質問編集履歴
1
プログラムを全文記載しました。
test
CHANGED
File without changes
|
test
CHANGED
@@ -28,13 +28,97 @@
|
|
28
28
|
|
29
29
|
```python3
|
30
30
|
|
31
|
+
def target_category_loss(x, category_index, nb_classes):
|
32
|
+
|
33
|
+
return tf.multiply(x, K.one_hot([category_index], nb_classes))
|
34
|
+
|
35
|
+
|
36
|
+
|
37
|
+
def target_category_loss_output_shape(input_shape):
|
38
|
+
|
39
|
+
return input_shape
|
40
|
+
|
41
|
+
|
42
|
+
|
43
|
+
def normalize(x):
|
44
|
+
|
45
|
+
return x / (K.sqrt(K.mean(K.square(x))) + 1e-5)
|
46
|
+
|
47
|
+
|
48
|
+
|
49
|
+
def load_image(path):
|
50
|
+
|
51
|
+
img_path = path
|
52
|
+
|
53
|
+
img = image.load_img(img_path, target_size=(224, 224))
|
54
|
+
|
55
|
+
x = image.img_to_array(img)
|
56
|
+
|
57
|
+
x = np.expand_dims(x, axis=0)
|
58
|
+
|
59
|
+
x = preprocess_input(x)
|
60
|
+
|
61
|
+
return x
|
62
|
+
|
63
|
+
|
64
|
+
|
65
|
+
|
66
|
+
|
67
|
+
def register_gradient():
|
68
|
+
|
69
|
+
if "GuidedBackProp" not in ops._gradient_registry._registry:
|
70
|
+
|
71
|
+
@ops.RegisterGradient("GuidedBackProp")
|
72
|
+
|
73
|
+
def _GuidedBackProp(op, grad):
|
74
|
+
|
75
|
+
dtype = op.inputs[0].dtype
|
76
|
+
|
77
|
+
return grad * tf.cast(grad > 0., dtype) * \
|
78
|
+
|
79
|
+
tf.cast(op.inputs[0] > 0., dtype)
|
80
|
+
|
81
|
+
|
82
|
+
|
83
|
+
|
84
|
+
|
85
|
+
def compile_saliency_function(model, activation_layer='block5_conv3'):
|
86
|
+
|
87
|
+
input_img = model.input
|
88
|
+
|
89
|
+
layer_dict = dict([(layer.name, layer) for layer in model.layers[1:]])
|
90
|
+
|
91
|
+
layer_output = layer_dict[activation_layer].output
|
92
|
+
|
93
|
+
max_output = K.max(layer_output, axis=3)
|
94
|
+
|
95
|
+
saliency = K.gradients(K.sum(max_output), input_img)[0]
|
96
|
+
|
97
|
+
return K.function([input_img, K.learning_phase()], [saliency])
|
98
|
+
|
99
|
+
|
100
|
+
|
101
|
+
|
102
|
+
|
103
|
+
def modify_backprop(model, name):
|
104
|
+
|
105
|
+
g = tf.get_default_graph()
|
106
|
+
|
107
|
+
with g.gradient_override_map({'Relu': name}):
|
108
|
+
|
109
|
+
new_model = VGG16(weights='imagenet')
|
110
|
+
|
111
|
+
return new_model
|
112
|
+
|
113
|
+
|
114
|
+
|
115
|
+
|
116
|
+
|
31
|
-
|
117
|
+
def deprocess_image(x):
|
32
118
|
|
33
119
|
if np.ndim(x) > 3:
|
34
120
|
|
35
|
-
x = np.squeeze(x)
|
121
|
+
x = np.squeeze(x)
|
36
|
-
|
37
|
-
# normalize tensor: center on 0., ensure std is 0.1
|
38
122
|
|
39
123
|
x -= x.mean()
|
40
124
|
|
@@ -44,16 +128,12 @@
|
|
44
128
|
|
45
129
|
|
46
130
|
|
47
|
-
# clip to [0, 1]
|
48
|
-
|
49
131
|
x += 0.5
|
50
132
|
|
51
133
|
x = np.clip(x, 0, 1)
|
52
134
|
|
53
135
|
|
54
136
|
|
55
|
-
# convert to RGB array
|
56
|
-
|
57
137
|
x *= 255
|
58
138
|
|
59
139
|
if K.image_dim_ordering() == 'tf':
|
@@ -66,6 +146,112 @@
|
|
66
146
|
|
67
147
|
|
68
148
|
|
149
|
+
|
150
|
+
|
151
|
+
def grad_cam(input_model, image, category_index, layer_name):
|
152
|
+
|
153
|
+
nb_classes = 1000
|
154
|
+
|
155
