Spaces:
Sleeping
Sleeping
File size: 10,903 Bytes
cb80c28 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 |
'''
results on CIFAR-100:
| Reported Resnet18 | Reproduced Resnet32
Protocols | Reported FC | Reported SVM | Reproduced FC | Reproduced SVM |
T = 5 | 64.7 | 66.3 | 65.775 | 65.375 |
T = 10 | 63.4 | 65.2 | 64.91 | 65.10 |
T = 60 | 50.8 | 59.8 | 62.09 | 61.72 |
'''
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,Dataset
from models.base import BaseLearner
from utils.inc_net import CosineIncrementalNet, FOSTERNet, IncrementalNet
from utils.toolkit import count_parameters, target2onehot, tensor2numpy
from sklearn.svm import LinearSVC
from torchvision import datasets, transforms
from utils.autoaugment import CIFAR10Policy,ImageNetPolicy
from utils.ops import Cutout
EPSILON = 1e-8
class FeTrIL(BaseLearner):
def __init__(self, args):
super().__init__(args)
self.args = args
self._network = IncrementalNet(args, False)
self._means = []
self._svm_accs = []
def after_task(self):
self._known_classes = self._total_classes
def incremental_train(self, data_manager):
self.data_manager = data_manager
self.data_manager._train_trsf = [
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=63/255),
ImageNetPolicy(),
Cutout(n_holes=1, length=16),
]
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)
self._network_module_ptr = self._network
logging.info(
'Learning on {}-{}'.format(self._known_classes, self._total_classes))
if self._cur_task > 0:
for p in self._network.convnet.parameters():
p.requires_grad = False
logging.info('All params: {}'.format(count_parameters(self._network)))
logging.info('Trainable params: {}'.format(
count_parameters(self._network, True)))
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=self.args["batch_size"], shuffle=True, num_workers=self.args["num_workers"], pin_memory=True)
test_dataset = data_manager.get_dataset(
np.arange(0, self._total_classes), source='test', mode='test')
self.test_loader = DataLoader(
test_dataset, batch_size=self.args["batch_size"], shuffle=False, num_workers=self.args["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)
if len(self._multiple_gpus) > 1:
self._network = self._network.module
def _train(self, train_loader, test_loader):
self._network.to(self._device)
if hasattr(self._network, "module"):
self._network_module_ptr = self._network.module
if self._cur_task == 0:
self._epoch_num = self.args["init_epochs"]
optimizer = optim.SGD(filter(lambda p: p.requires_grad, self._network.parameters(
)), momentum=0.9, lr=self.args["init_lr"], weight_decay=self.args["init_weight_decay"])
scheduler = optim.lr_scheduler.CosineAnnealingLR(
optimizer=optimizer, T_max=self.args["init_epochs"])
self._train_function(train_loader, test_loader, optimizer, scheduler)
self._compute_means()
self._build_feature_set()
else:
self._epoch_num = self.args["epochs"]
self._compute_means()
self._compute_relations()
self._build_feature_set()
train_loader = DataLoader(self._feature_trainset, batch_size=self.args["batch_size"], shuffle=True, num_workers=self.args["num_workers"], pin_memory=True)
optimizer = optim.SGD(self._network_module_ptr.fc.parameters(),momentum=0.9,lr=self.args["lr"],weight_decay=self.args["weight_decay"])
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer=optimizer,T_max = self.args["epochs"])
self._train_function(train_loader, test_loader, optimizer, scheduler)
self._train_svm(self._feature_trainset,self._feature_testset)
def _compute_means(self):
with torch.no_grad():
for class_idx in range(self._known_classes, self._total_classes):
data, targets, idx_dataset = self.data_manager.get_dataset(np.arange(class_idx, class_idx+1), source='train',
mode='test', ret_data=True)
idx_loader = DataLoader(idx_dataset, batch_size=self.args["batch_size"], shuffle=False, num_workers=4)
vectors, _ = self._extract_vectors(idx_loader)
class_mean = np.mean(vectors, axis=0)
self._means.append(class_mean)
