-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
254 lines (217 loc) · 12.6 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
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
import torch
from torch.utils.data import DataLoader
from torchinfo import summary
from torchvision import transforms
from tqdm import tqdm
from itertools import product
from pydantic import create_model
from json import dumps
from traceback import format_exc
from lib.writer import summary_writer, profile, mesh_dict, d42rgb
from lib.dataloader import ShapenetDataset
from lib.logger import logger, console_logger, file_logger
from lib.model import Pixel2Point
from lib.settings import Settings
from lib.utils import env_init, dataloader_init, show_result, show_3d
from lib.loss import ChamferDistance, EmdModule
from lib.notification import send_telegram
class MyProcess():
def __init__(self):
self.settings = Settings()
self.device = self.settings.device[0]
logger.debug(f"Using {self.device} device")
self.parameters = self.settings.dict(exclude={
'snapshot_path', 'output_path', 'model_path', 'current_time',
'train_dataset_path', 'val_dataset_path', 'test_dataset_path',
'telegram_token', 'telegram_chat_id', 'discord_webhook_url'
})
def train_loop(self):
self.loss_train = 0
self.pixel2point.train(mode=True)
self.loss_function.train_param(mode=True)
self.prof.start()
train_bar = tqdm(self.loader_train, unit='batch', leave=True, colour='#B8DA7E')
for i_batch, (self.pred, self.gt, index) in enumerate(train_bar):
self.global_step = self.i_epoch + 1
self.pred = self.pred.to(self.device)
self.gt = self.gt.to(self.device)
with torch.autocast(device_type='cuda', dtype=torch.float16, enabled=self.hparam.use_amp):
self.output = self.pixel2point.forward(self.pred)
self.output = self.output.type_as(self.gt).view(self.gt.shape[0], -1, 3)
loss, _ = self.loss_function.forward(self.output, self.gt)
self.optimizer.zero_grad()
# torch.autograd.backward(loss)
self.scaler.scale(loss).backward()
# self.optimizer.step()
self.scaler.step(self.optimizer)
self.scaler.update()
self.prof.step()
self.loss_train += loss.item()
train_bar.set_description(f'Epoch [{self.i_epoch + 1}/{self.hparam.epoch}]')
train_bar.set_postfix(loss=loss.item())
if i_batch + 1 == len(self.loader_train) - 1:
self.save_result(f'Training_{i_batch}', sample=100)
watch_index = 78
show_result(self.pred[watch_index], self.output[watch_index], self.gt[watch_index],
self.plotly_path, f'{self.global_step}_{watch_index}_train')
self.prof.stop()
def validation_loop(self):
self.loss_val = 0
self.pixel2point.train(mode=False)
self.loss_function.train_param(mode=False)
with torch.no_grad():
val_bar = tqdm(self.loader_validation, unit='batch', leave=True, colour='#7EA9DA')
for i_batch, (self.pred, self.gt, index) in enumerate(val_bar):
self.pred = self.pred.to(self.device)
self.gt = self.gt.to(self.device)
self.output = self.pixel2point.forward(self.pred)
self.output = self.output.type_as(self.gt).view(self.gt.shape[0], -1, 3)
loss, _ = self.loss_function.forward(self.output, self.gt)
self.loss_val += loss.item()
val_bar.set_description(f'Validating')
val_bar.set_postfix(loss=loss.item())
if i_batch + 1 == len(self.loader_validation) - 1:
self.save_result(f'validation_{i_batch}', sample=100)
watch_index = 3
show_result(self.pred[watch_index], self.output[watch_index], self.gt[watch_index],
self.plotly_path, f'{self.global_step}_{watch_index}_validation')
def transform_config(self):
preprocess = []
if 'grayscale' in self.hparam.preprocess:
preprocess += [transforms.Grayscale(1)]
if 'resize' in self.hparam.preprocess:
