-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
176 lines (152 loc) · 9 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
import math
import os
import matplotlib.pyplot as plt
import torch
import torch.nn as nn
from torch import cuda, optim
from torch.cuda.amp.grad_scaler import GradScaler
from torch.utils.data import BatchSampler, DataLoader, DistributedSampler
from config import get_args
from Dataset import RerankDataset_TopKSIM, collate_tuples
from network import RerankTransformer
from utils import MetricLogger, WarmupCos_Scheduler, create_optimizer, get_current_dir, init_distributed_mode, is_main_process
def main(args):
init_distributed_mode(args)
if args.device == 'cuda':
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
else:
device = torch.device('cpu')
if args.distributed:
ngpus_per_node = cuda.device_count()
args.batch_size = int(args.batch_size / ngpus_per_node)
args.num_workers = int((args.num_workers + ngpus_per_node - 1) / ngpus_per_node)
print('>> batch size per node:{}'.format(args.batch_size))
print('>> num workers per node:{}'.format(args.num_workers))
dataset_names = list(args.dataset_names)
output_dir = os.path.join(get_current_dir(), 'Experiment_{}_trainingInfo'.format(args.test_name))
if not os.path.exists(output_dir):
os.makedirs(output_dir, exist_ok=True)
train_dataset = RerankDataset_TopKSIM(names=dataset_names, mode='train', topk=args.topk, sim_len=args.sim_len)
if args.distributed:
train_sampler = DistributedSampler(train_dataset)
train_batch_sampler = BatchSampler(train_sampler, args.batch_size, drop_last=False)
train_loader = DataLoader(dataset=train_dataset, batch_sampler=train_batch_sampler, collate_fn=collate_tuples, num_workers=args.num_workers, pin_memory=True)
else:
train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True, sampler=None, drop_last=True, collate_fn=collate_tuples)
model = RerankTransformer(embed_dim=args.embed_dim,
topk_dim=args.sim_len,
depth=args.depth,
num_heads=args.num_heads,
mlp_ratio=args.mlp_ratio,
qkv_bias=args.qkv_bias,
drop_rate=args.drop_rate,
attn_drop_rate=args.attn_drop_rate,
drop_path_rate=args.drop_path_rate).to(device)
model_without_ddp = model
if args.distributed:
model = nn.SyncBatchNorm.convert_sync_batchnorm(model)
model = nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
model_without_ddp = model.module
n_parameters = sum(p.numel() for p in model_without_ddp.parameters() if p.requires_grad)
print('>> number of params:{:.2f}M'.format(n_parameters / 1e6))
# define optimizer
param_dicts = create_optimizer(args.weight_decay, model_without_ddp)
optimizer = optim.SGD(param_dicts, lr=args.base_lr * args.batch_size / 256, weight_decay=args.weight_decay, momentum=args.momentum)
start_epoch = 0
if args.resume is not None:
if os.path.isfile(args.resume):
print(">> Loading checkpoint:\n>> '{}'".format(args.resume))
checkpoint = torch.load(args.resume, map_location='cpu')
start_epoch = checkpoint['epoch']
model_without_ddp.load_state_dict(checkpoint['state_dict'], strict=False)
print(">>>> loaded checkpoint:\n>>>> '{}' (epoch {})".format(args.resume, checkpoint['epoch']))
else:
print(">> No checkpoint found at '{}'".format(args.resume))
lr_scheduler = WarmupCos_Scheduler(optimizer=optimizer,
warmup_epochs=args.warmup_epochs,
warmup_lr=args.warmup_lr * args.batch_size * args.update_every / 256,
num_epochs=args.num_epochs,
base_lr=args.base_lr * args.batch_size * args.update_every / 256,
final_lr=args.final_lr * args.batch_size * args.update_every / 256,
iter_per_epoch=int(len(train_loader) / args.update_every))
lr_scheduler.iter = max(int(len(train_loader) * start_epoch / args.update_every), 0)