def _compute_relations(self):
old_means = np.array(self._means[:self._known_classes])
new_means = np.array(self._means[self._known_classes:])
self._relations=np.argmax((old_means/np.linalg.norm(old_means,axis=1)[:,None])@(new_means/np.linalg.norm(new_means,axis=1)[:,None]).T,axis=1)+self._known_classes
def _build_feature_set(self):
self.vectors_train = []
self.labels_train = []
for class_idx in range(self._known_classes, self._total_classes):
data, targets, idx_dataset = self.data_manager.get_dataset(np.arange(class_idx, class_idx+1), source='train',
mode='test', ret_data=True)
idx_loader = DataLoader(idx_dataset, batch_size=self.args["batch_size"], shuffle=False, num_workers=4)
vectors, _ = self._extract_vectors(idx_loader)
self.vectors_train.append(vectors)
self.labels_train.append([class_idx]*len(vectors))
for class_idx in range(0,self._known_classes):
new_idx = self._relations[class_idx]
self.vectors_train.append(self.vectors_train[new_idx-self._known_classes]-self._means[new_idx]+self._means[class_idx])
self.labels_train.append([class_idx]*len(self.vectors_train[-1]))
self.vectors_train = np.concatenate(self.vectors_train)
self.labels_train = np.concatenate(self.labels_train)
self._feature_trainset = FeatureDataset(self.vectors_train,self.labels_train)
self.vectors_test = []
self.labels_test = []
for class_idx in range(0, self._total_classes):
data, targets, idx_dataset = self.data_manager.get_dataset(np.arange(class_idx, class_idx+1), source='test',
mode='test', ret_data=True)
idx_loader = DataLoader(idx_dataset, batch_size=self.args["batch_size"], shuffle=False, num_workers=4)
vectors, _ = self._extract_vectors(idx_loader)
self.vectors_test.append(vectors)
self.labels_test.append([class_idx]*len(vectors))
self.vectors_test = np.concatenate(self.vectors_test)
self.labels_test = np.concatenate(self.labels_test)
self._feature_testset = FeatureDataset(self.vectors_test,self.labels_test)
def _train_function(self, train_loader, test_loader, optimizer, scheduler):
prog_bar = tqdm(range(self._epoch_num))
for _, epoch in enumerate(prog_bar):
if self._cur_task == 0:
self._network.train()
else:
self._network.eval()
losses = 0.
correct, total = 0, 0
for i, _, inputs, targets in enumerate(train_loader):
inputs, targets = inputs.to(
self._device, non_blocking=True), targets.to(self._device, non_blocking=True)
if self._cur_task ==0:
logits = self._network(inputs)['logits']
else:
logits = self._network_module_ptr.fc(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:
info = 'Task {}, Epoch {}/{} => Loss {:.3f}, Train_accy {:.2f}'.format(
self._cur_task, epoch+1, self._epoch_num, losses/len(train_loader), train_acc)
else:
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, self._epoch_num, losses/len(train_loader), train_acc, test_acc)
prog_bar.set_description(info)
logging.info(info)
def _train_svm(self,train_set,test_set):
train_features = train_set.features.numpy()
train_labels = train_set.labels.numpy()
test_features = test_set.features.numpy()
test_labels = test_set.labels.numpy()
train_features = train_features/np.linalg.norm(train_features,axis=1)[:,None]
test_features = test_features/np.linalg.norm(test_features,axis=1)[:,None]
svm_classifier = LinearSVC(random_state=42)
svm_classifier.fit(train_features,train_labels)
logging.info("svm train: acc: {}".format(np.around(svm_classifier.score(train_features,train_labels)*100,decimals=2)))
acc = svm_classifier.score(test_features,test_labels)
self._svm_accs.append(np.around(acc*100,decimals=2))
logging.info("svm evaluation: acc_list: {}".format(self._svm_accs))
class FeatureDataset(Dataset):
def __init__(self, features, labels):
assert len(features) == len(labels), "Data size error!"
self.features = torch.from_numpy(features)
self.labels = torch.from_numpy(labels)
def __len__(self):
return len(self.features)
def __getitem__(self, idx):
feature = self.features[idx]
label = self.labels[idx]
return idx, feature, label
|