def validation(self, dataloader, task_n=''):
# this might possibly change for other incremental scenario
# This function doesn't distinguish tasks.
batch_timer = Timer()
acc = AverageMeter()
losses = AverageMeter()
acc = AverageMeter()
batch_timer.tic()
orig_mode = self.training
self.eval()
for i, (input, target, task) in enumerate(dataloader):
if self.gpu:
with torch.no_grad():
input = input.cuda()
target = target.cuda()
output = self.predict(input, task_n)
loss = self.criterion(output, target, task)
losses.update(loss, input.size(0))
# Summarize the performance of all tasks, or 1 task, depends on dataloader.
# Calculated by total number of data.
acc = accumulate_acc(output, target, task, acc)
self.train(orig_mode)
self.log(' * Val Acc {acc.avg:.3f}, Total time {time:.2f}'.format(
acc=acc, time=batch_timer.toc()))
return acc, losses
def validation(self, dataloader):
# This function doesn't distinguish tasks.
batch_timer = Timer()
acc = AverageMeter()
batch_timer.tic()
orig_mode = self.training
self.eval()
for i, (input, target, task) in enumerate(dataloader):
if self.gpu:
with torch.no_grad():
input = input.cuda()
target = target.cuda()
output = self.predict(input)
# Summarize the performance of all tasks, or 1 task, depends on dataloader.
# Calculated by total number of data.
acc = accumulate_acc(output, target, task, acc)
self.train(orig_mode)
self.log(' * Val Acc {acc.avg:.3f}, Total time {time:.2f}'
.format(acc=acc,time=batch_timer.toc()))
return acc.avg
def validation(self, dataloader, val_name=''):
# This function doesn't distinguish tasks.
batch_timer = Timer()
acc = AverageMeter()
batch_timer.tic()
losses = AverageMeter()
orth_losses = AverageMeter()
sp_losses = AverageMeter()
orig_mode = self.training
self.model.eval()
for i, (input, target, task) in enumerate(dataloader):
# print(task)
if self.gpu:
with torch.no_grad():
input = input.cuda()
target = target.cuda()
output = self.predict(input)
# this works only for the current SVD change it to be more genric
loss, orth_loss, sp_loss = self.criterion(output, target, task)
# Summarize the performance of all tasks, or 1 task, depends on dataloader.
# Calculated by total number of data.
losses.update(loss, input.size(0))
orth_losses.update(orth_loss, input.size(0))
sp_losses.update(sp_loss, input.size(0))
acc = accumulate_acc(output, target, task, acc)
self.model.train()
self.log(' * Val Acc {acc.avg:.3f}, Total time {time:.2f}'
.format(acc=acc,time=batch_timer.toc()))
return acc, losses, orth_losses, sp_losses
def validation(self, dataloader):
# This function doesn't distinguish tasks.
batch_timer = Timer()
val_acc = AverageMeter()
losses = AverageMeter()
batch_timer.tic()
# self.hypermodel.eval()
self.model.eval()
for i, (inputs, target, task) in enumerate(dataloader):
if self.config['gpu']:
with torch.no_grad():
inputs = inputs.cuda()
target = target.cuda()
output = self.model.forward(inputs)
loss = self.criterion(output,
target,
task,
regularization=False)
losses.update(loss, inputs.size(0))
for t in output.keys():
output[t] = output[t].detach()
