def get_data_iters(dataset, batch_size, num_workers=1, rank=0):
"""get dataset iterators"""
if dataset == 'mnist':
train_data, val_data = get_mnist_iterator(batch_size, (1, 28, 28),
num_parts=num_workers,
part_index=rank)
elif dataset == 'cifar10':
train_data, val_data = get_cifar10_iterator(batch_size, (3, 32, 32),
num_parts=num_workers,
part_index=rank)
elif dataset == 'imagenet':
if not opt.data_dir:
raise ValueError(
'Dir containing raw images in train/val is required for imagenet, plz specify "--data-dir"'
)
if model_name == 'inceptionv3':
train_data, val_data = get_imagenet_iterator(
opt.data_dir, batch_size, opt.num_workers, 299, opt.dtype)
else:
train_data, val_data = get_imagenet_iterator(
opt.data_dir, batch_size, opt.num_workers, 224, opt.dtype)
elif dataset == 'dummy':
if model_name == 'inceptionv3':
train_data, val_data = dummy_iterator(batch_size, (3, 299, 299))
else:
train_data, val_data = dummy_iterator(batch_size, (3, 224, 224))
return train_data, val_data
def get_data_iters(dataset, batch_size, num_workers=1, rank=0):
# get dataset iterators
if dataset == 'mnist':
train_data, val_data = get_mnist_iterator(batch_size, (1, 28, 28),
num_parts=num_workers,
part_index=rank)
elif dataset == 'cifar10':
train_data, val_data = get_cifar10_iterator(batch_size, (3, 32, 32),
num_parts=num_workers,
part_index=rank)
elif dataset == 'imagenet':
if model_name == 'inceptionv3':
train_data, val_data = get_imagenet_iterator(opt.train_data,
opt.val_data,
batch_size,
(3, 299, 299),
num_parts=num_workers,
part_index=rank)
else:
train_data, val_data = get_imagenet_iterator(opt.train_data,
opt.val_data,
batch_size,
(3, 224, 224),
num_parts=num_workers,
part_index=rank)
elif dataset == 'dummy':
if model_name == 'inceptionv3':
train_data, val_data = dummy_iterator(batch_size, (3, 299, 299))
else:
train_data, val_data = dummy_iterator(batch_size, (3, 224, 224))
return train_data, val_data
def get_data_iters(dataset, batch_size, opt):
"""get dataset iterators"""
if dataset == 'mnist':
train_data, val_data = get_mnist_iterator(batch_size, (1, 28, 28),
num_parts=kv.num_workers,
part_index=kv.rank)
elif dataset == 'cifar10':
train_data, val_data = get_cifar10_iterator(batch_size, (3, 32, 32),
num_parts=kv.num_workers,
part_index=kv.rank)
elif dataset == 'imagenet':
shape_dim = 299 if model_name == 'inceptionv3' else 224
if not opt.data_dir:
raise ValueError(
'Dir containing raw images in train/val is required for imagenet.'
