From d80ac2fd2df4e7fb8a28acfa512bb11472b5cc99 Mon Sep 17 00:00:00 2001
From: liugz18 <57401541+liugz18@users.noreply.github.com>
Date: 星期四, 18 七月 2024 21:34:55 +0800
Subject: [PATCH] Rename 'res' in line 514 to avoid with naming conflict with line 365
---
funasr/losses/label_smoothing_loss.py | 62 ++++++++++++++++++++++++++-----
1 files changed, 52 insertions(+), 10 deletions(-)
diff --git a/funasr/losses/label_smoothing_loss.py b/funasr/losses/label_smoothing_loss.py
index 28df73f..ffc38da 100644
--- a/funasr/losses/label_smoothing_loss.py
+++ b/funasr/losses/label_smoothing_loss.py
@@ -8,7 +8,7 @@
import torch
from torch import nn
-from funasr.modules.nets_utils import make_pad_mask
+from funasr.models.transformer.utils.nets_utils import make_pad_mask
class LabelSmoothingLoss(nn.Module):
@@ -50,8 +50,8 @@
"""
assert x.size(2) == self.size
batch_size = x.size(0)
- x = x.view(-1, self.size)
- target = target.view(-1)
+ x = x.contiguous().view(-1, self.size)
+ target = target.contiguous().view(-1)
with torch.no_grad():
true_dist = x.clone()
true_dist.fill_(self.smoothing / (self.size - 1))
@@ -65,17 +65,59 @@
class SequenceBinaryCrossEntropy(nn.Module):
- def __init__(
- self,
- normalize_length=False,
- criterion=nn.BCEWithLogitsLoss(reduction="none")
- ):
+ def __init__(self, normalize_length=False, criterion=nn.BCEWithLogitsLoss(reduction="none")):
super().__init__()
self.normalize_length = normalize_length
self.criterion = criterion
def forward(self, pred, label, lengths):
- pad_mask = make_pad_mask(lengths, maxlen=pred.shape[1])
+ pad_mask = make_pad_mask(lengths, maxlen=pred.shape[1]).to(pred.device)
loss = self.criterion(pred, label)
denom = (~pad_mask).sum() if self.normalize_length else pred.shape[0]
- return loss.masked_fill(pad_mask, 0).sum() / denom
+ return loss.masked_fill(pad_mask.unsqueeze(-1), 0).sum() / denom
+
+
+class NllLoss(nn.Module):
+ """Nll loss.
+
+ :param int size: the number of class
+ :param int padding_idx: ignored class id
+ :param bool normalize_length: normalize loss by sequence length if True
+ :param torch.nn.Module criterion: loss function
+ """
+
+ def __init__(
+ self,
+ size,
+ padding_idx,
+ normalize_length=False,
+ criterion=nn.NLLLoss(reduction="none"),
+ ):
+ """Construct an NllLoss object."""
+ super(NllLoss, self).__init__()
+ self.criterion = criterion
+ self.padding_idx = padding_idx
+ self.size = size
+ self.true_dist = None
+ self.normalize_length = normalize_length
+
+ def forward(self, x, target):
+ """Compute loss between x and target.
+
+ :param torch.Tensor x: prediction (batch, seqlen, class)
+ :param torch.Tensor target:
+ target signal masked with self.padding_id (batch, seqlen)
+ :return: scalar float value
+ :rtype torch.Tensor
+ """
+ assert x.size(2) == self.size
+ batch_size = x.size(0)
+ x = x.view(-1, self.size)
+ target = target.view(-1)
+ with torch.no_grad():
+ ignore = target == self.padding_idx # (B,)
+ total = len(target) - ignore.sum().item()
+ target = target.masked_fill(ignore, 0) # avoid -1 index
+ kl = self.criterion(x, target)
+ denom = total if self.normalize_length else batch_size
+ return kl.masked_fill(ignore, 0).sum() / denom
--
Gitblit v1.9.1