# Summarize the performance of all tasks, or 1 task, depends on dataloader.
# Calculated by total number of data.
val_acc = accumulate_acc(output, target, task, val_acc)
self.log(' * Val Acc {acc.avg:.3f}, Total time {time:.2f}'.format(
acc=val_acc, time=batch_timer.toc()))
self.log(' * Val loss {loss.avg:.3f}, Total time {time:.2f}'.format(
loss=losses, time=batch_timer.toc()))
return val_acc.avg
def train(epoch, train_loader, learner, args):
# This function optimize the objective
# Initialize all meters
data_timer = Timer()
batch_timer = Timer()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
confusion = Confusion(args.out_dim)
# Setup learner's configuration
print('\n\n==== Epoch:{0} ===='.format(epoch))
learner.train()
learner.step_schedule(epoch)
# The optimization loop
data_timer.tic()
batch_timer.tic()
if args.print_freq > 0: # Enable to print mini-log
print('Itr |Batch time |Data Time |Loss')
for i, (input, target) in enumerate(train_loader):
data_time.update(data_timer.toc()) # measure data loading time
# Prepare the inputs
if args.use_gpu:
input = input.cuda()
target = target.cuda()
train_target, eval_target = prepare_task_target(input, target, args)
# Optimization
loss, output = learner.learn(input, train_target)
# Update the performance meter
confusion.add(output, eval_target)
# Measure elapsed time
batch_time.update(batch_timer.toc())
data_timer.toc()
# Mini-Logs
losses.update(loss, input.size(0))
if args.print_freq > 0 and ((i % args.print_freq == 0) or
(i == len(train_loader) - 1)):
print('[{0:6d}/{1:6d}]\t'
'{batch_time.val:.4f} ({batch_time.avg:.4f})\t'
'{data_time.val:.4f} ({data_time.avg:.4f})\t'
'{loss.val:.3f} ({loss.avg:.3f})'.format(
i,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses))
# Loss-specific information
if args.loss == 'CE':
print('[Train] ACC: ', confusion.acc())
elif args.loss in ['KCL', 'MCL']:
args.cluster2Class = confusion.optimal_assignment(
train_loader.num_classes
) # Save the mapping in args to use in eval
if args.out_dim <= 20: # Avoid to print a large confusion matrix
confusion.show()
print('Clustering scores:', confusion.clusterscores())
print('[Train] ACC: ', confusion.acc())
elif args.loss == 'DPS':
confusion.show(width=15,
row_labels=['GT_dis-simi', 'GT_simi'],
column_labels=['Pred_dis-simi', 'Pred_simi'])
print('[Train] similar pair f1-score:',
confusion.f1score(1)) # f1-score for similar pair (label:1)
print('[Train] dissimilar pair f1-score:', confusion.f1score(0))
def learn_batch(self, train_loader, val_loader=None):
if self.reset_optimizer: # Reset optimizer before learning each task
self.log('Optimizer is reset !')
self.init_optimizer()
# for epoch in range(self.config['schedule'][-1]):
for epoch in range(self.config.nepoch):
data_timer = Timer()
batch_timer = Timer()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
acc = AverageMeter()
# Config the model and optimizer
self.log('Epoch:{0}'.format(epoch))
self.model.train()
if self.config.scheduler :
self.scheduler.step(epoch)
for param_group in self.optimizer.param_groups:
self.log('LR:',param_group['lr'])
# Learning with mini-batch
data_timer.tic()
batch_timer.tic()
self.log('Itr\t\tTime\t\t Data\t\t Loss\t\tAcc')
for i, (input, target, task) in enumerate(train_loader):
data_time.update(data_timer.toc()) # measure data loading time
if self.gpu:
input = input.cuda()
target = target.cuda()
loss, output = self.update_model(input, target, task)
input = input.detach()
target = target.detach()
# measure accuracy and record loss
acc = accumulate_acc(output, target, task, acc)
losses.update(loss, input.size(0))
batch_time.update(batch_timer.toc()) # measure elapsed time
data_timer.toc()
# Add wandb logging
# log_dict = dict()
# wandb.log(log_dict)
if ((self.config['print_freq']>0) and (i % self.config['print_freq'] == 0)) or (i+1)==len(train_loader):
self.log('[{0}/{1}]\t'
'{batch_time.val:.4f} ({batch_time.avg:.4f})\t'
'{data_time.val:.4f} ({data_time.avg:.4f})\t'
'{loss.val:.3f} ({loss.avg:.3f})\t'
'{acc.val:.2f} ({acc.avg:.2f})'.format(
i, len(train_loader), batch_time=batch_time,
data_time=data_time, loss=losses, acc=acc))
self.log(' * Train Acc {acc.avg:.3f}'.format(acc=acc))
# Evaluate the performance of current task
if val_loader != None:
self.validation(val_loader)
def learn_batch(self,
train_loader,
val_loader=None,
curr_global_decoder=None,
local_vae=None,
class_table=None,
global_classes_list=None,
task_id=None,
n_codes=None,
global_n_codes=None,
new_task_data_processing='original'):
if self.reset_optimizer: # Reset optimizer before learning each task
self.log('Optimizer is reset!')
