add more loss

setup.py

@@ -1,7 +1,7 @@
 from setuptools import setup, find_packages
 setup(
     name='deepke',  # 打包后的包文件名
-    version='0.2.90',    #版本号
+    version='0.2.91',    #版本号
     keywords=["pip", "RE","NER","AE"],    # 关键字
     description='DeepKE 是基于 Pytorch 的深度学习中文关系抽取处理套件。',  # 说明
     long_description="client",  #详细说明

src/deepke/relation_extraction/standard/tools/__init__.py

@@ -1,4 +1,5 @@
 from .dataset import *
+from .loss import *
 from .metrics import *
 from .preprocess import *
 from .serializer import *

src/deepke/relation_extraction/standard/tools/loss.py

@@ -0,0 +1,70 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class LabelSmoothSoftmaxCEV1(nn.Module):
+    def __init__(self, lb_smooth=0.1, reduction='mean', ignore_index=-100):
+        super(LabelSmoothSoftmaxCEV1, self).__init__()
+        self.lb_smooth = lb_smooth
+        self.reduction = reduction
+        self.lb_ignore = ignore_index
+        self.log_softmax = nn.LogSoftmax(dim=1)
+
+    def forward(self, logits, label):
+        
+        logits = logits.float() # use fp32 to avoid nan
+        with torch.no_grad():
+            num_classes = logits.size(1)
+            label = label.clone().detach()
+            ignore = label.eq(self.lb_ignore)
+            n_valid = ignore.eq(0).sum()
+            label[ignore] = 0
+            lb_pos, lb_neg = 1. - self.lb_smooth, self.lb_smooth / num_classes
+            lb_one_hot = torch.empty_like(logits).fill_(
+                lb_neg).scatter_(1, label.unsqueeze(1), lb_pos).detach()
+
+        logs = self.log_softmax(logits)
+        loss = -torch.sum(logs * lb_one_hot, dim=1)
+        loss[ignore] = 0
+        if self.reduction == 'mean':
+            loss = loss.sum() / n_valid
+        if self.reduction == 'sum':
+            loss = loss.sum()
+
+        return loss
+
+
+def taylor_softmax_v1(x, dim=1, n=4, use_log=False):
+    assert n % 2 == 0 and n > 0
+    fn = torch.ones_like(x)
+    denor = 1.
+    for i in range(1, n + 1):
+        denor *= i
+        fn = fn + x.pow(i) / denor
+    out = fn / fn.sum(dim=dim, keepdims=True)
+    if use_log: out = out.log()
+    return out
+
+class LogTaylorSoftmaxV1(nn.Module):
+    def __init__(self, dim=1, n=2):
+        super(LogTaylorSoftmaxV1, self).__init__()
+        assert n % 2 == 0
+        self.dim = dim
+        self.n = n
+    def forward(self, x):
+        return taylor_softmax_v1(x, self.dim, self.n, use_log=True)
+
+
+class TaylorCrossEntropyLossV1(nn.Module):
+    def __init__(self, n=2, ignore_index=-1, reduction='mean'):
+        super(TaylorCrossEntropyLossV1, self).__init__()
+        assert n % 2 == 0
+        self.taylor_softmax = LogTaylorSoftmaxV1(dim=1, n=n)
+        self.reduction = reduction
+        self.ignore_index = ignore_index
+
+    def forward(self, logits, labels):
+        log_probs = self.taylor_softmax(logits)
+        loss = F.nll_loss(log_probs, labels, reduction=self.reduction,
+                ignore_index=self.ignore_index)
+        return loss