'Please specify "--data-dir"')
train_data, val_data = get_imagenet_iterator(opt.data_dir, batch_size,
opt.num_workers,
shape_dim, opt.dtype)
elif dataset == 'caltech101':
train_data, val_data = get_caltech101_iterator(batch_size,
opt.num_workers,
opt.dtype)
elif dataset == 'dummy':
shape_dim = 299 if model_name == 'inceptionv3' else 224
train_data, val_data = dummy_iterator(batch_size,
(3, shape_dim, shape_dim))
return train_data, val_data
def get_test_data_iter(config, kv):
from mxnet.test_utils import get_mnist_iterator
from filelock import FileLock
with FileLock("data.lock"):
iters = get_mnist_iterator(config["batch_size"], (1, 28, 28),
num_parts=kv.num_workers, part_index=kv.rank)
return iters[1]
def get_data_iters(dataset, batch_size, num_workers=1, rank=0):
"""get dataset iterators"""
if dataset == 'mnist':
train_data, val_data = get_mnist_iterator(batch_size, (1, 28, 28),
num_parts=num_workers, part_index=rank)
elif dataset == 'cifar10':
train_data, val_data = get_cifar10_iterator(batch_size, (3, 32, 32),
num_parts=num_workers, part_index=rank)
elif dataset == 'imagenet':
if not opt.data_dir:
raise ValueError('Dir containing raw images in train/val is required for imagenet, plz specify "--data-dir"')
if model_name == 'inceptionv3':
train_data, val_data = get_imagenet_iterator(opt.data_dir, batch_size, opt.num_workers, 299, opt.dtype)
else:
train_data, val_data = get_imagenet_iterator(opt.data_dir, batch_size, opt.num_workers, 224, opt.dtype)
elif dataset == 'dummy':
if model_name == 'inceptionv3':
train_data, val_data = dummy_iterator(batch_size, (3, 299, 299))
else:
train_data, val_data = dummy_iterator(batch_size, (3, 224, 224))
return train_data, val_data
def get_data_iters(dataset, batch_size, opt):
"""get dataset iterators"""
if dataset == 'mnist':
train_data, val_data = get_mnist_iterator(batch_size, (1, 28, 28),
num_parts=kv.num_workers, part_index=kv.rank)
elif dataset == 'cifar10':
train_data, val_data = get_cifar10_iterator(batch_size, (3, 32, 32),
num_parts=kv.num_workers, part_index=kv.rank)
elif dataset == 'imagenet':
shape_dim = 299 if model_name == 'inceptionv3' else 224
if not opt.data_dir:
raise ValueError('Dir containing raw images in train/val is required for imagenet.'
'Please specify "--data-dir"')
train_data, val_data = get_imagenet_iterator(opt.data_dir, batch_size,
opt.num_workers, shape_dim, opt.dtype)
elif dataset == 'caltech101':
train_data, val_data = get_caltech101_iterator(batch_size, opt.num_workers, opt.dtype)
elif dataset == 'dummy':
shape_dim = 299 if model_name == 'inceptionv3' else 224
train_data, val_data = dummy_iterator(batch_size, (3, shape_dim, shape_dim))
return train_data, val_data
def get_data_iters(dataset, batch_size, num_workers=1, rank=0):
# get dataset iterators
if dataset == 'mnist':
train_data, val_data = get_mnist_iterator(batch_size, (1, 28, 28),
num_parts=num_workers, part_index=rank)
elif dataset == 'cifar10':
train_data, val_data = get_cifar10_iterator(batch_size, (3, 32, 32),
num_parts=num_workers, part_index=rank)
elif dataset == 'imagenet':
if model_name == 'inceptionv3':
train_data, val_data = get_imagenet_iterator(opt.train_data, opt.val_data,
batch_size, (3, 299, 299),
num_parts=num_workers, part_index=rank)
else:
train_data, val_data = get_imagenet_iterator(opt.train_data, opt.val_data,
batch_size, (3, 224, 224),
num_parts=num_workers, part_index=rank)
elif dataset == 'dummy':
if model_name == 'inceptionv3':
train_data, val_data = dummy_iterator(batch_size, (3, 299, 299))
else:
train_data, val_data = dummy_iterator(batch_size, (3, 224, 224))
return train_data, val_data
