def MultiLayerPerceptron(*args, **kwargs):
'''
positional arguments:
the number of hidden units of each layer
keyword arguments(optional):
activation: ReLU by default
affine_monitor: bool
activation_monitor: bool
storage: dictionary
'''
assert all(isinstance(arg, int) for arg in args)
try:
activation = kwargs.pop('activation', 'ReLU')
activation = getattr(builder, activation)()
except:
raise Exception('unsupported activation function')
affine_monitor = kwargs.pop('affine_monitor', False)
activation_monitor = kwargs.pop('activation_monitor', False)
if affine_monitor or activation_monitor:
try:
storage = kwargs['storage']
except:
raise Exception('storage required to monitor intermediate result')
network = builder.Sequential()
for i, arg in enumerate(args[:-1]):
network.append(builder.Affine(arg))
if affine_monitor:
network.append(builder.Export('affine%d' % i, storage))
network.append(activation)
if activation_monitor:
network.append(builder.Export('activation%d' % i, storage))
network.append(builder.Affine(args[-1]))
return network
def main(args):
# Define a convolutional neural network the same as above
net = builder.Sequential(
builder.Convolution((7, 7), 32),
builder.ReLU(),
builder.Pooling('max', (2, 2), (2, 2)),
builder.Reshape((flattened_input_size,))
builder.Affine(hidden_size),
builder.Affine(num_classes),
)
# Cast the definition to a model compatible with minpy solver
model = builder.Model(net, 'softmax', (3 * 32 * 32,))
data = get_CIFAR10_data(args.data_dir)
train_dataiter = NDArrayIter(data['X_train'],
data['y_train'],
batch_size=batch_size,
shuffle=True)
test_dataiter = NDArrayIter(data['X_test'],
data['y_test'],
batch_size=batch_size,
shuffle=False)
solver = Solver(model,
train_dataiter,
test_dataiter,
num_epochs=10,
init_rule='gaussian',
init_config={
'stdvar': 0.001
},
update_rule='sgd_momentum',
optim_config={
'learning_rate': 1e-3,
'momentum': 0.9
},
verbose=True,
print_every=20)
solver.init()
solver.train()
def main(args):
# Define a 2-layer perceptron
MLP = builder.Sequential(
builder.Affine(512),
builder.ReLU(),
builder.Affine(10)
)
# Cast the definition to a model compatible with minpy solver
model = builder.Model(MLP, 'softmax', (3 * 32 * 32,))
data = get_CIFAR10_data(args.data_dir)
data['X_train'] = data['X_train'].reshape([data['X_train'].shape[0], 3 * 32 * 32])
data['X_val'] = data['X_val'].reshape([data['X_val'].shape[0], 3 * 32 * 32])
data['X_test'] = data['X_test'].reshape([data['X_test'].shape[0], 3 * 32 * 32])
train_dataiter = NDArrayIter(data['X_train'],
data['y_train'],
batch_size=100,
shuffle=True)
test_dataiter = NDArrayIter(data['X_test'],
data['y_test'],
batch_size=100,
shuffle=False)
solver = Solver(model,
train_dataiter,
test_dataiter,
num_epochs=10,
init_rule='gaussian',
init_config={
'stdvar': 0.001
},
update_rule='sgd_momentum',
optim_config={
'learning_rate': 1e-5,
'momentum': 0.9
},
verbose=True,
print_every=20)
solver.init()
solver.train()
def residual_network(n):
'''
n: the network contains 6 * n + 2 layers, including 6 * n + 1 convolution layers and 1 affine layer
please refer to the paper for details
'''
def normalized_convolution(kernel_shape, kernel_number, stride, pad, activate=None):
module = builder.Sequential(
builder.Convolution(kernel_shape, kernel_number, stride, pad),
builder.SpatialBatchNormalization()
)
if activate:
module.append(getattr(builder, activate)())
return module
def residual(kernel_number, project=False):
if project:
module = builder.Add(
builder.Sequential(
normalized_convolution((3, 3), kernel_number, (2, 2), (1, 1), 'ReLU'),
