-
Notifications
You must be signed in to change notification settings - Fork 0
/
scorecam.py
298 lines (226 loc) · 9.12 KB
/
scorecam.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
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
"""
PyTorch implementation of Score-CAM.
"""
# Import standard libraries.
import copy
# Import third-party packages.
import numpy as np
import torch
class ScoreCAM:
"""
PyTorch implementation of Score-CAM [1].
Example:
>>> from torchvision.models import resnet18
>>> from scorecam import ScoreCAM
>>>
>>> # Load NN model.
>>> model = resnet18(weights="IMAGENET1K_V1")
>>>
>>> # Create Score-CAM instance.
>>> scorecam = ScoreCAM(model, actmap="layer4")
>>>
>>> # Compute visual explanation.
>>> L = scorecam.compute(x, coi=242)
>>> print(L)
References:
[1] H. Wang, Z. Wang, M. Du, F. Yang, Z. Zhang, S. Ding, P. Mardziel, and X. Hu,
"Score-CAM: Score-weighted visual explanations for convolutional neural networks",
CVPR, 2020.
"""
def __init__(self, model, actmap, device="cpu"):
"""
Constructor.
Args:
model (torch.nn.Module): Target NN model.
actmap (str) : Name of layer to extract activation maps.
device (str) : Device name ("cpu" or "cuda").
"""
# Copy the target NN model and prepare it.
self.model = copy.deepcopy(model.to("cpu"))
self.model.to(device)
self.model.eval()
# Register a hook function to extract activation maps.
getattr(self.model, actmap).register_forward_hook(self.hook)
def hook(self, module, x_in, x_out):
"""
Hook function to extract activation maps.
This function is assumed to be registered to NN layer.
Args:
module (torch.nn.Module): Target layer.
x_in (torch.Tensor) : Input tensor of the layer.
x_out (torch.Tensor) : Output tensor of the layer.
"""
self.activation_map = x_out.detach().to("cpu")
def compute(self, X, coi, batch_size=128, cskip=False, cskip_out=16):
"""
Compute visual explanation.
Args:
X (np.ndarray or torch.Tensor): Input image.
coi (int or Callable) : Class of interest.
batch_size (int) : Batch size.
cskip (bool) : Enable CSKIP optimization.
cskip_out (int) : Output channels of the CSKIP.
Returns:
(torch.Tensor): 2D array of visual explanation.
"""
# Define a scoring function.
if isinstance(coi, int):
self.scoring_fn = lambda output: output[:, coi]
elif hasattr(coi, "__call__"):
self.scoring_fn = coi
# Verify the data type of the input array.
if (type(X) != np.ndarray) and (type(X) != torch.Tensor):
raise TypeError("input array should be NumPy or PyTorch array.")
# Get device.
device = next(self.model.parameters()).device
# Convert the input tensor to torch.Tensor on CPU.
X = torch.Tensor(X).to("cpu")
# Reshape the input tensor to (B, C, H, W).
X, (B, C, H, W) = reshape_input_tensor(X)
# Run inference and get activation maps. The activation maps are
# acquired by the hook function registered in the __init__ function.
with torch.no_grad():
p = self.model(X.to(device))
# Get the reference score.
s_ref = self.scoring_fn(p.to("cpu"))
# Apply CSKIP if specified.
if cskip: A = channel_skipping(self.activation_map, cskip_out)
else : A = self.activation_map
# Get the channel number of the activation maps.
# Note that the activation maps are always on CPU.
K = A.shape[1]
# Upsample the activation maps and change the dimension order
# from [1, K, H, W] to [K, 1, H, W].
A = torch.nn.functional.interpolate(A, (H, W), mode="bicubic")
A = torch.permute(A, [1, 0, 2, 3])
# Normalize the activation maps.
self.A_normalized = normalize_activation_map(A)
# Compute the masked images.
self.M = X * self.A_normalized
# Initialize the list of predictions.
batch_p = list()
for batch_idx_bgn in range(0, K, batch_size):
# Get input batch.
M_batch = self.M[batch_idx_bgn:batch_idx_bgn+batch_size, :, :, :]
# Run inference to get the predictions.
with torch.no_grad():
p = self.model(M_batch.to(device))
# Add predictions to the list.
batch_p.append(p.to("cpu"))
p = torch.concat(batch_p, dim=0)
A = A.to("cpu")
X = X.to("cpu")
# Compute the CIC score for the activation maps.
s = self.scoring_fn(p) - s_ref
a = torch.nn.functional.softmax(s, dim=0)
a = a.reshape([-1, 1, 1])
# The tensor A should have the shape [C, 1, H, W] at this moment.
# Change the shape to [C, H, W] for the following summation.
A = torch.squeeze(A, dim=1)
# Compute the visual explanation.
L = torch.nn.functional.relu(torch.sum(a * A.reshape([K, H, W]), dim=0))
# Returns as NumPy array.
return L.numpy()
@staticmethod
def to_colormap(X, normalize=True):
"""
Convert input 2D array to a color heat map.
Args:
X (np.ndarray): Input array.
normalize (bool) : Normalize the input array if True.
Returns:
(np.ndarray): Color heat map.
"""
# Normalize if specified.
if normalize:
X = (X - np.min(X)) / (np.max(X) - np.min(X))
X = np.clip(255 * X, 0, 255).astype(np.uint8)
# Create JET color map.
CMAP = np.zeros([256, 3], dtype=np.uint8)
for i in range(256):
if i < 32: j = i - 0; CMAP[i, :] = ( 0, 0, 127+4*j)
elif i < 96: j = i - 32; CMAP[i, :] = ( 0, 4*j, 255)
elif i < 160: j = i - 96; CMAP[i, :] = ( 4*j, 255, 255-4*j)
elif i < 224: j = i - 160; CMAP[i, :] = ( 255, 255-4*j, 0)
else : j = i - 224; CMAP[i, :] = (255-4*j, 0, 0)
# Apply the JET color map.
Y = np.stack([CMAP[X, 0], CMAP[X, 1], CMAP[X, 2]], axis=2)
return Y
@staticmethod
def overlay(image, L):
"""
Overlay the given visual explanation to the given image.
Args:
image (np.ndarray or torch.Tensor): Input image with shape (H, W, C).
L (np.ndarray or torch.Tensor): Visual explanation with shape (H, W).
Returns:
(np.ndarray): Overlay image with shape (H, W, C).
"""
# Convert the image and the visual explanation to NumPy array.
if isinstance(image, torch.Tensor):
image = image.numpy()
if isinstance(L, torch.Tensor):
L = L.numpy()
M = ScoreCAM.to_colormap(L)
return np.clip(0.5 * image + 0.5 * M, 0, 255).astype(np.uint8)
def reshape_input_tensor(X):
"""
Reshape input tensor to shape (B, C, H, W) where B is always 1.
Args:
X (torch.Tensor): Input array.
Returns:
(tuple): A tuple of (reshaped array, shape of the array).
"""
# Case 1: X.shape == (H, W)
if len(X.shape) == 2:
# Get array shape.
H, W = X.shape
# Reshape to (B, C, H, W).
X = X.reshape((1, 1, H, W))
return (X, (1, 1, H, W))
# Case 2: X.shape == (H, W, C)
elif len(X.shape) == 3:
# Get array shape.
H, W, C = X.shape
# Reshape to (B, C, H, W).
X = torch.permute(X, [2, 0, 1])
X = X.reshape((1, C, H, W))
return (X, (1, C, H, W))
# Case 3: X.shape == (1, C, H, W)
elif len(X.shape) == 4:
# Verify the batch size is one.
if X.shape[0] != 1:
raise ValueError("Batch size should be 1")
return (X, X.shape)
else:
raise ValueError("Unexpected input shape: %s" % str(X.shape))
def normalize_activation_map(A, eps=1.0E-10):
"""
Normalize the given activation map.
Args:
A (np.ndarray): Activation map with shape (K, 1, H, W).
eps (float) : Small value to prevent zero division.
Returns:
(np.ndarray): Normalized activation map.
"""
# Compute min/max of each channel.
A_min, _ = torch.min(torch.flatten(A, start_dim=2), dim=2)
A_max, _ = torch.max(torch.flatten(A, start_dim=2), dim=2)
# Reshape the min/max array to the shape (K, 1, 1, 1).
A_min = A_min.reshape([-1, 1, 1, 1])
A_max = A_max.reshape([-1, 1, 1, 1])
# Normalize the activation map.
return (A - A_min) / (A_max - A_min + eps)
def channel_skipping(A, cskip_out=16):
"""
CSKIP optimization.
Args:
A (torch.Tensor): Activation maps.
cskip_out (int) : Number of output channels.
"""
# Compute max value of each channel.
A_max, _ = torch.max(torch.flatten(A, start_dim=2), dim=2)
# Sort channels by the max value.
idx_max = torch.argsort(A_max.reshape([-1]), descending=True)
return A[:, idx_max[:cskip_out], :, :]
# vim: expandtab tabstop=4 shiftwidth=4 fdm=marker