self.init_optimizer()
print("Classifier: learning new task in '{}' new data processing mode".
format(new_task_data_processing))
if new_task_data_processing == 'original':
process_through_local_vae = False
train_only_on_generated_data = False
elif new_task_data_processing == 'original_through_vae':
process_through_local_vae = True
train_only_on_generated_data = False
elif new_task_data_processing == 'generated':
process_through_local_vae = False
train_only_on_generated_data = True
else:
raise ValueError(
"'new_task_data_processing' argument is invalid: '{}'. "
"Valid values are: 'original', 'original_through_vae', 'generated."
)
if self.score_generated_images_by_freezed_classifier:
frozen_model = copy.deepcopy(self.model)
frozen_model.eval()
train_accs = []
val_accs = []
for epoch in range(self.config['base_schedule'][-1]):
data_timer = Timer()
batch_timer = Timer()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
acc = AverageMeter()
# Config the model and optimizer
self.log('Epoch:{0}'.format(epoch))
self.model.train()
for param_group in self.optimizer.param_groups:
self.log('LR:', param_group['lr'])
# Learning with mini-batch
data_timer.tic()
batch_timer.tic()
self.log('Itr\t\t Time\t\t\t Data\t\t\t Loss\t\t\t Acc')
current_start = 0
if train_only_on_generated_data:
n_tasks_to_generate = task_id + 1
else:
n_tasks_to_generate = task_id
if not train_only_on_generated_data and (task_id == 0):
starting_points_fixed = np.array([[0]])
else:
starting_points = []
for prev_task_id in range(n_tasks_to_generate):
starting_points.append(
np.random.permutation(
np.array(
range(
math.ceil(global_n_codes[prev_task_id] /
train_loader.batch_size)))))
max_len = max([len(repeats) for repeats in starting_points])
starting_points_fixed = []
for points in starting_points:
starting_points_fixed.append(
np.pad(points, [0, max_len - len(points)],
mode="reflect"))
starting_points_fixed = np.array(starting_points_fixed)
for i, (orig_input, orig_target,
orig_task) in enumerate(train_loader):
data_time.update(data_timer.toc()) # measure data loading time
batch_size = len(orig_task)
# generate data so every task is equally represented
with torch.no_grad():
if process_through_local_vae:
orig_input, orig_target, _ = vae_utils.generate_current_data(
local_vae.decoder, task_id, batch_size,
current_start, global_classes_list, n_codes,
global_n_codes)
generate_impl = vae_utils.generate_previous_data
if train_only_on_generated_data:
# generate data from previous tasks and the current one
generate_impl = vae_utils.generate_previous_and_current_data
# clear original data
orig_input, orig_target = torch.Tensor(), torch.Tensor(
)
if train_only_on_generated_data or (task_id > 0):
gen_input, gen_target_orig, _ = generate_impl(
curr_global_decoder, task_id, batch_size,
starting_points_fixed[:, current_start] *
batch_size, global_classes_list, n_codes,
global_n_codes)
current_start += 1
else:
gen_input = torch.Tensor()
gen_target_orig = torch.Tensor()
if self.score_generated_images_by_freezed_classifier:
if task_id > 0:
gen_target = frozen_model.forward(
gen_input[:-batch_size])
gen_target = gen_target['All']
gen_target = F.softmax(gen_target, 1)
if train_only_on_generated_data:
targets_orig = self.one_hot_targets(
gen_target_orig[-batch_size:]).to(
local_vae.device)
gen_target = torch.cat(
[gen_target, targets_orig])
else:
targets_orig = self.one_hot_targets(
orig_target).to(local_vae.device)
