forked from amrta-coder/FME-2021
-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_casme2.py
215 lines (202 loc) · 8.59 KB
/
train_casme2.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
import pandas as pd
import matplotlib.pyplot as plt
import os
from configparser import ConfigParser
from generator_vad import FeatruesSequence
from keras.optimizers import Adam, SGD
from sklearn.model_selection import train_test_split, KFold, cross_val_score, cross_validate, StratifiedKFold
import shutil
from datetime import datetime
from model import Conv1D_model, LSTM_model, GRU_model, ConcatCNN_model
from clr_callback import *
from callback import SaveMinLoss
import pickle
from sklearn.preprocessing import StandardScaler
from math import ceil
from pathlib import Path
from keras.utils import multi_gpu_model, plot_model
from model_resnet1d import Resnet1D
import tensorflow as tf
def set_sess_cfg():
gpu_devices = tf.config.experimental.list_physical_devices('GPU')
for device in gpu_devices:
tf.config.experimental.set_memory_growth(device, True)
def save_config(output):
if not os.path.exists(output):
os.makedirs(output)
dir = os.path.dirname(__file__)
output = os.path.join(output, str(datetime.now().strftime("%Y%m%d-%H%M%S")+'_casme2_uncrop_'+model_name+str(n_gru_units)))
output_dir_src = os.path.join(output, 'src')
os.makedirs(output_dir_src, exist_ok=True)
print(f"backup config file to {output_dir_src}")
shutil.copy("config.ini", os.path.join(output_dir_src, 'config.ini'))
shutil.copy("model.py", os.path.join(output_dir_src, 'model.py'))
train_file = os.path.basename(__file__)
shutil.copy(os.path.join(dir, train_file), os.path.join(output_dir_src, train_file))
return output
def scheduler(epoch, lr):
if epoch < 10:
lr = lr*0.1
elif epoch < 20:
lr = lr*0.01
elif epoch < 30:
lr = lr*0.001
elif epoch < 40:
lr = lr*0.0001
return lr
def main(output_dir, optimizer):
model_method = eval(f'{model_name}_model')
if model_name == 'ConcatCNN':
model = model_method(n_timesteps=n_timesteps, n_features=n_features)
else:
model = model_method(n_timesteps=n_timesteps, n_features=n_features, n_gru_units=n_gru_units)
print(model.summary())
skf = StratifiedKFold(n_splits=8, shuffle=True, random_state=0)
X = df_label['file_path']
Y = df_label['label']
best_train_mses = []
best_val_mses = []
for i, (train_index, test_index) in enumerate(skf.split(X, Y)):
# del model
K.clear_session()
# gc.collect()
output_dir_full = os.path.join(output_dir, str(i))
os.makedirs(output_dir_full, exist_ok=True)
# get train/dev sample counts
train_counts = len(train_index)
val_counts = len(test_index)
train_steps = ceil(train_counts / batch_size)
validation_steps = ceil(val_counts / batch_size)
train_csv = Path(output_dir_full).joinpath("train_{:02d}.csv".format(i))
val_csv = Path(output_dir_full).joinpath("val_{:02d}.csv".format(i))
df_label.loc[train_index].to_csv(str(train_csv), index=False, columns=["file_path", "label"])
df_label.loc[test_index].to_csv(str(val_csv), index=False, columns=["file_path", "label"])
train_sequence = FeatruesSequence(
dataset_csv_file=str(train_csv),
# csv_source_dir=csv_source_dir,
batch_size=batch_size,
random_state=seed,
)
validation_sequence = FeatruesSequence(
dataset_csv_file=str(val_csv),
# csv_source_dir=csv_source_dir,
batch_size=batch_size,
shuffle_on_epoch_end=False,
test=True,
)
plot_model(model, to_file=os.path.join(output_dir_full, 'model.png'))
if optimizer == 'adam':
optimizer = Adam(lr=initial_learning_rate)
base_lr = base_lr_adam
max_lr = max_lr_adam
elif optimizer == 'sgd':
optimizer = SGD(momentum=momentum, decay=decay)
base_lr = base_lr_sgd
max_lr = max_lr_sgd
model.compile(optimizer=optimizer, loss=loss_func, metrics=['mse', 'mae'])
earlystop = EarlyStopping(monitor='val_loss', min_delta=0, patience=36, verbose=0, mode='min')
