-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathlate_fusion_improved.py
207 lines (171 loc) · 9.48 KB
/
late_fusion_improved.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
#FOR MODIFYING IMAGES AND ARRAYS
from datetime import datetime
import os,cv2
#from cv2 import getRotationMatrix2D, warpAffine,getAffineTransform,resize,imread,BORDER_REFLECT
import numpy as np
#KERAS IMPORTS
from keras.applications.vgg16 import VGG16
from keras.callbacks import ProgbarLogger, EarlyStopping, ModelCheckpoint, TensorBoard
from keras.models import Model, Sequential
from keras.layers import Input, Convolution2D, MaxPooling2D, Conv2DTranspose, Conv2D, concatenate
from keras.layers.core import Reshape, Activation, Dropout
from keras.preprocessing.image import *
from keras.optimizers import SGD
#UTILITY GLOBAL VARIABLES
input_dim = (512,928)
dim_tup = (928,512)
num_class = 6
C = 4
index = [0, 1020, 1377 , 240, 735, 2380]
#HELPER FUNCTION OF SEGMENT_DATA_GENERATOR
# comprises of path and extension of images in a directory
class gen_args:
data_dir = None
data_ext = None
def __init__(self,dirr,ext):
self.data_dir = dirr
self.data_ext = ext
#CONVERTING MASKED IMAGES(image) TO A ARRAY OF PIXELWISE ONE-HOT VECTORS(of dimension 'no_class')
def fix_label(image, no_class):
width , height, depth = image.shape
#generating hashes for each pixel (index array above has the hash values for each class)
image = np.dot(image.reshape(width*height,depth)[:,],[1,4,9])
#converting hashes to indices of classes
for i in range(no_class):
image[image == index[i]] = i
#converting each index into one-hot vector of dim of classes(no_class)
image = (np.arange(no_class) == image[...,None])*1
return image
#====================================================data==augmentation==============================================================
'''class aug_state:
def __init__(self,flip_axis_index=0,zoom_range=(1.2,1.2)):
self.flip_axis_index=flip_axis_index
self.zoom_range=zoom_range
def data_augmentor(x,state,row_axis=1,col_axis=0,channel_axis=-1):
#dt = datetime.now()
#(int(str(dt).split('.')[1])%100)
t = np.random.randint(4,size=2)
temp =[0,0,0,0,0]
temp[t[0]] = 1
temp[t[1]] = 1
#print temp
if temp[0]:
x = flip_axis(x, state.flip_axis_index)
if temp[1]:
M = cv2.getRotationMatrix2D((x.shape[1]/2,x.shape[0]/2),np.random.randint(360),1) #last argument is scale in rotation
x = cv2.warpAffine(x,M,(x.shape[1],x.shape[0]), borderMode=cv2.BORDER_REFLECT)
#del M
if temp[2]:
M = np.float32([[1,0,np.random.randint(x.shape[0])],[0,1,np.random.randint(x.shape[1])]])
x = cv2.warpAffine(x,M,(x.shape[1],x.shape[0]), borderMode = cv2.BORDER_REFLECT)
#del M
if temp[3]:
pts1 = np.float32([[np.random.randint(x.shape[0]),np.random.randint(x.shape[1])],[np.random.randint(x.shape[0]),np.random.randint(x.shape[1])],[np.random.randint(x.shape[0]),np.random.randint(x.shape[1])]])
pts2 = np.float32([[np.random.randint(x.shape[0]),np.random.randint(x.shape[1])],[np.random.randint(x.shape[0]),np.random.randint(x.shape[1])],[np.random.randint(x.shape[0]),np.random.randint(x.shape[1])]])
M = cv2.getAffineTransform(pts1,pts2)
x = cv2.warpAffine(x,M,(x.shape[1],x.shape[0]),borderMode = cv2.BORDER_REFLECT)
#del M
#del pts1
#:del pts2
if 0:
x = random_zoom(x, state.zoom_range, row_axis, col_axis, channel_axis,fill_mode='reflect')
x = np.swapaxes(x,0,1)
x = np.swapaxes(x,1,2)
return x
'''
#=====================================================================================================================
#DATAGENERATOR FOR MULTIMODAL SEMANTIC SEGMENTATION
def Segment_datagen(file_path, rgb_args, nir_args, label_args, batch_size, input_size,val_flag):
# Create MEMORY enough for one batch of input(s) + augmentation & labels + augmentation
data = np.zeros((2,batch_size,input_size[0],input_size[1],3), dtype=np.uint8)
labels = np.zeros((batch_size,input_size[0]*input_size[1],6), dtype=np.uint8)
# Read the file names
files = open(file_path)
names = files.readlines()
files.close()
# Enter the indefinite loop of generator
while True:
dt = datetime.now()
np.random.seed(int(str(dt).split('.')[1])%100)
rand_inds = np.random.random_integers(0,len(names)-1, size=batch_size)
for i in range(batch_size):
flag = np.random.randint(4)
print flag
if flag or val_flag:
print names[rand_inds[i]].strip('\n')
data[0][i] = cv2.resize(cv2.imread(rgb_args.data_dir+names[rand_inds[i]].strip('\n')+rgb_args.data_ext), dim_tup)
data[1][i]= cv2.resize(cv2.imread(nir_args.data_dir+names[rand_inds[i]].strip('\n')+nir_args.data_ext), dim_tup)
labels[i] = fix_label(cv2.resize(cv2.imread(label_args.data_dir+names[rand_inds[i]].strip('\n')+label_args.data_ext), dim_tup),num_class)
print 'done'
else:
