-
Notifications
You must be signed in to change notification settings - Fork 1
/
runner.py
318 lines (263 loc) · 11.7 KB
/
runner.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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
import os
import copy
import time
from glob import glob
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributed as dist
from torch.utils.tensorboard import SummaryWriter
from utils import *
IAGENET_IMAGES_NUM_TEST = 50000
IAGENET_IMAGES_NUM_TRAIN = 1281166
class Runner():
def __init__(self, arg, net, optim, rank, loss, logger, scheduler=None, world_size=1):
self.arg = arg
self.save_dir = arg.save_dir
self.logger = logger
self.rank = rank
self.world_size = world_size
self.net = net
if self.arg.ema:
self.ema = copy.deepcopy(net.module).cpu()
self.ema.eval()
self.ema_has_module = hasattr(self.ema, 'module')
for p in self.ema.parameters():
p.requires_grad_(False)
self.ema_decay = arg.ema_decay
self.loss = loss
self.optim = optim
self.scheduler = scheduler
self.start_epoch = 0
self.best_metric = -1
if self.rank == 0:
self.writter = SummaryWriter()
self.load()
def save(self, epoch, filename="train"):
"""Save current epoch model
Save Elements:
model_type : arg.model
start_epoch : current epoch
network : network parameters
optimizer: optimizer parameters
best_metric : current best score
Parameters:
epoch : current epoch
filename : model save file name
"""
if self.arg.ema:
torch.save({"model_type": self.arg.model,
"start_epoch": epoch + 1,
"network": self.net.module.state_dict(),
"ema": self.ema.state_dict(),
"optimizer": self.optim.state_dict(),
"best_metric": self.best_metric,
"scheduler": self.scheduler.state_dict()
}, self.save_dir + "/%s.pth.tar" % (filename))
else:
torch.save({"model_type": self.arg.model,
"start_epoch": epoch + 1,
"network": self.net.module.state_dict(),
"optimizer": self.optim.state_dict(),
"best_metric": self.best_metric,
"scheduler": self.scheduler.state_dict()
}, self.save_dir + "/%s.pth.tar" % (filename))
print("Model saved %d epoch" % (epoch))
def load(self, filename=""):
""" Model load. same with save"""
if filename == "":
# load last epoch model
filenames = sorted(glob(self.save_dir + "/*.pth.tar"))
if len(filenames) == 0:
print("Not Load")
return
else:
filename = os.path.basename(filenames[-1])
file_path = self.save_dir + "/" + filename
if os.path.exists(file_path) is True:
print("Load %s to %s File" % (self.save_dir, filename))
dist.barrier()
map_location = {'cuda:%d' % 0: 'cuda:%d' % self.rank}
ckpoint = torch.load(file_path, map_location=map_location)
if ckpoint["model_type"] != self.arg.model:
raise ValueError("Ckpoint Model Type is %s" %
(ckpoint["model_type"]))
self.net.module.load_state_dict(ckpoint['network'])
if self.arg.ema:
self.ema.load_state_dict(ckpoint['ema'])
self.optim.load_state_dict(ckpoint['optimizer'])
self.start_epoch = ckpoint['start_epoch']
self.best_metric = ckpoint["best_metric"]
self.scheduler.load_state_dict(ckpoint['scheduler'])
print("Load Model Type : %s, epoch : %d acc : %f" %
(ckpoint["model_type"], self.start_epoch, self.best_metric))
else:
print("Load Failed, not exists file")
def update_ema(self):
needs_module = hasattr(self.net, 'module')
with torch.no_grad():
msd = self.net.state_dict()
for k, ema_v in self.ema.state_dict().items():
if needs_module:
k = 'module.' + k
model_v = msd[k].detach().cpu()
# if self.rank:
# model_v = model_v.to(device=self.rank)
ema_v.copy_(ema_v * self.ema_decay + (1. - self.ema_decay) * model_v)
def train(self, train_loader, val_loader=None):
# print("Model:\n{}".format(self.net))
train_num = IAGENET_IMAGES_NUM_TRAIN if self.arg.dali else len(train_loader.dataset)
print("\nStart Train len :", train_num)
all_iters = train_num // (self.arg.batch_size * self.world_size)
self.net.train()
if self.arg.amp:
scaler = torch.cuda.amp.GradScaler()
for epoch in range(self.start_epoch, self.arg.epoch):
self.net.train()
if train_loader.sampler:
train_loader.sampler.set_epoch(epoch)
epoch_start = time.time()
start_time = time.time()
for i, (input_, target_) in enumerate(train_loader):
target_ = target_.to(self.rank, non_blocking=True)
self.optim.zero_grad()
if self.arg.amp:
with torch.cuda.amp.autocast():
out = self.net(input_)
loss = self.loss(out, target_)
