-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathCoRE-ATAC-ModelTrainer-SigSeq.py
262 lines (191 loc) · 8.55 KB
/
CoRE-ATAC-ModelTrainer-SigSeq.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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
from __future__ import print_function
import keras
from keras.models import Model
from keras.layers import Dense, Dropout, Flatten, Input
from keras.layers import Conv1D, Conv2D, MaxPooling1D, MaxPooling2D, BatchNormalization, Activation, SeparableConv1D
from keras.callbacks import ModelCheckpoint
from keras import backend as K
from tensorloader import TensorLoader as tl
import partitioning_util as part
import sys
import numpy as np
import pandas as pd
import pickle
import h5py
from sklearn.utils import class_weight
import argparse
import os
batch_size = 32
epochs = 20
args = sys.argv
sig_shape = (600,10)
num_classes = 4
featurefile = "/CoRE-ATAC/PEAS/features.txt"
labelencoder = "/CoRE-ATAC/PEAS/labelencoder.txt"
#Step 0: Process arguments
parser = argparse.ArgumentParser(description='CoRE-ATAC Model Trainer - SigSeq')
parser.add_argument("basenames")
parser.add_argument("datadirectories")
parser.add_argument("trainchrs")
parser.add_argument("outputmodel")
args = parser.parse_args()
basef = args.basenames
dirf = args.datadirectories
trainchrf = args.trainchrs
outmodelf = args.outputmodel
def readList(file):
return list(pd.read_csv(file, header=None).values.flatten())
def getTrainChr(file):
data = pd.read_csv(file, header=None, sep="\t").values
trainchr = []
valchr = []
testchr = []
for i in range(len(data)):
if(data[i,0] == "train"):
trainchr.extend(data[i,1].split(","))
if(data[i,0] == "val"):
valchr.extend(data[i,1].split(","))
if(data[i,0] == "test"):
testchr.extend(data[i,1].split(","))
return trainchr, valchr, testchr
datanames = readList(basef)
datadirectories = readList(dirf)
trainchr, valchr, testchr = getTrainChr(trainchrf)
bestmodelsaver = ModelCheckpoint(outmodelf, monitor='val_output_loss', verbose=1, save_best_only=True, mode='min')
callbacks = [bestmodelsaver]
###################
#Step 1: Load Data#
###################
seqdata,sigdata,annot,summitpeaks,peaks = tl.readTensors(datanames[0], datadirectories[0], 600, sequence=True, signal=True)
testindices = part.chrSelect(peaks, valchr)
ftestindices = part.chrSelect(peaks, testchr)
trainindices = part.chrSelect(peaks, trainchr)
sigdata = tl.getSeqSigTensor(seqdata, sigdata)
x_train_sig = sigdata[trainindices]
y_train_0 = annot[trainindices]
x_test_sig = sigdata[testindices]
y_test_0 = annot[testindices]
x_ftest_sig = sigdata[ftestindices]
y_ftest_0 = annot[ftestindices]
y_train = y_train_0
y_test = y_test_0
y_ftest = y_ftest_0
print(datanames[0]+" loaded.")
for curindex in range(1, len(datadirectories)):
print("Begin loading "+datanames[curindex]+".")
seqdata,sigdata,annot,summitpeaks,peaks = tl.readTensors(datanames[curindex], datadirectories[curindex], 600, sequence=True, signal=True)
testindices = part.chrSelect(peaks, valchr)
ftestindices = part.chrSelect(peaks, testchr)
trainindices = part.chrSelect(peaks, trainchr)
sigdata = tl.getSeqSigTensor(seqdata, sigdata)
x_train_sig = np.concatenate((x_train_sig, sigdata[trainindices]), axis=0)
y_train_0 = annot[trainindices]
x_test_sig = np.concatenate((x_test_sig, sigdata[testindices]), axis=0)
y_test_0 = annot[testindices]
x_ftest_sig = np.concatenate((x_ftest_sig, sigdata[ftestindices]), axis=0)
y_ftest_0 = annot[ftestindices]
y_train = np.concatenate((y_train, y_train_0), axis=0)
y_test = np.concatenate((y_test, y_test_0), axis=0)
y_ftest = np.concatenate((y_ftest, y_ftest_0), axis=0)
print(datanames[curindex]+" loaded.")
