-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy path3_lodocv_on_GDSC.py
185 lines (151 loc) · 8.72 KB
/
3_lodocv_on_GDSC.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
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 LeaveOneGroupOut, train_test_split
from refdnn.model import REFDNN
from refdnn.dataset import DATASET, read_cancertypeFile
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_3', help="A directory path for saving outputs (default:'output_3')")
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('-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()
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] 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
########################################################
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. Train RefDNN using the best hyperparameters
########################################################
batchsize = args.b
numtrainingsteps = args.t
## 2-1) init lists for metrics
ACCURACY_outer = []
AUCROC_outer = []
AUCPR_outer = []
DRUG_outer = []
kf = LeaveOneGroupOut()
n_splits = kf.get_n_splits(groups=dataset.get_drugs())
print("LeaveOneGroupOut.get_n_splits: {}".format(n_splits))
for k, (idx_train, idx_test) in enumerate(kf.split(X=np.zeros(len(dataset)), groups=dataset.get_drugs())):
## 2-2) Check a cancer type in test
test_drug = np.unique(dataset.get_drugs()[idx_test])[0]
DRUG_outer.append(test_drug)
print('[{}/{}] TEST_DRUG: {}'.format(k+1, n_splits, test_drug))
## 2-3) Set the best values of hyperparameters
BEST_HIDDEN_UNITS = 49
BEST_LEARNING_RATE_FTRL = 7.94581095185585e-06
BEST_LEARNING_RATE_ADAM = 0.0004067851789088527
BEST_L1_REGULARIZATION_STRENGTH = 0.001
BEST_L2_REGULARIZATION_STRENGTH = 66.7516541409175
## 2-4) 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)
## 2-5) Create a model using the best parameters
if verbose > 0:
print('[{}/{}] NOW TRAINING THE MODEL WITH BEST PARAMETERS...'.format(k+1, n_splits))
checkpoint_path = "RefDNN_lodocv_{}.ckpt".format(test_drug)
checkpoint_path = os.path.join(checkpointdir, checkpoint_path)
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)
## 2-6) Fit a model
history = clf.fit(X_train, S_train, I_train, Y_train,
X_valid, S_valid, I_valid, Y_valid,
verbose=verbose)
## 2-7) 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) if np.unique(Y_test).shape[0] == 2 else -0.999 # if a current test set contains only single label, then the calculate of AUCROC is skipped
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('[{}/{}] BEST_TEST_ACCURACY : {:.3f}'.format(k+1, n_splits, ACCURACY_outer_k))
print('[{}/{}] BEST_TEST_AUCROC : {:.3f}'.format(k+1, n_splits, AUCROC_outer_k))
print('[{}/{}] BEST_TEST_AUCPR : {:.3f}'.format(k+1, n_splits, AUCPR_outer_k))
## time log
if verbose > 0:
print(timeformat.format(3, datetime.now()))
#######################################################
## 3. Save the results
########################################################
res = pd.DataFrame.from_dict({'DRUGNAME':DRUG_outer,
'ACCURACY':ACCURACY_outer,
'AUCROC':AUCROC_outer,
'AUCPR':AUCPR_outer,
'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})
res = res[['DRUGNAME', '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]')
if __name__=="__main__":
main()