質問編集履歴

1

内容の加筆

2020/07/15 07:46

投稿

satoUmino
satoUmino

スコア19

test CHANGED
File without changes
test CHANGED
@@ -6,80 +6,418 @@
6
6
 
7
7
  下記コードの”heatmap”という関数をグラフ化し、x軸、z軸を1から224、y軸を計算される値にしたいです。
8
8
 
9
-
9
+ [参考にしているコードの説明](https://qiita.com/MuAuan/items/cbd739808c64501a1024)
10
+
10
-
11
+ [コードDLリンク](https://github.com/MuAuan/cheating_DL)
12
+
11
-
13
+ grad-cam_5category.pyというプログラムです。
14
+
12
-
15
+ 3Dマップで出したいのは下記のコードのheatmap部分です
16
+
13
-
17
+ 一番下に全体のコードを載せておきます。
18
+
14
-
19
+ printした際、普通にheatmapを出力した時は1列224行でしか出なかったのでfor文で224回繰り返すことで224列分出力しています。
20
+
21
+ この224行224列の行列式を3Dプロットしたいです。
22
+
23
+ ![1列分の出力結果](0bc4b0140c1ecb7057b249da11b3fb62.png)
24
+
15
- ### 該当のソースコード
25
+ ###該当コード
16
-
17
-
18
26
 
