def __init__(self, args):
super().__init__()
self._network = IncrementalNet(args['convnet_type'], False)
self._device = args['device']
self.convnet_type = args['convnet_type']
self.expert = None
def _train_expert(self, train_loader, test_loader):
self.expert = IncrementalNet(self.convnet_type, False)
self.expert.update_fc(self.task_size)
if len(self._multiple_gpus) > 1:
self.expert = nn.DataParallel(self.expert, self._multiple_gpus)
self.expert.to(self._device)
optimizer = optim.SGD(self.expert.parameters(),
lr=lrate_expert,
momentum=0.9,
weight_decay=weight_decay)
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer, milestones=milestones_expert, gamma=lrate_decay_expert)
prog_bar = tqdm(range(epochs_expert))
for _, epoch in enumerate(prog_bar):
self.expert.train()
losses = 0.
for i, (_, inputs, targets) in enumerate(train_loader):
inputs, targets = inputs.to(
self._device), (targets - self._known_classes).to(
self._device)
logits = self.expert(inputs)['logits']
loss = F.cross_entropy(logits, targets)
losses += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
train_acc = self._compute_accuracy(self.expert, train_loader,
self._known_classes)
test_acc = self._compute_accuracy(self.expert, test_loader,
self._known_classes)
info = 'Expert CNN => Epoch {}/{}, Loss {:.3f}, Train accy {:.2f}, Test accy {:.2f}'.format(
epoch + 1, epochs_expert, losses / len(train_loader),
train_acc, test_acc)
prog_bar.set_description(info)
logging.info(info)
if len(self._multiple_gpus) > 1:
self.expert = self.expert.module
def __init__(self, args):
super().__init__(args)
self._network = IncrementalNet(args['convnet_type'], False)
self._seen_classes = []
class End2End(BaseLearner):
def __init__(self, args):
super().__init__(args)
self._network = IncrementalNet(args['convnet_type'], False)
self._seen_classes = []
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.task_size = data_manager.get_task_size(self._cur_task)
self._total_classes = self._known_classes + self.task_size
self._network.update_fc(self._total_classes)
self._seen_classes.append(self.task_size)
logging.info('Learning on {}-{}'.format(self._known_classes,
self._total_classes))
# Loader
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)
# Procedure
if len(self._multiple_gpus) > 1:
self._network = nn.DataParallel(self._network, self._multiple_gpus)
self._train(data_manager, 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, data_manager, 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(),
lr=lrate_init,
momentum=0.9,
weight_decay=weight_decay)
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer=optimizer,
milestones=milestones_init,
gamma=lrate_decay)
self._is_finetuning = False
self._run(self.train_loader, self.test_loader, epochs_init,
optimizer, scheduler, 'Training')
return
# New + exemplars
optimizer = optim.SGD(self._network.parameters(),
lr=lrate,
momentum=0.9,
weight_decay=weight_decay)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer=optimizer,
milestones=milestones,
gamma=lrate_decay)
self._is_finetuning = False
self._run(self.train_loader, self.test_loader, epochs, optimizer,
scheduler, 'Training')
# Finetune
if self._fixed_memory:
finetune_samples_per_class = self._memory_per_class
self._construct_exemplar_unified(data_manager,
finetune_samples_per_class)
else:
finetune_samples_per_class = self._memory_size // self._known_classes
self._reduce_exemplar(data_manager, finetune_samples_per_class)
self._construct_exemplar(data_manager, finetune_samples_per_class)
if len(self._multiple_gpus) > 1:
self._old_network = self._network.module.copy().freeze()
else:
self._old_network = self._network.copy().freeze()
finetune_train_dataset = data_manager.get_dataset(
[], source='train', mode='train', appendent=self._get_memory())
finetune_train_loader = DataLoader(finetune_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
# Update all weights or only the weights of FC layer?
# According to the experiment results, fine-tuning all weights is slightly better.
optimizer = optim.SGD(self._network.parameters(),
lr=lrate_finetune,
momentum=0.9,
weight_decay=weight_decay)
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer=optimizer,
milestones=milestones_finetune,
gamma=lrate_decay)
self._is_finetuning = True
self._run(finetune_train_loader, self.test_loader, epochs_finetune,
optimizer, scheduler, 'Finetuning')
# Remove the temporary exemplars of new classes
if self._fixed_memory:
self._data_memory = self._data_memory[:-self._memory_per_class *
self.task_size]
self._targets_memory = self._targets_memory[:-self.