def get_data_iters(opt, num_workers=1, rank=0):
"""get dataset iterators"""
if opt.dry_run:
return None, None
if opt.dataset == 'mnist':
train_data, val_data = get_mnist_iterator(opt.batch_size, (1, 28, 28),
num_parts=num_workers,
part_index=rank)
elif opt.dataset == 'cifar10':
train_data, val_data = get_cifar10_iterator(
opt.batch_size, (3, 32, 32),
num_parts=num_workers,
part_index=rank,
dir=opt.data_path,
aug_level=opt.augmentation_level,
mean_subtraction=opt.mean_subtraction)
elif opt.dataset == 'imagenet':
if not opt.data_dir:
raise ValueError(
'Dir containing rec files is required for imagenet, please specify "--data-dir"'
)
if opt.model == 'inceptionv3':
train_data, val_data = get_imagenet_iterator(
opt.data_dir, opt.batch_size, opt.num_workers, 299, opt.dtype)
else:
train_data, val_data = get_imagenet_iterator(
opt.data_dir, opt.batch_size, opt.num_workers, 224, opt.dtype)
elif dataset == 'dummy':
if opt.model == 'inceptionv3':
train_data, val_data = dummy_iterator(opt.batch_size,
(3, 299, 299))
else:
train_data, val_data = dummy_iterator(opt.batch_size,
(3, 224, 224))
return train_data, val_data
x = F.relu(self.conv2(x))
x = self.pool(x)
x = self.dropout1(x)
x = self.flatten(x)
x = F.relu(self.dense1(x))
x = self.dropout2(x)
x = self.dense2(x)
return x
mx.random.seed(42)
random.seed(42)
# get data
input_shape = (1, 28, 28)
train_data, test_data = get_mnist_iterator(input_shape=input_shape,
batch_size=BATCH_SIZE)
# build nodel
model = Model()
# hybridize for speed
model.hybridize(static_alloc=True, static_shape=True)
# pin GPU
ctx = mx.gpu()
# optimizer
opt_params={'learning_rate':0.001, 'beta1':0.9, 'beta2':0.999, 'epsilon':1e-08}
opt = mx.optimizer.create('adam', **opt_params)
# initialize parameters
model.initialize(force_reinit=True, ctx=ctx)
# fetch and broadcast parameters
params = model.collect_params()
return Softmax()
# define mlp
data = mx.symbol.Variable('data')
fc1 = mx.symbol.FullyConnected(data = data, name='fc1', num_hidden=128)
act1 = mx.symbol.Activation(data = fc1, name='relu1', act_type="relu")
fc2 = mx.symbol.FullyConnected(data = act1, name = 'fc2', num_hidden = 64)
act2 = mx.symbol.Activation(data = fc2, name='relu2', act_type="relu")
fc3 = mx.symbol.FullyConnected(data = act2, name='fc3', num_hidden=10)
#mlp = mx.symbol.Softmax(data = fc3, name = 'softmax')
mlp = mx.symbol.Custom(data=fc3, name='softmax', op_type='softmax')
# data
train, val = get_mnist_iterator(batch_size=100, input_shape = (784,))
# train
logging.basicConfig(level=logging.DEBUG)
# MXNET_CPU_WORKER_NTHREADS must be greater than 1 for custom op to work on CPU
context=mx.cpu()
# Uncomment this line to train on GPU
# context=mx.gpu(0)
mod = mx.mod.Module(mlp, context=context)
mod.fit(train_data=train, eval_data=val, optimizer='sgd',
optimizer_params={'learning_rate':0.1, 'momentum': 0.9, 'wd': 0.00001},
num_epoch=10, batch_end_callback=mx.callback.Speedometer(100, 100))
def get_data_iters(config, kv):
return get_mnist_iterator(config["batch_size"], (1, 28, 28),
num_parts=kv.num_workers,
part_index=kv.rank)
x = F.relu(self.dense1(x))
x = self.dropout2(x)
x = self.dense2(x)
return x
mx.random.seed(42)
random.seed(42)
# initialize Horovod
hvd.init()
# get data
input_shape = (1, 28, 28)
train_data, test_data = get_mnist_iterator(input_shape=input_shape,
batch_size=BATCH_SIZE,
num_parts=hvd.size(),
part_index=hvd.rank())
# build model
model = Model()
# hybridize for speed
model.hybridize(static_alloc=True, static_shape=True)
# pin GPU to be used to process local rank
ctx = mx.gpu(hvd.local_rank())
# optimizer
opt_params = {
'learning_rate': 0.001,
'beta1': 0.9,
'beta2': 0.999,
'epsilon': 1e-08
