forked from deepinsight/insightface
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
303 lines (242 loc) · 10.6 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
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
import argparse
import logging
import os
import time
import timm
import glob
import numpy as np
import os.path as osp
import torch
import torch.distributed as dist
from torch import nn
import torch.nn.functional as F
import torch.utils.data.distributed
from torch.nn.utils import clip_grad_norm_
from dataset import FaceDataset, DataLoaderX, MXFaceDataset, get_tris
from backbones import get_network
from utils.utils_amp import MaxClipGradScaler
from utils.utils_callbacks import CallBackVerification, CallBackLogging, CallBackModelCheckpoint
from utils.utils_logging import AverageMeter, init_logging
from utils.utils_config import get_config
from lr_scheduler import get_scheduler
from timm.optim.optim_factory import create_optimizer
def main(args):
cfg = get_config(args.config)
if not cfg.tf32:
torch.backends.cuda.matmul.allow_tf32 = False
torch.backends.cudnn.allow_tf32 = False
else:
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
try:
world_size = int(os.environ['WORLD_SIZE'])
rank = int(os.environ['RANK'])
dist.init_process_group('nccl')
except KeyError:
world_size = 1
rank = 0
dist.init_process_group(backend='nccl', init_method="tcp://127.0.0.1:12584", rank=rank, world_size=world_size)
local_rank = args.local_rank
torch.cuda.set_device(local_rank)
if not os.path.exists(cfg.output) and rank is 0:
os.makedirs(cfg.output)
else:
time.sleep(2)
log_root = logging.getLogger()
init_logging(log_root, rank, cfg.output)
if rank==0:
logging.info(args)
logging.info(cfg)
print(cfg.flipindex.shape, cfg.flipindex[400:410])
train_set = MXFaceDataset(cfg=cfg, is_train=True, local_rank=local_rank)
cfg.num_images = len(train_set)
cfg.world_size = world_size
total_batch_size = cfg.batch_size * cfg.world_size
epoch_steps = cfg.num_images // total_batch_size
cfg.warmup_steps = epoch_steps * cfg.warmup_epochs
if cfg.max_warmup_steps>0:
cfg.warmup_steps = min(cfg.max_warmup_steps, cfg.warmup_steps)
cfg.total_steps = epoch_steps * cfg.num_epochs
if cfg.lr_epochs is not None:
cfg.lr_steps = [m*epoch_steps for m in cfg.lr_epochs]
else:
cfg.lr_steps = None
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_set, shuffle=True)
train_loader = torch.utils.data.DataLoader(
dataset=train_set, batch_size=cfg.batch_size,
sampler=train_sampler, num_workers=4, pin_memory=False, drop_last=True)
net = get_network(cfg).to(local_rank)
if cfg.resume:
try:
ckpts = list(glob.glob(osp.join(cfg.resume_path, "backbone*.pth")))
backbone_pth = sorted(ckpts)[-1]
backbone_ckpt = torch.load(backbone_pth, map_location=torch.device(local_rank))
net.load_state_dict(backbone_ckpt['model'])
if rank==0:
logging.info("backbone resume successfully! %s"%backbone_pth)
except (FileNotFoundError, KeyError, IndexError, RuntimeError):
logging.info("resume fail!!")
raise RuntimeError
net = torch.nn.parallel.DistributedDataParallel(
module=net, broadcast_buffers=False, device_ids=[local_rank])
net.train()
if cfg.opt=='sgd':
opt = torch.optim.SGD(
params=[
{"params": net.parameters()},
],
lr=cfg.lr, momentum=0.9, weight_decay=cfg.weight_decay)
elif cfg.opt=='adam':
opt = torch.optim.Adam(
params=[
{"params": net.parameters()},
],
lr=cfg.lr)
elif cfg.opt=='adamw':
opt = torch.optim.AdamW(
params=[
{"params": net.parameters()},
],
lr=cfg.lr, weight_decay=cfg.weight_decay)
scheduler = get_scheduler(opt, cfg)
if cfg.resume:
if rank==0:
logging.info(opt)
if cfg.resume:
for g in opt_pfc.param_groups:
for key in ['lr', 'initial_lr']:
g[key] = cfg.lr
start_epoch = 0
total_step = cfg.total_steps
if rank==0:
logging.info(opt)
logging.info("Total Step is: %d" % total_step)
loss = {
'Loss': AverageMeter(),
}
global_step = 0
grad_amp = None
if cfg.fp16>0:
if cfg.fp16==1:
grad_amp = torch.cuda.amp.grad_scaler.GradScaler(growth_interval=100)
elif cfg.fp16==2:
grad_amp = MaxClipGradScaler(64, 1024, growth_interval=200)
elif cfg.fp16==3:
grad_amp = MaxClipGradScaler(4, 8, growth_interval=200)
else:
assert 'fp16 mode not set'
callback_checkpoint = CallBackModelCheckpoint(rank, cfg)
callback_checkpoint(global_step, net, opt)
