forked from idrugLab/FG-BERT
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Reg_hyperopt.py
230 lines (192 loc) · 8.39 KB
/
Reg_hyperopt.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
import tensorflow as tf
import tensorflow.keras as keras
import tensorflow.keras.layers as layers
import os
import pandas as pd
import numpy as np
from tensorflow.keras.constraints import max_norm
from dataset_scaffold_random import Graph_Regression_Dataset
from sklearn.metrics import roc_auc_score
from model import PredictModel,BertModel
from hyperopt import fmin, tpe, hp
os.environ["TF_FORCE_GPU_ALLOW_GROWTH"] = "true"
os.environ['TF_DETERMINISTIC_OPS'] = '1'
keras.backend.clear_session()
def count_parameters(model):
total_params = 0
for variable in model.trainable_variables:
shape = variable.shape
params = 1
for dim in shape:
params *= dim
total_params += params
return total_params
def main(seed,args):
# tasks = ['ESOL', 'FreeSolv', 'Lipophilicity', 'Malaria', 'cep']
task = 'ESOL'
print(task)
if task == 'ESOL':
label = ['measured log solubility in mols per litre']
elif task == 'FreeSolv':
label = ['expt']
elif task == 'Lipophilicity':
label = ['exp']
elif task == 'Malaria':
label = ['PCE']
elif task == 'cep':
label = ['activity']
vocab_size = 18
trained_epoch = 20
num_layers = 6
d_model = 256
addH = True
dff = d_model * 2
seed = seed
arch = {'name': 'Medium', 'path': 'medium3_weights'}
dense_dropout = args['dense_dropout']
learning_rate = args['learning_rate']
batch_size = args['batch_size']
num_heads = args['num_heads']
pretraining = True
pretraining_str = 'pretraining' if pretraining else ''
np.random.seed(seed=seed)
tf.random.set_seed(seed=seed)
graph_dataset = Graph_Regression_Dataset('ESOL.csv', smiles_field='smiles',
label_field=label,normalize=True,seed=seed,batch_size=batch_size,a=len(label),max_len=500,addH=True)
train_dataset, test_dataset,val_dataset = graph_dataset.get_data()
x, adjoin_matrix, y = next(iter(train_dataset.take(1)))
seq = tf.cast(tf.math.equal(x, 0), tf.float32)
mask = seq[:, tf.newaxis, tf.newaxis, :]
model = PredictModel(num_layers=num_layers, d_model=d_model, dff=dff, num_heads=num_heads, vocab_size=vocab_size,a=len(label),
dense_dropout = dense_dropout)
if pretraining:
temp = BertModel(num_layers=num_layers, d_model=d_model, dff=dff, num_heads=num_heads, vocab_size=vocab_size)
pred = temp(x, mask=mask, training=True, adjoin_matrix=adjoin_matrix)
temp.load_weights(arch['path']+'/bert_weights{}_{}.h5'.format(arch['name'],trained_epoch))
temp.encoder.save_weights(arch['path']+'/bert_weights_encoder{}_{}.h5'.format(arch['name'],trained_epoch))
del temp
pred = model(x, mask=mask, training=True, adjoin_matrix=adjoin_matrix)
model.encoder.load_weights(arch['path']+'/bert_weights_encoder{}_{}.h5'.format(arch['name'],trained_epoch))
print('load_wieghts')
total_params = count_parameters(model)
print('*'*100)
print("Total Parameters:", total_params)
print('*'*100)
class CustomSchedule(tf.keras.optimizers.schedules.LearningRateSchedule):
def __init__(self, d_model, total_steps=4000):
super(CustomSchedule, self).__init__()
self.d_model = d_model
self.d_model = tf.cast(self.d_model, tf.float32)
self.total_step = total_steps
self.warmup_steps = total_steps*0.10
def __call__(self, step):
arg1 = step/self.warmup_steps
arg2 = 1-(step-self.warmup_steps)/(self.total_step-self.warmup_steps)
return 10e-5* tf.math.minimum(arg1, arg2)
steps_per_epoch = len(train_dataset)
learning_rate = CustomSchedule(128,100*steps_per_epoch)
optimizer = tf.keras.optimizers.Adam(learning_rate = learning_rate)
value_range = graph_dataset.value_range
