forked from facebookresearch/deit
-
Notifications
You must be signed in to change notification settings - Fork 0
/
engine.py
115 lines (89 loc) · 4.36 KB
/
engine.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
# Copyright (c) 2015-present, Facebook, Inc.
# All rights reserved.
"""
Train and eval functions used in main.py
"""
import math
import logging
import sys
from typing import Iterable, Optional
from timm.data import Mixup
from timm.utils import ModelEma, accuracy
import torch
from losses import DistillationLoss
import utils
logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
def train_one_epoch(model: torch.nn.Module, criterion: DistillationLoss,
data_loader: Iterable, optimizer: torch.optim.Optimizer,
device: torch.device, epoch: int, loss_scaler, max_norm: float = 0,
model_ema: Optional[ModelEma] = None, mixup_fn: Optional[Mixup] = None,
set_training_mode=True, args = None):
model.train(set_training_mode)
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
header = 'Epoch: [{}]'.format(epoch)
print_freq = 50
logging.info(f'Epoch {epoch} starts training!')
if args.cosub:
criterion = torch.nn.BCEWithLogitsLoss()
for samples, targets in metric_logger.log_every(data_loader, print_freq, header):
samples = samples.to(device, non_blocking=True)
targets = targets.to(device, non_blocking=True)
if mixup_fn is not None:
samples, targets = mixup_fn(samples, targets)
if args.cosub:
samples = torch.cat((samples,samples),dim=0)
if args.bce_loss:
targets = targets.gt(0.0).type(targets.dtype)
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
outputs = model(samples)
if not args.cosub:
loss = criterion(samples, outputs, targets)
else:
outputs = torch.split(outputs, outputs.shape[0]//2, dim=0)
loss = 0.25 * criterion(outputs[0], targets)
loss = loss + 0.25 * criterion(outputs[1], targets)
loss = loss + 0.25 * criterion(outputs[0], outputs[1].detach().sigmoid())
loss = loss + 0.25 * criterion(outputs[1], outputs[0].detach().sigmoid())
loss_value = loss.item()
if not math.isfinite(loss_value):
logging.error("Loss is {}, stopping training".format(loss_value))
sys.exit(1)
optimizer.zero_grad()
# this attribute is added by timm on one optimizer (adahessian)
is_second_order = hasattr(optimizer, 'is_second_order') and optimizer.is_second_order
loss_scaler(loss, optimizer, clip_grad=max_norm,
parameters=model.parameters(), create_graph=is_second_order)
torch.cuda.synchronize()
if model_ema is not None:
model_ema.update(model)
metric_logger.update(loss=loss_value)
metric_logger.update(lr=optimizer.param_groups[0]["lr"])
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
@torch.no_grad()
def evaluate(data_loader, model, device):
criterion = torch.nn.CrossEntropyLoss()
metric_logger = utils.MetricLogger(delimiter=" ")
header = 'Test:'
# switch to evaluation mode
model.eval()
for images, target in metric_logger.log_every(data_loader, 10, header):
images = images.to(device, non_blocking=True)
target = target.to(device, non_blocking=True)
# compute output
with torch.cuda.amp.autocast(dtype=torch.bfloat16):
output = model(images)
loss = criterion(output, target)
acc1, acc5 = accuracy(output, target, topk=(1, 5))
batch_size = images.shape[0]
metric_logger.update(loss=loss.item())
metric_logger.meters['acc1'].update(acc1.item(), n=batch_size)
metric_logger.meters['acc5'].update(acc5.item(), n=batch_size)
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print('* Acc@1 {top1.global_avg:.3f} Acc@5 {top5.global_avg:.3f} loss {losses.global_avg:.3f}'
.format(top1=metric_logger.acc1, top5=metric_logger.acc5, losses=metric_logger.loss))
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}