def main(controller, cosine_annealing_scheduler):
running_reward = 0
child = Child(dropout_rate=float(args.dropout),
use_auxiliary=args.use_auxiliary).to(device)
child_optimizer = optim.SGD(child.parameters(),
lr=0.05,
momentum=0.9,
weight_decay=1e-4,
nesterov=True)
cosine_lr_scheduler = cosine_annealing_scheduler(child_optimizer, lr=0.05)
for epoch in range(150):
cosine_lr_scheduler.step()
child = train_child(epoch, controller, child, child_optimizer)
torch.save(child.state_dict(), './save_model/child.pth')
torch.save(controller.state_dict(), './save_model/controller.pth')
if epoch < 150 - 1:
controller_model, running_reward = train_controller(
controller, child, running_reward)
final_model = model
final_acc = test_final_model(final_model)
print('training is done. Final accuarcy is ', final_acc)
if batch_idx % 10 == 0:
print('train child epoch : {} [{}/{}]| loss: {:.3f} | acc: {:.3f}'.
format(epoch, batch_idx, len(train_loader),
train_loss / (batch_idx + 1), 100. * correct / total))
child.eval()
test_loss = 0
correct = 0
total = 0
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(test_loader):
inputs = inputs.to(device)
targets = targets.to(device)
outputs, aux_outs = child(inputs, normal_arc, reduction_arc)
loss = criterion(outputs, targets)
test_loss += loss.item()
_, predicted = outputs.max(1)
total += targets.size(0)
correct += predicted.eq(targets).sum().item()
acc = 100 * correct / total
print('epoch : [{}/310]| acc: {:.3f}'.format(epoch, acc))
if acc >= best_acc:
print('best accuracy is ', best_acc)
best_acc = acc
torch.save(child.state_dict(), './save_model/best_child.pt')