-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy path1_nested_cv_RefDNN.py
356 lines (291 loc) · 17.3 KB
/
1_nested_cv_RefDNN.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
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
import argparse
import os
import math
import numpy as np
import pandas as pd
import skopt
from datetime import datetime
from sklearn.metrics import accuracy_score, average_precision_score, roc_auc_score
from sklearn.model_selection import StratifiedKFold, train_test_split
from refdnn.model import REFDNN
from refdnn.dataset import DATASET
def get_args():
parser = argparse.ArgumentParser()
## positional
parser.add_argument('responseFile', type=str, help="A filepath of drug response data for TRAINING")
parser.add_argument('expressionFile', type=str, help="A filepath of gene expression data for TRAINING")
parser.add_argument('fingerprintFile', type=str, help="A filepath of fingerprint data for TRAINING")
## optional
parser.add_argument('-o', metavar='outputdir', type=str, default='output_1', help="A directory path for saving outputs (default:'output_1')")
parser.add_argument('-b', metavar='batchsize', type=int, default=64, help="A size of batch on training process. The small size is recommended if an available size of RAM is small (default: 64)")
parser.add_argument('-t', metavar='numtrainingsteps', type=int, default=5000, help="Number of training steps on training process. It is recommended that the steps is larger than (numpairs / batchsize) (default: 5000)")
parser.add_argument('-s', metavar='numbayesiansearch', type=int, default=20, help="Number of bayesian search for hyperparameter tuning (default: 20)")
parser.add_argument('-k', metavar='outerkfold', type=int, default=5, help="K for outer k-fold cross validation (default: 5)")
parser.add_argument('-l', metavar='innerkfold', type=int, default=3, help="L for inner l-fold cross validation (default: 3)")
parser.add_argument('-v', metavar='verbose', type=int, default=1, help="0:No logging, 1:Basic logging to check process, 2:Full logging for debugging (default:1)")
return parser.parse_args()
def main():
args = get_args()
global outputdir
global checkpointdir
global verbose
outputdir = args.o
verbose = args.v
if verbose > 0:
print('[START]')
if verbose > 1:
print('[ARGUMENT] RESPONSEFILE: {}'.format(args.responseFile))
print('[ARGUMENT] EXPRESSIONFILE: {}'.format(args.expressionFile))
print('[ARGUMENT] FINGERPRINTFILE: {}'.format(args.fingerprintFile))
print('[ARGUMENT] OUTPUTDIR: {}'.format(args.o))
print('[ARGUMENT] NUMBAYESIANSEARCH: {}'.format(args.s))
print('[ARGUMENT] OUTERKFOLD: {}'.format(args.k))
print('[ARGUMENT] INNERKFOLD: {}'.format(args.l))
print('[ARGUMENT] VERBOSE: {}'.format(args.v))
## output directory
if not os.path.exists(outputdir):
os.mkdir(outputdir)
checkpointdir = os.path.join(outputdir, "checkpoint")
if not os.path.exists(checkpointdir):
os.mkdir(checkpointdir)
########################################################
## 1. Read data
########################################################
global dataset
responseFile = args.responseFile
expressionFile = args.expressionFile
fingerprintFile = args.fingerprintFile
dataset = DATASET(responseFile, expressionFile, fingerprintFile)
if verbose > 0:
print('[DATA] NUM_PAIRS: {}'.format(len(dataset)))
print('[DATA] NUM_DRUGS: {}'.format(len(dataset.get_drugs(unique=True))))
print('[DATA] NUM_CELLS: {}'.format(len(dataset.get_cells(unique=True))))
print('[DATA] NUM_GENES: {}'.format(len(dataset.get_genes())))
print('[DATA] NUM_SENSITIVITY: {}'.format(np.count_nonzero(dataset.get_labels()==0)))
print('[DATA] NUM_RESISTANCE: {}'.format(np.count_nonzero(dataset.get_labels()==1)))
## time log
timeformat = '[TIME] [{0}] {1.year}-{1.month}-{1.day} {1.hour}:{1.minute}:{1.second}'
if verbose > 0:
print(timeformat.format(1, datetime.now()))
########################################################
## 2. Define the space of hyperparameters
########################################################