gen_target = torch.cat(
[gen_target, targets_orig])
else:
gen_target = gen_target_orig
gen_target = self.one_hot_targets(
gen_target, self.model.n_classes)
else:
gen_target = self.one_hot_targets(
gen_target, self.model.n_classes)
orig_target = self.one_hot_targets(orig_target,
self.model.n_classes)
if self.gpu:
orig_input = orig_input.cuda()
orig_target = orig_target.cuda()
gen_input = gen_input.cuda()
gen_target = gen_target.cuda()
# merge original and generated data
multi_input = torch.cat((orig_input, gen_input), 0)
multi_target = torch.cat((orig_target, gen_target), 0)
# zip and shuffle
multibatch = list(zip(multi_input, multi_target))
random.shuffle(multibatch)
# iterate over batches in multibatch
multibatch_parted = zip(*(iter(multibatch), ) * batch_size)
for part in multibatch_parted:
input, target = zip(*part)
# convert tuples of tensors into one tensor
input = torch.stack(input)
target = torch.stack(target)
loss, output = self.update_model(input, target, None)
input = input.detach()
target = target.detach()
# measure accuracy and record loss
acc = accumulate_acc(output, target, None, acc)
losses.update(loss, input.size(0))
batch_time.update(
batch_timer.toc()) # measure elapsed time
data_timer.toc()
if ((self.config['base_print_freq'] > 0) and
(i % self.config['base_print_freq']
== 0)) or (i + 1) == len(train_loader):
self.log(
'[{0}/{1}]\t'
'{batch_time.val:.4f} ({batch_time.avg:.4f})\t'
'{data_time.val:.4f} ({data_time.avg:.4f})\t'
'{loss.val:.3f} ({loss.avg:.3f})\t'
'{acc.val:.2f} ({acc.avg:.2f})'.format(
i,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=acc))
train_accs.append(acc.avg)
self.log(
' * Train on {} original batches, Acc {acc.avg:.3f}'.format(
len(train_loader), acc=acc))
# Evaluate the performance of current task
if val_loader != None:
val_accs.append(self.validation(val_loader))
print("All epochs ended")
def train_(self, epochs, finetune=False):
str_ = 'pretrain'
self.str_ = str_
if finetune:
self.switch_finetune()
str_ = 'finetune'
self.str_ = str_
for epoch in range(epochs):
data_timer = Timer()
batch_timer = Timer()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
acc = AverageMeter()
self.model.train()
self.scheduler.step(epoch)
if self.config['train_between']:
if epoch == self.config['schedule'][0]:
for param in self.model.parameters():
param.requires_grad = True
#self.config['lr'] = 0.01
self.config['weight_decay'] = 5e-4
self.init_optimizer()
if self.config['switch_all']:
if epoch == self.config['switch_all']:
self.config['weight_decay'] = 5e-3
for param in self.model.parameters():
param.requires_grad = True
self.init_optimizer()
#self.config['lr'] = 0.01
for param_group in self.optimizer.param_groups:
self.log('LR:', param_group['lr'])
self.log('Itr\t\tTime\t\t Data\t\t Loss\t\tAcc')
self.log('{0} Epoch:{1}'.format(str_, epoch))
data_timer.tic()
batch_timer.tic()
for i, (input, target) in enumerate(self.train_loader):
self.model.train()
data_time.update(data_timer.toc()) # measure data loading time
if self.gpu:
input = input.cuda()
target = target.cuda()
loss, output = self.update_model(input, target)
input = input.detach()
target = target.detach()
# measure accuracy and record loss
acc = self.accumulate_acc(output, target, acc)
losses.update(loss, input.size(0))
batch_time.update(batch_timer.toc()) # measure elapsed time
data_timer.toc()
self.n_iter = (epoch) * len(self.train_loader) + i
self.writer.add_scalar(str_ + '/Loss_train', losses.avg,
