/
train.py
246 lines (222 loc) · 10.7 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
import argparse
import logging
import os
import torch
import torch.distributed as dist
import torch.nn.functional as F
import torch.utils.data.distributed
from torch.nn.utils import clip_grad_norm_
import losses
from backbones import get_model
from dataset import MXFaceDataset, SyntheticDataset, DataLoaderX
from partial_fc import PartialFC
from utils.utils_amp import MaxClipGradScaler
from utils.utils_callbacks import CallBackVerification, CallBackLogging, CallBackModelCheckpoint
from utils.utils_config import get_config
from utils.utils_logging import AverageMeter, init_logging
import wandb
from memory import LatentMemory
from momentum_head import MomentumCalcHead
import random
import numpy as np
torch.manual_seed(1337)
torch.cuda.manual_seed(1337)
torch.cuda.manual_seed_all(1337)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
random.seed(1337)
np.random.seed(1337)
def _init_fn():
np.random.seed(1337)
def gather(tensor, tensor_list=None, root=0, group=None):
"""
Sends tensor to root process, which store it in tensor_list.
"""
rank = dist.get_rank()
if group is None:
group = dist.group.WORLD
if rank == root:
assert(tensor_list is not None)
dist.gather(tensor, gather_list=tensor_list, group=group)
else:
dist.gather(tensor, dst=root, group=group)
def main(args):
cfg = get_config(args.config)
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)
os.makedirs(cfg.output, exist_ok=True)
init_logging(rank, cfg.output)
if cfg.rec == "synthetic":
train_set = SyntheticDataset(local_rank=local_rank)
else:
train_set = MXFaceDataset(root_dir=cfg.rec, local_rank=local_rank, n_classes=cfg.num_classes)
train_sampler = torch.utils.data.distributed.DistributedSampler(train_set, shuffle=True)
train_loader = DataLoaderX(
local_rank=local_rank, dataset=train_set, batch_size=cfg.batch_size,
sampler=train_sampler, num_workers=2, pin_memory=True, drop_last=True)
backbone = get_model(cfg.network, dropout=0.0, fp16=cfg.fp16, num_features=cfg.embedding_size).to(local_rank)
# memory
SAMPLE_NUMS = train_set.get_sample_num_of_each_class()
state = None
if rank == 0:
if not cfg.momentum:
calc_head = LatentMemory(n_data=train_set.__len__(), feat_dim=512, cls_positive=SAMPLE_NUMS, T=cfg.T, gamma=cfg.gamma).to(local_rank)
else:
calc_head = MomentumCalcHead(feat_dim=512, cls_positive=SAMPLE_NUMS, T=cfg.T, gamma=cfg.gamma)
wandber = wandb.init(config=cfg, project="research-face", name=cfg.instance)
wandber.watch(backbone)
if cfg.resume:
try:
backbone_pth = os.path.join(cfg.output, "backbone.pth")
backbone.load_state_dict(torch.load(backbone_pth, map_location=torch.device(local_rank)))
if rank == 0:
logging.info("backbone resume successfully!")
except (FileNotFoundError, KeyError, IndexError, RuntimeError):
if rank == 0:
logging.info("resume fail, backbone init successfully!")
if cfg.momentum:
logging.info("Initializing momentum encoder!")
momentum_encoder = get_model(cfg.network, dropout=0.0, fp16=cfg.fp16, num_features=cfg.embedding_size).to(local_rank)
for param_q, param_k in zip(backbone.parameters(), momentum_encoder.parameters()):
param_k.data.copy_(param_q.data) # initialize
param_k.requires_grad = False # not update by gradient
momentum_encoder.eval()
backbone = torch.nn.parallel.DistributedDataParallel(
module=backbone, broadcast_buffers=False, device_ids=[local_rank])
backbone.train()
margin_softmax = losses.get_loss(cfg.loss, cfg)
module_partial_fc = PartialFC(
rank=rank, local_rank=local_rank, world_size=world_size, resume=cfg.resume,
batch_size=cfg.batch_size, margin_softmax=margin_softmax, num_classes=cfg.num_classes,
sample_rate=cfg.sample_rate, embedding_size=cfg.embedding_size, prefix=cfg.output)
opt_backbone = torch.optim.SGD(
params=[{'params': backbone.parameters()}],
lr=cfg.lr / 512 * cfg.batch_size * world_size,
momentum=0.9, weight_decay=cfg.weight_decay)
opt_pfc = torch.optim.SGD(
params=[{'params': module_partial_fc.parameters()}],
lr=cfg.lr / 512 * cfg.batch_size * world_size,
momentum=0.9, weight_decay=cfg.weight_decay)
num_image = len(train_set)
total_batch_size = cfg.batch_size * world_size
cfg.warmup_step = num_image // total_batch_size * cfg.warmup_epoch
cfg.total_step = num_image // total_batch_size * cfg.num_epoch
def lr_step_func(current_step):
cfg.decay_step = [x * num_image // total_batch_size for x in cfg.decay_epoch]
if current_step < cfg.warmup_step:
return current_step / cfg.warmup_step