scaler.scale(loss).backward()
scaler.step(self.optim)
scaler.update()
else:
out = self.net(input_)
loss = self.loss(out, target_)
loss.backward()
self.optim.step()
if self.scheduler:
self.scheduler.step()
if self.arg.ema:
self.update_ema()
if (i % self.arg.print_freq) == 0:
duration = time.time() - start_time
if self.rank == 0:
lr = self.optim.param_groups[0]['lr']
self.logger.log_write("train", epoch=epoch, iters=str(i) + "/" + str(all_iters), loss=loss.item(), time=duration, lr=lr)
self.writter.add_scalar('train_loss', loss, epoch*train_num//self.arg.batch_size + i)
start_time = time.time()
if (val_loader is not None) and self.rank == 0:
self.valid(epoch, val_loader, self.arg.ema)
epoch_time = time.time() - epoch_start
print('epoch_time: %.4f' % (epoch_time))
self.logger.log_write("valid", epoch_time=epoch_time)
if self.rank == 0:
self.writter.close()
def _get_acc(self, loader, ema=True):
total_num = IAGENET_IMAGES_NUM_TEST if self.arg.dali else len(loader.dataset)
acc1, acc5, loss = 0, 0, 0
if not ema:
self.net.eval()
with torch.no_grad():
for input_, target_ in loader:
target_ = target_.to(self.rank, non_blocking=True)
input_ = input_.to(self.rank, non_blocking=True)
if ema:
self.ema.to(self.rank)
out = self.ema(input_)
else:
if self.arg.amp:
with torch.cuda.amp.autocast():
out = self.net(input_)
else:
out = self.net(input_)
loss = self.loss(out, target_)
# out = F.softmax(out, dim=1)
# _, idx = out.max(dim=1)
# correct += (target_ == idx).sum().item()
_acc1, _acc5 = self.accuracy(out, target_, topk=(1,5))
acc1 += _acc1[0]
acc5 += _acc5[0]
loss += loss
acc1 /= total_num
acc5 /= total_num
loss = loss.item() / total_num * self.arg.batch_size
acc1 = acc1.item()
acc5 = acc5.item()
if ema:
self.ema.to('cpu')
# return correct / IAGENET_IMAGES_NUM_TEST
return acc1, acc5, loss
def valid(self, epoch, val_loader, ema=True):
start = time.time()
acc, acc5, loss = self._get_acc(val_loader, ema)
val_time = time.time() - start
self.logger.log_write("valid", epoch=epoch, acc=acc, acc5=acc5, loss=loss, time=val_time)
self.writter.add_scalar('valid_acc_top1', acc, epoch)
self.writter.add_scalar('valid_acc_top5', acc5, epoch)
if acc > self.best_metric:
start = time.time()
self.best_metric = acc
self.save(epoch, "epoch[%05d]_acc[%.4f]_top5[%.4f]" % (
epoch, acc, acc5))
save_time = time.time() - start
print('save_time: %.4f s' % (save_time))
self.save(epoch, 'last')
def test(self, train_loader, val_loader, ema=True):
print("\n Start Test")
self.load()
#, _, _ = self._get_acc(train_loader, ema=ema)
valid_acc, acc5, _ = self._get_acc(val_loader, ema=ema)
self.logger.log_write("test", fname="test", valid_acc=valid_acc)
return acc5, valid_acc
def accuracy(self, output, target, topk=(1,)):
"""Computes the accuracy over the k top predictions for the specified values of k"""
with torch.no_grad():
maxk = max(topk)
# batch_size = target.size(0)
_, pred = output.topk(maxk, 1, True, True)
pred = pred.t()
correct = pred.eq(target.view(1, -1).expand_as(pred))
res = []
for k in topk:
correct_k = correct[:k].reshape(-1).float().sum(0, keepdim=True)
res.append(correct_k)
return res
def profiler(self, train_loader, val_loader, trainsampler=None):
train_num = IAGENET_IMAGES_NUM_TRAIN if self.arg.dali else len(train_loader.dataset)
print("\nStart Train len :", train_num)
self.net.train()
if trainsampler:
trainsampler.set_epoch(0)
if self.arg.amp:
scaler = torch.cuda.amp.GradScaler()
with torch.profiler.profile(
activities=[
torch.profiler.ProfilerActivity.CPU,
torch.profiler.ProfilerActivity.CUDA],
schedule=torch.profiler.schedule(wait=2, warmup=2, active=6, repeat=1),
on_trace_ready=torch.profiler.tensorboard_trace_handler('./profiler'),
with_stack=True
) as profiler:
for i, (input_, target_) in enumerate(train_loader):
target_ = target_.to(self.rank, non_blocking=True)
self.optim.zero_grad()
if self.arg.amp:
with torch.cuda.amp.autocast():
out = self.net(input_)
loss = self.loss(out, target_)
scaler.scale(loss).backward()
scaler.step(self.optim)
scaler.update()
else:
out = self.net(input_)
loss = self.loss(out, target_)
loss.backward()
self.optim.step()
if self.scheduler:
self.scheduler.step()
print(loss)
if self.arg.ema:
self.update_ema()
profiler.step()
if i + 1 >= 11:
break