self.n_iter)
self.writer.add_scalar(str_ + '/Acc_train', acc.avg,
self.n_iter)
# if ((self.config['print_freq']>0) and (i % self.config['print_freq'] == 0)) or (i+1)==len(train_loader):
self.log('[{0}/{1}]\t'
'{batch_time.val:.4f} ({batch_time.avg:.4f})\t'
'{data_time.val:.4f} ({data_time.avg:.4f})\t'
'{loss.val:.3f} ({loss.avg:.3f})\t'
'{acc.val:.2f} ({acc.avg:.2f})'.format(
i,
len(self.train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=acc))
acc_v, loss_v = self.validation(self.test_loader)
self.writer.add_scalar(str_ + '/Loss_test', loss_v.avg,
self.n_iter)
self.writer.add_scalar(str_ + '/Acc_test', acc_v.avg, self.n_iter)
if epoch % self.save_after == 0 and epoch != 0:
self.save_model(str_ + str(epoch))
def validation(self, test_loader, from_train=1):
# this might possibly change for other incremental scenario
# This function doesn't distinguish tasks.
batch_timer = Timer()
acc = AverageMeter()
losses = AverageMeter()
acc_5 = AverageMeter()
acc_class = [
AverageMeter()
for i in range(len(self.train_loader.dataset.class_list))
] #[AverageMeter()] * len(self.train_loader.dataset.class_list)
acc_class_5 = [
AverageMeter()
for i in range(len(self.train_loader.dataset.class_list))
]
batch_timer.tic()
orig_mode = self.training
self.eval()
for i, (input, target) in enumerate(test_loader):
if self.gpu:
with torch.no_grad():
input = input.cuda()
target = target.cuda()
output = self.forward(input)
loss = self.criterion(output, target)
losses.update(loss, input.size(0))
# Summarize the performance of all tasks, or 1 task, depends on dataloader.
# Calculated by total number of data.
t_acc, acc_class = accuracy(
output, target, topk=(1, ), avg_meters=acc_class
) #self.accumulate_acc(output, target, acc)
t_acc_5, acc_class_5 = accuracy(output,
target,
topk=(5, ),
avg_meters=acc_class_5)
# import pdb; pdb.set_trace()
acc.update(t_acc, len(target))
acc_5.update(t_acc_5, len(target))
class_list = self.train_loader.dataset.class_list.inverse
acc_cl_1 = {}
acc_cl_5 = {}
#from accuracies obtained create inst size based accuracies
inst_clss_lst = self.train_loader.dataset.class_inst_list
# import pdb; pdb.set_trace()
for ins_clss_, insts in inst_clss_lst.items():
cls_sum = sum([acc_class[inst].sum for inst in insts])
cls_cnt = sum([acc_class[inst].count for inst in insts])
if cls_cnt == 0:
import pdb
pdb.set_trace()
inst_avg = cls_sum / cls_cnt
self.writer.add_scalar(self.str_ + '/Acc_1_{}'.format(ins_clss_),
inst_avg, self.n_iter)
cls_sum_5 = sum([acc_class_5[inst].sum for inst in insts])
cls_cnt_5 = sum([acc_class_5[inst].count for inst in insts])
inst_avg_5 = cls_sum_5 / cls_cnt_5
self.writer.add_scalar(self.str_ + '/Acc_5_{}'.format(ins_clss_),
inst_avg_5, self.n_iter)
for idx, cl_ in class_list.items():
acc_cl_1[cl_] = [
acc_class[idx].avg, acc_class[idx].sum, acc_class[idx].count
]
acc_cl_5[cl_] = [
acc_class_5[idx].avg, acc_class_5[idx].sum,
acc_class_5[idx].count
]
# self.log(' * Val Acc {acc.avg:.3f} for class {cls}, {acc.sum} / {acc.count} '
# .format(acc=acc_class[idx], cls=cl_))
self.train(orig_mode)
self.log(' * Val Acc {acc.avg:.3f}, Total time {time:.2f}'.format(
acc=acc, time=batch_timer.toc()))
if from_train:
return acc, losses
else:
return acc, acc_5, acc_cl_1, acc_cl_5, losses
def learn_batch(self, train_loader, val_loader=None):
if self.reset_optimizer: # Reset optimizer before learning each task
self.log('Optimizer is reset!')