num = bin(np.random.randint(1,64))[2:]
num = '0'*(6-len(num))+num
print names[rand_inds[i]].strip('\n')+'_'+num
data[0][i] = cv2.resize(cv2.imread(rgb_args.data_dir+'Augmented/'+names[rand_inds[i]].strip('\n')+'_'+num+rgb_args.data_ext), dim_tup)
data[1][i] =cv2.resize(cv2.imread(nir_args.data_dir+'Augmented/'+names[rand_inds[i]].strip('\n')+'_'+num+nir_args.data_ext), dim_tup)
labels[i] = fix_label(cv2.resize(cv2.imread(label_args.data_dir+'Augmented/'+names[rand_inds[i]].strip('\n')+'_'+num+label_args.data_ext), dim_tup),num_class)
print 'done'
yield [data[0],data[1]],[labels]
#ARGUMENTS FOR DATA_GENERATOR
#state_aug = aug_state()
train_RGB_args = gen_args ('/home/krishna/freiburg_forest_dataset/train/rgb/','.jpg')
train_NIR_args = gen_args ('/home/krishna/freiburg_forest_dataset/train/nir_color/','.png')
train_Label_args = gen_args ('/home/krishna/freiburg_forest_dataset/train/GT_color/','.png')
train_generator = Segment_datagen(
file_path = '/home/krishna/freiburg_forest_dataset/train/train.txt',
rgb_args = train_RGB_args,
nir_args = train_NIR_args,
label_args = train_Label_args,
batch_size= 1,
input_size=input_dim,
val_flag = False)
valid_RGB_args = gen_args ('/home/krishna/freiburg_forest_dataset/valid/rgb/','.jpg')
valid_NIR_args = gen_args ('/home/krishna/freiburg_forest_dataset/valid/nir_color/','.png')
valid_Label_args = gen_args ('/home/krishna/freiburg_forest_dataset/valid/GT_color/','.png')
valid_generator = Segment_datagen(
file_path = '/home/krishna/freiburg_forest_dataset/valid/valid.txt',
rgb_args = valid_RGB_args,
nir_args = valid_NIR_args,
label_args = valid_Label_args,
batch_size= 1,
input_size=input_dim,
val_flag = True)
#================================================MODEL_ARCHITECTURE============================================================
# RGB MODALITY BRANCH OF CNN
inputs_rgb = Input(shape=(input_dim[0],input_dim[1],3))
vgg_model_rgb = VGG16(weights='imagenet', include_top = False,modality_num=0)
conv_model_rgb = vgg_model_rgb(inputs_rgb)
#conv_model_rgb = Conv2D(32, (3,3), strides=(1, 1), padding = 'same', activation='tanh',data_format="channels_last") (conv_model_rgb)
#conv_model_rgb = Conv2D(64, (3,3), strides=(1, 1), padding = 'same', activation='tanh',data_format="channels_last") (conv_model_rgb)
conv_model_rgb = Conv2D(128, (3,3), strides=(1, 1), padding = 'same', activation='tanh',data_format="channels_last") (conv_model_rgb)
conv_model_rgb = Conv2D(256, (3,3), strides=(1, 1), padding = 'same', activation='tanh',data_format="channels_last") (conv_model_rgb)
dropout_rgb = Dropout(0.2)(conv_model_rgb)
# NIR MODALITY BRANCH OF CNN
inputs_nir = Input(shape=(input_dim[0],input_dim[1],3))
vgg_model_nir = VGG16(weights='imagenet', include_top= False,modality_num=1)
conv_model_nir = vgg_model_nir(inputs_nir)
#conv_model_nir = Conv2D(32, (3,3), strides=(1, 1), padding = 'same', activation='tanh',data_format="channels_last") (conv_model_nir)
#conv_model_nir = Conv2D(64, (3,3), strides=(1, 1), padding = 'same', activation='tanh',data_format="channels_last") (conv_model_nir)
conv_model_nir = Conv2D(128, (3,3), strides=(1, 1), padding = 'same', activation='tanh',data_format="channels_last") (conv_model_nir)
conv_model_nir = Conv2D(256, (3,3), strides=(1, 1), padding = 'same', activation='tanh',data_format="channels_last") (conv_model_nir)
dropout_nir = Dropout(0.2)(conv_model_nir)
# CONACTENATE the ends of RGB & NIR
merge_rgb_nir = concatenate([dropout_rgb, dropout_nir], axis=-1)
# DECONVOLUTION Layers
deconv_last = Conv2DTranspose(num_class, (64,64), strides=(32, 32), padding='same', data_format="channels_last", activation='tanh',kernel_initializer='glorot_normal') (merge_rgb_nir)
#VECTORIZING OUTPUT
out_reshape = Reshape((input_dim[0]*input_dim[1],num_class))(deconv_last)
out = Activation('softmax')(out_reshape)
# MODAL [INPUTS , OUTPUTS]
model = Model(inputs=[inputs_rgb, inputs_nir], outputs=[out])
print 'compiling'
model.compile(optimizer=SGD(lr=0.008, decay=1e-6, momentum=0.9, nesterov=True),
loss='categorical_crossentropy',
metrics=['accuracy'])
model.summary()
# Save the model according to the conditions
progbar = ProgbarLogger(count_mode='steps')
checkpoint = ModelCheckpoint("late_fusion_new_arc_{epoch:02d}.hdf5", monitor='val_loss', verbose=1, save_best_only=False, save_weights_only=False, mode='auto', period=1)
early = EarlyStopping(monitor='val_acc', min_delta=0, patience=1, verbose=1, mode='auto')
board = TensorBoard(log_dir='./logs_training', histogram_freq=2, write_graph=True)
model.fit_generator(train_generator,steps_per_epoch=100,epochs=200, callbacks=[progbar,checkpoint,board],validation_data = valid_generator, validation_steps = 2, max_q_size=4, pickle_safe = True)