_memory_per_class *
self.task_size]
# Check
assert len(
np.setdiff1d(self._targets_memory,
np.arange(
0,
self._known_classes))) == 0, 'Exemplar error!'
def _run(self, train_loader, test_loader, epochs_, optimizer, scheduler,
process):
prog_bar = tqdm(range(epochs_))
for _, epoch in enumerate(prog_bar, start=1):
self._network.train()
losses = 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']
# CELoss
clf_loss = F.cross_entropy(logits, targets)
if self._cur_task == 0:
distill_loss = torch.zeros(1, device=self._device)
else:
finetuning_task = (
self._cur_task +
1) if self._is_finetuning else self._cur_task
distill_loss = 0.
old_logits = self._old_network(inputs)['logits']
for i in range(1, finetuning_task + 1):
lo = sum(self._seen_classes[:i - 1])
hi = sum(self._seen_classes[:i])
task_prob_new = F.softmax(logits[:, lo:hi], dim=1)
task_prob_old = F.softmax(old_logits[:, lo:hi], dim=1)
task_prob_new = task_prob_new**(1 / T)
task_prob_old = task_prob_old**(1 / T)
task_prob_new = task_prob_new / task_prob_new.sum(
1).view(-1, 1)
task_prob_old = task_prob_old / task_prob_old.sum(
1).view(-1, 1)
distill_loss += F.binary_cross_entropy(
task_prob_new, task_prob_old)
distill_loss *= 1 / finetuning_task
loss = clf_loss + distill_loss
losses += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
# acc
_, preds = torch.max(logits, dim=1)
correct += preds.eq(targets.expand_as(preds)).cpu().sum()
total += len(targets)
scheduler.step()
# train_acc = self._compute_accuracy(self._network, train_loader)
train_acc = np.around(tensor2numpy(correct) * 100 / total,
decimals=2)
test_acc = self._compute_accuracy(self._network, test_loader)
info1 = '{} => '.format(process)
info2 = '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)
prog_bar.set_description(info1 + info2)
logging.info(info1 + info2)
class iCaRL(BaseLearner):
def __init__(self, args):
super().__init__(args)
self._network = IncrementalNet(args['convnet_type'], 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))
# Loader
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)
# Procedure
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)
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 _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.
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']
onehots = target2onehot(targets, self._total_classes)
if self._old_network is None:
loss = F.binary_cross_entropy_with_logits(logits, onehots)
else:
old_onehots = torch.sigmoid(
self._old_network(inputs)['logits'].detach())
new_onehots = onehots.clone()
new_onehots[:, :self._known_classes] = old_onehots
loss = F.binary_cross_entropy_with_logits(
logits, new_onehots)
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses += loss.item()
# acc
_, preds = torch.max(logits, dim=1)
correct += preds.eq(targets.expand_as(preds)).cpu().sum()
total += len(targets)
scheduler.step()
# train_acc = self._compute_accuracy(self._network, train_loader)
train_acc = np.around(tensor2numpy(correct) * 100 / total,
decimals=2)
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)
prog_bar.set_description(info)
logging.info(info)
class End2End(BaseLearner):
def __init__(self, args):
super().__init__()
self._network = IncrementalNet(args['convnet_type'], False)
self._device = args['device']
self._seen_classes = []
def after_task(self):
self._old_network = self._network.copy().freeze()
self._known_classes = self._total_classes
def incremental_train(self, data_manager):
self._cur_task += 1
task_size = data_manager.get_task_size(self._cur_task)
self._total_classes = self._known_classes + task_size
self._network.update_fc(self._total_classes)
self._seen_classes.append(task_size)
logging.info('Learning on {}-{}'.format(self._known_classes,
self._total_classes))
# Loader
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=4)
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=4)
# Procedure
self._train(data_manager, self.train_loader, self.test_loader)
self._reduce_exemplar(data_manager, memory_size // self._total_classes)