preprocess += [transforms.Resize(self.hparam.resize)]
if 'totensor' in self.hparam.preprocess:
preprocess += [transforms.ToTensor()]
return transforms.Compose(preprocess)
def loss_config(self):
if self.hparam.loss_function == 'CD':
return ChamferDistance()
elif self.hparam.loss_function == 'EMD':
return EmdModule()
def shapenet_config(self, dataset_path):
return ShapenetDataset(
dataset_path=dataset_path, snapshot_path=self.settings.snapshot_path,
transforms=self.preprocess, only=self.hparam.only,
mode=self.hparam.mode, remake=self.hparam.dataset_remake
)
def loader_config(self, datset):
return DataLoader(
dataset=datset, batch_size=self.hparam.batch_size, shuffle=self.hparam.shuffle,
num_workers=self.hparam.num_workers, pin_memory=self.hparam.pin_memory,
worker_init_fn=self.worker_init_fn, generator=self.generator
)
def save_model(self, key, data):
model_path = self.settings.output_path.joinpath('model')
model_path.mkdir(parents=True, exist_ok=True)
torch.save({
'epoch': self.hparam.epoch,
'model_state_dict': self.pixel2point.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
'scaler_state_dict': self.scaler.state_dict(),
'criterion': self.loss_function,
}, model_path.joinpath(f"{key}_param.pt"))
with open(model_path.joinpath('hparam.txt'), 'a') as f:
f.write(f'{key}, {data}\n')
def save_weight(self):
self.writer.add_histogram('Layer1_Conv/weight', self.pixel2point.layer1[0].weight, global_step=self.global_step)
self.writer.add_histogram('Layer1_Conv/bais', self.pixel2point.layer1[0].bias, global_step=self.global_step)
self.writer.add_histogram('Layer2_Conv/weight', self.pixel2point.layer2[0].weight, global_step=self.global_step)
self.writer.add_histogram('Layer2_Conv/bais', self.pixel2point.layer2[0].bias, global_step=self.global_step)
self.writer.add_histogram('Layer3_Conv/weight', self.pixel2point.layer3[0].weight, global_step=self.global_step)
self.writer.add_histogram('Layer3_Conv/bais', self.pixel2point.layer3[0].bias, global_step=self.global_step)
self.writer.add_histogram('Layer4_Conv/weight', self.pixel2point.layer4[0].weight, global_step=self.global_step)
self.writer.add_histogram('Layer4_Conv/bais', self.pixel2point.layer4[0].bias, global_step=self.global_step)
self.writer.add_histogram('Layer5_Conv/weight', self.pixel2point.layer5[0].weight, global_step=self.global_step)
self.writer.add_histogram('Layer5_Conv/bais', self.pixel2point.layer5[0].bias, global_step=self.global_step)
self.writer.add_histogram('Layer6_Conv/weight', self.pixel2point.layer6[0].weight, global_step=self.global_step)
self.writer.add_histogram('Layer6_Conv/bais', self.pixel2point.layer6[0].bias, global_step=self.global_step)
self.writer.add_histogram('Layer7_Conv/weight', self.pixel2point.layer7[0].weight, global_step=self.global_step)
self.writer.add_histogram('Layer7_Conv/bais', self.pixel2point.layer7[0].bias, global_step=self.global_step)
self.writer.add_histogram('fc1/weight', self.pixel2point.fc1[0].weight, global_step=self.global_step)
self.writer.add_histogram('fc1/bais', self.pixel2point.fc1[0].bias, global_step=self.global_step)
self.writer.add_histogram('fc2/weight', self.pixel2point.fc2[0].weight, global_step=self.global_step)
self.writer.add_histogram('fc2/bais', self.pixel2point.fc2[0].bias, global_step=self.global_step)
self.writer.add_histogram('fc3/weight', self.pixel2point.fc3[0].weight, global_step=self.global_step)
self.writer.add_histogram('fc3/bais', self.pixel2point.fc3[0].bias, global_step=self.global_step)
self.writer.add_histogram('fc4/weight', self.pixel2point.fc4.weight, global_step=self.global_step)
