# In[45]:
train_loss = []
train_acc = []
val_loss = []
val_acc = []
best_acc = 0
best_model = None
best_generator = None
for epoch in range(num_epochs):
print("Epoch: {}/{}".format(epoch, num_epochs))
# train
model.train()
trange = tqdm(train_dataloader)
for batch in trange:
optim.zero_grad()
x, y = batch
x, y = x.to(device), y.to(device)
model_output = model(x)
prototype, query_samples = get_proto_query(model_output.cpu(),
y.cpu(),
n_aug=sample_M,
n_support=N_shot)
# loss, acc = loss_fn(query_samples, prototype, n_aug=sample_M)
loss, acc = cal_loss(query_samples, prototype, n_aug=sample_M)
loss.backward()
old_lr = optimizer.param_groups[0]['lr']
scheduler.step()
new_lr = optimizer.param_groups[0]['lr']
if epoch == 0:
print_log('\tInitial lr is {:.8f}\n'.format(old_lr), opts.log_file)
if new_lr != old_lr:
print_log(
'\tLR changes from {:.8f} to {:.8f} at epoch {:d}\n'.format(
old_lr, new_lr, epoch), opts.log_file)
tr_iter = iter(train_db)
for batch in tqdm(tr_iter):
net.train()
x, y = batch[0].to(opts.device), batch[1].to(opts.device)
# TODO use k_query or not?
loss, acc = loss_fn(net(x),
target=y,
n_support=opts.k_shot,
distance=opts.distance,
device=opts.device)
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_loss.append(loss.item())
train_acc.append(acc.item())
# ONE EPOCH ENDS