19
27
  ```python
20
28
 
21
- for i in range(1):
22
-
23
- i=10#順位指定
24
-
25
- top_1 = decode_predictions(predictions,i)[0]
26
-
27
- #print('label番号',predictions.argsort()[0][::-1][i])
28
-
29
- print(predictions.argsort()[0][::-1][i],predictions[0][::-1][i])
30
-
31
- #print(top_1,i[::-1][i])
32
-
33
- predicted_class = predictions.argsort()[0][::-1][i]
29
+ def grad_cam(input_model, image, category_index, layer_name):
30
+
31
+ nb_classes = 1000
32
+
33
+ target_layer = lambda x: target_category_loss(x, category_index, nb_classes)
34
+
35
+ x = Lambda(target_layer, output_shape = target_category_loss_output_shape)(input_model.output)
36
+
37
+ model = Model(inputs=input_model.input, outputs=x)
38
+
39
+ #model.summary()
40
+
41
+ loss = K.sum(model.output)
42
+
43
+ conv_output = [l for l in model.layers if l.name == layer_name][0].output #is
44
+
45
+ grads = normalize(_compute_gradients(loss, [conv_output])[0])
46
+
47
+ gradient_function = K.function([model.input], [conv_output, grads])
48
+
49
+
50
+
51
+ output, grads_val = gradient_function([image])
52
+
53
+ output, grads_val = output[0, :], grads_val[0, :, :, :]
54
+
55
+
56
+
57
+ weights = np.mean(grads_val, axis = (0, 1))
58
+
59
+ cam = np.ones(output.shape[0 : 2], dtype = np.float32)
60
+
61
+
62
+
63
+ for i, w in enumerate(weights):
64
+
65
+ cam += w * output[:, :, i]
66
+
67
+
68
+
69
+ cam = cv2.resize(cam, (224,224)) #299,299)) #224, 224))
70
+
71
+ cam = np.maximum(cam, 0)
72
+
73
+ heatmap = cam / np.max(cam)
74
+
75
+ for x in range(224):
76
+
77
+ print (heatmap[x]) #3マップで出力したい部分
78
+
79
+
80
+
81
+ #Return to BGR [0..255] from the preprocessed image
82
+
83
+ image = image[0, :]
84
+
85
+ image -= np.min(image)
86
+
87
+ image = np.minimum(image, 255)
88
+
89
+
90
+
91
+ cam = cv2.applyColorMap(np.uint8(255*heatmap), cv2.COLORMAP_JET)
92
+
93
+ cam = np.float32(cam) + np.float32(image)
94
+
95
+ cam = 255 * cam / np.max(cam)
96
+
97
+ return np.uint8(cam), heatmap
98
+
99
+ ```
100
+
101
+
102
+
103
+
104
+
105
+ ###全体のコード
106
+
107
+ ```python
108
+
109
+ from keras.applications.vgg16 import (VGG16, preprocess_input, decode_predictions)
110
+
111
+ from keras.models import Model
112
+
113
+ from keras.preprocessing import image
114
+
115
+ from keras.layers.core import Lambda
116
+
117
+ from keras.models import Sequential
118
+
119
+ from tensorflow.python.framework import ops
120
+
121
+ import keras.backend as K
122
+
123
+ import tensorflow as tf
124
+
125
+ import numpy as np
126
+
127
+ import keras
128
+
129
+ import sys
130
+
131
+ import cv2
132
+
133
+ #from keras.applications.resnet50 import ResNet50, preprocess_input, decode_predictions
134
+
135
+ #from keras.applications.vgg19 import VGG19, preprocess_input, decode_predictions
136
+
137
+ #from keras.applications.inception_v3 import InceptionV3, preprocess_input, decode_predictions
138
+
139
+
140
+
141
+ def target_category_loss(x, category_index, nb_classes):
142
+
143
+ return tf.multiply(x, K.one_hot([category_index], nb_classes))
144
+
145
+
146
+
147
+ def target_category_loss_output_shape(input_shape):
148
+
149
+ return input_shape
150
+
151
+
152
+
153
+ def normalize(x):
154
+
155
+ # utility function to normalize a tensor by its L2 norm
156
+
157
+ return x / (K.sqrt(K.mean(K.square(x))) + 1e-5)
158
+
159
+
160
+
161
+ def load_image(path):
162
+
163
+ img_path = sys.argv[1]
164
+
165
+ img = image.load_img(img_path, target_size=(224,224)) #299,299)) #224, 224))
166
+
167
+ x = image.img_to_array(img)
168
+
169
+ x = np.expand_dims(x, axis=0)
170
+
171
+ x = preprocess_input(x)
172
+
173
+ return x
174
+
175
+
176
+
177
+ def register_gradient():
178
+
179
+ if "GuidedBackProp" not in ops._gradient_registry._registry:
180
+
181
+ @ops.RegisterGradient("GuidedBackProp")
182
+
183
+ def _GuidedBackProp(op, grad):
184
+
185
+ dtype = op.inputs[0].dtype
186
+
187
+ return grad * tf.cast(grad > 0., dtype) * \
188
+
189
+ tf.cast(op.inputs[0] > 0., dtype)
190
+
191
+
192
+
193
+ def compile_saliency_function(model, activation_layer='block5_conv3'): #mixed10 'activation_49' add_16 add_32 activation_98
194
+
195
+ input_img = model.input
196
+
197
+ layer_dict = dict([(layer.name, layer) for layer in model.layers[1:]])
198
+
199
+ #print(layer_dict)
200
+
201
+ layer_output = layer_dict[activation_layer].output
202
+
203
+ max_output = K.max(layer_output, axis=3)
204
+
205
+ saliency = K.gradients(K.sum(max_output), input_img)[0]
206
+
207
+ return K.function([input_img, K.learning_phase()], [saliency])
208
+
209
+
210
+
211
+ def modify_backprop(model, name):
212
+
213
+ g = tf.get_default_graph()
214
+
215
+ with g.gradient_override_map({'Relu': name}):
216
+
217
+
218
+
219
+ # get layers that have an activation
220
+
221
+ layer_dict = [layer for layer in model.layers[1:]
222
+
223
+ if hasattr(layer, 'activation')]
224
+
225
+
226
+
227
+ # replace relu activation
228
+
229
+ for layer in layer_dict:
230
+
231
+ if layer.activation == keras.activations.relu:
232
+
233
+ layer.activation = tf.nn.relu
