/
train.py
149 lines (127 loc) · 6.56 KB
/
train.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
"""
Train I3D and save network model files(.ckpt)
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import time
import random
import numpy as np
import mindspore
from mindspore import context, Model
from mindspore.nn import SoftmaxCrossEntropyWithLogits
from mindspore.communication.management import init
from mindspore.context import ParallelMode
from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, LossMonitor, TimeMonitor, LearningRateScheduler
from mindspore.train.loss_scale_manager import DynamicLossScaleManager
import src.data_factory as data_factory
import src.model_factory as model_factory
from src.transforms.spatial_transforms import Compose, RandomHorizontalFlip, RandomCrop, CenterCrop
from src.transforms.target_transforms import ClassLabel
from src.transforms.temporal_transforms import TemporalRandomCrop
from src.utils import print_config, write_config, prepare_output_dirs, get_optimizer
from config import parse_opts
def run():
os.environ['HCCL_CONNECT_TIMEOUT'] = "6000"
tic = time.time()
config = parse_opts()
if config.dataset == 'ucf101':
config.finetune_num_classes = 101
mindspore.set_seed(2022)
random.seed(2022)
np.random.seed(2022)
mindspore.dataset.config.set_seed(2022)
if config.distributed:
config.save_dir = './output_distribute/'
if config.openI:
import moxing as mox
obs_data_url = config.data_url
config.data_url = 'cache/user-job-dir/inputs/data/'
obs_train_url = config.train_url
config.train_url = 'cache/user-job-dir/outputs/model/'
mox.file.copy_parallel(obs_data_url, config.data_url)
print("Successfully Download {} to {}".format(obs_data_url, config.data_url))
config.save_dir = config.train_url
config.video_path = os.path.join(config.data_url, config.dataset, 'jpg')
if config.mode == 'rgb':
config.checkpoint_path = os.path.join(os.path.abspath(__file__).replace('train.py', ''),
'src/pretrained/rgb_imagenet.ckpt')
if config.mode == 'flow':
config.checkpoint_path = os.path.join(os.path.abspath(__file__).replace('train.py', ''),
'src/pretrained/flow_imagenet.ckpt')
if config.dataset == 'ucf101':
config.annotation_path = os.path.join(config.data_url, config.dataset, 'annotation/ucf101_01.json')
if config.dataset == 'hmdb51':
config.annotation_path = os.path.join(config.data_url, config.dataset, 'annotation/hmdb51_1.json')
config = prepare_output_dirs(config)
print_config(config)
write_config(config, os.path.join(config.save_dir, 'config.json'))
assert config.context in ['py', 'gr']
if config.distributed:
if config.context == 'py':
context.set_context(mode=context.PYNATIVE_MODE, device_target=config.device_target,
device_id=int(os.environ["DEVICE_ID"]))
else:
context.set_context(mode=context.GRAPH_MODE, device_target=config.device_target,
device_id=int(os.environ["DEVICE_ID"]))
config.device_id = int(os.environ["DEVICE_ID"])
init()
context.set_auto_parallel_context(device_num=config.device_num, parallel_mode=ParallelMode.DATA_PARALLEL,
gradients_mean=True)
else:
if config.context == 'py':
context.set_context(mode=context.PYNATIVE_MODE, device_target=config.device_target,
device_id=config.device_id)
else:
context.set_context(mode=context.GRAPH_MODE, device_target=config.device_target,
device_id=config.device_id)
train_transforms = {'spatial': Compose([RandomCrop(config.spatial_size), RandomHorizontalFlip()]),
'temporal': TemporalRandomCrop(config.train_sample_duration),
'target': ClassLabel()}
validation_transforms = {'spatial': Compose([CenterCrop(config.spatial_size)]),
'temporal': TemporalRandomCrop(config.test_sample_duration),
'target': ClassLabel()}
model, parameters = model_factory.get_model(config)
model.set_train()
optimizer = get_optimizer(config, parameters, config.lr)
criterion = SoftmaxCrossEntropyWithLogits(sparse=True, reduction="mean")
dataset = data_factory.get_dataset(config, train_transforms, validation_transforms)
step_size = dataset['train'].get_dataset_size()
print('step size per epoch:', step_size)
lr_de_steps = step_size * config.lr_de_epochs
def learning_rate_function(lr, cur_step_num):
if not config.has_back:
if config.mode == 'flow' and cur_step_num % lr_de_steps == 0 and lr <= 1e-5 * 5:
lr = 0.001
config.has_back = True
if config.mode == 'flow' and cur_step_num % lr_de_steps == 0 and lr > 1e-5 * 5:
lr = lr * config.lr_de_rate
if config.mode == 'flow' and cur_step_num % lr_de_steps == 0 and config.has_back and lr > 1e-5 * 5:
lr = lr * config.lr_de_rate
if config.mode == 'rgb' and cur_step_num % lr_de_steps == 0 and lr > 1e-5:
lr = lr * config.lr_de_rate
return lr
if config.mode == 'rgb':
config.checkpoints_num_keep = config.checkpoints_num_keep / 2
loss_scale_manager = DynamicLossScaleManager()
time_cb = TimeMonitor(data_size=step_size)
loss_cb = LossMonitor()
lr_cb = LearningRateScheduler(learning_rate_function)
config_ck = CheckpointConfig(save_checkpoint_steps=config.checkpoint_frequency * step_size,
keep_checkpoint_max=int(config.checkpoints_num_keep))
ckpt_cb = ModelCheckpoint(prefix="i3d", directory=config.checkpoint_dir, config=config_ck)
cb = [time_cb, loss_cb, lr_cb, ckpt_cb]
model = Model(network=model, loss_fn=criterion, optimizer=optimizer, amp_level=config.amp_level,
loss_scale_manager=loss_scale_manager)
model.train(epoch=config.num_epochs, train_dataset=dataset['train'], callbacks=cb,
dataset_sink_mode=config.sink_mode)
toc = time.time()
total_time = toc - tic
print('total_time:', total_time)
print('Finished training.')
if config.openI:
mox.file.copy_parallel(config.train_url, obs_train_url)
print("Successfully Upload {} to {}".format(config.train_url, obs_train_url))
if __name__ == '__main__':
run()