# Start training
metric_logger = MetricLogger(delimiter=" ")
print_freq = 10
model_path = None
scaler = GradScaler()
Gradient_Logger = {'topk projecter': [], 'last transformer block': []}
Loss_logger = {'Contrastive loss': [], 'MSE loss': []}
for epoch in range(start_epoch, args.stop_at_epoch):
header = '>> Train Epoch: [{}]'.format(epoch)
optimizer.zero_grad()
for idx, (sim_list, targets) in enumerate(metric_logger.log_every(train_loader, print_freq, header)):
model.train()
targets = targets.to(device, non_blocking=True)
exp_sim, MSE_loss = model(sim_list.to(device, non_blocking=True).half())
Contrastive_loss = torch.mean(-1.0 * torch.log(torch.sum(exp_sim * targets, dim=-1) / torch.sum(exp_sim, dim=-1)))
loss = Contrastive_loss + args.alpha * MSE_loss
if not math.isfinite(Contrastive_loss.item()):
print(">> Contrastive loss is nan, skip to the next iteration")
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
continue
if not math.isfinite(MSE_loss.item()):
print(">> MSE loss is nan, skip to the next iteration")
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
continue
scaler.scale(loss).backward()
metric_logger.meters['Contrastive loss'].update(Contrastive_loss.item())
metric_logger.meters['MSE loss'].update(MSE_loss.item())
with torch.no_grad():
topk_project_gradient_norm = max(p.grad.detach().abs().max().cpu() for p in model_without_ddp.topk_proj.parameters())
transformer_gradient_norm = max(p.grad.detach().abs().max().cpu() for p in model_without_ddp.blocks[-1].parameters())
metric_logger.meters['topk gradient'].update(topk_project_gradient_norm.item())
metric_logger.meters['last layer gradient'].update(transformer_gradient_norm.item())
if (idx + 1) % args.update_every == 0:
if args.clip_max_norm > 0:
scaler.unscale_(optimizer)
torch.nn.utils.clip_grad_norm_(model.parameters(), args.clip_max_norm)
lr = lr_scheduler.step()
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
if (idx + 1) % 50 == 0:
if is_main_process():
Gradient_Logger['topk projecter'].append(metric_logger.meters['topk gradient'].avg)
Gradient_Logger['last transformer block'].append(metric_logger.meters['last layer gradient'].avg)
Loss_logger['Contrastive loss'].append(metric_logger.meters['Contrastive loss'].avg)
Loss_logger['MSE loss'].append(metric_logger.meters['MSE loss'].avg)
fig, axes = plt.subplots(nrows=2, ncols=1, figsize=(8, 6))
fig.tight_layout()
axes = axes.flatten()
for (key, value) in Gradient_Logger.items():
axes[0].plot(value, 'o-', label=key, linewidth=1, markersize=2)
axes[0].legend(loc='upper center', shadow=True, fontsize='medium')
axes[0].grid(b=True, which='major', color='gray', linestyle='-', alpha=0.1)
axes[0].grid(b=True, which='minor', color='gray', linestyle='-', alpha=0.1)
axes[0].set_xlabel('iter')
axes[0].set_ylabel("gradient,inf-norm")
axes[0].minorticks_on()
for (key, value) in Loss_logger.items():
axes[1].plot(value, 'o-', label=key, linewidth=1, markersize=2)
axes[1].legend(loc='upper right', shadow=True, fontsize='medium')
axes[1].grid(b=True, which='major', color='gray', linestyle='-', alpha=0.1)
axes[1].grid(b=True, which='minor', color='gray', linestyle='-', alpha=0.1)
axes[1].set_xlabel('iter')
axes[1].set_ylabel("loss")
axes[1].minorticks_on()
plt.savefig(os.path.join(output_dir, 'training_logger.png'))
plt.close()
if is_main_process():
# Save checkpoint
model_path = os.path.join(output_dir, 'checkpoint-epoch{}.pth'.format(epoch + 1))
torch.save({'epoch': epoch + 1, 'state_dict': model_without_ddp.state_dict()}, model_path)
if __name__ == "__main__":
args = get_args()
main(args=args)