-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathengine_cls.py
197 lines (153 loc) · 8.01 KB
/
engine_cls.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
# Copyright (c) 2015-present, Facebook, Inc.
# All rights reserved.
"""
Train and eval functions used in main.py
"""
import math
import sys
from typing import Iterable, Optional
import torch
import numpy as np
from timm.data import Mixup
from timm.utils import accuracy, ModelEma
from losses import DistillationLoss
import utils
WARM_UP_EPOCH = 10
TOTAL_DECAY_EPOCH = 100
def adjust_keep_rate(iters, epoch, warmup_epochs, total_epochs,
iter_per_epoch, base_keep_rate=0.7, max_keep_rate=1):
# reference: https://github.com/youweiliang/evit/blob/master/helpers.py#L7
# Token-Level Function
if epoch < warmup_epochs:
return 1
if epoch >= total_epochs:
return base_keep_rate
total_decay_iters = iter_per_epoch * (total_epochs - warmup_epochs)
iters = iters - iter_per_epoch * warmup_epochs
keep_rate = base_keep_rate + (max_keep_rate - base_keep_rate) \
* (math.cos(iters / total_decay_iters * math.pi) + 1) * 0.5
return keep_rate
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,
example_stats_train=None, train_example_idx=None, # additional
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 = 10
#i=0
# Example-level variables
correct = 0.
total = 0.
# Token-Level variables
cur_iter = epoch * len(data_loader)
iter_per_epoch = len(data_loader)
ite_step = 0
for samples, targets, idx in metric_logger.log_every(data_loader, print_freq, header):
samples = samples.to(device, non_blocking=True)
targets = targets.to(device, non_blocking=True)
idx = idx.to(device, non_blocking=True)
#print('idx',idx) #tensor([536147, 536147, 536147, 706066], device='cuda:0')
targets_org=targets
# >>>>>>>> Dynamic Keep Ratio (Token Sparse) >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
token_ratio = adjust_keep_rate(cur_iter + ite_step, epoch, warmup_epochs=WARM_UP_EPOCH,
total_epochs=WARM_UP_EPOCH + TOTAL_DECAY_EPOCH,
iter_per_epoch=iter_per_epoch, base_keep_rate=args.keep_ratio) # get current keep ratio, gradually decrease from 1
attn_ratio = adjust_keep_rate(cur_iter + ite_step, epoch, warmup_epochs=WARM_UP_EPOCH,
total_epochs=WARM_UP_EPOCH + TOTAL_DECAY_EPOCH,
iter_per_epoch=iter_per_epoch, base_keep_rate=args.attn_ratio) # get current keep ratio, gradually decrease from 1
ite_step += 1
# <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
if mixup_fn is not None:
samples, targets = mixup_fn(samples, targets)
with torch.cuda.amp.autocast():#(enabled=False):
outputs, cls_attn = model(samples, ratio=token_ratio, attn_ratio=attn_ratio)
loss = criterion(samples, outputs, targets)
# >>>>>>>> data sparse >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
outputs_pred = outputs
#print('outputs_pred',outputs_pred)
_, predicted = torch.max(outputs_pred.data, 1) #max value in each row (token), total of B values,
cls_var = torch.var(cls_attn,1) #add cls var
#print('cls_var',cls_var)
#print('predicted',predicted)
#print('predicted.shape',predicted.shape)
acc = predicted == targets_org
correct += predicted.eq(targets_org.data).cpu().sum()
total += targets_org.size(0)
#<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
loss_value = loss.item()
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
sys.exit(1)
optimizer.zero_grad()
# >>>>>>>> data sparse >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
for j, index in enumerate(idx):
# Get index in original dataset (not sorted by forgetting)
index_in_original_dataset = train_example_idx[index]
# Compute missclassification margin
output_correct_class = outputs_pred.data[j, targets_org[j].item()]
sorted_output, _ = torch.sort(outputs_pred.data[j, :])
if acc[j]:
# Example classified correctly, highest incorrect class is 2nd largest output
output_highest_incorrect_class = sorted_output[-2]
else:
# Example misclassified, highest incorrect class is max output
output_highest_incorrect_class = sorted_output[-1]
margin = output_correct_class.item(
) - output_highest_incorrect_class.item()
# Add the statistics of the current training example to dictionary
index_stats = example_stats_train.get(index_in_original_dataset, [[], [], [] , []])
index_stats[0].append(loss.item())
index_stats[1].append(acc[j].sum().item())
index_stats[2].append(margin)
index_stats[3].append(cls_var[j].sum().item())
example_stats_train[index_in_original_dataset] = index_stats
# Add training accuracy to dict
index_stats = example_stats_train.get('train', [[], []])
index_stats[1].append(100. * correct.item() / float(total))
example_stats_train['train'] = index_stats
#<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
#exit()
# 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)
#i+=1
#if i > 20: break
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, args):
# fix keep ratio in inference >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
criterion = torch.nn.CrossEntropyLoss()
metric_logger = utils.MetricLogger(delimiter=" ")
header = 'Test:'
# switch to evaluation mode
model.eval()
for images, target, idx 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, cls_attn = model(images, ratio=args.keep_ratio, attn_ratio=args.attn_ratio)
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()}