-
Notifications
You must be signed in to change notification settings - Fork 0
/
engine.py
executable file
·173 lines (143 loc) · 6.78 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
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
"""
Train and eval functions used in main.py
"""
import math
import sys
from typing import Iterable, Optional
import torch
from timm.data import Mixup
from timm.utils import accuracy, ModelEma
import utils
import h5py
from tqdm import tqdm
def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module,
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):
# TODO fix this for finetuning
model.train(set_training_mode)
criterion.train()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
header = 'Epoch: [{}]'.format(epoch)
print_freq = 10
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)
with torch.cuda.amp.autocast():
outputs = model(samples)
if isinstance(outputs, list):
loss_list = [criterion(o, targets) / len(outputs) for o in outputs]
loss = sum(loss_list)
else:
loss = criterion(outputs, targets)
loss_value = loss.item()
print("learning_rate :", optimizer.param_groups[0]["lr"])
print("loss_value :", loss_value)
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
sys.exit(1)
print("1")
optimizer.zero_grad()
print("2")
# 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)
if isinstance(outputs, list):
metric_logger.update(loss_0=loss_list[0].item())
metric_logger.update(loss_1=loss_list[1].item())
else:
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():
output = model(images)
# Conformer
if isinstance(output, list):
loss_list = [criterion(o, target) / len(output) for o in output]
loss = sum(loss_list)
# others
else:
loss = criterion(output, target)
if isinstance(output, list):
# Conformer
acc1_head1 = accuracy(output[0], target, topk=(1,))[0]
acc1_head2 = accuracy(output[1], target, topk=(1,))[0]
acc1_total = accuracy(output[0] + output[1], target, topk=(1,))[0]
else:
# others
acc1, acc5 = accuracy(output, target, topk=(1, 5))
batch_size = images.shape[0]
if isinstance(output, list):
metric_logger.update(loss=loss.item())
metric_logger.update(loss_0=loss_list[0].item())
metric_logger.update(loss_1=loss_list[1].item())
metric_logger.meters['acc1'].update(acc1_total.item(), n=batch_size)
metric_logger.meters['acc1_head1'].update(acc1_head1.item(), n=batch_size)
metric_logger.meters['acc1_head2'].update(acc1_head2.item(), n=batch_size)
else:
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)
if isinstance(output, list):
print('* Acc@heads_top1 {heads_top1.global_avg:.3f} Acc@head_1 {head1_top1.global_avg:.3f} Acc@head_2 {head2_top1.global_avg:.3f} '
'loss@total {losses.global_avg:.3f} loss@1 {loss_0.global_avg:.3f} loss@2 {loss_1.global_avg:.3f} '
.format(heads_top1=metric_logger.acc1, head1_top1=metric_logger.acc1_head1, head2_top1=metric_logger.acc1_head2,
losses=metric_logger.loss, loss_0=metric_logger.loss_0, loss_1=metric_logger.loss_1))
else:
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()}
@torch.no_grad()
def evaluate_for_extract(data_loader, model, device):
criterion = torch.nn.CrossEntropyLoss()
header = 'Test:'
# switch to evaluation mode
model.eval()
# ft = h5py.File('/data2/CONFORMER_COCO2014_t.hdf5','w')
#f = h5py.File('CONFORMER_COCO2014_total.hdf5', 'w')
f = h5py.File('CONFORMER_clevr_change.hdf5', 'w')
for images, label, name in tqdm(data_loader):
images = images.to(device, non_blocking=True)
# compute output
with torch.cuda.amp.autocast():
output, output_t = model(images, True)
# for i, (val, val_t) in enumerate(zip(output, output_t)):
# val = val.squeeze()
# val_t = val_t.squeeze()
# feat = val.cpu().detach().numpy()
# feat_t = val_t.cpu().detach().numpy()
# f.create_dataset(name[0] + "_feat" + str(i), data=feat)
# f.create_dataset(name[0] + "_feat_t" + str(i), data=feat_t)
#print(len(name))
output_t = output_t.squeeze()
output = output.squeeze()
feat = output.cpu().detach().numpy() # [1, 1536, 7, 7]
#feat_t = output_t.cpu().detach().numpy() # [1, 197, 576]
# print("@@@@@@@@@@@@@@@@@@@@@@@@@@@@")
# print(feat_t.shape)
# print("@@@@@@@@@@@@@@@@@@@@@@@@@@@@")
f.create_dataset(name[0] + "_feat", data=feat)
f.close()
return True