self._construct_exemplar(data_manager,
memory_size // self._total_classes)
def _train(self, data_manager, 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(),
lr=lrate_init,
momentum=0.9,
weight_decay=1e-3)
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer=optimizer,
milestones=milestones_init,
gamma=lrate_decay)
self._is_finetuning = False
self._run(self.train_loader, self.test_loader, epochs_init,
optimizer, scheduler, 'Training')
return
# New + exemplars
optimizer = optim.SGD(self._network.parameters(),
lr=lrate,
momentum=0.9,
weight_decay=1e-3)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer=optimizer,
milestones=milestones,
gamma=lrate_decay)
self._is_finetuning = False
self._run(self.train_loader, self.test_loader, epochs, optimizer,
scheduler, 'Training')
# Finetune
samples_per_class = memory_size // self._known_classes
self._reduce_exemplar(data_manager, samples_per_class)
self._construct_exemplar(data_manager, samples_per_class)
self._old_network = self._network.copy().freeze()
finetune_train_dataset = data_manager.get_dataset(
[], source='train', mode='train', appendent=self._get_memory())
finetune_train_loader = DataLoader(finetune_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
optimizer = optim.SGD(self._network.parameters(),
lr=lrate_finetune,
momentum=0.9,
weight_decay=1e-3)
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer=optimizer,
milestones=milestones_finetune,
gamma=lrate_decay)
self._is_finetuning = True
self._run(finetune_train_loader, self.test_loader, epochs_finetune,
optimizer, scheduler, 'Finetuning')
def _run(self, train_loader, test_loader, epochs_, optimizer, scheduler,
process):
prog_bar = tqdm(range(epochs_))
for _, epoch in enumerate(prog_bar, start=1):
self._network.train()
losses = 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)
# CELoss
clf_loss = F.cross_entropy(logits, targets)
if self._cur_task == 0:
distill_loss = torch.zeros(1, device=self._device)
else:
finetuning_task = (
self._cur_task +
1) if self._is_finetuning else self._cur_task
distill_loss = 0.
old_logits = self._old_network(inputs)
for i in range(1, finetuning_task + 1):
lo = sum(self._seen_classes[:i - 1])
hi = sum(self._seen_classes[:i])
distill_loss += F.binary_cross_entropy(
F.softmax(logits[:, lo:hi] / T, dim=1),
F.softmax(old_logits[:, lo:hi] / T, dim=1))
loss = clf_loss + distill_loss
losses += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
# acc
_, preds = torch.max(logits, dim=1)
correct += preds.eq(targets.expand_as(preds)).cpu().sum()
total += len(targets)
scheduler.step()
# train_acc = self._compute_accuracy(self._network, train_loader)
train_acc = np.around(tensor2numpy(correct) * 100 / total,
decimals=2)
test_acc = self._compute_accuracy(self._network, test_loader)
info1 = '{} => '.format(process)
info2 = '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)
prog_bar.set_description(info1 + info2)
logging.info(info1 + info2)
def __init__(self, args):
super().__init__(args)
self._network = IncrementalNet(args['convnet_type'],
pretrained=False,
gradcam=True)
class LwM(BaseLearner):
def __init__(self, args):
super().__init__(args)
self._network = IncrementalNet(args['convnet_type'],
pretrained=False,
gradcam=True)
def after_task(self):
self._network.zero_grad()
self._network.unset_gradcam_hook()
self._old_network = self._network.copy().eval()
self._network.set_gradcam_hook()
self._old_network.set_gradcam_hook()
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))
# Loader
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)
# Procedure
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 self._old_network is not None:
self._old_network.to(self._device)
optimizer = optim.SGD(self._network.parameters(),
lr=lrate,
momentum=0.9,
weight_decay=weight_decay)
# optimizer = optim.Adam(self._network.parameters(), lr=lrate, weight_decay=1e-5)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer=optimizer,
milestones=milestones,
gamma=lrate_decay)
self._run(train_loader, test_loader, optimizer, scheduler)