# if cyclicLR_mode == 'exp_range':
# gamma = 0.99994
# else:
# gamma = 1.
# clr = CyclicLR(mode=cyclicLR_mode, step_size=train_steps*2, base_lr=base_lr, max_lr=max_lr, gamma=gamma)
# change_lr = LearningRateScheduler(scheduler, verbose=1)
reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=8, verbose=1)
model_save_path = os.path.join(output_dir_full, output_model_name)
checkpoint = ModelCheckpoint(
model_save_path,
save_weights_only=False,
save_best_only=True,
verbose=1,
)
csv_logger = CSVLogger(os.path.join(output_dir_full, 'training.csv'))
save_min_loss = SaveMinLoss(filepath=output_dir_full)
tensor_board = TensorBoard(log_dir=os.path.join(output_dir_full, "logs"), batch_size=batch_size)
callbacks = [
checkpoint,
tensor_board,
csv_logger,
# clr,
reduce_lr,
save_min_loss,
earlystop,
]
print("** start training **")
# Training.
history = model.fit_generator(
generator=train_sequence,
steps_per_epoch=train_steps,
epochs=epochs,
validation_data=validation_sequence,
validation_steps=validation_steps,
callbacks=callbacks,
# class_weight='auto',
workers=generator_workers,
shuffle=False,
)
# dump history
print("** dump history **")
with open(os.path.join(output_dir_full, "history.pkl"), "wb") as f:
pickle.dump({
"history": history.history,
}, f)
print("** done! **")
mse = history.history["mse"]
val_mse = history.history["val_mse"]
best_val_mses.append(min(val_mse))
best_train_mses.append(min(mse))
# h = clr.history
# lr = h['lr']
# iterations = h['iterations']
# plt.xlabel('Training Iterations')
# plt.ylabel('Learning Rate')
# plt.title("CLR - 'exp_range' Policy")
# plt.plot(iterations, lr)
# plt.savefig(os.path.join(output_dir_full, 'iterations_lr.png'))
results_mean = np.array(best_val_mses).mean()
results_std = np.array(best_val_mses).std()
print(f'best_train_mses:{best_train_mses}')
print(f'best_val_mses:{best_val_mses}')
print(f'best_val_mses mean:{results_mean}')
print(f'best_val_mses std:{results_std}')
results_file = os.path.join(output_dir, "results.log")
with open(results_file, 'a') as f:
f.write(f'best_train_mses:{best_train_mses}')
f.write(f'best_val_mses:{best_val_mses}')
f.write(f'best_val_mses mean:{results_mean}')
f.write(f'best_val_mses std:{results_std}')
if __name__ == "__main__":
set_sess_cfg()
# parser config
config_file = "./config.ini"
cp = ConfigParser()
cp.read(config_file)
fme_dir = cp["DEFAULT"].get("fme_dir")
features_engineered = cp["DEFAULT"].get("features_engineered")
features_subdir_vad = cp["DEFAULT"].get("features_subdir_vad")
label_croped_file_vad = cp["DEFAULT"].get("label_croped_file_vad")
batch_size = cp["TRAIN"].getint("batch_size")
seed = cp["TRAIN"].getint("seed")
output_fold = cp["DEFAULT"].get("output_fold")
n_timesteps = cp["DEFAULT"].getint("n_timesteps_vad")
n_features = cp["DEFAULT"].getint("n_features_vad")
initial_learning_rate = cp["TRAIN"].getfloat("initial_learning_rate")
output_model_name = cp["TRAIN"].get("output_model_name")
base_lr_adam = cp["TRAIN"].getfloat("base_lr_adam")
max_lr_adam = cp["TRAIN"].getfloat("max_lr_adam")
momentum = cp["TRAIN"].getfloat("momentum")
decay = cp["TRAIN"].getfloat("decay")
base_lr_sgd = cp["TRAIN"].getfloat("base_lr_sgd")
max_lr_sgd = cp["TRAIN"].getfloat("max_lr_sgd")
loss_func = cp["TRAIN"].get("loss_func")
cyclicLR_mode = cp["TRAIN"].get("cyclicLR_mode")
epochs = cp["TRAIN"].getint("epochs")
generator_workers = cp["TRAIN"].getint("generator_workers")
optimizer = cp["TRAIN"].get("optimizer")
model_name = cp["TRAIN"].get("model_name")
n_gru_units = cp["TRAIN"].getint("n_gru_units")
features_engineered_root = os.path.join(fme_dir, features_engineered)
features_engineered_dir = os.path.join(features_engineered_root, features_subdir_vad)
df_label = pd.read_csv(os.path.join(features_engineered_root, label_croped_file_vad))
output_dir = save_config(output_fold)
main(output_dir, optimizer)