self.init_optimizer()
self.model.zero_grad()
# epoch_iterator = tqdm(train_loader, desc="Iteration", disable=False)
# global_step = 0
# epochs_trained = 0
# steps_trained_in_current_epoch = 0
# train_iterator = trange(
# epochs_trained, int(self.config['schedule'][-1]), desc="Epoch", disable=False,
# )
# for _ in train_iterator:
for epoch in range(self.config['schedule'][-1]):
data_timer = Timer()
batch_timer = Timer()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
acc = AverageMeter()
# Config the model and optimizer
self.log('Epoch:{0}'.format(epoch))
self.model.train()
self.scheduler.step(epoch)
for param_group in self.optimizer.param_groups:
self.log('LR:', param_group['lr'])
# Learning with mini-batch
data_timer.tic()
batch_timer.tic()
self.log('Itr\t\tTime\t\t Data\t\t Loss\t\tAcc')
for i, (inputs_1, inputs_2, inputs_3,
target) in enumerate(train_loader):
#changed here for creating 2d tensor
input = torch.stack([inputs_1, inputs_2, inputs_3]).reshape(
(8, -1))
task = 'mrpc'
data_time.update(data_timer.toc()) # measure data loading time
if self.gpu:
input = input.cuda()
target = target.cuda()
loss, output = self.update_model(input, target, task)
input = input.detach()
target = target.detach()
# measure accuracy and record loss
acc = accumulate_acc(output, target, task, acc)
losses.update(loss, input.size(0))
batch_time.update(batch_timer.toc()) # measure elapsed time
data_timer.toc()
if ((self.config['print_freq'] > 0) and
(i % self.config['print_freq']
== 0)) or (i + 1) == len(train_loader):
self.log('[{0}/{1}]\t'
'{batch_time.val:.4f} ({batch_time.avg:.4f})\t'
'{data_time.val:.4f} ({data_time.avg:.4f})\t'
'{loss.val:.3f} ({loss.avg:.3f})\t'
'{acc.val:.2f} ({acc.avg:.2f})'.format(
i,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=acc))
self.log(' * Train Acc {acc.avg:.3f}'.format(acc=acc))
# Evaluate the performance of current task
if val_loader != None:
self.validation(val_loader)
#from regularization
# 2.Backup the weight of current task
task_param = {}
for n, p in self.params.items():
task_param[n] = p.clone().detach()
# 3.Calculate the importance of weights for current task
importance = self.calculate_importance(train_loader)
# Save the weight and importance of weights of current task
self.task_count += 1
if self.online_reg and len(self.regularization_terms) > 0:
# Always use only one slot in self.regularization_terms
self.regularization_terms[1] = {
'importance': importance,
'task_param': task_param
}
else:
# Use a new slot to store the task-specific information
self.regularization_terms[self.task_count] = {
'importance': importance,
'task_param': task_param
}
def learn_batch(self, train_loader, val_loader=None):
if self.reset_optimizer: # Reset optimizer before learning each task
self.log('Optimizer is reset!')