def _run(self, train_loader, test_loader, optimizer, scheduler):
for epoch in range(1, epochs + 1):
self._network.train()
clf_losses = 0. # cross entropy
distill_losses = 0. # distillation
attention_losses = 0. # attention distillation
correct, total = 0, 0
for i, (_, inputs, targets) in enumerate(train_loader):
inputs, targets = inputs.to(self._device), targets.to(
self._device)
outputs = self._network(inputs)
logits = outputs['logits']
optimizer.zero_grad() # Same effect as nn.Module.zero_grad()
if self._old_network is None:
clf_loss = F.cross_entropy(logits, targets)
clf_losses += clf_loss.item()
loss = clf_loss
else:
self._old_network.zero_grad()
old_outputs = self._old_network(inputs)
old_logits = old_outputs['logits']
# Classification loss
# if no old samples saved, only calculate loss for new logits
clf_loss = F.cross_entropy(logits[:, self._known_classes:],
targets - self._known_classes)
clf_losses += clf_loss.item()
# Distillation loss
# if no old samples saved, only calculate distillation loss for old logits
'''
distill_loss = F.binary_cross_entropy_with_logits(
logits[:, :self._known_classes], torch.sigmoid(old_logits.detach())
) * distill_ratio
'''
distill_loss = _KD_loss(logits[:, :self._known_classes],
old_logits.detach(),
T=2) * distill_ratio
distill_losses += distill_loss.item()
# Attention distillation loss
top_base_indices = logits[:, :self._known_classes].argmax(
dim=1)
onehot_top_base = target2onehot(
top_base_indices, self._known_classes).to(self._device)
logits[:, :self._known_classes].backward(
gradient=onehot_top_base, retain_graph=True)
old_logits.backward(gradient=onehot_top_base)
attention_loss = gradcam_distillation(
outputs['gradcam_gradients'][0],
old_outputs['gradcam_gradients'][0].detach(),
outputs['gradcam_activations'][0],
old_outputs['gradcam_activations']
[0].detach()) * attention_ratio
attention_losses += attention_loss.item()
# Integration
loss = clf_loss + distill_loss + attention_loss
self._old_network.zero_grad()
self._network.zero_grad()
optimizer.zero_grad() # Same effect as nn.Module.zero_grad()
loss.backward()
optimizer.step()
# acc
_, preds = torch.max(logits, dim=1)
correct += preds.eq(targets.expand_as(preds)).cpu().sum()
total += len(targets)
scheduler.step()
# train_acc = self._compute_accuracy(self._network, train_loader)
train_acc = np.around(tensor2numpy(correct) * 100 / total,
decimals=2)
test_acc = self._compute_accuracy(self._network, test_loader)
info1 = 'Task {}, Epoch {}/{} => clf_loss {:.2f}, '.format(
self._cur_task, epoch, epochs, clf_losses / (i + 1))
info2 = 'distill_loss {:.2f}, attention_loss {:.2f}, Train_accy {:.2f}, Test_accy {:.2f}'.format(
distill_losses / (i + 1), attention_losses / (i + 1),
train_acc, test_acc)
logging.info(info1 + info2)
class iCaRL(BaseLearner):
def __init__(self, args):
super().__init__()
self._network = IncrementalNet(args['convnet_type'], False)
self._device = args['device']
def after_task(self):
self._old_network = self._network.copy().freeze()
self._known_classes = self._total_classes
def eval_task(self):
y_pred, y_true = self._eval_ncm(self.test_loader, self._class_means)
accy = accuracy(y_pred, y_true, self._known_classes)
return accy
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))
# Loader
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=4)
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=4)
# Procedure
self._train(self.train_loader, self.test_loader) #训练集 测试集训练
self._reduce_exemplar(data_manager,
memory_size // self._total_classes) #范例集精简
self._construct_exemplar(data_manager, memory_size //
self._total_classes) #为新类样本创建范例集
def _train(self, train_loader, test_loader):
self._network.to(self._device)
optimizer = optim.SGD(self._network.parameters(),
lr=lrate,
momentum=0.9,
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 _update_representation(self, train_loader, test_loader, optimizer,
scheduler): #特征表示更新
prog_bar = tqdm(range(epochs))
for _, epoch in enumerate(prog_bar):
self._network.train()
losses = 0.