normalized_convolution((3, 3), kernel_number, (1, 1), (1, 1))
),
builder.Sequential(
builder.Pooling('avg', (2, 2), (2, 2)),
builder.Convolution((1, 1), kernel_number)
)
)
else:
module = builder.Add(
builder.Sequential(
normalized_convolution((3, 3), kernel_number, (1, 1), (1, 1), 'ReLU'),
normalized_convolution((3, 3), kernel_number, (1, 1), (1, 1))
),
builder.Identity()
)
return module
network = builder.Sequential(
builder.Reshape((3, 32, 32)),
normalized_convolution((3, 3), 16, (1, 1), (1, 1), 'ReLU')
)
for i in range(n):
network.append(residual(16))
network.append(residual(32, project=True))
for i in range(n-1):
network.append(residual(32))
network.append(residual(64, project=True))
for i in range(n-1):
network.append(residual(64))
network.append(builder.Pooling('avg', (8, 8)))
network.append(builder.Reshape((64,)))
network.append(builder.Affine(10))
return network
from facility import *
from solver_primitives import *
sys.path.append('../')
from utilities.data_utility import load_cifar10
data = load_cifar10(path='../utilities/cifar/', center=True, rescale=True)
X = data[0][:16]
hidden_layers = 4
shapes = (1024, ) * hidden_layers + (10, )
activation = builder.ReLU
storage = {}
mlp = builder.Sequential()
for i, shape in enumerate(shapes[:-1]):
mlp.append(builder.Affine(shape))
mlp.append(builder.Export('affine%d' % i, storage))
mlp.append(activation())
mlp.append(builder.Affine(shapes[-1]))
model = builder.Model(mlp, 'softmax', (3072, ))
'''
for key, value in model.param_configs.items():
if 'weight' in key:
value['init_rule'] = 'gaussian'
value['init_config'] = {'stdvar' : 1}
'''
initialize(model)
for key, value in model.params.items():
if 'weight' in key:
from solver_primitives import *
from utilities.data_utility import load_cifar10
from GPU_utility import GPU_availability
from minpy.context import set_context, gpu
set_context(gpu(GPU_availability()[0]))
ACTIVATION = 'ReLU'
SHAPE = (1024, ) * 3 + (10, )
BATCH_SIZE = 64
X_SHAPE = (3072, )
activation = getattr(builder, ACTIVATION)
mlp = builder.Sequential()
for shape in SHAPE[:-1]:
mlp.append(builder.Affine(shape))
mlp.append(activation())
mlp.append(builder.Affine(SHAPE[-1]))
model = builder.Model(mlp, 'softmax', X_SHAPE)
initialize(model)
updater = Updater(model, 'sgd', {'learning_rate': 0.01})
training_X, training_Y, validation_X, validation_Y, test_X, test_Y, = \
load_cifar10(path='../../cifar10/utilities/cifar/', center=True, rescale=True)
X_batches = Batches(training_X, BATCH_SIZE)
Y_batches = Batches(training_Y, BATCH_SIZE)
ITERATIONS = 20000
LOGGING_INTERVAL = 10
VALIDATION_INTERVAL = 50
loss_table = []
parameters = list(dmodel.params.values()) + list(gmodel.params.values())
return gl(*parameters)
N, D = 50000, 16
data, p = generate_data(N, D)
BATCH_SIZE = 100
X_batches = Batches(data, BATCH_SIZE)
p_batches = Batches(p.reshape((N, 1)), BATCH_SIZE)
ACTIVATION = 'ReLU'
activation = getattr(builder, ACTIVATION)
DSHAPE = (16,) * 4 + (1,)
dmlp = builder.Sequential()
for shape in DSHAPE[:-1]:
dmlp.append(builder.Affine(shape))
dmlp.append(activation())
dmlp.append(builder.Affine(DSHAPE[-1]))
dmodel = builder.Model(dmlp, 'l2', (D,))
initialize(dmodel)
dupdater = Updater(dmodel, 'sgd', {'learning_rate' : -0.01})
GSHAPE = (16,) * 4 + (D,)
gmlp = builder.Sequential()
for shape in GSHAPE[:-1]:
gmlp.append(builder.Affine(shape))
gmlp.append(activation())
gmlp.append(builder.Affine(GSHAPE[-1]))
gmodel = builder.Model(gmlp, 'l2', (D,))
initialize(gmodel)
gupdater = Updater(gmodel, 'sgd', {'learning_rate' : 0.01})
