-
Notifications
You must be signed in to change notification settings - Fork 138
/
icarl.py
205 lines (178 loc) · 7.15 KB
/
icarl.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
import logging
import numpy as np
from tqdm import tqdm
import torch
from torch import nn
from torch import optim
from torch.nn import functional as F
from torch.utils.data import DataLoader
from models.base import BaseLearner
from utils.inc_net import IncrementalNet
from utils.inc_net import CosineIncrementalNet
from utils.toolkit import target2onehot, tensor2numpy
EPSILON = 1e-8
init_epoch = 200
init_lr = 0.1
init_milestones = [60, 120, 170]
init_lr_decay = 0.1
init_weight_decay = 0.0005
epochs = 170
lrate = 0.1
milestones = [80, 120]
lrate_decay = 0.1
batch_size = 128
weight_decay = 2e-4
num_workers = 8
T = 2
class iCaRL(BaseLearner):
def __init__(self, args):
super().__init__(args)
self._network = IncrementalNet(args, False)
def after_task(self):
self._old_network = self._network.copy().freeze()
self._known_classes = self._total_classes
logging.info("Exemplar size: {}".format(self.exemplar_size))
def incremental_train(self, data_manager):
self._cur_task += 1
self._total_classes = self._known_classes + data_manager.get_task_size(
self._cur_task
)
self._network.update_fc(self._total_classes)
logging.info(
"Learning on {}-{}".format(self._known_classes, self._total_classes)
)
train_dataset = data_manager.get_dataset(
np.arange(self._known_classes, self._total_classes),
source="train",
mode="train",
appendent=self._get_memory(),
)
self.train_loader = DataLoader(
train_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers
)
test_dataset = data_manager.get_dataset(
np.arange(0, self._total_classes), source="test", mode="test"
)
self.test_loader = DataLoader(
test_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers
)
if len(self._multiple_gpus) > 1:
self._network = nn.DataParallel(self._network, self._multiple_gpus)
self._train(self.train_loader, self.test_loader)
self.build_rehearsal_memory(data_manager, self.samples_per_class)
if len(self._multiple_gpus) > 1:
self._network = self._network.module
def _train(self, train_loader, test_loader):
self._network.to(self._device)
if self._old_network is not None:
self._old_network.to(self._device)
if self._cur_task == 0:
optimizer = optim.SGD(
self._network.parameters(),
momentum=0.9,
lr=init_lr,
weight_decay=init_weight_decay,
)
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer=optimizer, milestones=init_milestones, gamma=init_lr_decay
)
self._init_train(train_loader, test_loader, optimizer, scheduler)
else:
optimizer = optim.SGD(
self._network.parameters(),
lr=lrate,
momentum=0.9,
weight_decay=weight_decay,
) # 1e-5
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer=optimizer, milestones=milestones, gamma=lrate_decay
)
self._update_representation(train_loader, test_loader, optimizer, scheduler)
def _init_train(self, train_loader, test_loader, optimizer, scheduler):
prog_bar = tqdm(range(init_epoch))
for _, epoch in enumerate(prog_bar):
self._network.train()
losses = 0.0
correct, total = 0, 0
for i, (_, inputs, targets) in enumerate(train_loader):
inputs, targets = inputs.to(self._device), targets.to(self._device)
logits = self._network(inputs)["logits"]
loss = F.cross_entropy(logits, targets)
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses += loss.item()
_, preds = torch.max(logits, dim=1)
correct += preds.eq(targets.expand_as(preds)).cpu().sum()
total += len(targets)
scheduler.step()
train_acc = np.around(tensor2numpy(correct) * 100 / total, decimals=2)
if epoch % 5 == 0:
test_acc = self._compute_accuracy(self._network, test_loader)
info = "Task {}, Epoch {}/{} => Loss {:.3f}, Train_accy {:.2f}, Test_accy {:.2f}".format(
self._cur_task,
epoch + 1,
init_epoch,
losses / len(train_loader),
train_acc,
test_acc,
)
else:
info = "Task {}, Epoch {}/{} => Loss {:.3f}, Train_accy {:.2f}".format(
self._cur_task,
epoch + 1,
init_epoch,
losses / len(train_loader),
train_acc,
)
prog_bar.set_description(info)
logging.info(info)
def _update_representation(self, train_loader, test_loader, optimizer, scheduler):
prog_bar = tqdm(range(epochs))
for _, epoch in enumerate(prog_bar):
self._network.train()
losses = 0.0
correct, total = 0, 0
for i, (_, inputs, targets) in enumerate(train_loader):
inputs, targets = inputs.to(self._device), targets.to(self._device)
logits = self._network(inputs)["logits"]
loss_clf = F.cross_entropy(logits, targets)
loss_kd = _KD_loss(
logits[:, : self._known_classes],
self._old_network(inputs)["logits"],
T,
)
loss = loss_clf + loss_kd
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses += loss.item()
_, preds = torch.max(logits, dim=1)
correct += preds.eq(targets.expand_as(preds)).cpu().sum()
total += len(targets)
scheduler.step()
train_acc = np.around(tensor2numpy(correct) * 100 / total, decimals=2)
if epoch % 5 == 0:
test_acc = self._compute_accuracy(self._network, test_loader)
info = "Task {}, Epoch {}/{} => Loss {:.3f}, Train_accy {:.2f}, Test_accy {:.2f}".format(
self._cur_task,
epoch + 1,
epochs,
losses / len(train_loader),
train_acc,
test_acc,
)
else:
info = "Task {}, Epoch {}/{} => Loss {:.3f}, Train_accy {:.2f}".format(
self._cur_task,
epoch + 1,
epochs,
losses / len(train_loader),
train_acc,
)
prog_bar.set_description(info)
logging.info(info)
def _KD_loss(pred, soft, T):
pred = torch.log_softmax(pred / T, dim=1)
soft = torch.softmax(soft / T, dim=1)
return -1 * torch.mul(soft, pred).sum() / pred.shape[0]