-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathdata_parallel.py
196 lines (153 loc) · 7.75 KB
/
data_parallel.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
########################################################################################################################
# @author Oriol Aranda (https://github.com/oriolaranda/)
# @date Oct 2021
########################################################################################################################
from os import environ, path
import time
import argparse
from datetime import datetime, timedelta
import json
import tensorflow as tf
from auxiliary_brain_seg import *
import ray
from itertools import product
from ray.util.sgd.tf.tf_trainer import TFTrainer, TFTrainable
from utils import loss_metric, data_generator
def data_creator(config):
files_tr = [config['dataset_path'] + f"train_{i}.tfrecord" for i in range(config['train_shard'])]
files_val = [config['dataset_path'] + f"valid_{i}.tfrecord" for i in range(config['valid_shard'])]
options = tf.data.Options()
options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF
train_g = data_generator(files_tr, config, config['train_size'] // 10, train=True)
valid_g = data_generator(files_val, config, config['valid_size'] // 10)
return train_g.with_options(options), valid_g.with_options(options)
def model_creator(config):
loss, metric = loss_metric(*config['loss_metric'])
model = model_unet_tune(input_shape=(4, 240, 240, 152), filter_start=8, loss_function=loss, metrics=[metric],
initial_learning_rate=config['lr'], amsgrad=config['amsgrad'], b2=config['b2'],
norm=config['norm'])
return model
def create_config(config):
train_size, valid_size = config['train_size'], config['valid_size']
tstamp = "{}".format(datetime.now().strftime("%m_%d_%Y-%H:%M:%S"))
logdir = f"{config['log_dir']}/logs/{tstamp}"
tb_callback = tf.keras.callbacks.TensorBoard(log_dir=logdir)
callbacks = [tb_callback]
config['fit_config'] = {
'steps_per_epoch': train_size // config['batch_size'],
'epochs': config['num_epochs'],
'callbacks': callbacks,
'verbose': 1 if config['debug'] else 2,
}
config['evaluate_config'] = {
'steps': valid_size // config['batch_size'],
'verbose': 0 # there is only 1, 0 possible values
}
return config
def multi_node_training(config):
print("Iniciando ray multi-nodo...")
ray.init(address='auto', _node_ip_address=environ["ip_head"].split(":")[0],
_redis_password=environ["redis_password"])
assert ray.is_initialized(), "ERROR: Ray is not initialized!"
print('''This cluster consists of
{} nodes in total
{} CPU resources in total
{} GPU resources in total
'''.format(len(ray.nodes()), ray.cluster_resources()['CPU'], ray.cluster_resources()['GPU']))
trainer = TFTrainer(
model_creator=model_creator,
data_creator=data_creator,
num_replicas=config['num_replicas'],
use_gpu=True,
verbose=True,
config=create_config(config)
)
trainer.train() # train with train_config
eval_stats = trainer.validate() # validate with validate_config
ray.shutdown() # Shutdown ray to prevent an error of ports
assert not ray.is_initialized(), "ERROR: Ray is still running!"
return tuple(eval_stats.values())
def distributed_training(config):
loss, metric = loss_metric(*config['loss_metric'])
if config['num_replicas'] > 1:
gpus = tf.config.list_physical_devices('GPU')
gpus_name = [g.name.split('e:')[1] for g in gpus]
assert len(gpus_name) == config['num_replicas'], f"Number of GPU available {len(gpus_name)}"
strategy = tf.distribute.MirroredStrategy(gpus_name[:config['num_replicas']])
print("Number of GPUs:", strategy.num_replicas_in_sync) # sanity check
with strategy.scope():
model = model_unet_tune(input_shape=(4, 240, 240, 152), filter_start=8, loss_function=loss,
metrics=[metric], initial_learning_rate=config['lr'], amsgrad=config['amsgrad'],
b2=config['b2'], norm=config['norm'])
else:
print("Number of GPUs:", 1)
model = model_unet_tune(input_shape=(4, 240, 240, 152), filter_start=8, loss_function=loss, metrics=[metric],
initial_learning_rate=config['lr'], amsgrad=config['amsgrad'], b2=config['b2'],
norm=config['norm'])
# model.summary()
train_size, valid_size = config['train_size'], config['valid_size']
steps_per_epoch = train_size // config['batch_size']
validation_steps = valid_size // config['batch_size']
files_tr = [config['dataset_path'] + f"train_{i}.tfrecord" for i in range(config['train_shard'])]
files_val = [config['dataset_path'] + f"valid_{i}.tfrecord" for i in range(config['valid_shard'])]
train_g = data_generator(files_tr, config, config['train_size'] // 10, train=True)
valid_g = data_generator(files_val, config, config['valid_size'] // 10, )
tstamp = "{}".format(datetime.now().strftime("%m_%d_%Y-%H:%M:%S"))
logdir = f"{config['log_dir']}/logs/{tstamp}"
tb_callback = tf.keras.callbacks.TensorBoard(log_dir=logdir)
callbacks = [] if config['debug'] else [tb_callback]
model.fit(x=train_g,
steps_per_epoch=steps_per_epoch,
epochs=config['num_epochs'],
callbacks=callbacks,
verbose=1 if config['debug'] else 2)
results = model.evaluate(valid_g, steps=validation_steps, verbose=1 if config['debug'] else 2)
return tuple(results) # val_loss, val_metric
def main(args):
with open(args.config) as f:
config = json.load(f)
config['batch_size'] = config['batch_size_per_replica'] * config['num_replicas']
config['log_dir'] = args.config.split('/config.json')[0]
config['dataset_path'] = "/gpfs/projects/bsc31/bsc31654/dataset/preprocessed/"
with open(config['dataset_path'] + "info.json") as f:
info = json.load(f)
config.update(info) # add info to config
config['prefetch_buffer_size'] = tf.data.experimental.AUTOTUNE
config['map_num_parallel_calls'] = tf.data.experimental.AUTOTUNE
config['tfrecord_buffer_size'] = tf.data.experimental.AUTOTUNE
config['tfrecord_num_parallel_reads'] = tf.data.experimental.AUTOTUNE
config['loss_metric'] = (dice_loss, dice_coeff)
multi = config['num_replicas'] > 4
s = time.time()
#####################################
# DEFINE HYPERPARAMETERS TUNE
#####################################
lr = [x * config['num_replicas'] for x in [1e-04, 5e-04, 1e-03, 5e-03]]
norm = ['bn', 'gn']
amsgrad = [True, False]
b2 = [0.99, 0.999]
tune = [lr, norm, b2, amsgrad]
keys = ['lr', 'norm', 'b2', 'amsgrad']
######################################
######################################
best_config = {}
trials = list(product(*tune))
for num, trial in enumerate(trials, 1):
config.update(dict(zip(keys, trial)))
print(f"Trial {num}/{len(trials)} : {trial}")
s_ = time.time()
if multi:
res = multi_node_training(config)
else:
res = distributed_training(config)
print("Trial time:", time.time() - s_, "s")
best_config[res] = trial
best = max(best_config.keys(), key=lambda x: x[1])
print("Best:", best_config[best])
print("Elapsed time:", timedelta(seconds=(time.time() - s)), "s")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--config", type=str, required=True, help="Path: Json file configuration")
_args, _ = parser.parse_known_args()
assert path.exists(_args.config), f"Config file doesn't exist: {_args.config}"
main(_args)