default='rmsprop',
help="Name of optimizer (rmsprop or sgd)")
parser.set_defaults(batch_size=4, num_iterations=10000000, iter_interval=2000)
args = parser.parse_args()
# Set the random seed
np.random.seed(1)
# Number of outputs of last layer.
ax.Y.length = 1000
ax.N.length = args.batch_size
# Build AEON data loader objects
train_set, valid_set = make_aeon_loaders(
train_manifest=args.train_manifest_file,
valid_manifest=args.valid_manifest_file,
batch_size=args.batch_size,
train_iterations=args.num_iterations,
dataset='i1k',
datadir=args.image_dir)
inputs = train_set.make_placeholders(include_iteration=True)
# Input size is 299 x 299 x 3
image_size = 299
# Build the network
inception = inception.Inception(mini=args.mini)
# Declare the optimizer
if args.optimizer_name == 'sgd':
learning_rate_policy = {
'name': 'schedule',
'--stage_depth',
type=int,
default=2,
help='depth of each stage (network depth will be 9n+2)')
parser.add_argument('--use_aeon',
action='store_true',
help='whether to use aeon dataloader')
args = parser.parse_args()
np.random.seed(args.rng_seed)
# Create the dataloader
if args.use_aeon:
from data import make_aeon_loaders
train_set, valid_set = make_aeon_loaders(args.data_dir,
args.batch_size,
args.num_iterations)
else:
from ngraph.frontends.neon import ArrayIterator # noqa
from ngraph.frontends.neon import CIFAR10 # noqa
train_data, valid_data = CIFAR10(args.data_dir).load_data()
train_set = ArrayIterator(train_data,
args.batch_size,
total_iterations=args.num_iterations)
valid_set = ArrayIterator(valid_data, args.batch_size)
# we need to ask the dataset to create an iteration
# placeholder for our learning rate schedule
inputs = train_set.make_placeholders(include_iteration=True)
ax.Y.length = 10
if device_backend == 'hetr':
if 'HETR_AEON_IP' not in os.environ or 'HETR_AEON_PORT' not in os.environ:
raise ValueError(
'To run hetr with more than one device, you need to set an ip address \
for the HETR_AEON_IP environment variable and a port number for the HETR_AEON_PORT \
environment variable')
aeon_address = os.environ['HETR_AEON_IP']
aeon_port = int(os.environ['HETR_AEON_PORT'])
# Create training and validation dataset objects
train_set, valid_set = make_aeon_loaders(args.data_dir,
args.batch_size,
args.num_iterations,
dataset=args.dataset,
num_devices=args.num_devices,
device=device_backend,
split_batch=True,
address=aeon_address,
port=aeon_port)
print("Completed loading " + args.dataset + " dataset")
# Make input_ops or placeholders depending on single device or multi device compute
input_ops_train = train_set.make_input_ops(aeon_address, aeon_port, ax.N,
device_hetr, device_id)
input_ops_valid = valid_set.make_input_ops(aeon_address, aeon_port, ax.N,
device_hetr, device_id)
with ng.metadata(device=device_hetr, device_id=device_id, parallel=ax.N):
# Build the network
resnet = BuildResnet(args.dataset,