callback_logging = CallBackLogging(50, rank, total_step, cfg.batch_size, world_size, None)
l1loss = nn.L1Loss()
tris = get_tris(cfg)
tri_index = torch.tensor(tris, dtype=torch.long).to(local_rank)
use_eyes = cfg.eyes is not None
for epoch in range(start_epoch, cfg.num_epochs):
train_sampler.set_epoch(epoch)
for step, value in enumerate(train_loader):
global_step += 1
img = value['img_local'].to(local_rank)
dloss = {}
assert cfg.task==0
label_verts = value['verts'].to(local_rank)
label_points2d = value['points2d'].to(local_rank)
#need_eyes = 'eye_verts' in value
preds = net(img)
if use_eyes:
pred_verts, pred_points2d, pred_eye_verts, pred_eye_points = preds.split([1220*3, 1220*2, 481*2*3, 481*2*2], dim=1)
pred_eye_verts = pred_eye_verts.view(cfg.batch_size, 481*2, 3)
pred_eye_points = pred_eye_points.view(cfg.batch_size, 481*2, 2)
else:
pred_verts, pred_points2d = preds.split([1220*3, 1220*2], dim=1)
pred_verts = pred_verts.view(cfg.batch_size, 1220, 3)
pred_points2d = pred_points2d.view(cfg.batch_size, 1220, 2)
if not cfg.use_rtloss:
loss1 = F.l1_loss(pred_verts, label_verts)
else:
label_Rt = value['rt'].to(local_rank)
_ones = torch.ones([pred_verts.shape[0], 1220, 1], device=pred_verts.device)
pred_verts = torch.cat([pred_verts/10, _ones], dim=2)
pred_verts = torch.bmm(pred_verts,label_Rt) * 10.0
label_verts = torch.cat([label_verts/10, _ones], dim=2)
label_verts = torch.bmm(label_verts,label_Rt) * 10.0
loss1 = F.l1_loss(pred_verts, label_verts)
loss2 = F.l1_loss(pred_points2d, label_points2d)
loss3d = loss1 * cfg.lossw_verts3d
loss2d = loss2 * cfg.lossw_verts2d
dloss['Loss'] = loss3d + loss2d
dloss['Loss3D'] = loss3d
dloss['Loss2D'] = loss2d
if use_eyes:
label_eye_verts = value['eye_verts'].to(local_rank)
label_eye_points = value['eye_points'].to(local_rank)
loss3 = F.l1_loss(pred_eye_verts, label_eye_verts)
loss4 = F.l1_loss(pred_eye_points, label_eye_points)
loss3 = loss3 * cfg.lossw_eyes3d
loss4 = loss4 * cfg.lossw_eyes2d
dloss['Loss'] += loss3
dloss['Loss'] += loss4
dloss['LossEye3d'] = loss3
dloss['LossEye2d'] = loss4
if cfg.loss_bone3d:
bone_losses = []
for i in range(3):
pred_verts_x = pred_verts[:,tri_index[:,i%3],:]
pred_verts_y = pred_verts[:,tri_index[:,(i+1)%3],:]
label_verts_x = label_verts[:,tri_index[:,i%3],:]
label_verts_y = label_verts[:,tri_index[:,(i+1)%3],:]
dist_pred = torch.norm(pred_verts_x - pred_verts_y, p=2, dim=-1, keepdim=False)
dist_label = torch.norm(label_verts_x - label_verts_y, p=2, dim=-1, keepdim=False)
bone_losses.append(F.l1_loss(dist_pred, dist_label) * cfg.lossw_bone3d)
_loss = sum(bone_losses)
dloss['Loss'] += _loss
dloss['LossBone3d'] = _loss
if cfg.loss_bone2d:
bone_losses = []
for i in range(3):
pred_points2d_x = pred_points2d[:,tri_index[:,i%3],:]
pred_points2d_y = pred_points2d[:,tri_index[:,(i+1)%3],:]
label_points2d_x = label_points2d[:,tri_index[:,i%3],:]
label_points2d_y = label_points2d[:,tri_index[:,(i+1)%3],:]
dist_pred = torch.norm(pred_points2d_x - pred_points2d_y, p=2, dim=-1, keepdim=False)
dist_label = torch.norm(label_points2d_x - label_points2d_y, p=2, dim=-1, keepdim=False)
bone_losses.append(F.l1_loss(dist_pred, dist_label) * cfg.lossw_bone2d)
_loss = sum(bone_losses)
dloss['Loss'] += _loss
dloss['LossBone2d'] = _loss
iter_loss = dloss['Loss']
if cfg.fp16>0:
grad_amp.scale(iter_loss).backward()
grad_amp.unscale_(opt)
if cfg.fp16<2:
torch.nn.utils.clip_grad_norm_(net.parameters(), 5)
grad_amp.step(opt)
grad_amp.update()
else:
iter_loss.backward()
opt.step()
opt.zero_grad()
if cfg.lr_func is None:
scheduler.step()
with torch.no_grad():
loss['Loss'].update(iter_loss.item(), 1)
for k in dloss:
if k=='Loss':
continue
v = dloss[k].item()
if k not in loss:
loss[k] = AverageMeter()
loss[k].update(v, 1)
callback_logging(global_step, loss, epoch, cfg.fp16, grad_amp, opt)
if cfg.lr_func is not None:
scheduler.step()
callback_checkpoint(9999, net, opt)
dist.destroy_process_group()
if __name__ == "__main__":
torch.backends.cudnn.benchmark = True
parser = argparse.ArgumentParser(description='JMLR Training')
parser.add_argument('config', type=str, help='config file')
parser.add_argument('--local_rank', type=int, default=0, help='local_rank')
args_ = parser.parse_args()
main(args_)