mse = 10000
stopping_monitor = 0
for epoch in range(200):
mse_object = tf.keras.metrics.MeanSquaredError()
for x,adjoin_matrix,y in train_dataset:
with tf.GradientTape() as tape:
seq = tf.cast(tf.math.equal(x, 0), tf.float32)
mask = seq[:, tf.newaxis, tf.newaxis, :]
preds = model(x,mask=mask,training=True,adjoin_matrix=adjoin_matrix)
loss = tf.reduce_mean(tf.square(y-preds))
grads = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
mse_object.update_state(y,preds)
print('epoch: ',epoch,'loss: {:.4f}'.format(loss.numpy().item()))
y_true = []
y_preds = []
for x, adjoin_matrix, y in val_dataset:
seq = tf.cast(tf.math.equal(x, 0), tf.float32)
mask = seq[:, tf.newaxis, tf.newaxis, :]
preds = model(x,mask=mask,adjoin_matrix=adjoin_matrix,training=False)
y_true.append(y.numpy())
y_preds.append(preds.numpy())
y_true = np.concatenate(y_true,axis=0).reshape(-1)
y_preds = np.concatenate(y_preds,axis=0).reshape(-1)
mse_new = keras.metrics.MSE(y_true, y_preds).numpy() * (value_range**2)
val_mse = keras.metrics.MSE(y_true, y_preds).numpy() * (value_range**2)
print(f'val mse: {val_mse.item()}')
if mse_new.item() < mse:
mse = mse_new.item()
stopping_monitor = 0
np.save('{}/{}{}{}{}{}'.format(arch['path'], task, seed, arch['name'], trained_epoch, trained_epoch,pretraining_str),
[y_true, y_preds])
model.save_weights('regression_weights/{}.h5'.format(task))
else:
stopping_monitor +=1
print('best mse: {:.4f}'.format(mse))
if stopping_monitor>0:
print('stopping_monitor:',stopping_monitor)
if stopping_monitor>30:
break
y_true = []
y_preds = []
model.load_weights('regression_weights/{}.h5'.format(task, seed))
for x, adjoin_matrix, y in test_dataset:
seq = tf.cast(tf.math.equal(x, 0), tf.float32)
mask = seq[:, tf.newaxis, tf.newaxis, :]
preds = model(x, mask=mask, adjoin_matrix=adjoin_matrix, training=False)
y_true.append(y.numpy())
y_preds.append(preds.numpy())
y_true = np.concatenate(y_true, axis=0).reshape(-1,len(label))
y_preds = np.concatenate(y_preds, axis=0).reshape(-1,len(label))
test_mse = keras.metrics.MSE(y_true.reshape(-1), y_preds.reshape(-1)).numpy() * (value_range**2)
test_RMSE = np.sqrt(test_mse.item())
print('test rmse:{:.4f}'.format(test_RMSE))
return mse, test_RMSE
space = {"dense_dropout": hp.quniform("dense_dropout", 0, 0.5, 0.05),
"learning_rate": hp.loguniform("learning_rate", np.log(3e-5), np.log(15e-5)),
"batch_size": hp.choice("batch_size", [16,32,48,64]),
"num_heads": hp.choice("num_heads", [4, 8])
}
def hy_main(arch):
x = 0
mse_list = []
RMSE_list = []
for seed in [1,2,3]:
print(seed)
mse, test_RMSE = main(seed,arch)
x+= test_RMSE
mse_list.append(mse)
RMSE_list.append(test_RMSE)
mse_list.append(np.mean(mse_list))
RMSE_list.append(np.mean(RMSE_list))
print(mse_list)
print(RMSE_list)
print(arch["dense_dropout"])
print(arch["learning_rate"])
print(arch["batch_size"])
print(arch["num_heads"])
return x/3
best = fmin(hy_main, space, algo = tpe.suggest, max_evals = 30)
print(best)
best_dict = {}
a = [16,32,48,64]
b = [4, 8]
best_dict["dense_dropout"] = best["dense_dropout"]
best_dict["learning_rate"] = best["learning_rate"]
best_dict["batch_size"] = a[best["batch_size"]]
best_dict["num_heads"] = b[best["num_heads"]]
print(best_dict)
print(hy_main(best_dict))
# if __name__ == "__main__":
# mse_list = []
# RMSE_list = []
# for seed in [1,2,3]:
# print(seed)
# args = {"dense_dropout":0.05, "learning_rate":0.0000636859, "batch_size":16, "num_heads":8}
# MSE, test_RMSE= main(seed, args)
# mse_list.append(MSE)
# RMSE_list.append(test_RMSE)
# mse_list.append(np.mean(mse_list))
# RMSE_list.append(np.mean(RMSE_list))
# print(mse_list)
# print(RMSE_list)