#####################
#Step 2: Build Model#
#####################
#Inputs
sig_input = Input(shape=sig_shape, dtype='float', name='sig_input')
#Signal
sig_conv1 = SeparableConv1D(name='sig_conv_1',
filters=256, kernel_size=19, activation='relu',
data_format='channels_last')(sig_input)
sig_batch1 = BatchNormalization(name='sig_batch_normalization1', axis=-1)(sig_conv1)
sig_conv2 = SeparableConv1D(name="sig_conv_2",
filters=256, kernel_size=19, activation='relu',
data_format='channels_last')(sig_batch1)
sig_maxpool1 = MaxPooling1D(pool_size=2, name='sig_maxpool1',
data_format='channels_last')(sig_conv2)
sig_batch2 = BatchNormalization(name='sig_batch_normalization2', axis=-1)(sig_maxpool1)
sig_conv3 = SeparableConv1D(name='sig_conv_3',
filters=256, kernel_size=19, activation='relu',
data_format='channels_last')(sig_batch2)
sig_batch3 = BatchNormalization(name='sig_batch_normalization3', axis=-1)(sig_conv3)
sig_conv4 = SeparableConv1D(name="sig_conv_4",
filters=256, kernel_size=19, activation='relu',
data_format='channels_last')(sig_batch3)
sig_maxpool2 = MaxPooling1D(pool_size=2, name='sig_maxpool2',
data_format='channels_last')(sig_conv4)
sig_batch4 = BatchNormalization(name='sig_batch_normalization4', axis=-1)(sig_maxpool2)
sig_conv5 = SeparableConv1D(name='sig_conv_5',
filters=512, kernel_size=19, activation='relu',
data_format='channels_last')(sig_batch4)
sig_batch5 = BatchNormalization(name='sig_batch_normalization5', axis=-1)(sig_conv5)
sig_conv6 = SeparableConv1D(name="sig_conv_6",
filters=512, kernel_size=19, activation='relu',
data_format='channels_last')(sig_batch5)
sig_maxpool3 = MaxPooling1D(pool_size=2, name='sig_maxpool3',
data_format='channels_last')(sig_conv6)
sig_batch6 = BatchNormalization(name='sig_batch_normalization6', axis=-1)(sig_maxpool3)
#
sig_conv7 = SeparableConv1D(name='sig_conv_7',
filters=512, kernel_size=19, activation='relu',
data_format='channels_last')(sig_batch6)
sig_batch7 = BatchNormalization(name='sig_batch_normalization7', axis=-1)(sig_conv7)
sig_conv8 = SeparableConv1D(name="sig_conv_8",
filters=512, kernel_size=19, activation='relu',
data_format='channels_last')(sig_batch7)
sig_maxpool4 = MaxPooling1D(pool_size=2, name='sig_maxpool4',
data_format='channels_last')(sig_conv8)
sig_batch8 = BatchNormalization(name='sig_batch_normalization8', axis=-1)(sig_maxpool4)
sig_flatten = Flatten(name='sig_flatten')(sig_batch8)
sig_dense = Dense(2048, name='sig_dense', activation='relu')(sig_flatten)
output = Dense(num_classes, activation='softmax', name='output')(sig_dense)
model = Model(inputs=[sig_input], outputs=[output])
model.compile(loss=keras.losses.sparse_categorical_crossentropy,
optimizer=keras.optimizers.Adam(),
metrics=['accuracy'])
######################################################
#Step 3: Create generator for reading sparse matrices#
######################################################
def listToDense(positions, nrows, ncols):
num_samples = len(positions)
rv = np.zeros((num_samples, nrows, ncols,1), dtype=bool)
for i in range(num_samples):
for coord in positions[i]:
if len(coord) == 2:
rv[i,coord[0],coord[1],0] = 1
return rv
def sparse_to_dense_generator(X, y, batch_size, shuffle, rstate=None):
num_samples = X[0].shape[0]
mat_len = X[0].shape[2]
num_batches = int(num_samples/batch_size) + 1 - int(num_samples % batch_size == 0)
idx = np.arange(num_samples)
rs = np.random.RandomState(seed=rstate)
if shuffle:
rs.shuffle(idx)
curbatch = 0
while True:
subset = idx[batch_size*curbatch:batch_size*(curbatch+1)]
#cuts_to_dense = listToDense(X[3][subset], mat_len, mat_len)
rv_X = [X[0][subset]]
rv_y = y[subset]
yield rv_X, rv_y
curbatch += 1
if curbatch == num_batches:
if shuffle:
rs.shuffle(idx)
curbatch = 0
#####################
#Step 4: Train Model#
#####################
x_train_sig = np.moveaxis(x_train_sig, 1, -1)
x_test_sig = np.moveaxis(x_test_sig, 1, -1)
steps_per_epoch = int(np.ceil(x_train_sig.shape[0]/batch_size))
generator = sparse_to_dense_generator([x_train_sig], y_train, batch_size=batch_size, shuffle=True, rstate=929)
class_weights = class_weight.compute_class_weight('balanced', np.unique(y_train), y_train)
class_weights_dict = dict()
for i in range(num_classes):
class_weights_dict[i] = class_weights[i]
model.fit_generator(generator,
steps_per_epoch=steps_per_epoch,
epochs=epochs, callbacks=callbacks,
verbose=1, class_weight=class_weights_dict,
validation_data=([x_test_sig], y_test))