-
Notifications
You must be signed in to change notification settings - Fork 0
/
camVizualINCEPTIONV3.py
215 lines (191 loc) · 9.51 KB
/
camVizualINCEPTIONV3.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
from keras.models import load_model
from keras import backend as K
from keras.preprocessing import image
import numpy as np
import cv2
from PIL import Image
import json
import os
import dense_crf
import tensorflow as tf
def openJson(file):
with open(file) as File:
dict = json.load(File)
return dict
def getLayerOutput(model, layer):
conv_layer = model.get_layer(layer)
out = conv_layer.output
return out
def CAM(model, output, img_path, layer, rangeNo, outputName):
"""
This function is the one that creates and returns heatmap from the output of the given convolution layer
"""
last_conv_layer = model.get_layer(layer)
grads = K.gradients(output, last_conv_layer.output)[0] # grads: tensor that has 4 axes,
# like the output of the conv layer, but it has the gradients of the output tensor
pooled_grads = K.mean(grads, axis=(0, 1, 2)) # the K.mean returns
# a tensor with the mean of elements of grads. So a tensor like this (512,). 512 channels with the means of the grads.
# axis=(0,1,2): means that the mean runs through the 0 to 2 axes of the tensor
iterate = K.function([model.input], [pooled_grads, last_conv_layer.output[0]]) # K.function returns: output values as Numpy arrays.
# first argument is input to this model
# and second is set of 2 outputs - one for convolution and second for softmax output at the last layer.
pooled_grads_value, conv_layer_output_value = iterate([x]) # both arrays
# print (conv_layer_output_value.shape)
# print (pooled_grads_value.shape)
# We multiply each channel in the feature map array by "how important this channel is" with regard to the class
for i in range(rangeNo):
conv_layer_output_value[:, :, i] *= pooled_grads_value[i]
# plotting the heatmap
heatmap = np.mean(conv_layer_output_value, axis=-1)
heatmap = np.maximum(heatmap, 0)
if (np.max(heatmap)!= 0):
heatmap /= np.max(heatmap)
img = cv2.imread(img_path)
img = cv2.resize(img, (299, 299))
# heatmap or CAM or localization map
heatmap = cv2.resize(heatmap, (img.shape[1], img.shape[0]))
heatmap1 = heatmap
# heatmap = np.uint8(255 * heatmap) # Convert heatmap to RGB (224,224)
# heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET) # Apply heatmap to original Image
# hif = .8
# superimposed_img = heatmap * hif + img # Apply heatmap intensity factor
# cv2.imwrite('food11_output/' + outputName, superimposed_img) # produce the output.jpeg to vizualize the map
return heatmap1
def localizationMap(model, output, img_path, s):
heatmap1 = CAM(model, output, img_path, 'conv2d_88', 384, str(s)+'_output1.jpeg')
heatmap2 = CAM(model, output, img_path, 'conv2d_89', 384, str(s)+'_output2.jpeg')
heatmap3 = CAM(model, output, img_path, 'conv2d_92', 384, str(s)+'_output3.jpeg')
heatmap4 = CAM(model, output, img_path, 'conv2d_93', 384, str(s)+'_output4.jpeg')
heatmap5 = CAM(model, output, img_path, 'conv2d_86', 320, str(s)+'_output5.jpeg')
heatmap6 = CAM(model, output, img_path, 'conv2d_94', 190, str(s)+'_output6.jpeg')
superimposed = np.average((np.array([heatmap1, heatmap2, heatmap3, heatmap4, heatmap5, heatmap6])), axis=0)
return superimposed
def foregroundMap(model, output, img_path, s):
heatmap5 = CAM(model, output, img_path, 'conv2d_95', 2048, str(s)+'_output-dilat1.jpeg')
heatmap6 = CAM(model, output, img_path, 'conv2d_96', 2048, str(s)+'_output5_dilat2.jpeg')
heatmap7 = CAM(model, output, img_path, 'conv2d_97', 2048, str(s) + '_output5_dilat3.jpeg')
heatmap8 = CAM(model, output, img_path, 'conv2d_98', 2048, str(s) + '_output5_dilat4.jpeg')
finalMap = np.add(heatmap6, heatmap5)
finalMap1 = np.add(finalMap, heatmap7)
finalMap2 = np.add(finalMap1,heatmap8)
# scale the final map between 0 and 1 and get a foreground probability map Pf
finalMap2 = finalMap2 / np.ptp(finalMap2)
return finalMap2
def plotHeatmap(heatmap, outputName, img_path):
"""
This function is used like CAM but it does not take the output of a convolution layer.
