atom-predict/msunet/core/metrics.py

60 lines
1.5 KiB
Python
Executable File

import torch
import torch.nn as nn
class DiceLoss(nn.Module):
def __init__(self):
super(DiceLoss, self).__init__()
def forward(self, x, y, eps=1):
N = y.size(0)
x_flat = x.view(N, -1)
y_flat = y.view(N, -1)
inter = (x_flat * y_flat).sum(1)
union = x_flat.sum(1) + y_flat.sum(1)
dice_coefficient = 2. * (inter + eps) / (union + eps)
loss = 1. - dice_coefficient.sum() / N
return loss
class MyLoss(nn.Module):
def __init__(self, weights):
super().__init__()
self.bceloss = nn.BCELoss()
self.diceloss = DiceLoss()
self.weights = weights
def forward(self, x, y):
loss = 0.
for i in range(len(x)):
loss += (self.bceloss(x[i], y[i]) + self.diceloss(x[i], y[i])) * self.weights[i]
return loss
def mce(y_pred, y_true):
x = torch.sum(y_pred[3], dim=[1, 2, 3])
y = torch.sum(y_true[3], dim=[1, 2, 3])
return torch.mean(torch.abs(x - y) / 100.)
def iou(y_pred, y_true, eps=1.):
y_pred = y_pred[-1] > 0.5
y_true = y_true[-1]
iou_score = torch.mean((y_pred * y_true).sum((1, 2, 3)) / (torch.Tensor((y_true + y_pred) != 0.).sum((1, 2, 3)) + eps))
return iou_score
def dice(y_pred, y_true, eps=1.):
y_pred = y_pred[-1] > 0.5
y_true = y_true[-1]
dice_score = 2 * torch.mean((y_pred * y_true).sum((1, 2, 3)) / (y_true.sum((1, 2, 3)) + y_pred.sum((1, 2, 3)) + eps))
return dice_score