class Trainer:
def __init__(self):
self._setup_logger()
torch.backends.cudnn.deterministic = True
self.device = torch.device(
"cuda:%d" % (args.gpu) if torch.cuda.is_available() else "cpu")
torch.cuda.set_device(args.gpu)
self.dataset = Cifar10()
self.saver = Saver(self.save_dir)
if args.arch == 'auto':
self.Net = lambda: AutoNetwork(args.init_channels, self.dataset.
num_classes, args.layers,
nn.CrossEntropyLoss())
elif args.arch == 'VGG11':
self.Net = lambda: VGG('VGG11', self.dataset.num_classes,
nn.CrossEntropyLoss())
elif args.arch == 'mobilenet':
self.Net = lambda: MobileNetV2(self.dataset.num_classes,
nn.CrossEntropyLoss())
elif args.arch == 'simple':
self.Net = lambda: SimpleNetwork(self.dataset.num_classes,
nn.CrossEntropyLoss())
else:
raise Exception('Net not defined!')
def _setup_logger(self):
self.save_dir = os.path.join('./checkpoints', args.logdir)
if not os.path.isdir(self.save_dir):
os.makedirs(self.save_dir)
log_path = os.path.join(
self.save_dir,
datetime.datetime.now().strftime('%Y-%m-%d-%H-%M-%S'))
self.logger = L.create_logger(args.logdir, log_path)
for arg in vars(args):
self.logger.info("%-25s %-20s" % (arg, getattr(args, arg)))
def test_beta(self, optimizee):
self.logger.info('Testing beta')
torch.manual_seed(2 * args.seed)
optimizee.reset_model_parameters()
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizee.optimizer,
float(args.max_epoch),
eta_min=args.learning_rate_min)
optimizee.optimizer.print_beta_greedy()
for epoch in range(0, args.max_test_epoch):
scheduler.step()
lr = scheduler.get_lr()[0]
self.train_epoch(epoch, lr, optimizee, False)
if (epoch + 1) % args.test_freq == 0:
acc = self.eval(optimizee.model)
def train(self):
optimizee = Optimizee(self.Net)
start_episode, start_epoch, best_test_acc = 0, 0, 0.
if args.checkpoint != '':
start_episode, start_epoch = self.saver.load_checkpoint(
optimizee, args.checkpoint)
# Hao: in my implementation, the optimizers for alpha and beta will preserve their states across episodes
for episode in range(start_episode, args.max_episodes):
torch.manual_seed(args.seed + episode)
optimizee.reset_model_parameters()
if args.arch == 'auto':
optimizee.reset_arch_parameters()
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizee.optimizer,
float(args.max_epoch),
eta_min=args.learning_rate_min,
last_epoch=start_epoch - 1)
for epoch in range(
start_epoch,
args.max_epoch): # loop over the dataset multiple times
scheduler.step()
lr = scheduler.get_lr()[0]
training_status = self.train_epoch(epoch, lr, optimizee)
if (epoch + 1) % args.test_freq == 0:
acc = self.eval(optimizee.model)
# if accuracy not above random, discard this example and start a new one
if acc <= 0.11 or training_status == False:
self.logger.warning(
'training_status false or acc too low, break')
break
checkpoint_path = self.save_dir + '/episode_{}_epoch_{}_acc_{}'.format(
str(episode), str(epoch), str(acc))
self.saver.save_checkpoint(optimizee, epoch, episode,
checkpoint_path)
return optimizee
def train_epoch(self, epoch, lr, optimizee, train_beta=True):
if train_beta:
optimizee.optimizer.beta_temp = np.interp(
epoch + 1, [0, args.max_epoch],
[args.min_beta_temp, args.max_beta_temp])
status = self._train_epoch(epoch, lr, optimizee)
self.saver.write_beta_embedding()
optimizee.optimizer.print_beta_greedy()
else:
optimizee.optimizer.beta_temp = args.max_beta_temp
status = self._train_epoch_fix_beta(epoch, lr, optimizee)
return status
def eval(self, model):
correct, total = 0, 0
with torch.no_grad():
for images, labels in self.dataset.test_loader:
images, labels = images.to(self.device), labels.to(self.device)
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
self.logger.info(
'Accuracy of the network on the 10000 test images: %d %%' %
(100 * correct / total))
return float(correct) / total
def _train_epoch(self, epoch, lr, optimizee):
losses, next_step_losses, meta_update_losses = 0.0, 0.0, 0.0
model_grad_norms, beta_grad_norms = 0., 0.
