-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathpercept_loss.py
127 lines (107 loc) · 4.23 KB
/
percept_loss.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
import os
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from MedicalNet.model import generate_model
from MedicalNet.setting import parse_opts
from collections import OrderedDict
class Variables():
def __init__(self):
self.gpu_id = [0]
self.n_seg_classes = 1
self.img_list = ''
self.n_epochs = 1
self.no_cuda = True
self.data_root = ''
self.pretrain_path = './MedicalNet/pretrain/resnet_10_23dataset.pth'
self.batch_size = 1
self.num_workers = 0
self.model_depth = 10
self.resnet_shortcut = 'B'
self.input_D = 32
self.input_H = 32
self.input_W = 32
self.model = 'resnet'
self.phase = 'test'
class MedPerceptualLoss(torch.nn.Module):
def __init__(self, model, resize=True):
super(MedPerceptualLoss, self).__init__()
blocks = []
blocks.append(model.conv1.eval())
blocks.append(model.bn1.eval())
blocks.append(model.relu.eval()) ####
blocks.append(model.maxpool.eval())
blocks.append(model.layer1.eval()) ####
blocks.append(model.layer2.eval()) ####
# blocks.append(model.layer3.eval())
# blocks.append(model.layer4.eval())
for bl in blocks:
for p in bl.parameters():
p.requires_grad = False
self.blocks = torch.nn.ModuleList(blocks)
self.transform = torch.nn.functional.interpolate
self.resize = resize
self.mean = 271.64814106698583
self.std = 377.117173547721
self.min = (0 - self.mean) / self.std
def denorm(self, volume):
return volume * self.std + self.mean
def norm(self, volume):
"""
normalize the itensity of an nd volume based on the mean and std of nonzeor region
inputs:
volume: the input nd volume
outputs:
out: the normalized nd volume
"""
# volume = self.denorm(volume)
pixels = volume[volume > self.min]
mean = pixels.mean()
std = pixels.std()
out = (volume - mean)/std
out_random = torch.normal(0, 1, size = volume.shape).cuda()
out[volume == 0.] = out_random[volume == 0.]
return out
def forward(self, input, target, feature_layers=[0, 1], style_layers=[0, 1]):
# input = self.norm2(input)
# target = self.norm2(target)
feature_layers = list(np.array(feature_layers) + 4)
feature_layers.append(2)
if self.resize:
orig_size = input.shape[-1]
input = self.transform(input, mode='trilinear', size=(orig_size*2), align_corners=False)
target = self.transform(target, mode='trilinear', size=(orig_size*2), align_corners=False)
loss = 0.0
x = input
y = target
for i, block in enumerate(self.blocks):
x = block(x)
y = block(y)
if i in feature_layers:
loss += torch.nn.functional.l1_loss(x, y)
if i in style_layers:
act_x = x.reshape(x.shape[0], x.shape[1], -1)
act_y = y.reshape(y.shape[0], y.shape[1], -1)
gram_x = act_x @ act_x.permute(0, 2, 1)
gram_y = act_y @ act_y.permute(0, 2, 1)
loss += torch.nn.functional.l1_loss(gram_x, gram_y)
return loss
class MedPercept(nn.Module):
def __init__(self, resize=True):
super(MedPercept, self).__init__()
self.resize = resize
self.sets = Variables()
self.m = generate_model(self.sets)
self.x = torch.load('./MedicalNet/pretrain/resnet_10_23dataset.pth')['state_dict']
self.x2 = OrderedDict()
for k, v in self.x.items():
self.x2[k[7:]] = v
self.model = self.m[0]
self.model.load_state_dict(self.x2, strict=False)
self.percept = MedPerceptualLoss(self.model.cuda(), resize=self.resize)
def forward(self, input, target, feature_layers=[0, 1], style_layers=[]):
if len(input.shape) == 3:
input = torch.unsqueeze(torch.unsqueeze(input,0),0)
target = torch.unsqueeze(torch.unsqueeze(target,0),0)
return self.percept(input, target, feature_layers, style_layers)