#------ Train Classifier-------#
if i==0:
print('\n--------------------------------------')
print('Run #{} Task #{} --> Train Classifier'.format(
run, task))
print('--------------------------------------\n')
#---------------
# Iteration Loop
for it in range(args.disc_iters):
model = retrieve_replay_update(args,
model, opt, data, target, buffer, task, tr_loader,rehearse=task>0)
buffer.add_reservoir(data.cpu(), target.cpu(), None, task)
# ------------------------ eval ------------------------ #
model = model.eval()
eval_loaders = [('valid', val_loader), ('test', test_loader)]
for mode, loader_ in eval_loaders:
for task_t, te_loader in enumerate(loader_):
if task_t > task: break
LOG_temp = get_temp_logger(None, ['cls_loss', 'acc'])
# iterate over samples from task
for i, (data, target) in enumerate(te_loader):
if args.unit_test and i > 10: break
if args.cuda:
LOG_temp['acc'] += [
pred.eq(input_y.view_as(pred)).sum().item() /
pred.size(0)
]
LOG_temp['cls_loss'] += [loss.item()]
LOG_temp['gen_loss'] += [gen_loss.item()]
# add to buffer only during last epoch of task to get the best reconstructions
if epoch == args.n_epochs - 1:
# buffer.add_reservoir(data, target, None, task)
split = data.size(0)
if args.store_latents:
buffer.add_reservoir(hid[:split], target, None, task)
else:
buffer.add_reservoir(x_recon[:split], target, None,
task)
LOG_temp.print_('train', task)
# ------------------------ eval ------------------------ #
# We need to update the buffer representations
with torch.no_grad():
if prev_gen is not None and args.store_latents and args.update_buffer_hid:
# sample from buffer
for i in range(buffer.bx.data.size(0) // 64):
#------ Train Classifier-------#
if i==0:
print('\n--------------------------------------')
print('Run #{} Task #{} --> Train Classifier'.format(
run, task))
print('--------------------------------------\n')
#---------------
# Iteration Loop
for it in range(args.disc_iters):
model = retrieve_replay_update(args,
model, opt, data, target, buffer, task, tr_loader,rehearse=task>0)
buffer.add_reservoir(data, target, None, task)
# ------------------------ eval ------------------------ #
model = model.eval()
eval_loaders = [('valid', val_loader), ('test', test_loader)]
for mode, loader_ in eval_loaders:
for task_t, te_loader in enumerate(loader_):
if task_t > task: break
LOG_temp = get_temp_logger(None, ['cls_loss', 'acc'])
# iterate over samples from task
for i, (data, target) in enumerate(te_loader):
if args.unit_test and i > 10: break
if args.cuda:
