146 lines
5.2 KiB
Python
146 lines
5.2 KiB
Python
import torch
|
||
import torch.nn as nn
|
||
import torch.nn.functional as F
|
||
|
||
|
||
class BinaryCrossEntropy(nn.Module):
|
||
"""
|
||
基本的二元交叉熵損失函數
|
||
"""
|
||
def __init__(self, reduction='mean'):
|
||
"""
|
||
初始化
|
||
|
||
Args:
|
||
reduction (str): 'mean', 'sum' 或 'none'，指定如何減少損失
|
||
"""
|
||
super(BinaryCrossEntropy, self).__init__()
|
||
|
||
def forward(self, predictions, targets):
|
||
"""
|
||
計算二元交叉熵損失
|
||
|
||
Args:
|
||
predictions (torch.Tensor): 模型的預測輸出，形狀為 [batch_size, ...]
|
||
targets (torch.Tensor): 目標標籤，形狀與 predictions 相同
|
||
|
||
Returns:
|
||
torch.Tensor: 計算得到的損失值
|
||
"""
|
||
# 確保輸入是張量
|
||
predictions = torch.as_tensor(predictions, dtype=torch.float32)
|
||
targets = torch.as_tensor(targets, dtype=torch.float32)
|
||
Loss = nn.BCELoss()
|
||
|
||
return Loss(predictions, targets)
|
||
|
||
# # 檢查輸出和標籤的維度是否匹配
|
||
# if predictions.shape[1] != targets.shape[1]:
|
||
# # 如果維度不匹配，使用交叉熵損失函數
|
||
# # 對於交叉熵損失，標籤需要是類別索引而不是one-hot編碼
|
||
# # 將one-hot編碼轉換為類別索引
|
||
# _, targets_indices = torch.max(targets, dim=1)
|
||
# return F.cross_entropy(predictions, targets_indices, reduction=self.reduction)
|
||
# else:
|
||
# # 如果維度匹配，使用二元交叉熵損失函數
|
||
# # 使用 PyTorch 內建的 binary_cross_entropy_with_logits 函數
|
||
# # 它會自動應用 sigmoid 函數，避免輸入值超出 [0,1] 範圍
|
||
# return F.binary_cross_entropy_with_logits(predictions, targets, reduction=self.reduction)
|
||
|
||
|
||
class WeightedBinaryCrossEntropy(nn.Module):
|
||
"""
|
||
帶權重的二元交叉熵損失函數
|
||
"""
|
||
def __init__(self, pos_weight=1.0, neg_weight=1.0, reduction='mean'):
|
||
"""
|
||
初始化
|
||
|
||
Args:
|
||
pos_weight (float): 正樣本的權重
|
||
neg_weight (float): 負樣本的權重
|
||
reduction (str): 'mean', 'sum' 或 'none'，指定如何減少損失
|
||
"""
|
||
super(WeightedBinaryCrossEntropy, self).__init__()
|
||
self.pos_weight = pos_weight
|
||
self.neg_weight = neg_weight
|
||
self.reduction = reduction
|
||
|
||
def forward(self, predictions, targets):
|
||
"""
|
||
計算帶權重的二元交叉熵損失
|
||
|
||
Args:
|
||
predictions (torch.Tensor): 模型的預測輸出，形狀為 [batch_size, ...]
|
||
targets (torch.Tensor): 目標標籤，形狀與 predictions 相同
|
||
|
||
Returns:
|
||
torch.Tensor: 計算得到的損失值
|
||
"""
|
||
# 確保輸入是張量
|
||
predictions = torch.as_tensor(predictions, dtype=torch.float32)
|
||
targets = torch.as_tensor(targets, dtype=torch.float32)
|
||
|
||
# 使用 sigmoid 確保預測值在 [0,1] 範圍內
|
||
predictions = torch.sigmoid(predictions)
|
||
|
||
# 計算帶權重的二元交叉熵損失
|
||
loss = -self.pos_weight * targets * torch.log(predictions + 1e-7) - \
|
||
self.neg_weight * (1 - targets) * torch.log(1 - predictions + 1e-7)
|
||
|
||
# 根據 reduction 方式返回損失
|
||
if self.reduction == 'mean':
|
||
return loss.mean()
|
||
elif self.reduction == 'sum':
|
||
return loss.sum()
|
||
else: # 'none'
|
||
return loss
|
||
|
||
|
||
class LabelSmoothingBCE(nn.Module):
|
||
"""
|
||
帶標籤平滑的二元交叉熵損失函數
|
||
"""
|
||
def __init__(self, smoothing=0.1, reduction='mean'):
|
||
"""
|
||
初始化
|
||
|
||
Args:
|
||
smoothing (float): 標籤平滑係數，範圍 [0, 1]
|
||
reduction (str): 'mean', 'sum' 或 'none'，指定如何減少損失
|
||
"""
|
||
super(LabelSmoothingBCE, self).__init__()
|
||
self.smoothing = smoothing
|
||
self.reduction = reduction
|
||
|
||
def forward(self, predictions, targets):
|
||
"""
|
||
計算帶標籤平滑的二元交叉熵損失
|
||
|
||
Args:
|
||
predictions (torch.Tensor): 模型的預測輸出，形狀為 [batch_size, ...]
|
||
targets (torch.Tensor): 目標標籤，形狀與 predictions 相同
|
||
|
||
Returns:
|
||
torch.Tensor: 計算得到的損失值
|
||
"""
|
||
# 確保輸入是張量
|
||
predictions = torch.as_tensor(predictions, dtype=torch.float32)
|
||
targets = torch.as_tensor(targets, dtype=torch.float32)
|
||
|
||
# 應用標籤平滑
|
||
targets = targets * (1 - self.smoothing) + 0.5 * self.smoothing
|
||
|
||
# 使用 sigmoid 確保預測值在 [0,1] 範圍內
|
||
predictions = torch.sigmoid(predictions)
|
||
|
||
# 計算二元交叉熵損失
|
||
loss = -targets * torch.log(predictions + 1e-7) - (1 - targets) * torch.log(1 - predictions + 1e-7)
|
||
|
||
# 根據 reduction 方式返回損失
|
||
if self.reduction == 'mean':
|
||
return loss.mean()
|
||
elif self.reduction == 'sum':
|
||
return loss.sum()
|
||
else: # 'none'
|
||
return loss |