It takes an already created heatmap and saves it in a folder for its vizualization
and also returns the heatmap
:param heatmap: the map that we want to plot
:param outputName: the name of the map we want to plot
:param img_path: the path to the original image
:return: the heatmap
"""
heatmap = (heatmap - heatmap.min())/ np.ptp(heatmap)
img = cv2.imread(img_path)
img = cv2.resize(img, (299, 299))
# plotting the heatmap
heatmap = np.uint8(255 * heatmap)
heatmap1 = heatmap
# heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET) # Apply heatmap to original Image
# hif = .8
# superimposed_img = heatmap * hif + img # Apply heatmap intensity factor
# cv2.imwrite('food11_output_inc/'+outputName, superimposed_img) # produce the output.jpeg
return heatmap1
def loadImage(img_path):
img = image.load_img(img_path, target_size=(299, 299, 3))
# plt.imshow(img)
# plt.show()
return img
def editImage(img):
img = image.img_to_array(img)
img = img / 255
X = np.array(img)
x = np.expand_dims(X, axis=0)
return x
def test_dense_crf(img_path,heatmaps, name, prediction):
"""
:param img_path: the path of the original images for which it wants to segment
:param heatmaps: the final heatmap created by localization and fg/bg map
:param name: name of the image in order to give the same name to the mask
:param prediction: the class to which the image belongs
:return: the final ground truth mask
"""
I = Image.open(img_path)
I = I.resize((299, 299), Image.ANTIALIAS)
Iq = np.asarray(I.convert('L'), np.uint8)
P = np.asarray(heatmaps, np.float32)
fP = P/255.0
totalLab = []
w1 = 5.0 # weight of bilateral term
alpha = 10 # spatial std
beta = 10 # rgb std
w2 = 3.0 # weight of spatial term
gamma = 10 # spatial std
it = 5.0 # iteration
param = (w1, alpha, beta, w2, gamma, it)
lab = dense_crf.dense_crf(Iq, fP, param)
# superimposed_img = P + Iq
# fig = plt.figure()
# ax1 = fig.add_subplot(231); ax1.imshow(I); ax1.set_title('Image')
# ax2 = fig.add_subplot(232); ax2.imshow(P); ax2.set_title('Heatmap')
# ax3 = fig.add_subplot(233); ax3.imshow(superimposed_img); ax3.set_title('Superimosed')
P[P < (0.3 * P.max())] = 0
P[P >= (0.3*P.max())] = 1
# ax4 = fig.add_subplot(234); ax4.imshow(P); ax4.set_title('Segmentation')
# mask3d = np.where(P[:,:,None]==1,I,[0,0,255])
# ax5 = fig.add_subplot(235); ax5.imshow(mask3d); ax5.set_title('Cutted Image')
# fig.subplots_adjust(hspace=2.0, wspace=0.3)
# trueclass = name.split('_')
# fig.text(0.8, 0.25, 'Prediction:'+str(prediction)+'\nTrue class:'+trueclass[0], ha='center', va='center')
# name = str(name.replace('.jpg', '.png'))
# plt.tight_layout()
# plt.savefig('food11_output_inc/'+ name)
cv2.imwrite('food-11/segmentation/train_masks_INCEPTIONV3/' + name, P)
return lab
def concatenations(model, output, img_path, s):
heatmap6 = CAM(model, output, img_path, 'mixed6', 768, str(s) + '_output_conc6.jpeg')
heatmap7 = CAM(model, output, img_path, 'mixed7', 768, str(s) + '_output_conc7.jpeg')
heatmap8 = CAM(model, output, img_path, 'mixed8', 1280, str(s) + '_output_conc8.jpeg')
heatmap9 = CAM(model, output, img_path, 'mixed9', 2048, str(s) + '_output_conc9.jpeg')
conc = np.average((np.array([heatmap6, heatmap7, heatmap8, heatmap9])), axis=0)
return conc
if __name__ == "__main__":
with tf.device('/gpu:1'):
K.clear_session()
model = load_model('models/InceptionV3.h5')
trainFile = 'food-11/train.json'
train = os.listdir("""food-11/training""")
predictions = []
truth = []
expected = []
nb_test_samples = len(openJson(trainFile))
test_true = openJson(trainFile)
values = (test_true.values())
keys = test_true.keys()
categories = ['Bread', 'Dairy product', 'Dessert', 'Egg', 'Fried food', 'Meat', 'Noodles/Pasta', 'Rice', 'Seafood', 'Soup', 'Vegetable/Fruit']
for t in range(len(train)):
print(t)
img_path = 'food-11/training/'+str(train[t])
img = loadImage(img_path)
x = editImage(img)
a = 0.5
s = str(train[t]).replace('.jpg', ' ')
s1 = str(train[t]).replace('.jpg', '.png')
folderElements = os.listdir("""food-11/segmentation/train_masks_INCEPTIONV3""")
# print (len(folderElements))
if str(s1) not in folderElements:
classes = categories
preds = model.predict(x)
argmax = np.argmax(preds[0])
ar = model.output[:, argmax]
concat = concatenations(model, ar, img_path,s)
# plotHeatmap(concat, str(s) + 'cocnats.jpeg', img_path)
Mc = localizationMap(model, ar, img_path,s)
# plotHeatmap(Mc, str(s) + 'local.jpeg', img_path)
Qc = np.multiply(Mc, concat)
# plotHeatmap(Qc, str(s) + 'Qc.jpeg', img_path)
newMc = localizationMap(model, ar, img_path,s)
Pc = a * Qc + (1-a) * newMc
Ptest = plotHeatmap(Pc, str(s) + 'heatmap.png', img_path)
Dcrf = test_dense_crf(img_path, Ptest, s1, argmax)