self.init_optimizer()
schedule = self.schedule_stack.pop()
for epoch in range(schedule):
data_timer = Timer()
batch_timer = Timer()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
acc = AverageMeter()
robust_err, robust_loss = -1, -1
# Config the model and optimizer
self.log('Epoch:{0}'.format(epoch))
self.model.train()
for param_group in self.optimizer.param_groups:
self.log('LR:', param_group['lr'])
# Learning with mini-batch
data_timer.tic()
batch_timer.tic()
for i, (inputs, target, task) in enumerate(train_loader):
data_time.update(data_timer.toc()) # measure data loading time
if self.gpu:
inputs = inputs.cuda()
target = target.cuda()
loss, robust_err, robust_loss, output = self.update_model(
inputs, target, task)
inputs = inputs.detach()
target = target.detach()
self.tb.add_scalar(f"Loss/train - task {self.current_task}",
loss, epoch)
self.tb.add_scalar(
f"Robust error/train - task {self.current_task}",
robust_err, epoch)
# measure accuracy and record loss
acc = accumulate_acc(output, target, task, acc)
losses.update(loss, inputs.size(0))
batch_time.update(batch_timer.toc()) # measure elapsed time
data_timer.toc()
self.log(' * Train Acc {acc.avg:.3f}, Loss {loss.avg:.3f}'.format(
loss=losses, acc=acc))
self.log(
f" * robust loss: {robust_loss:.10f} robust error: {robust_err:.10f}"
)
# self.log(f" * model: {self.model.features_loss_term}")
# Evaluate the performance of current task
if val_loader is not None:
self.validation(val_loader)
self.scheduler.step()
def learn_batch(self,
train_loader,
val_loader=None,
epochs=[0, 40],
task_n=''):
itrs = 0
if epochs[
0] == 0: # Only for the first epoch of each task or classReset optimizer before incrementally learning
self.task_num += 1
if self.reset_optimizer:
self.log('Optimizer is reset!')
self.freeze(task_n)
self.init_optimizer(params=filter(lambda p: p.requires_grad,
self.model.parameters()))
data_timer = Timer()
batch_timer = Timer()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
acc = AverageMeter()
for epoch in range(epochs[0], epochs[1]):
self.writer = SummaryWriter(log_dir="runs/" + self.exp_name)
self.scheduler.step(epoch)
# # # Config the model and optimizer
self.log('Epoch:{0}'.format(epoch))
self.model.train()
for param_group in self.optimizer.param_groups:
self.log('LR:', param_group['lr'])
# # Learning with mini-batch
data_timer.tic()
batch_timer.tic()
self.log('Itr\t\tTime\t\t Data\t\t Loss\t\tAcc')
for i, (input, target, task) in enumerate(train_loader):
self.n_iter = (epoch) * len(train_loader) + i + 1
data_time.update(data_timer.toc()) # measure data loading time
if self.gpu:
input = input.cuda()
target = target.cuda()
loss, output = self.update_model(input, target, task, task_n)
input = input.detach()
target = target.detach()
# measure accuracy and record loss
acc = accumulate_acc(output, target, task, acc)
losses.update(loss, input.size(0))
self.writer.add_scalar(
'Run' + str(self.config['run_num']) + '/Loss/train' +
task_n, losses.avg, self.n_iter)
self.writer.add_scalar(
'Run' + str(self.config['run_num']) + '/Accuracy/train' +
task_n, acc.avg, self.n_iter)
batch_time.update(batch_timer.toc()) # measure elapsed time
data_timer.toc()
if ((self.config['print_freq'] > 0) and
(i % self.config['print_freq']
== 0)) or (i + 1) == len(train_loader):
self.log('[{0}/{1}]\t'
'{batch_time.val:.4f} ({batch_time.avg:.4f})\t'
'{data_time.val:.4f} ({data_time.avg:.4f})\t'
'{loss.val:.3f} ({loss.avg:.3f})\t'
'{acc.val:.2f} ({acc.avg:.2f})'.format(
i,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=acc))
self.log(' * Train Acc {acc.avg:.3f}'.format(acc=acc))
# Evaluate the performance of current task
if val_loader != None:
acc_val, loss_val = self.validation(val_loader, task_n)
self.writer.add_scalar(
'Run' + str(self.config['run_num']) + 'Loss/test' + task_n,
loss_val.avg, self.n_iter)
self.writer.add_scalar(
'Run' + str(self.config['run_num']) + 'Accuracy/test' +
task_n, acc_val.avg, self.n_iter)
self.writer.close()
def learn_batch(self, train_loader, val_loader=None, task_name=None):
if self.reset_optimizer: # Reset optimizer before learning each task
self.log('Optimizer is reset!')