## 2-1) Set the range of hyperparameters
space_hidden_units = skopt.space.Integer(low=4, high=128, name='hidden_units')
space_learning_rate_ftrl = skopt.space.Real(low=1e-6, high=1e-1, prior='log-uniform', name='learning_rate_ftrl')
space_learning_rate_adam = skopt.space.Real(low=1e-6, high=1e-1, prior='log-uniform', name='learning_rate_adam')
space_l1_regularization_strength = skopt.space.Real(low=1e-3, high=1e+2, prior='log-uniform', name='l1_regularization_strength')
space_l2_regularization_strength = skopt.space.Real(low=1e-3, high=1e+2, prior='log-uniform', name='l2_regularization_strength')
## 2-2) Define hyperparmeter space
dimensions_hyperparameters = [space_hidden_units,
space_learning_rate_ftrl,
space_learning_rate_adam,
space_l1_regularization_strength,
space_l2_regularization_strength]
## time log
if verbose > 0:
print(timeformat.format(2, datetime.now()))
#######################################################
## 3. Start the hyperparameter tuning jobs
########################################################
global fitness_step
global fitness_idx_train
global fitness_idx_test
global innerkfold
global batchsize
global numtrainingsteps
outerkfold = args.k
innerkfold = args.l
numbayesiansearch = args.s
batchsize = args.b
numtrainingsteps = args.t
## 3-1) init lists for metrics
ACCURACY_outer = []
AUCROC_outer = []
AUCPR_outer = []
## 3-2) init lists for hyperparameters
Hidden_units_outer = []
Learning_rate_ftrl_outer = []
Learning_rate_adam_outer = []
L1_strength_outer = []
L2_strength_outer = []
kf = StratifiedKFold(n_splits=outerkfold, shuffle=True)
for k, (idx_train, idx_test) in enumerate(kf.split(X=np.zeros(len(dataset)), y=dataset.get_drugs())):
fitness_step = 1
fitness_idx_train = idx_train
fitness_idx_test = idx_test
## 3-3) Bayesian optimization with gaussian process
if verbose > 0:
print('[OUTER] [{}/{}] NOW TUNING THE MODEL USING BAYESIAN OPTIMIZATION...'.format(k+1, kf.get_n_splits()))
search_result = skopt.gp_minimize(func=fitness,
dimensions=dimensions_hyperparameters,
n_calls=numbayesiansearch,
n_initial_points=3, # 'n_random_starts' is deprecated in skopt 0.8 and replaced by 'n_initial_points'
acq_func='EI',
noise=1e-10,
verbose=0)
BEST_HIDDEN_UNITS = search_result.x[0]
BEST_LEARNING_RATE_FTRL = search_result.x[1]
BEST_LEARNING_RATE_ADAM = search_result.x[2]
BEST_L1_REGULARIZATION_STRENGTH = search_result.x[3]
BEST_L2_REGULARIZATION_STRENGTH = search_result.x[4]
BEST_TRAINING_ACCURACY = search_result.fun
configs_path = os.path.join(checkpointdir, "{:03d}_configs.csv".format(k))
with open(configs_path, 'w') as fout:
fout.write("HIDDEN_UNITS,{:d}\n".format(BEST_HIDDEN_UNITS))
fout.write("LEARNING_RATE_FTRL,{:.6f}\n".format(BEST_LEARNING_RATE_FTRL))
fout.write("LEARNING_RATE_ADAM,{:.6f}\n".format(BEST_LEARNING_RATE_ADAM))
fout.write("L1_REGULARIZATION_STRENGTH,{:.6f}\n".format(BEST_L1_REGULARIZATION_STRENGTH))
fout.write("L2_REGULARIZATION_STRENGTH,{:.6f}\n".format(BEST_L2_REGULARIZATION_STRENGTH))
Hidden_units_outer.append(BEST_HIDDEN_UNITS)
Learning_rate_ftrl_outer.append(BEST_LEARNING_RATE_FTRL)
Learning_rate_adam_outer.append(BEST_LEARNING_RATE_ADAM)
L1_strength_outer.append(BEST_L1_REGULARIZATION_STRENGTH)
L2_strength_outer.append(BEST_L2_REGULARIZATION_STRENGTH)
if verbose > 0:
print('[OUTER] [{}/{}] BEST_HIDDEN_UNITS : {}'.format(k+1, kf.get_n_splits(), BEST_HIDDEN_UNITS))
print('[OUTER] [{}/{}] BEST_LEARNING_RATE_FTRL : {:.3e}'.format(k+1, kf.get_n_splits(), BEST_LEARNING_RATE_FTRL))
print('[OUTER] [{}/{}] BEST_LEARNING_RATE_ADAM : {:.3e}'.format(k+1, kf.get_n_splits(), BEST_LEARNING_RATE_ADAM))
print('[OUTER] [{}/{}] BEST_L1_REGULARIZATION_STRENGTH : {:.3e}'.format(k+1, kf.get_n_splits(), BEST_L1_REGULARIZATION_STRENGTH))
print('[OUTER] [{}/{}] BEST_L2_REGULARIZATION_STRENGTH : {:.3e}'.format(k+1, kf.get_n_splits(), BEST_L2_REGULARIZATION_STRENGTH))
print('[OUTER] [{}/{}] BEST_TRAINING_ACCURACY : {:.3f}'.format(k+1, kf.get_n_splits(), BEST_TRAINING_ACCURACY))
## 3-4) Dataset
idx_train_train, idx_train_valid = train_test_split(idx_train, test_size=0.2, stratify=dataset.get_drugs()[idx_train])