for i, data in enumerate(self.dataset.train_loader, 0):
x_train, y_train = data[0].to(self.device), data[1].to(self.device)
val_data = next(iter(self.dataset.val_loader))
x_val, y_val = val_data[0].to(self.device), val_data[1].to(
self.device)
# Derive \frac{\partial L}{\partial \theta}
optimizee.model.zero_grad()
loss = optimizee.model.forward_pass(x_train, y_train)
loss.backward()
model_grad_norms += nn.utils.clip_grad_norm_(
optimizee.model.parameters(), args.model_grad_norm)
param_updates = optimizee.optimizer.step()
# do a differentiable step of update over optimizee.symbolic_model
optimizee.differentiable_update(param_updates)
# check the next-step loss after update using the symbolic model
# so that it is on the computational graph
# TODO use validation data? not sure
next_step_loss = optimizee.symbolic_model.forward_pass(
x_train, y_train)
if (math.isnan(next_step_loss.item())):
self.logger.error('next step loss becomes NaN, break')
raise Exception
next_step_losses += next_step_loss.item()
losses += loss.item()
meta_update_losses += next_step_loss
# do a non-differentiable update over optimizee.model if next_step_loss is smaller
if (next_step_loss.item() <
loss.item()): # This is still important for performance
optimizee.update(param_updates)
# forsee bptt_steps then update beta
if (i + 1) % args.bptt_step == 0:
beta_grad_norms += optimizee.beta_step(
meta_update_losses).item()
# let saver save the grads for beta
with torch.no_grad():
self.saver.add_beta_grads([
b.grad.abs().mean() for b in optimizee.optimizer.beta
])
meta_update_losses = 0
optimizee.sync_symbolic_model()
if args.arch_training and (i + 1) % args.update_alpha_step == 0:
optimizee.alpha_step(
x_train, y_train, x_val, y_val,
lr) # TODO: make sure alpha step won't change weights
optimizee.sync_symbolic_model(skip_weights=True)
if (i + 1) % args.log_freq == 0: # print every 2000 mini-batches
beta_out = optimizee.optimizer.beta[-1].data.cpu().numpy()
alpha_out = optimizee.model.alphas_normal[-1].data.cpu().numpy(
) if args.arch_training else ''
self.logger.info(
'[%d, %5d] loss: %.3f/%.3f, next step loss: %.3f, beta[-1]/L2(g): %s/%s, alpha[-1]: %s'
% (epoch + 1, i + 1, losses / args.log_freq,
model_grad_norms / args.log_freq,
next_step_losses / args.log_freq, beta_out,
beta_grad_norms / args.log_freq, alpha_out))
losses, next_step_losses, model_grad_norms, beta_grad_norms = 0., 0., 0., 0.
self.saver.write_beta(optimizee.optimizer.beta)
return True
def _train_epoch_fix_beta(self, epoch, lr, optimizee):
losses, model_grad_norms = 0.0, 0.0
for i, data in enumerate(self.dataset.train_loader, 0):
x_train, y_train = data[0].to(self.device), data[1].to(self.device)
val_data = next(iter(self.dataset.val_loader))
x_val, y_val = val_data[0].to(self.device), val_data[1].to(
self.device)
optimizee.model.zero_grad()
loss = optimizee.model.forward_pass(x_train, y_train)
loss.backward()
model_grad_norms += nn.utils.clip_grad_norm_(
optimizee.model.parameters(), args.model_grad_norm)
param_updates = optimizee.optimizer.step(do_update=True)
losses += loss.item()
if args.use_darts_arch and args.train_alpha:
optimizee.alpha_step(x_train, y_train, x_val, y_val, lr)
if (i + 1) % args.log_freq == 0: # print every 2000 mini-batches
beta_out = optimizee.optimizer.beta[-1].data.cpu().numpy()
alpha_out = optimizee.model.alphas_normal[-1].data.cpu().numpy(
) if args.arch_training else 'None'
self.logger.info(
'[%d, %5d] loss: %.3f/%.3f, beta[-1]: %s, alpha[-1]: %s' %
(epoch + 1, i + 1, losses / args.log_freq,
model_grad_norms / args.log_freq, beta_out, alpha_out))
losses, model_grad_norms = 0., 0.
return True
