class TestDarts(unittest.TestCase):
def setUp(self):
data_path = os.path.join(os.path.dirname(__file__), 'mock_data.darts')
self._ins = Darts(data_path)
def test_common_prefix_search(self):
eq_(['star'], self._ins.common_prefix_search('star'))
eq_(['star', 'star wars'], self._ins.common_prefix_search('star wars'))
def test_longest_match(self):
eq_('star', self._ins.longest_match('star'))
eq_('star wars', self._ins.longest_match('star wars'))
def training_phase(args, device, genotype):
cudnn.benchmark = True
cudnn.enabled = True
torch.manual_seed(args.manual_seed)
torch.cuda.manual_seed(args.manual_seed)
C = args.init_channels
num_cells = args.num_cells
num_nodes = args.num_nodes
num_classes = args.num_classes
criterion = nn.CrossEntropyLoss()
logging.info("args = %s", args)
if args.gpus == "all":
gpus = None
else:
gpus = [int(gpu.strip()) for gpu in args.gpus.split(",")]
train_transform, test_transform = utils.data_transform_cifar10()
train_data = torchvision.datasets.CIFAR10(root=args.data, train=True, transform=train_transform, download=True)
test_data = torchvision.datasets.CIFAR10(root=args.data, train=False, transform=test_transform)
darts = Darts(C, num_cells, num_nodes, num_classes, criterion, found_genotype=genotype)
darts = nn.DataParallel(darts, device_ids=gpus).to(device)
weights_optimizer = torch.optim.SGD(darts.parameters(),
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay)
darts.module.set_optimizers(None, weights_optimizer)
train_queue = torch.utils.data.DataLoader(
train_data, batch_size=args.batch_size,
shuffle=True, pin_memory=True)
test_queue = torch.utils.data.DataLoader(test_data, batch_size=args.batch_size, shuffle=True, pin_memory=True)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(weights_optimizer, float(args.epochs), eta_min=args.learning_rate_min)
for epoch in range(args.epochs):
logging.info("EPOCH %d", epoch)
lr = weights_optimizer.param_groups[0]["lr"]
logging.info('Learning rate %e', lr)
# validation
valid_acc, valid_loss = evaluate(test_queue, darts, device)
logging.info('Validation Accuracy: %.2f', valid_acc)
logging.info("Validation Loss: %.4f", valid_loss)
# training
train_acc, train_loss = train(train_queue, darts, device)
logging.info('Train Accuracy: %.2f', train_acc)
logging.info("Train Loss: %.4f", train_loss)
utils.save(darts, os.path.join(args.save, 'weights%d.pt' % epoch))
scheduler.step()
logging.info("-------------------------------------------------------------------")
def setUp(self):
data_path = os.path.join(os.path.dirname(__file__), 'mock_data.darts')
self._ins = Darts(data_path)