for i, (_, inputs, targets) in enumerate(train_loader): #新类
inputs, targets = inputs.to(self._device), targets.to(
self._device)
logits = self._network(inputs) #新类的输入和forward
onehots = target2onehot(targets,
self._total_classes) #新类标签->onehot
if self._old_network is None: #没有原始网络 单纯分类
loss = F.binary_cross_entropy_with_logits(logits, onehots)
else:
old_onehots = torch.sigmoid(
self._old_network(
inputs).detach()) #原来的网络对新样本的预测（用于计算蒸馏损失）
new_onehots = onehots.clone() #新的onehot
new_onehots[:, :self.
_known_classes] = old_onehots #由于gt这个onehot向量的label位肯定在 :_known_classes后面
loss = F.binary_cross_entropy_with_logits(
logits, new_onehots) #因此新网络的输出既可以与新样本的新位置的gt算CELoss
#也可以与原来的网络的预测结果算CELoss 以求不要忘记原来网络的输出
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses += loss.item()
scheduler.step()
train_acc = self._compute_accuracy(self._network, train_loader)
test_acc = self._compute_accuracy(self._network, test_loader)
info = 'Task {}, Epoch {}/{} => Loss {:.3f}, Train_accy {:.3f}, Test_accy {:.3f}'.format(
self._cur_task, epoch + 1, epochs, losses / len(train_loader),
train_acc, test_acc)
prog_bar.set_description(info)
logging.info(info)
class DR(BaseLearner):
def __init__(self, args):
super().__init__(args)
self._network = IncrementalNet(args['convnet_type'], False)
self.convnet_type = args['convnet_type']
self.expert = None
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.task_size = data_manager.get_task_size(self._cur_task)
self._total_classes = self._known_classes + self.task_size
self._network.update_fc(self._total_classes)
logging.info('Learning on {}-{}'.format(self._known_classes,
self._total_classes))
# Loader
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)
expert_train_dataset = data_manager.get_dataset(np.arange(
self._known_classes, self._total_classes),
source='train',
mode='train')
self.expert_train_loader = DataLoader(expert_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
expert_test_dataset = data_manager.get_dataset(np.arange(
self._known_classes, self._total_classes),
source='test',
mode='test')
self.expert_test_loader = DataLoader(expert_test_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers)
# Procedure
logging.info('Training the expert CNN...')
self._train_expert(self.expert_train_loader, self.expert_test_loader)
if self._cur_task == 0:
self._network = self.expert.copy()
else:
self.expert = self.expert.freeze()
logging.info('Training the updated CNN...')