self.init_optimizer()
for epoch in range(self.config['schedule'][-1]):
data_timer = Timer()
batch_timer = Timer()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
acc = AverageMeter()
orth_losses = AverageMeter()
sp_losses = AverageMeter()
# Config the model and optimizer
self.log('Epoch:{0}'.format(epoch))
self.model.train()
#self.model.no_tbn()
self.scheduler.step(epoch)
for param_group in self.optimizer.param_groups:
self.log('LR:',param_group['lr'])
# Learning with mini-batch
data_timer.tic()
batch_timer.tic()
self.log('Itr\t\tTime\t\t Data\t\t Loss\t\tAcc')
task_n = 'blah'
for i, (input, target, task) in enumerate(train_loader):
# import pdb; pdb.set_trace()
self.n_iter = (epoch) * len(train_loader) + i + 1
task_n = task[0]
data_time.update(data_timer.toc()) # measure data loading time
if self.gpu:
input = input.cuda()
target = target.cuda()
loss, output, orth_loss, sp_loss = self.update_model(input, target, task, i)
# print(f' Loss is {loss.data}')
input = input.detach()
target = target.detach()
# measure accuracy and record loss
acc = accumulate_acc(output, target, task, acc)
losses.update(loss, input.size(0))
sp_losses.update(sp_loss, input.size(0))
orth_losses.update(orth_loss, input.size(0))
# print(f'updating loss, {losses.avg} {self.n_iter}')
self.writer.add_scalar( '/All_Losses/train' + task[0], losses.avg, self.n_iter)
self.writer.add_scalar( '/Orth_Loss/train' + task[0], orth_losses.avg, self.n_iter)
self.writer.add_scalar( '/Sparsity_Loss/train' + task[0], sp_losses.avg, self.n_iter)
self.writer.add_scalar('/Accuracy/train' + task[0], acc.avg, self.n_iter)
batch_time.update(batch_timer.toc()) # measure elapsed time
data_timer.toc()
# self.writer.add_scalar('Loss_train', losses.avg, self.n_iter)
# self.writer.add_scalar('Acc_train' + task_n, acc.avg, self.n_iter)
if ((self.config['print_freq']>0) and (i % self.config['print_freq'] == 0)) or (i+1)==len(train_loader):
self.log('[{0}/{1}]\t'
'{batch_time.val:.4f} ({batch_time.avg:.4f})\t'
'{data_time.val:.4f} ({data_time.avg:.4f})\t'
'{loss.val:.3f} ({loss.avg:.3f})\t'
'{acc.val:.2f} ({acc.avg:.2f})'.format(
i, len(train_loader), batch_time=batch_time,
data_time=data_time, loss=losses, acc=acc))
self.log(' * Train Acc {acc.avg:.3f}'.format(acc=acc))
# Evaluate the performance of current task
if val_loader != None:
v_acc, v_losses, v_orth_losses, v_sp_losses = self.validation(val_loader)
self.writer.add_scalar( '/All_Losses/val' + task[0], v_losses.avg, self.n_iter)
self.writer.add_scalar( '/Orth_Loss/val' + task[0], v_orth_losses.avg, self.n_iter)
self.writer.add_scalar( '/Sparsity_Loss/val' + task[0], v_sp_losses.avg, self.n_iter)
self.writer.add_scalar('/Accuracy/val' + task_n, v_acc.avg, self.n_iter)