-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtraining_function.py
executable file
·156 lines (132 loc) · 6.87 KB
/
training_function.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
import time
import torch
import numpy as np
import datasets.datasetfactory as df
from datasets.task_sampler import get_sampler
from pathlib import Path
from logging_utils import Metrics, setup_loggers
from training_utils import iterator_trainer_statistics, to_numpy
def log_dict(logger, dictionary, prefix=''):
for k, v in dictionary.items():
logger.log_key_val(prefix + k, v)
def continual_meta_training(dataset_name,
trainer,
training_steps,
classes_per_task=1,
task_batch_size=20,
all_tasks_batch_size=64,
meta_batch_size=1,
separate_validation_data=False,
all_data_meta=False,
task_validation_batch_size=5,
detect_anomaly=False,
device=None,
evaluation_batch_size=128,
log_training_every=40,
log_evaluation_every=2000,
save_weights_every=10000):
if detect_anomaly:
print('WARNING: running with activated anomaly detection')
torch.autograd.set_detect_anomaly(True)
work_dir = Path.cwd()
print('workspace: {}'.format(work_dir))
train_logger, test_logger = setup_loggers(d='data',
namespaces=['train', 'test'],
colors=['white', 'yellow'],)
dataset_train = df.DatasetFactory.get_dataset(
dataset_name, background=True, train=True,
all=not separate_validation_data)
if all_data_meta:
assert separate_validation_data
if all_data_meta:
dataset_test = df.DatasetFactory.get_dataset(
dataset_name, background=True, train=False,
all=True)
else:
dataset_test = df.DatasetFactory.get_dataset(
dataset_name, background=True, train=False,
all=not separate_validation_data)
classes_idxs = list(range(np.max(dataset_train.targets)))
# Iterators used for evaluation
iterator_test = torch.utils.data.DataLoader(dataset_test, batch_size=evaluation_batch_size,
shuffle=True, num_workers=1)
iterator_train = torch.utils.data.DataLoader(dataset_train, batch_size=evaluation_batch_size,
shuffle=True, num_workers=1)
sampler = get_sampler(dataset_name, classes_idxs, trainset=dataset_train,
testset=dataset_test, batch_size=task_batch_size,
validation_batch_size=task_validation_batch_size,
all_tasks_batch_size=all_tasks_batch_size)
if device is None:
if torch.cuda.is_available():
device = torch.device('cuda')
else:
device = torch.device('cpu')
trainer.set_device(device=device)
training_metrics = Metrics('training_accuracy', 'meta_loss',
'training_time')
start_training_time = time.time()
test_logger.log_key_val('step', 0)
step_start_eval_time = time.time()
training_data_stats = iterator_trainer_statistics(trainer=trainer,
iterator=iterator_train)
test_data_stats = iterator_trainer_statistics(trainer=trainer,
iterator=iterator_test)
log_dict(test_logger, training_data_stats, prefix='train_data_')
log_dict(test_logger, test_data_stats, prefix='test_data_')
step_end_eval_time = time.time()
evaluation_time = step_end_eval_time - step_start_eval_time
test_logger.log_key_val('evaluation_time', evaluation_time)
test_logger.log_iteration()
for step in range(training_steps):
step_start_training_time = time.time()
task_iterators = []
if separate_validation_data:
validation_task_iterators = []
else:
validation_task_iterators = None
all_tasks_iterators = []
for i in range(meta_batch_size):
sampled_classes = np.random.choice(classes_idxs,
size=classes_per_task,
replace=False)
for c in sampled_classes:
task_iterators.append(sampler.sample_task([c], train=True))
if separate_validation_data:
validation_task_iterators.append(
sampler.sample_task([c], train=False))
all_tasks_iterator = sampler.sample_all_tasks()
all_tasks_iterators.append(all_tasks_iterator)
training_accuracy, meta_loss = trainer.meta_train(
task_iterators=task_iterators,
validation_task_iterators=validation_task_iterators, # todo
all_tasks_iterators=all_tasks_iterators)
if trainer.meta_lr_scheduler is not None:
trainer.meta_lr_scheduler.step()
step_end_training_time = time.time()
training_time = step_end_training_time - step_start_training_time
training_metrics.update(training_accuracy=training_accuracy,
meta_loss=meta_loss, training_time=np.around(training_time, 3))
if (step + 1) % log_training_every == 0:
train_logger.log_key_val('step', step + 1)
current_training_metrics = training_metrics.get()
training_metrics.reset()
log_dict(logger=train_logger, dictionary=current_training_metrics)
log_dict(logger=train_logger, dictionary=trainer.logging_stats())
train_logger.log_iteration()
if (step + 1) % log_evaluation_every == 0 or step == training_steps:
test_logger.log_key_val('step', step + 1)
step_start_eval_time = time.time()
training_data_stats = iterator_trainer_statistics(trainer=trainer,
iterator=iterator_train)
test_data_stats = iterator_trainer_statistics(trainer=trainer,
iterator=iterator_test)
log_dict(test_logger, training_data_stats, prefix='train_data_')
log_dict(test_logger, test_data_stats, prefix='test_data_')
step_end_eval_time = time.time()
evaluation_time = step_end_eval_time - step_start_eval_time
test_logger.log_key_val('evaluation_time', evaluation_time)
test_logger.log_iteration()
if (step + 1) % save_weights_every == 0 or step == training_steps:
print('saving model for iteration {}'.format(step+1))
torch.save(trainer.learner, 'learner_{}.net'.format(step + 1))
print('Total execution time: {}'.format(time.time() - start_training_time))