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 and self._cur_task > 0:
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)
optimizer = optim.SGD(self._network.parameters(),
lr=lrate,
momentum=0.9,
weight_decay=weight_decay)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer,
milestones=milestones,
gamma=lrate_decay)
prog_bar = tqdm(range(epochs))
for _, epoch in enumerate(prog_bar):
self._network.train()
losses = 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']
exp_logits = self.expert(inputs)['logits']
old_logits = self._old_network(inputs)['logits']
# Distillation
dist_term = _KD_loss(logits[:, self._known_classes:],
exp_logits, T1)
# Retrospection
retr_term = _KD_loss(logits[:, :self._known_classes],
old_logits, T2)
loss = dist_term + retr_term
losses += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
# acc
_, preds = torch.max(logits, dim=1)
correct += preds.eq(targets.expand_as(preds)).cpu().sum()
total += len(targets)
scheduler.step()
# train_acc = self._compute_accuracy(self._network, train_loader)
train_acc = np.around(tensor2numpy(correct) * 100 / total,
decimals=2)
test_acc = self._compute_accuracy(self._network, test_loader)
info = 'Updated CNN => Epoch {}/{}, Loss {:.3f}, Train accy {:.2f}, Test accy {:.2f}'.format(
epoch + 1, epochs, losses / len(train_loader), train_acc,
test_acc)
prog_bar.set_description(info)
logging.info(info)
def _train_expert(self, train_loader, test_loader):
self.expert = IncrementalNet(self.convnet_type, False)
self.expert.update_fc(self.task_size)
if len(self._multiple_gpus) > 1:
self.expert = nn.DataParallel(self.expert, self._multiple_gpus)
self.expert.to(self._device)
optimizer = optim.SGD(self.expert.parameters(),
lr=lrate_expert,
momentum=0.9,
weight_decay=weight_decay)
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer, milestones=milestones_expert, gamma=lrate_decay_expert)
prog_bar = tqdm(range(epochs_expert))
for _, epoch in enumerate(prog_bar):
self.expert.train()
losses = 0.
for i, (_, inputs, targets) in enumerate(train_loader):
inputs, targets = inputs.to(
self._device), (targets - self._known_classes).to(
self._device)
logits = self.expert(inputs)['logits']
loss = F.cross_entropy(logits, targets)
losses += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
train_acc = self._compute_accuracy(self.expert, train_loader,
self._known_classes)
test_acc = self._compute_accuracy(self.expert, test_loader,
self._known_classes)
info = 'Expert CNN => Epoch {}/{}, Loss {:.3f}, Train accy {:.2f}, Test accy {:.2f}'.format(
epoch + 1, epochs_expert, losses / len(train_loader),
train_acc, test_acc)
prog_bar.set_description(info)
logging.info(info)
if len(self._multiple_gpus) > 1:
self.expert = self.expert.module
def _compute_accuracy(self, model, loader, offset=0):
model.eval()
correct, total = 0, 0
for i, (_, inputs, targets) in enumerate(loader):
inputs = inputs.to(self._device)
targets -= offset
with torch.no_grad():
outputs = model(inputs)['logits']
predicts = torch.max(outputs, dim=1)[1]
correct += (predicts.cpu() == targets).sum()
total += len(targets)
return np.around(tensor2numpy(correct) * 100 / total, decimals=2)
class DR(BaseLearner):
def __init__(self, args):
super().__init__()
self._network = IncrementalNet(args['convnet_type'], False) #网络
self._device = args['device']
self.convnet_type = args['convnet_type']
self.expert = None #专家网络
def after_task(self):
self._old_network = self._network.copy().freeze() #上一次迭代的网络
self._known_classes = self._total_classes
def eval_task(self):
y_pred, y_true = self._eval_ncm(self.test_loader, self._class_means)
accy = accuracy(y_pred, y_true, self._known_classes)
return accy
def incremental_train(self, data_manager):
self._cur_task += 1
self.task_size = data_manager.get_task_size(self._cur_task)
self._total_classes = self._known_classes + self.task_size
self._network.update_fc(self._total_classes)
logging.info('Learning on {}-{}'.format(self._known_classes,
self._total_classes))
# Loader
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=4)
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=4)
expert_train_dataset = data_manager.get_dataset(np.arange(
self._known_classes, self._total_classes),
source='train',
mode='train')
self.expert_train_loader = DataLoader(expert_train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
expert_test_dataset = data_manager.get_dataset(np.arange(
self._known_classes, self._total_classes),
source='test',
mode='test')
self.expert_test_loader = DataLoader(expert_test_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=4)
# Procedure
logging.info('Training the expert CNN...')