base_drugs = np.unique(dataset.get_drugs()[idx_train_train])
X_train = dataset.make_xdata(idx_train_train)
S_train = dataset.make_sdata(base_drugs, idx_train_train)
I_train = dataset.make_idata(base_drugs, idx_train_train)
Y_train = dataset.make_ydata(idx_train_train)
X_valid = dataset.make_xdata(idx_train_valid)
S_valid = dataset.make_sdata(base_drugs, idx_train_valid)
I_valid = dataset.make_idata(base_drugs, idx_train_valid)
Y_valid = dataset.make_ydata(idx_train_valid)
X_test = dataset.make_xdata(idx_test)
S_test = dataset.make_sdata(base_drugs, idx_test)
Y_test = dataset.make_ydata(idx_test)
## 3-5) Create a model using the best parameters
if verbose > 0:
print('[OUTER] [{}/{}] NOW TRAINING THE MODEL WITH BEST PARAMETERS...'.format(k+1, kf.get_n_splits()))
checkpoint_path = os.path.join(checkpointdir, "{:03d}_RefDNN_cv_outer.ckpt".format(k))
clf = REFDNN(hidden_units=BEST_HIDDEN_UNITS,
learning_rate_ftrl=BEST_LEARNING_RATE_FTRL,
learning_rate_adam=BEST_LEARNING_RATE_ADAM,
l1_regularization_strength=BEST_L1_REGULARIZATION_STRENGTH,
l2_regularization_strength=BEST_L2_REGULARIZATION_STRENGTH,
batch_size=batchsize,
training_steps=numtrainingsteps,
checkpoint_path=checkpoint_path)
## 3-6) Fit a model
history = clf.fit(X_train, S_train, I_train, Y_train,
X_valid, S_valid, I_valid, Y_valid,
verbose=verbose)
## 3-7) Compute the metric
Pred_test = clf.predict(X_test, S_test, verbose=verbose)
Prob_test = clf.predict_proba(X_test, S_test, verbose=verbose)
ACCURACY_outer_k = accuracy_score(Y_test, Pred_test)
ACCURACY_outer.append(ACCURACY_outer_k)
AUCROC_outer_k = roc_auc_score(Y_test, Prob_test)
AUCROC_outer.append(AUCROC_outer_k)
AUCPR_outer_k = average_precision_score(Y_test, Prob_test)
AUCPR_outer.append(AUCPR_outer_k)
if verbose > 0:
print('[OUTER] [{}/{}] BEST_TEST_ACCURACY : {:.3f}'.format(k+1, kf.get_n_splits(), ACCURACY_outer_k))
print('[OUTER] [{}/{}] BEST_TEST_AUCROC : {:.3f}'.format(k+1, kf.get_n_splits(), AUCROC_outer_k))
print('[OUTER] [{}/{}] BEST_TEST_AUCPR : {:.3f}'.format(k+1, kf.get_n_splits(), AUCPR_outer_k))
## 3-8) Save meta data
drugnames_path = os.path.join(checkpointdir, "{:03d}_drugnames.csv".format(k))
with open(drugnames_path, 'w') as fout:
for drugname in base_drugs:
fout.write("{}\n".format(drugname))
genenames_path = os.path.join(checkpointdir, "{:03d}_genenames.csv".format(k))
with open(genenames_path, 'w') as fout:
for genename in dataset.get_genes():
fout.write("{}\n".format(genename))
## time log
if verbose > 0:
print(timeformat.format(3, datetime.now()))
#######################################################
## 4. Save the results
########################################################
res = pd.DataFrame.from_dict({'ACCURACY':ACCURACY_outer,
'AUCROC':AUCROC_outer,
'AUCPR':AUCPR_outer,
'Hidden_units':Hidden_units_outer,
'Learning_rate_ftrl':Learning_rate_ftrl_outer,
'Learning_rate_adam':Learning_rate_adam_outer,
'L1_regularization_strength':L1_strength_outer,
'L2_regularization_strength':L2_strength_outer})
res = res[['ACCURACY', 'AUCROC', 'AUCPR', 'Hidden_units', 'Learning_rate_ftrl', 'Learning_rate_adam', 'L1_regularization_strength', 'L2_regularization_strength']]
res.to_csv(os.path.join(outputdir, 'metrics_hyperparameters.csv'), sep=',')
## time log
if verbose > 0:
print(timeformat.format(4, datetime.now()))
if verbose > 0:
print('[FINISH]')
def fitness(hyperparameters):
global outputdir
global checkpointdir
global verbose
global dataset
global fitness_step
global fitness_idx_train
global fitness_idx_test
global innerkfold
global batchsize
global numtrainingsteps
## 1. Hyperparameters
HIDDEN_UNITS = hyperparameters[0]
LEARNING_RATE_FTRL = hyperparameters[1]
LEARNING_RATE_ADAM = hyperparameters[2]
L1_REGULARIZATION_STRENGTH = hyperparameters[3]
L2_REGULARIZATION_STRENGTH = hyperparameters[4]
## 2. 2-fold Cross Validation
if verbose > 1:
print('[INNER] [{}/{}] NOW EVALUATING PARAMETERS IN THE INNER LOOP...'.format(fitness_step, innerkfold))
objective_metrics = 0.
