Adaptation by Backpropagation (Ganin & Lempitsky, 2015)
|
||||
|
||||
Forward pass is the identity function. In the backward pass,
|
||||
the upstream gradients are multiplied by -lambda (i.e. gradient is reversed)
|
||||
"""
|
||||
|
||||
def __init__(self, lambda_=1):
|
||||
super(GradientReversalLayer, self).__init__()
|
||||
self.lambda_ = lambda_
|
||||
|
||||
def forward(self, x):
|
||||
"""Forward in networks
|
||||
"""
|
||||
return GradientReversalFunction.apply(x, self.lambda_)
|
||||
@ -0,0 +1,55 @@
|
||||
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# Modified from Cross-Lingual-Voice-Cloning(https://github.com/deterministic-algorithms-lab/Cross-Lingual-Voice-Cloning)
|
||||
import paddle
|
||||
from paddle import nn
|
||||
from typeguard import check_argument_types
|
||||
|
||||
|
||||
class SpeakerClassifier(nn.Layer):
|
||||
def __init__(
|
||||
self,
|
||||
idim: int,
|
||||
hidden_sc_dim: int,
|
||||
spk_num: int, ):
|
||||
assert check_argument_types()
|
||||
super().__init__()
|
||||
# store hyperparameters
|
||||
self.idim = idim
|
||||
self.hidden_sc_dim = hidden_sc_dim
|
||||
self.spk_num = spk_num
|
||||
|
||||
self.model = nn.Sequential(
|
||||
nn.Linear(self.idim, self.hidden_sc_dim),
|
||||
nn.Linear(self.hidden_sc_dim, self.spk_num))
|
||||
|
||||
def parse_outputs(self, out, text_lengths):
|
||||
mask = paddle.arange(out.shape[1]).expand(
|
||||
[out.shape[0], out.shape[1]]) < text_lengths.unsqueeze(1)
|
||||
out = paddle.transpose(out, perm=[2, 0, 1])
|
||||
out = out * mask
|
||||
out = paddle.transpose(out, perm=[1, 2, 0])
|
||||
return out
|
||||
|
||||
def forward(self, encoder_outputs, text_lengths):
|
||||
"""
|
||||
encoder_outputs = [batch_size, seq_len, encoder_embedding_size]
|
||||
text_lengths = [batch_size]
|
||||
|
||||
log probabilities of speaker classification = [batch_size, seq_len, spk_num]
|
||||
"""
|
||||
|
||||
out = self.model(encoder_outputs)
|
||||
out = self.parse_outputs(out, text_lengths)
|
||||
return out
|
||||
Loading…
Reference in new issue