forked from kazuto1011/grad-cam-pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
grad_cam.py
216 lines (171 loc) · 6.66 KB
/
grad_cam.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
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
#!/usr/bin/env python
# coding: utf-8
#
# Author: Kazuto Nakashima
# URL: http://kazuto1011.github.io
# Created: 2017-05-26
from collections import Sequence
import numpy as np
import torch
import torch.nn as nn
from torch.nn import functional as F
from tqdm import tqdm
class _BaseWrapper(object):
def __init__(self, model):
super(_BaseWrapper, self).__init__()
self.device = next(model.parameters()).device
self.model = model
self.handlers = [] # a set of hook function handlers
def _encode_one_hot(self, ids):
one_hot = torch.zeros_like(self.logits).to(self.device)
one_hot.scatter_(1, ids, 1.0)
return one_hot
def forward(self, image):
self.image_shape = image.shape[2:]
self.logits = self.model(image)
self.probs = F.softmax(self.logits, dim=1)
return self.probs.sort(dim=1, descending=True) # ordered results
def backward(self, ids):
"""
Class-specific backpropagation
"""
one_hot = self._encode_one_hot(ids)
self.model.zero_grad()
self.logits.backward(gradient=one_hot, retain_graph=True)
def generate(self):
raise NotImplementedError
def remove_hook(self):
"""
Remove all the forward/backward hook functions
"""
for handle in self.handlers:
handle.remove()
class BackPropagation(_BaseWrapper):
def forward(self, image):
self.image = image.requires_grad_()
return super(BackPropagation, self).forward(self.image)
def generate(self):
gradient = self.image.grad.clone()
self.image.grad.zero_()
return gradient
class GuidedBackPropagation(BackPropagation):
"""
"Striving for Simplicity: the All Convolutional Net"
https://arxiv.org/pdf/1412.6806.pdf
Look at Figure 1 on page 8.
"""
def __init__(self, model):
super(GuidedBackPropagation, self).__init__(model)
def backward_hook(module, grad_in, grad_out):
# Cut off negative gradients
if isinstance(module, nn.ReLU):
return (F.relu(grad_in[0]),)
for module in self.model.named_modules():
self.handlers.append(module[1].register_backward_hook(backward_hook))
class Deconvnet(BackPropagation):
"""
"Striving for Simplicity: the All Convolutional Net"
https://arxiv.org/pdf/1412.6806.pdf
Look at Figure 1 on page 8.
"""
def __init__(self, model):
super(Deconvnet, self).__init__(model)
def backward_hook(module, grad_in, grad_out):
# Cut off negative gradients and ignore ReLU
if isinstance(module, nn.ReLU):
return (F.relu(grad_out[0]),)
for module in self.model.named_modules():
self.handlers.append(module[1].register_backward_hook(backward_hook))
class GradCAM(_BaseWrapper):
"""
"Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization"
https://arxiv.org/pdf/1610.02391.pdf
Look at Figure 2 on page 4
"""
def __init__(self, model, candidate_layers=None):
super(GradCAM, self).__init__(model)
self.fmap_pool = {}
self.grad_pool = {}
self.candidate_layers = candidate_layers # list
def save_fmaps(key):
def forward_hook(module, input, output):
self.fmap_pool[key] = output.detach()
return forward_hook
def save_grads(key):
def backward_hook(module, grad_in, grad_out):
self.grad_pool[key] = grad_out[0].detach()
return backward_hook
# If any candidates are not specified, the hook is registered to all the layers.
for name, module in self.model.named_modules():
if self.candidate_layers is None or name in self.candidate_layers:
self.handlers.append(module.register_forward_hook(save_fmaps(name)))
self.handlers.append(module.register_backward_hook(save_grads(name)))
def _find(self, pool, target_layer):
if target_layer in pool.keys():
return pool[target_layer]
else:
raise ValueError("Invalid layer name: {}".format(target_layer))
def generate(self, target_layer):
fmaps = self._find(self.fmap_pool, target_layer)
grads = self._find(self.grad_pool, target_layer)
weights = F.adaptive_avg_pool2d(grads, 1)
gcam = torch.mul(fmaps, weights).sum(dim=1, keepdim=True)
gcam = F.relu(gcam)
gcam = F.interpolate(
gcam, self.image_shape, mode="bilinear", align_corners=False
)
B, C, H, W = gcam.shape
gcam = gcam.view(B, -1)
gcam -= gcam.min(dim=1, keepdim=True)[0]
gcam /= gcam.max(dim=1, keepdim=True)[0]
gcam = gcam.view(B, C, H, W)
return gcam
def occlusion_sensitivity(
model, images, ids, mean=None, patch=35, stride=1, n_batches=128
):
"""
"Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization"
https://arxiv.org/pdf/1610.02391.pdf
Look at Figure A5 on page 17
Originally proposed in:
"Visualizing and Understanding Convolutional Networks"
https://arxiv.org/abs/1311.2901
"""
torch.set_grad_enabled(False)
model.eval()
mean = mean if mean else 0
patch_H, patch_W = patch if isinstance(patch, Sequence) else (patch, patch)
pad_H, pad_W = patch_H // 2, patch_W // 2
# Padded image
images = F.pad(images, (pad_W, pad_W, pad_H, pad_H), value=mean)
B, _, H, W = images.shape
new_H = (H - patch_H) // stride + 1
new_W = (W - patch_W) // stride + 1
# Prepare sampling grids
anchors = []
grid_h = 0
while grid_h <= H - patch_H:
grid_w = 0
while grid_w <= W - patch_W:
grid_w += stride
anchors.append((grid_h, grid_w))
grid_h += stride
# Baseline score without occlusion
baseline = model(images).detach().gather(1, ids)
# Compute per-pixel logits
scoremaps = []
for i in tqdm(range(0, len(anchors), n_batches), leave=False):
batch_images = []
batch_ids = []
for grid_h, grid_w in anchors[i : i + n_batches]:
images_ = images.clone()
images_[..., grid_h : grid_h + patch_H, grid_w : grid_w + patch_W] = mean
batch_images.append(images_)
batch_ids.append(ids)
batch_images = torch.cat(batch_images, dim=0)
batch_ids = torch.cat(batch_ids, dim=0)
scores = model(batch_images).detach().gather(1, batch_ids)
scoremaps += list(torch.split(scores, B))
diffmaps = torch.cat(scoremaps, dim=1) - baseline
diffmaps = diffmaps.view(B, new_H, new_W)
return diffmaps