本文记录自己的学习过程,内容包括:
代码解读:Pytorch-UNet
深度学习编程基础:Pytorch-深度学习(新手友好)
UNet论文解读:医学图像分割:U_Net 论文阅读
数据:https://hackernoon.com/hacking-gta-v-for-carvana-kaggle-challenge-6d0b7fb4c781
完整代码解读详见:U-Net代码复现--更新中
utils
dice_score.py
在train.py中:
loss = criterion(masks_pred.squeeze(1), true_masks.float())
loss += dice_loss(F.sigmoid(masks_pred.squeeze(1)), true_masks.float(), multiclass=False)
其中:
- 交叉信息熵:
criterion = nn.CrossEntropyLoss() if model.n_classes > 1 else nn.BCEWithLogitsLoss()
. - Dice 系数(Dice coefficient) 损失函数:
dice_loss(F.sigmoid(masks_pred.squeeze(1)), true_masks.float(), multiclass=False)
Dice coefficient也被称为F1-score,是一种用于衡量两个集合相似度的指标,通常用于图像分割任务中。在医学图像分割中,Dice coefficient常用于衡量模型预测的分割掩码与真实标签之间的相似度。Dice coefficient越接近1,表示预测结果与真实标签越相似。
dice coefficient 源于二分类,本质上是衡量两个样本的重叠部分。该指标范围从0到1,其中"1"表示完整的重叠。 其计算公式为:
D i c e = 2 ∣ A ∩ B ∣ ∣ A ∣ + ∣ B ∣ Dice=\frac{2|A∩B|}{|A|+|B|} Dice=∣A∣+∣B∣2∣A∩B∣
其中 ∣ A ∩ B ∣ |A∩B| ∣A∩B∣ 表示集合A、B 之间的共同元素,|A| 表示 A 中的元素的个数,B也用相似的表示方法。
关于inter = 2 * (input * target).sum(dim=sum_dim)
中sum
的理解参考:Pytorch:torch.sum()函数用法
python
def dice_coeff(input: Tensor, target: Tensor, reduce_batch_first: bool = False, epsilon: float = 1e-6):
# Average of Dice coefficient for all batches, or for a single mask
# 确保input和target数量一致
assert input.size() == target.size()
# 确保imput维度为3,或者reduce_batch_first为False(这里reduce_batch_first为True,所以input.dim() == 3)
assert input.dim() == 3 or not reduce_batch_first
# sum_dim = (-1, -2, -3)
sum_dim = (-1, -2) if input.dim() == 2 or not reduce_batch_first else (-1, -2, -3)
# input 和 target像素为[0,1],相乘则得到均为1的部分,即A∩B
inter = 2 * (input * target).sum(dim=sum_dim) # 对所有维度求和
# 计算|A|+|B|
sets_sum = input.sum(dim=sum_dim) + target.sum(dim=sum_dim)
sets_sum = torch.where(sets_sum == 0, inter, sets_sum)
dice = (inter + epsilon) / (sets_sum + epsilon)
return dice.mean()
python
def multiclass_dice_coeff(input: Tensor, target: Tensor, reduce_batch_first: bool = False, epsilon: float = 1e-6):
# Average of Dice coefficient for all classes
return dice_coeff(input.flatten(0, 1), target.flatten(0, 1), reduce_batch_first, epsilon)
python
def dice_loss(input: Tensor, target: Tensor, multiclass: bool = False):
# Dice loss (objective to minimize) between 0 and 1
fn = multiclass_dice_coeff if multiclass else dice_coeff
return 1 - fn(input, target, reduce_batch_first=True)