else:
return 0.1 ** len([m for m in cfg.decay_step if m <= current_step])
scheduler_backbone = torch.optim.lr_scheduler.LambdaLR(
optimizer=opt_backbone, lr_lambda=lr_step_func)
scheduler_pfc = torch.optim.lr_scheduler.LambdaLR(
optimizer=opt_pfc, lr_lambda=lr_step_func)
for key, value in cfg.items():
num_space = 25 - len(key)
logging.info(": " + key + " " * num_space + str(value))
print(len(train_loader))
val_target = cfg.val_targets
callback_verification = CallBackVerification(2000, rank, val_target, cfg.rec)
callback_logging = CallBackLogging(50, rank, cfg.total_step, cfg.batch_size, world_size, None)
callback_checkpoint = CallBackModelCheckpoint(rank, cfg.output)
loss = AverageMeter()
start_epoch = 0
global_step = 0
weights = torch.ones(cfg.num_classes).to(local_rank)
grad_amp = MaxClipGradScaler(cfg.batch_size, 128 * cfg.batch_size, growth_interval=100) if cfg.fp16 else None
best_results = {v:0 for v in val_target}
if rank == 0:
for v in val_target:
if not os.path.exists(cfg.output + '/' + v):
os.mkdir(cfg.output + '/' + v)
for epoch in range(start_epoch, cfg.num_epoch):
train_sampler.set_epoch(epoch)
for step, (img, label, indexes) in enumerate(train_loader):
global_step += 1
features = F.normalize(backbone(img))
x_grad, loss_v = module_partial_fc.forward_backward(label, features, opt_pfc, weights)
if cfg.fp16:
features.backward(grad_amp.scale(x_grad))
grad_amp.unscale_(opt_backbone)
clip_grad_norm_(backbone.parameters(), max_norm=5, norm_type=2)
grad_amp.step(opt_backbone)
grad_amp.update()
else:
features.backward(x_grad)
clip_grad_norm_(backbone.parameters(), max_norm=5, norm_type=2)
opt_backbone.step()
opt_pfc.step()
module_partial_fc.update()
opt_backbone.zero_grad()
opt_pfc.zero_grad()
loss.update(loss_v, 1)
callback_logging(global_step, loss, epoch, cfg.fp16, scheduler_backbone.get_last_lr()[0], grad_amp)
scheduler_backbone.step()
scheduler_pfc.step()
# update memory
if cfg.momentum:
with torch.no_grad():
for param_q, param_k in zip(backbone.parameters(), momentum_encoder.parameters()):
param_k.data = param_k.data * 0.9 + param_q.data * (1. - 0.9)
features = F.normalize(momentum_encoder(img))
total_features = torch.zeros(
size=[cfg.batch_size * world_size, cfg.embedding_size], device=local_rank)
dist.all_gather(list(total_features.chunk(world_size, dim=0)), features.data)
total_label = torch.zeros(
size=[cfg.batch_size * world_size], device=local_rank, dtype=torch.long)
dist.all_gather(list(total_label.chunk(world_size, dim=0)), label)
total_indexes = torch.zeros(
size=[cfg.batch_size * world_size], device=local_rank, dtype=torch.long)
dist.all_gather(list(total_indexes.chunk(world_size, dim=0)), indexes)
total_features.requires_grad = False
if rank == 0:
calc_head(total_features, total_label, total_indexes)
report = callback_verification(global_step, backbone)
to_save = []
if report is not None:
for dataset, res in report.items():
if best_results[dataset] < res:
best_results[dataset] = res
to_save.append(dataset)
if to_save.__len__() > 0:
callback_checkpoint(global_step, backbone, module_partial_fc, to_save, None)
report = callback_verification(2000, backbone)
to_save = []
if report is not None:
for dataset, res in report.items():
if best_results[dataset] < res:
best_results[dataset] = res
to_save.append(dataset)
state = {}
if rank == 0:
kappas = calc_head.update_kappa()
weights = calc_head.update_weights()
state = {'kappas': kappas, 'samples': SAMPLE_NUMS, 'weights': weights}
kappa_report = {'max_kappa': kappas.max().item(), 'min_kappa': kappas.min().item(), 'mean_kappa': kappas.mean().item()}
report = dict({'epoch': epoch, 'train_loss': loss.avg}, **report)
report = dict(report, **kappa_report)
wandber.log(report)
else:
weights = torch.ones(cfg.num_classes).to(local_rank) * -1.
#dist.broadcast(weights, src=0)
total_weights = torch.zeros(
size=[cfg.num_classes * world_size], device=local_rank)
dist.all_gather(list(total_weights.chunk(world_size, dim=0)), weights)
weights = total_weights[total_weights != -1]
assert weights.__len__() == cfg.num_classes
callback_checkpoint(global_step, backbone, module_partial_fc, to_save, state)
dist.destroy_process_group()
if __name__ == "__main__":
torch.backends.cudnn.benchmark = True
parser = argparse.ArgumentParser(description='PyTorch KappaFace Training')
parser.add_argument('config', type=str, help='py config file')
parser.add_argument('--local_rank', type=int, default=0, help='local_rank')
main(parser.parse_args())