# Clone model
net = meta_net.clone()
optimizer = get_optimizer(net, state)
# load state of base optimizer?
# Sample base task from Meta-Train
train = meta_train.get_random_task(args.classes, args.train_shots
or args.shots)
train_iter = make_infinite(DataLoader(train, args.batch, shuffle=True))
# Update fast net
loss = do_learning(net, optimizer, train_iter, args.iterations)
state = optimizer.state_dict() # save optimizer state
# Update slow net
meta_net.point_grad_to(net)
meta_optimizer.step()
# Meta-Evaluation
if meta_iteration % args.validate_every == 0:
print '\n\nMeta-iteration', meta_iteration
print '(started at {})'.format(args.start_meta_iteration)
print 'Meta LR', meta_lr
for (meta_dataset, mode) in [(meta_train, 'train'),
(meta_test, 'val')]:
train, test = meta_dataset.get_random_task_split(
args.classes, train_K=args.shots, test_K=5) # is that 5 ok?
train_iter = make_infinite(
DataLoader(train, args.batch, shuffle=True))
DataLoader(train, args.train_batch, shuffle=True))
# Update fast net
loss = do_learning(net, optimizer, train_iter, args.train_iters)
state = optimizer.state_dict() # save optimizer state
batch_nets.append(net)
batch_params = [list(net.parameters()) for net in batch_nets]
batch_params = zip(*batch_params)
for mean_param, batch_param in zip(mean_net.parameters(), batch_params):
np_mean_param = np.mean([param.data.numpy() for param in batch_param],
axis=0)
mean_param.data = torch.FloatTensor(np_mean_param)
# Update slow net
meta_net.point_grad_to(mean_net)
meta_optimizer.step()
# Meta-Evaluation
if meta_iteration % args.validate_every == 0:
print '\n\nMeta-iteration', meta_iteration
print '(started at {})'.format(args.start_meta_iteration)
print 'Meta LR', meta_lr
for (meta_dataset, mode) in [(meta_train, 'train'),
(meta_test, 'val')]:
net = meta_net.clone()
optimizer = get_optimizer(net,
state) # do not save state of optimizer
train, test = meta_dataset.get_random_task_split(