|
+
target_layer = lambda x: target_category_loss(x, category_index, nb_classes)
|
156
|
+
|
157
|
+
|
158
|
+
|
159
|
+
x = input_model.layers[-1].output
|
160
|
+
|
161
|
+
x = Lambda(target_layer, output_shape=target_category_loss_output_shape)(x)
|
162
|
+
|
163
|
+
model = keras.models.Model(input_model.layers[0].input, x)
|
164
|
+
|
165
|
+
|
166
|
+
|
167
|
+
loss = K.sum(model.layers[-1].output)
|
168
|
+
|
169
|
+
conv_output = [l for l in model.layers if l.name is layer_name][0].output
|
170
|
+
|
171
|
+
|
172
|
+
|
173
|
+
grads = normalize(K.gradients(loss, conv_output)[0])
|
174
|
+
|
175
|
+
gradient_function = K.function([model.layers[0].input], [conv_output, grads])
|
176
|
+
|
177
|
+
|
178
|
+
|
179
|
+
output, grads_val = gradient_function([image])
|
180
|
+
|
181
|
+
output, grads_val = output[0, :], grads_val[0, :, :, :]
|
182
|
+
|
183
|
+
|
184
|
+
|
185
|
+
weights = np.mean(grads_val, axis = (0, 1))
|
186
|
+
|
187
|
+
cam = np.ones(output.shape[0 : 2], dtype = np.float32)
|
188
|
+
|
189
|
+
|
190
|
+
|
191
|
+
for i, w in enumerate(weights):
|
192
|
+
|
193
|
+
cam += w * output[:, :, i]
|
194
|
+
|
195
|
+
|
196
|
+
|
197
|
+
cam = cv2.resize(cam, (224, 224))
|
198
|
+
|
199
|
+
cam = np.maximum(cam, 0)
|
200
|
+
|
201
|
+
heatmap = cam / np.max(cam)
|
202
|
+
|
203
|
+
|
204
|
+
|
205
|
+
image = image[0, :]
|
206
|
+
|
207
|
+
image -= np.min(image)
|
208
|
+
|
209
|
+
image = np.minimum(image, 255)
|
210
|
+
|
211
|
+
|
212
|
+
|
213
|
+
cam = cv2.applyColorMap(np.uint8(255*heatmap), cv2.COLORMAP_JET)
|
214
|
+
|
215
|
+
cam = np.float32(cam) + np.float32(image)
|
216
|
+
|
217
|
+
cam = 255 * cam / np.max(cam)
|
218
|
+
|
219
|
+
return np.uint8(cam), heatmap
|
220
|
+
|
221
|
+
|
222
|
+
|
223
|
+
preprocessed_input = load_image("./dog_cat.jpg")
|
224
|
+
|
225
|
+
|
226
|
+
|
227
|
+
model = VGG16(weights='imagenet')
|
228
|
+
|
229
|
+
|
230
|
+
|
231
|
+
predictions = model.predict(preprocessed_input)
|
232
|
+
|
233
|
+
top_1 = decode_predictions(predictions)[0][0]
|
234
|
+
|
235
|
+
print('Predicted class:')
|
236
|
+
|
237
|
+
print('%s (%s) with probability %.2f' % (top_1[1], top_1[0], top_1[2]))
|
238
|
+
|
239
|
+
|
240
|
+
|
241
|
+
predicted_class = np.argmax(predictions)
|
242
|
+
|
243
|
+
print(predicted_class)
|
244
|
+
|
245
|
+
|
246
|
+
|
247
|
+
cam, heatmap = grad_cam(model, preprocessed_input, predicted_class, "block5_conv3")
|
248
|
+
|
249
|
+
|
250
|
+
|
251
|
+
cv2.imwrite("gradcam.jpg", cam)
|
252
|
+
|
253
|
+
|
254
|
+
|
69
255
|
register_gradient()
|
70
256
|
|
71
257
|
guided_model = modify_backprop(model, 'GuidedBackProp')
|
@@ -76,11 +262,7 @@
|
|
76
262
|
|
77
263
|
gradcam = saliency[0] * heatmap[..., np.newaxis]
|
78
264
|
|
79
|
-
|
80
|
-
|
81
|
-
cv2.imwrite("guied_gradcam.jpg", deprocess_image(gradcam))
|
265
|
+
cv2.imwrite("guided_gradcam.jpg", deprocess_image(gradcam))
|
82
|
-
|
83
|
-
|
84
266
|
|
85
267
|
|
86
268
|
|