You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
PaddleSpeech/deepspeech/modules/loss.py

145 lines
5.2 KiB

# 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.
import paddle
from paddle import nn
from paddle.nn import functional as F
from deepspeech.utils.log import Log
logger = Log(__name__).getlog()
__all__ = ['CTCLoss', "LabelSmoothingLoss"]
class CTCLoss(nn.Layer):
def __init__(self, blank=0, reduction='sum', batch_average=False):
super().__init__()
# last token id as blank id
self.loss = nn.CTCLoss(blank=blank, reduction=reduction)
self.batch_average = batch_average
def forward(self, logits, ys_pad, hlens, ys_lens):
"""Compute CTC loss.
Args:
logits ([paddle.Tensor]): [B, Tmax, D]
ys_pad ([paddle.Tensor]): [B, Tmax]
hlens ([paddle.Tensor]): [B]
ys_lens ([paddle.Tensor]): [B]
Returns:
[paddle.Tensor]: scalar. If reduction is 'none', then (N), where N = \text{batch size}.
"""
B = paddle.shape(logits)[0]
# warp-ctc need logits, and do softmax on logits by itself
# warp-ctc need activation with shape [T, B, V + 1]
# logits: (B, L, D) -> (L, B, D)
logits = logits.transpose([1, 0, 2])
# (TODO:Hui Zhang) ctc loss does not support int64 labels
ys_pad = ys_pad.astype(paddle.int32)
loss = self.loss(
logits, ys_pad, hlens, ys_lens, norm_by_times=self.batch_average)
if self.batch_average:
# Batch-size average
loss = loss / B
return loss
class LabelSmoothingLoss(nn.Layer):
"""Label-smoothing loss.
In a standard CE loss, the label's data distribution is:
[0,1,2] ->
[
[1.0, 0.0, 0.0],
[0.0, 1.0, 0.0],
[0.0, 0.0, 1.0],
]
In the smoothing version CE Loss,some probabilities
are taken from the true label prob (1.0) and are divided
among other labels.
e.g.
smoothing=0.1
[0,1,2] ->
[
[0.9, 0.05, 0.05],
[0.05, 0.9, 0.05],
[0.05, 0.05, 0.9],
]
"""
def __init__(self,
size: int,
padding_idx: int,
smoothing: float,
normalize_length: bool=False):
"""Label-smoothing loss.
Args:
size (int): the number of class
padding_idx (int): padding class id which will be ignored for loss
smoothing (float): smoothing rate (0.0 means the conventional CE)
normalize_length (bool):
True, normalize loss by sequence length;
False, normalize loss by batch size.
Defaults to False.
"""
super().__init__()
self.size = size
self.padding_idx = padding_idx
self.smoothing = smoothing
self.confidence = 1.0 - smoothing
self.normalize_length = normalize_length
self.criterion = nn.KLDivLoss(reduction="none")
def forward(self, x: paddle.Tensor, target: paddle.Tensor) -> paddle.Tensor:
"""Compute loss between x and target.
The model outputs and data labels tensors are flatten to
(batch*seqlen, class) shape and a mask is applied to the
padding part which should not be calculated for loss.
Args:
x (paddle.Tensor): prediction (batch, seqlen, class)
target (paddle.Tensor):
target signal masked with self.padding_id (batch, seqlen)
Returns:
loss (paddle.Tensor) : The KL loss, scalar float value
"""
B, T, D = paddle.shape(x)
assert D == self.size
x = x.reshape((-1, self.size))
target = target.reshape([-1])
# use zeros_like instead of torch.no_grad() for true_dist,
# since no_grad() can not be exported by JIT
true_dist = paddle.full_like(x, self.smoothing / (self.size - 1))
ignore = target == self.padding_idx # (B,)
# target = target * (1 - ignore) # avoid -1 index
target = target.masked_fill(ignore, 0) # avoid -1 index
# true_dist.scatter_(1, target.unsqueeze(1), self.confidence)
target_mask = F.one_hot(target, self.size)
true_dist *= (1 - target_mask)
true_dist += target_mask * self.confidence
kl = self.criterion(F.log_softmax(x, axis=1), true_dist)
#TODO(Hui Zhang): sum not support bool type
#total = len(target) - int(ignore.sum())
total = len(target) - int(ignore.type_as(target).sum())
denom = total if self.normalize_length else B
#numer = (kl * (1 - ignore)).sum()
numer = kl.masked_fill(ignore.unsqueeze(1), 0).sum()
return numer / denom