-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
201 lines (153 loc) · 6.42 KB
/
main.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
import os
from pytorch_lightning import Trainer, seed_everything
from pytorch_lightning.callbacks.early_stopping import EarlyStopping
# from pl_bolts.callbacks import PrintTableMetricsCallback
from pytorch_lightning.callbacks import ModelCheckpoint
from pytorch_lightning.callbacks.lr_monitor import LearningRateMonitor
# from pytorch_lightning.callbacks import RichProgressBar
from pytorch_lightning.loggers import MLFlowLogger
# from continuum import ClassIncremental
# from continuum.tasks import split_train_val
# from continuum.datasets import MNIST, CIFAR10
from data.incremental_datamodule import IncrementalDataModule
from data.incremental_scenario import incremental_scenario
from learner.Base_learner import Base_learner
from args.args_trainer import args_trainer
from args.args_model import args_model
# print(args_model)
# changhong code
from incremental_net.inc_net import IncrementalNet
inc_network = IncrementalNet("resnet32", pretrained=False, gradcam=False)
import sys
seed_everything(42, workers=True)
# model
learner = Base_learner(
inc_network,
args_model
)
exp_name = "incremental_learning"
mlflow_logger = MLFlowLogger(experiment_name=exp_name, tracking_uri="http://localhost:10500")
run_id = mlflow_logger.run_id
# callbacks
# print_table_metrics_callback = PrintTableMetricsCallback()
monitor_metric = 'loss_epoch'
mode = 'min'
early_stop_callback = EarlyStopping(
monitor=monitor_metric,
min_delta=0.00,
patience=3,
verbose=False,
mode=mode,
strict = True)
# checkpoint_callback = ModelCheckpoint(
# dirpath=f'saved_models/{run_id}/',
# monitor=monitor_metric,
# filename='sample-mnist-{epoch:02d}-{val_acc:.2f}',
# save_top_k=3,
# mode=mode,
# save_last=True
# )
learning_rate_monitor_callback = LearningRateMonitor(
logging_interval='epoch'
)
trainer = Trainer(
accelerator=args_trainer.accelerator,
gpus = args_trainer.gpus, # [0,1,7,8,9] / -1
# gpus = "1",
max_epochs=args_trainer.max_epochs,
progress_bar_refresh_rate=args_trainer.progress_bar_refresh_rate,
check_val_every_n_epoch = args_trainer.check_val_every_n_epoch,
weights_summary=args_trainer.weights_summary,
callbacks = [learning_rate_monitor_callback], # early_stop_callback, checkpoint_callback
log_every_n_steps = args_trainer.log_every_n_steps, # default: 50
logger = mlflow_logger,
sync_batchnorm = args_trainer.sync_batchnorm,
fast_dev_run = args_trainer.fast_dev_run,
num_sanity_val_steps = args_trainer.num_sanity_val_steps
) # precision=16 [checked]
# PATH_DATASETS = "/data/Public/Datasets"
# increment=2
# initial_increment=2
if trainer.is_global_zero:
# print(mlflow_logger.run_id)
print(f' \
batch_size: {args_model.batch_size},\n \
learning_rate: {args_model.learning_rate},\n \
dataset: {args_model.dataset},\n \
initial_increment: {args_model.initial_increment}, \n \
increment: {args_model.increment} \
')
ckpt_save_root_dir = 'saved_models/'
ckpt_save_dir = os.path.join(ckpt_save_root_dir, run_id)
if not os.path.exists(ckpt_save_dir):
os.makedirs(ckpt_save_dir)
# mlflow_logger.log_hyperparams(args_model) [checked]
inc_scenario = incremental_scenario(
dataset_name = args_model.dataset,
train_additional_transforms = [],
test_additional_transforms = [],
initial_increment = args_model.initial_increment,
increment = args_model.increment,
datasets_dir = args_model.datasets_dir
)
# inc_scenario.setup()
train_scenario, test_scenario = inc_scenario.get_incremental_scenarios()
# print(inc_scenario.class_order)
nb_seen_classes = args_model.initial_increment
for task_id, taskset in enumerate(train_scenario):
try:
learner.model.update_fc(nb_seen_classes)
dm = IncrementalDataModule(
task_id = task_id,
train_taskset = taskset,
test_taskset = test_scenario[:task_id+1],
dims = inc_scenario.dims,
nb_total_classes = inc_scenario.nb_total_classes,
batch_size = args_model.batch_size,
num_workers = args_model.num_workers,
val_split_ratio = args_model.val_split_ratio)
trainer.fit(learner, datamodule=dm)
# [checked]
# dm.setup('test')
# trainer.test(learner, dataloaders=dm.test_dataloader())
if trainer.is_global_zero: # control that only one device can print.
print('*'*100)
print(f'nb_seen_classes have been seen is: {nb_seen_classes}')
trainer.save_checkpoint(os.path.join(ckpt_save_dir, f"{nb_seen_classes}-{learner.last_incremental_acc.item()}.ckpt"))
learner.update_nb_seen_classes(nb_seen_classes)
nb_seen_classes += args_model.increment
del dm
del trainer
trainer = Trainer(
accelerator=args_trainer.accelerator,
gpus = args_trainer.gpus, # [0,1,7,8,9] / -1
# gpus = "1",
max_epochs=args_trainer.max_epochs,
progress_bar_refresh_rate=args_trainer.progress_bar_refresh_rate,
check_val_every_n_epoch = args_trainer.check_val_every_n_epoch,
weights_summary=args_trainer.weights_summary,
callbacks = [learning_rate_monitor_callback], # early_stop_callback, checkpoint_callback RichProgressBar()
log_every_n_steps = args_trainer.log_every_n_steps, # default: 50
logger = mlflow_logger, # use the same logger, otherwise it will create new runs in mlflow.
sync_batchnorm = args_trainer.sync_batchnorm,
fast_dev_run = args_trainer.fast_dev_run,
num_sanity_val_steps = args_trainer.num_sanity_val_steps
) # precision=16 [checked]
except KeyboardInterrupt:
try:
break
sys.exit(0)
except SystemExit:
os._exit(0)
# test [worked]
# dm.setup('test')
# trainer.test(classifier, dataloaders=dm.test_dataloader())
# test [not worked]
# trainer.test(classifier, datamodule=dm)
# 复现
'''
查看iCaRL源代码，看看超参数设置
查看pl中有关lr adjust的callback https://pytorch-lightning.readthedocs.io/en/latest/extensions/callbacks.html 都看看
'''
# 提问，循环程序停止的问题，准备一个minimal example
# 发现trainer.fit 在训练第二个增量开始，只会训练一个epoch，这猜想是和trainer保持不变，存在状态记忆有关。