self.writer.add_histogram('fc4/bais', self.pixel2point.fc4.bias, global_step=self.global_step)
def save_hparam(self):
self.writer.add_hparams(
{
'use_amp': self.hparam.use_amp,
'reproducibility': self.hparam.reproducibility,
'mode': self.hparam.mode,
'dataset_remake': self.hparam.dataset_remake,
'batch_size': self.hparam.batch_size,
'shuffle': self.hparam.shuffle,
'num_workers': self.hparam.num_workers,
'pin_memory': self.hparam.pin_memory,
'initial_point': self.hparam.initial_point,
'loss_function': self.hparam.loss_function,
'epoch': self.hparam.epoch,
'learning_rate': self.hparam.learning_rate,
},
{
'Loss/train': self.loss_train / len(self.loader_train),
'Loss/validation': self.loss_val / len(self.loader_validation)
},
'./',
global_step=self.global_step
)
def save_mesh(self, tag, coordinate, global_step):
self.writer.add_mesh(tag, coordinate, config_dict=mesh_dict(coordinate), global_step=global_step)
def save_result(self, data_type, sample=None):
self.writer.add_images(f'Input/{data_type}', d42rgb(self.pred[:sample]), self.global_step)
self.save_mesh(f'Output_{data_type}', self.output[:sample, :, :3], self.global_step)
self.save_mesh(f'GT_{data_type}', self.gt[:sample, :, :3], self.global_step)
def train_validation(self):
for key, data in enumerate(product(*[v for v in self.parameters.values()])):
self.writer = summary_writer(logdir=f'./runs/{self.settings.current_time}/{key}_tb')
self.prof = profile(dir_name=self.writer.logdir)
hparam_dict = dict(zip(list(self.parameters.keys()), data))
self.hparam = create_model('HyperParameter', **hparam_dict)()
logger.debug(f"Hyper Parameter: {dumps(hparam_dict, indent=2)}")
self.writer.add_text('Hyper Parameter', text_string=f"<pre>{dumps(hparam_dict, indent=2)}", global_step=0)
env_init(self.hparam.reproducibility, self.hparam.seed)
self.worker_init_fn, self.generator = dataloader_init(self.hparam.loader_reproducibility, self.hparam.seed)
self.preprocess = self.transform_config()
self.dataset_train = self.shapenet_config(self.settings.train_dataset_path)
self.dataset_validation = self.shapenet_config(self.settings.val_dataset_path)
self.loader_train = self.loader_config(self.dataset_train)
self.loader_validation = self.loader_config(self.dataset_validation)
self.pixel2point = Pixel2Point(initial_point=self.hparam.initial_point).to(self.device)
input_size = [self.hparam.batch_size, 1] + self.hparam.resize
logger.debug(
summary(self.pixel2point, input_size, col_names=["input_size", "output_size", "num_params"], verbose=0)
)
self.writer.add_graph(self.pixel2point, torch.rand(input_size).to(self.device))
self.save_mesh('Initial_Point', self.pixel2point.initial_point.unsqueeze(0), global_step=0)
self.loss_function = self.loss_config()
self.optimizer = torch.optim.Adam(self.pixel2point.parameters(), lr=self.hparam.learning_rate)
self.scaler = torch.cuda.amp.GradScaler(enabled=self.hparam.use_amp)
self.plotly_path = self.settings.output_path.joinpath(f'{key}_plotly')
self.plotly_path.mkdir(parents=True, exist_ok=True)
show_3d(self.pixel2point.initial_point, path=self.plotly_path.joinpath('initial_point.html'))
for self.i_epoch in range(self.hparam.epoch):
self.train_loop()
self.validation_loop()
self.save_weight()
self.save_hparam()
if self.hparam.save_model is True:
self.save_model(key, hparam_dict)
self.writer.close()
if __name__ == '__main__':
console_logger()
file_logger()
my_process = MyProcess()
message = None
try:
my_process.train_validation()
except Exception as e:
message = f'🔴例外訊息：{e}'
logger.debug(format_exc())
else:
message = f'🟢訓練結束'
finally:
send_telegram(my_process.settings.telegram_token, my_process.settings.telegram_chat_id, message)