self._train_expert(
self.expert_train_loader,
self.expert_test_loader) #专家网络训练 专家网络在每次训练时 专门用于学习新类
if self._cur_task == 0:
self._network = self.expert.copy() #刚刚开始 网络=专家网络
else:
self.expert = self.expert.freeze() #专家网络.freeze()
logging.info('Training the updated CNN...')
self._train(self.train_loader,
self.test_loader) #针对专家网络和上一轮的旧网络进行增量学习
self._reduce_exemplar(data_manager,
memory_size // self._total_classes) #范例集精简
self._construct_exemplar(data_manager,
memory_size // self._total_classes) #新类构建范例集
def _train(self, train_loader, test_loader):
self._network.to(self._device)
optimizer = optim.SGD(self._network.parameters(),
lr=lrate,
momentum=0.9,
weight_decay=1e-5)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer,
milestones=milestones,
gamma=lrate_decay)
prog_bar = tqdm(range(epochs))
for _, epoch in enumerate(prog_bar): #epoch进度条
self._network.train()
losses = 0.
for i, (_, inputs, targets) in enumerate(train_loader): #训练集
inputs, targets = inputs.to(self._device), targets.to(
self._device) #数据 标签
logits = self._network(inputs) #当前网络的特征向量的softmax置信度输出
exp_logits = self.expert(inputs) #专家网络的分类置信度输出
old_logits = self._old_network(inputs) #旧网络的分类置信度输出
# Distillation
dist_term = _KD_loss(logits[:,
self._known_classes:], exp_logits,
T1) #蒸馏损失 当前网络对新类别的分类结果与专家网络贴近
# Retrospection
retr_term = _KD_loss(logits[:, :self._known_classes],
old_logits, T2) #记忆损失 当前网络对旧类别的分类结果与旧网络贴近
#旧网络是k1分类器 专家网络是k2分类器 当前网络是k1+k2分类器 损失分为两部分
loss = dist_term + retr_term
losses += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
train_acc = self._compute_accuracy(self._network, train_loader)
test_acc = self._compute_accuracy(self._network, test_loader)
info = 'Updated CNN => Epoch {}/{}, Loss {:.3f}, Train accy {:.3f}, Test accy {:.3f}'.format(
epoch + 1, epochs, losses / len(train_loader), train_acc,
test_acc)
prog_bar.set_description(info)
logging.info(info)
def _train_expert(self, train_loader, test_loader):
self.expert = IncrementalNet(self.convnet_type, False) #专家网络
self.expert.update_fc(self.task_size)
self.expert.to(self._device)
optimizer = optim.SGD(self.expert.parameters(),
lr=lrate_expert,
momentum=0.9,
weight_decay=1e-5)
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer, milestones=milestones_expert, gamma=lrate_decay_expert)
prog_bar = tqdm(range(epochs_expert))
for _, epoch in enumerate(prog_bar):
self.expert.train()
losses = 0.
for i, (_, inputs, targets) in enumerate(train_loader):
inputs, targets = inputs.to(
self._device), (targets - self._known_classes).to(
self._device)
#专家网络只分k2种类，这里有k1+k2种，因此下标要减去k1
logits = self.expert(inputs)
loss = F.cross_entropy(logits, targets)
losses += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
train_acc = self._compute_accuracy(self.expert, train_loader,
self._known_classes)
test_acc = self._compute_accuracy(self.expert, test_loader,
self._known_classes)
info = 'Expert CNN => Epoch {}/{}, Loss {:.3f}, Train accy {:.3f}, Test accy {:.3f}'.format(
epoch + 1, epochs_expert, losses / len(train_loader),
train_acc, test_acc)
prog_bar.set_description(info)
logging.info(info)
def _compute_accuracy(self, model, loader, offset=0):
model.eval()
correct, total = 0, 0
for i, (_, inputs, targets) in enumerate(loader):
inputs = inputs.to(self._device)
targets -= offset
with torch.no_grad():
outputs = model(inputs)
predicts = torch.max(outputs, dim=1)[1]
correct += (predicts.cpu() == targets).sum()
total += len(targets)
return np.around(tensor2numpy(correct) / total, decimals=3)