kf = StratifiedKFold(n_splits=innerkfold, shuffle=True)
for k, (idx_construction, idx_validation) in enumerate(kf.split(X=np.zeros_like(fitness_idx_train), y=dataset.get_drugs()[fitness_idx_train])):
## 2-1) dataset
idx_construction = fitness_idx_train[idx_construction]
idx_validation = fitness_idx_train[idx_validation]
base_drugs = np.unique(dataset.get_drugs()[idx_construction])
X_construction = dataset.make_xdata(idx_construction)
S_construction = dataset.make_sdata(base_drugs, idx_construction)
I_construction = dataset.make_idata(base_drugs, idx_construction)
Y_construction = dataset.make_ydata(idx_construction)
X_validation = dataset.make_xdata(idx_validation)
S_validation = dataset.make_sdata(base_drugs, idx_validation)
I_validation = dataset.make_idata(base_drugs, idx_validation)
Y_validation = dataset.make_ydata(idx_validation)
## 2-2) Create a model
checkpoint_path = "RefDNN_cv_inner.ckpt"
checkpoint_path = os.path.join(checkpointdir, checkpoint_path)
clf = REFDNN(hidden_units=HIDDEN_UNITS,
learning_rate_ftrl=LEARNING_RATE_FTRL,
learning_rate_adam=LEARNING_RATE_ADAM,
l1_regularization_strength=L1_REGULARIZATION_STRENGTH,
l2_regularization_strength=L2_REGULARIZATION_STRENGTH,
batch_size=batchsize,
training_steps=numtrainingsteps,
checkpoint_path=checkpoint_path)
## 2-3) Fit a model
history = clf.fit(X_construction, S_construction, I_construction, Y_construction,
X_validation, S_validation, I_validation, Y_validation,
verbose=verbose)
## 2-4) Compute the metric
Pred_validation = clf.predict(X_validation, S_validation, verbose=verbose)
objective_metrics += accuracy_score(Y_validation, Pred_validation)
training_accuracy = objective_metrics / kf.get_n_splits()
if verbose > 1:
print('[INNER] [{}/{}] HIDDEN_UNITS: {}'.format(fitness_step, innerkfold, HIDDEN_UNITS))
print('[INNER] [{}/{}] LEARNING_RATE_FTRL: {:.3e}'.format(fitness_step, innerkfold, LEARNING_RATE_FTRL))
print('[INNER] [{}/{}] LEARNING_RATE_ADAM: {:.3e}'.format(fitness_step, innerkfold, LEARNING_RATE_ADAM))
print('[INNER] [{}/{}] L1_REGULARIZATION_STRENGTH: {:.3e}'.format(fitness_step, innerkfold, L1_REGULARIZATION_STRENGTH))
print('[INNER] [{}/{}] L2_REGULARIZATION_STRENGTH: {:.3e}'.format(fitness_step, innerkfold, L2_REGULARIZATION_STRENGTH))
print('[INNER] [{}/{}] TRAINING_ACCURACY: {:.3f}'.format(fitness_step, innerkfold, training_accuracy))
## delete temporary checkpoints
os.system("rm {}.*".format(checkpoint_path))
fitness_step += 1
return -training_accuracy
if __name__=="__main__":
main()