234
+
235
+
236
+
237
+ # re-instanciate a new model
238
+
239
+ new_model = VGG16(weights='imagenet')
240
+
241
+ #new_model = ResNet50(weights='imagenet')
242
+
243
+ new_model.summary()
244
+
245
+ return new_model
246
+
247
+
248
+
249
+ def deprocess_image(x):
250
+
251
+ '''
252
+
253
+ Same normalization as in:
254
+
255
+ https://github.com/fchollet/keras/blob/master/examples/conv_filter_visualization.py
256
+
257
+ '''
258
+
259
+ if np.ndim(x) > 3:
260
+
261
+ x = np.squeeze(x)
262
+
263
+ # normalize tensor: center on 0., ensure std is 0.1
264
+
265
+ x -= x.mean()
266
+
267
+ x /= (x.std() + 1e-5)
268
+
269
+ x *= 0.1
270
+
271
+
272
+
273
+ # clip to [0, 1]
274
+
275
+ x += 0.5
276
+
277
+ x = np.clip(x, 0, 1)
278
+
279
+
280
+
281
+ # convert to RGB array
282
+
283
+ x *= 255
284
+
285
+ if K.image_dim_ordering() == 'th':
286
+
287
+ x = x.transpose((1, 2, 0))
288
+
289
+ x = np.clip(x, 0, 255).astype('uint8')
290
+
291
+ return x
292
+
293
+
294
+
295
+ def _compute_gradients(tensor, var_list):
296
+
297
+ grads = tf.gradients(tensor, var_list)
298
+
299
+ return [grad if grad is not None else tf.zeros_like(var) for var, grad in zip(var_list, grads)]
300
+
301
+
302
+
303
+ def grad_cam(input_model, image, category_index, layer_name):
304
+
305
+ nb_classes = 1000
306
+
307
+ target_layer = lambda x: target_category_loss(x, category_index, nb_classes)
308
+
309
+ x = Lambda(target_layer, output_shape = target_category_loss_output_shape)(input_model.output)
310
+
311
+ model = Model(inputs=input_model.input, outputs=x)
312
+
313
+ #model.summary()
314
+
315
+ loss = K.sum(model.output)
316
+
317
+ conv_output = [l for l in model.layers if l.name == layer_name][0].output #is
318
+
319
+ grads = normalize(_compute_gradients(loss, [conv_output])[0])
320
+
321
+ gradient_function = K.function([model.input], [conv_output, grads])
322
+
323
+
324
+
325
+ output, grads_val = gradient_function([image])
326
+
327
+ output, grads_val = output[0, :], grads_val[0, :, :, :]
328
+
329
+
330
+
331
+ weights = np.mean(grads_val, axis = (0, 1))
332
+
333
+ cam = np.ones(output.shape[0 : 2], dtype = np.float32)
334
+
335
+
336
+
337
+ for i, w in enumerate(weights):
338
+
339
+ cam += w * output[:, :, i]
340
+
341
+
342
+
343
+ cam = cv2.resize(cam, (224,224)) #299,299)) #224, 224))
344
+
345
+ cam = np.maximum(cam, 0)
346
+
347
+ heatmap = cam / np.max(cam)
348
+
349
+ for x in range(224):
350
+
351
+ print (heatmap[x])
352
+
353
+
354
+
355
+ #Return to BGR [0..255] from the preprocessed image
356
+
357
+ image = image[0, :]
358
+
359
+ image -= np.min(image)
360
+
361
+ image = np.minimum(image, 255)
362
+
363
+
364
+
365
+ cam = cv2.applyColorMap(np.uint8(255*heatmap), cv2.COLORMAP_JET)
366
+
367
+ cam = np.float32(cam) + np.float32(image)
368
+
369
+ cam = 255 * cam / np.max(cam)
370
+
371
+ return np.uint8(cam), heatmap
372
+
373
+
374
+
375
+ preprocessed_input = load_image(sys.argv[1])
376
+
377
+ model = VGG16(weights='imagenet')
378
+
379
+ #model = VGG19(weights='imagenet')
380
+
381
+ #model = InceptionV3(weights='imagenet')
382
+
383
+ #model = ResNet50(weights = 'imagenet')
384
+
385
+ #model.summary()
386
+
387
+ target_layer = 'block5_conv3' #'activation_49' add_16 "block5_conv3"
388
+
389
+
390
+
391
+ predictions = model.predict(preprocessed_input)
392
+
393
+ register_gradient()
394
+
395
+ guided_model = modify_backprop(model, 'GuidedBackProp')
396
+
397
+ guided_model.summary()
398
+
399
+ for i in range(5):
400
+
401
+ top_1 = decode_predictions(predictions)[0][i]
402
+
403
+ print(predictions.argsort()[0][::-1][i])
404
+
405
+ print('Predicted class:')
406
+
407
+ print('%s (%s) with probability %.2f' % (top_1[1], top_1[0], top_1[2]))
408
+
409
+ predicted_class = predictions.argsort()[0][::-1][i] #np.argmax(predictions)
34
410
 
35
411
  cam, heatmap = grad_cam(model, preprocessed_input, predicted_class, target_layer)
36
412
 
37
- #print(cam[0])
38
-
39
- for x in range(224):
40
-
41
- print (heatmap[x])
42
-
43
- #print("---------------------")
44
-
45
- #print ("\n".join([str(x) for x in heatmap[x]]))
413
+ cv2.imwrite("gradcam"+str(top_1[1])+".jpg", cam)
414
+
46
-
415
+ saliency_fn = compile_saliency_function(guided_model)
47
-
48
-
416
+
49
- import matplotlib.pyplot as plt
417
+ saliency = saliency_fn([preprocessed_input, 0])
50
-
51
- from mpl_toolkits.mplot3d import Axes3D
418
+
52
-
53
- # 3D散布図でプロットするデータを生成する為にnumpyを使用
54
-
55
- X = np.array([heatmap for heatmap in range(224)]) # 自然数の配列
419
+ gradcam = saliency[0] * heatmap[..., np.newaxis]
56
-
57
- Y = np.sin(X) # 特に意味のない正弦
420
+
58
-
59
- Z = np.sin(Y) # 特に意味のない正弦
60
-
61
-
62
-
63
- # グラフの枠を作成
64
-
65
- fig = plt.figure()
66
-
67
- ax = Axes3D(fig)
68
-
69
-
70
-
71
- # X,Y,Z軸にラベルを設定
72
-
73
- ax.set_xlabel("X")
74
-
75
- ax.set_ylabel("Y")
76
-
77
- ax.set_zlabel("Z")
78
-
79
-
80
-
81
- # .plotで描画
82
-
83
- ax.plot(X,Y,Z,marker="o",linestyle='None')
421
+ cv2.imwrite("guided_gradcam"+str(top_1[1])+".jpg", deprocess_image(gradcam))
84
422
 
85
423
  ```