-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathBinaryRuns_10models_100epochs.py
executable file
·207 lines (170 loc) · 8.28 KB
/
BinaryRuns_10models_100epochs.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
#!/usr/bin/env python
#!/usr/bin/env python
import os,sys,json
import keras
from keras import backend as K
from keras.models import Model,load_model
from keras.layers import Input,Dense,Dropout,Flatten
from keras.utils import plot_model
from keras.callbacks import EarlyStopping,ModelCheckpoint,ReduceLROnPlateau
from keras.layers.convolutional import Conv2D
from keras.layers.pooling import MaxPooling2D
from keras.layers.merge import concatenate
from sklearn import model_selection as md
from keras import regularizers
import numpy as np
def save_model_history(model,history_filename):
with open(history_filename,'w') as f:
json.dump(model.history.history, f)
######################################################################
def HSC_Subaru_CNN(args):
dx,dy,nChannels = args
# convolution model
inputs = Input(shape=(dx,dy,nChannels), name='main_input')
# first conv layer
x = Conv2D(32, kernel_size=(5,5),activation='relu',
padding='same',strides=(1, 1),name='Conv_1')(inputs)
x = MaxPooling2D(pool_size=(2,2),name='MP_C1')(x)
#x = Dropout(0.5,name='Drop_C1')(x)
# second conv layer
x = Conv2D(64, kernel_size=(3,3),activation='relu',
padding='same',strides=(1, 1),name='Conv_2')(x)
x = MaxPooling2D(pool_size=(2,2),name='MP_C2')(x)
#x = Dropout(0.25,name='Drop_C2')(x)
# third conv layer
x = Conv2D(128, kernel_size=(2,2),activation='relu',
padding='same',strides=(1, 1),name='Conv_3')(x)
x = MaxPooling2D(pool_size=(2,2),name='MP_C3')(x)
#x = Dropout(0.25,name='Drop_C3')(x)
# fourth conv layer
x = Conv2D(128, kernel_size=(3,3),activation='relu',
padding='same',strides=(1, 1),name='Conv_4')(x)
# x = Dropout(0.25,name='Drop_C4')(x)
# flatten for fully connected layers
x = Flatten(name='Flatten')(x)
# Fully Connected Layer
x = Dense(64,activation='relu',name='Dense_1')(x)
x = Dropout(0.5,name='DropFCL_1')(x)
x = Dense(16,activation='relu',name='Dense_2')(x)
x = Dropout(0.5,name='DropFCL_2')(x)
x = Dense(1,activation='sigmoid',name='Dense_3')(x)
# connect and compile
model = Model(inputs=inputs,outputs=x)
model.compile(optimizer=keras.optimizers.Adadelta(),
loss='binary_crossentropy',metrics=['accuracy'])
return model
######################################################################
def split_data(inp,tar,cat,n_negOrig,n_neg,n_posOrig,n_pos,
train_valid_split=0.3,valid_test_split=0.5,randomState=0):
f_negAug = int(n_neg/n_negOrig)-1
f_posAug = int(n_pos/n_posOrig)-1
# indices of the original pre-augmentation images
indices_neg0 = np.arange(n_negOrig).astype(int)
indices_pos0 = np.arange(n_posOrig).astype(int)+n_neg
# this randomstate ensures we would have the same images if we did this again
randomState = 0
indices_neg_train0,indices_neg_valid0 = md.train_test_split(indices_neg0,test_size=train_valid_split, random_state=randomState)
indices_neg_valid0,indices_neg_test0 = md.train_test_split(indices_neg_valid0,test_size=valid_test_split, random_state=randomState)
indices_pos_train0,indices_pos_valid0 = md.train_test_split(indices_pos0,test_size=train_valid_split, random_state=randomState)
indices_pos_valid0,indices_pos_test0 = md.train_test_split(indices_pos_valid0,test_size=valid_test_split, random_state=randomState)
indices_neg_train = np.array([]).astype(int)
indices_neg_valid = np.array([]).astype(int)
indices_neg_test = np.array([]).astype(int)
indices_pos_train = np.array([]).astype(int)
indices_pos_valid = np.array([]).astype(int)
indices_pos_test = np.array([]).astype(int)
# now incorporate augmented images into each dataset
for i in range(f_negAug+1):
indices_neg_train = np.concatenate([indices_neg_train,indices_neg_train0+i*n_negOrig])
indices_neg_valid = np.concatenate([indices_neg_valid,indices_neg_valid0+i*n_negOrig])
indices_neg_test = np.concatenate([indices_neg_test,indices_neg_test0+i*n_negOrig])
for i in range(f_posAug+1):
indices_pos_train = np.concatenate([indices_pos_train,indices_pos_train0+i*n_posOrig])
indices_pos_valid = np.concatenate([indices_pos_valid,indices_pos_valid0+i*n_posOrig])
indices_pos_test = np.concatenate([indices_pos_test,indices_pos_test0+i*n_posOrig])
print('Negatives:', len(indices_neg_train),len(indices_neg_valid),len(indices_neg_test))
print('Positives:', len(indices_pos_train),len(indices_pos_valid),len(indices_pos_test))
# join
indices_train = np.concatenate([indices_neg_train,indices_pos_train])
indices_valid = np.concatenate([indices_neg_valid,indices_pos_valid])
indices_test = np.concatenate([indices_neg_test,indices_pos_test])
# # Now apply to full data matrix
tar_train = tar[indices_train]
inp_train = inp[indices_train]
cat_train = cat[indices_train]
tar_valid = tar[indices_valid]
inp_valid = inp[indices_valid]
cat_valid = cat[indices_valid]
tar_test = tar[indices_test]
inp_test = inp[indices_test]
cat_test = cat[indices_test]
return inp_train,tar_train,cat_train,inp_valid,tar_valid,cat_valid,inp_test,tar_test,cat_test
classIDs = ['Negative','Positive']
modelDir = '/home/bottrell/scratch/Subaru/HyperSuprime/Models/Binary/'
dataDir = '/home/bottrell/scratch/Subaru/HyperSuprime/Data/Binary/'
fileNames = ['{}{}_i_Images-RZSAugNorm.npy'.format(dataDir,classID) for classID in classIDs]
catNames = ['{}{}_i_Images-RZSAug_cat.npy'.format(dataDir,classID) for classID in classIDs]
np.random.seed(0)
negative_subsample = 10000
tar = np.array([])
inp = np.array([])
# Negatives
#indices_neg = np.random.choice(np.arange(negative_total),negative_samples,replace=False)
inp_neg = np.load(fileNames[0])[:negative_subsample] # [indices_neg]
print('Negatives:',inp_neg.shape)
n_neg = inp_neg.shape[0]
n_negOrig = 5000
f_negAug = int(n_neg/n_negOrig)-1
print('Number of augmentations per original image:', f_negAug)
tar_neg = np.append(tar,np.ones(n_neg)*0)
cat_neg = np.load(catNames[0])[:negative_subsample]
# Positives
inp_pos = np.load(fileNames[1])
print('Positives:',inp_pos.shape)
n_pos = inp_pos.shape[0]
n_posOrig = 1201
f_posAug = int(n_pos/n_posOrig)-1
print('Number of augmentations per original image:', f_posAug)
tar_pos = np.ones(n_pos)
cat_pos = np.load(catNames[1])
# Combining
inp = np.concatenate([inp_neg,inp_pos],axis=0).reshape(-1,128,128,1)
tar = np.append(tar_neg,tar_pos)
cat = np.concatenate([cat_neg,cat_pos],axis=0)
nRecords,dx,dy,nChannels = inp.shape
nClasses=len(classIDs)
#del new_inp,new_cat
batch_size = 32
epochs=30
train_valid_split = 0.3
valid_test_split = 0.5
update = 0
version = 3
patience = 999
label = 'HSC-Subaru'
randomStates = np.arange(10)
for randomState in randomStates:
modelFile = modelDir+'{}_Binary_RS-{}_v{}_{}.h5'
historyFile = modelDir+'{}_Binary_RS-{}_v{}_{}_history.json'
if os.access(modelFile.format(label,randomState,version,update),0): continue
modelFile = modelFile.format(label,randomState,version,update)
print('Model will save as: {}'.format(modelFile))
historyFile = historyFile.format(label,randomState,version,update)
print('History will save as: {}'.format(historyFile))
mcp_file = modelFile.replace('.h5','_wgts.hdf5')
early_stopping = EarlyStopping(monitor='val_acc', patience=patience, verbose=1, mode='max')
mcp_save = ModelCheckpoint(mcp_file, save_best_only=True, monitor='val_acc', mode='max')
reduce_lr_loss = ReduceLROnPlateau(monitor='val_acc', factor=0.1, patience=7, verbose=1, min_delta=1e-4, mode='max')
args = (dx,dy,nChannels)
model = HSC_Subaru_CNN(args)
inp_train,tar_train,cat_train,inp_valid,tar_valid,cat_valid,inp_test,tar_test,cat_test = split_data(inp,tar,cat,
5000,10000,1201,9608,randomState=randomState)
model.summary()
model.fit(inp_train,tar_train,shuffle=True,epochs=epochs,
batch_size=batch_size,validation_data=[inp_valid,tar_valid],
callbacks=[early_stopping, mcp_save])
save_model_history(model=model,history_filename=historyFile)
if os.access(mcp_file,0):
model.load_weights(mcp_file)
os.remove(mcp_file)
model.save(modelFile)