def _createLayers(self, num_actions):
# create network
init_xavier_conv = Xavier(local=True)
init_xavier_affine = Xavier(local=False)
# init_uniform_conv = Uniform(low=-.01, high=.01)
# init_uniform_affine = Uniform(low=-.01, high=.01)
layers = []
# The first hidden layer convolves 32 filters of 8x8 with stride 4 with the input image and applies a rectifier nonlinearity.
# layers.append(Conv((8, 8, 32), strides=4, init=init_xavier_conv, activation=Rectlin(), batch_norm=self.batch_norm))
layers.append(
Conv((5, 5, 32),
strides=2,
init=init_xavier_conv,
activation=Rectlin(),
batch_norm=self.batch_norm))
# The second hidden layer convolves 64 filters of 4x4 with stride 2, again followed by a rectifier nonlinearity.
# layers.append(Conv((4, 4, 64), strides=2, init=init_xavier_conv, activation=Rectlin(), batch_norm=self.batch_norm))
layers.append(
Conv((5, 5, 32),
strides=2,
init=init_xavier_conv,
activation=Rectlin(),
batch_norm=self.batch_norm))
# This is followed by a third convolutional layer that convolves 64 filters of 3x3 with stride 1 followed by a rectifier.
# layers.append(Conv((3, 3, 64), strides=1, init=init_xavier_conv, activation=Rectlin(), batch_norm=self.batch_norm))
# The final hidden layer is fully-connected and consists of 512 rectifier units.
layers.append(
Affine(nout=512,
init=init_xavier_affine,
activation=Rectlin(),
batch_norm=self.batch_norm))
# The output layer is a fully-connected linear layer with a single output for each valid action.
layers.append(Affine(nout=num_actions, init=init_xavier_affine))
return layers
def create_network():
# weight initialization
g1 = Gaussian(scale=0.01)
g5 = Gaussian(scale=0.005)
c0 = Constant(0)
c1 = Constant(1)
# model initialization
padding = {'pad_d': 1, 'pad_h': 1, 'pad_w': 1}
strides = {'str_d': 2, 'str_h': 2, 'str_w': 2}
layers = [
Conv((3, 3, 3, 64), padding=padding, init=g1, bias=c0, activation=Rectlin()),
Pooling((1, 2, 2), strides={'str_d': 1, 'str_h': 2, 'str_w': 2}),
Conv((3, 3, 3, 128), padding=padding, init=g1, bias=c1, activation=Rectlin()),
Pooling((2, 2, 2), strides=strides),
Conv((3, 3, 3, 256), padding=padding, init=g1, bias=c1, activation=Rectlin()),
Pooling((2, 2, 2), strides=strides),
Conv((3, 3, 3, 256), padding=padding, init=g1, bias=c1, activation=Rectlin()),
Pooling((2, 2, 2), strides=strides),
Conv((3, 3, 3, 256), padding=padding, init=g1, bias=c1, activation=Rectlin()),
Pooling((2, 2, 2), strides=strides),
Affine(nout=2048, init=g5, bias=c1, activation=Rectlin()),
Dropout(keep=0.5),
Affine(nout=2048, init=g5, bias=c1, activation=Rectlin()),
Dropout(keep=0.5),
Affine(nout=101, init=g1, bias=c0, activation=Softmax())
]
return Model(layers=layers)
def _createLayers(self, num_actions):
init_xavier_conv = Xavier(local=True)
init_xavier_affine = Xavier(local=False)
layers = []
layers.append(
Conv((8, 8, 32),
strides=4,
init=init_xavier_conv,
activation=Rectlin(),
batch_norm=self.batch_norm))
layers.append(
Conv((4, 4, 64),
strides=2,
init=init_xavier_conv,
activation=Rectlin(),
batch_norm=self.batch_norm))
layers.append(
Conv((2, 2, 128),
strides=1,
init=init_xavier_conv,
activation=Rectlin(),
batch_norm=self.batch_norm))
layers.append(
Affine(nout=256,
init=init_xavier_affine,
activation=Rectlin(),
batch_norm=self.batch_norm))
layers.append(Affine(nout=num_actions, init=init_xavier_affine))
return layers
def build(self):
# setup model layers
layers = [Affine(nout=100, init=self.init, bias=self.init, activation=Rectlin()),
Affine(nout=2, init=self.init, bias=self.init, activation=Softmax())]
# initialize model object
self.model = Model(layers=layers)
def _createLayers(self, num_actions):
# create network
init_norm = Gaussian(loc=0.0, scale=0.01)
layers = []
if self.screen_dim == (84, 84):
# The first hidden layer convolves 32 filters of 8x8 with stride 4 with the input image and applies a rectifier nonlinearity.
logger.info("Using 8x8 filter for first Conv")
layers.append(
Conv((8, 8, 32),
strides=4,
init=init_norm,
activation=Rectlin()))
elif self.screen_dim == (52, 40):
# The first hidden layer convolves 32 filters of 5x4 with stride 2 with the input image and applies a rectifier nonlinearity.
logger.info("Using 5x4 filter for first Conv")
layers.append(
Conv((5, 4, 32),
strides=2,
init=init_norm,
activation=Rectlin()))
else:
raise NotImplementedError("Unsupported screen dim.")
# The second hidden layer convolves 64 filters of 4x4 with stride 2, again followed by a rectifier nonlinearity.
layers.append(
Conv((4, 4, 64), strides=2, init=init_norm, activation=Rectlin()))
# This is followed by a third convolutional layer that convolves 64 filters of 3x3 with stride 1 followed by a rectifier.
layers.append(
Conv((3, 3, 64), strides=1, init=init_norm, activation=Rectlin()))
# The final hidden layer is fully-connected and consists of 512 rectifier units.
layers.append(Affine(nout=512, init=init_norm, activation=Rectlin()))
# The output layer is a fully-connected linear layer with a single output for each valid action.
layers.append(Affine(nout=num_actions, init=init_norm))
return layers
def layers(self):
init_uni = Uniform(low=-0.1, high=0.1)
bn = False
return [
DOG((5.0, 4.0, 3.0, 1.6), 1.8),
Conv((5, 5, 16),
init=init_uni,
activation=Rectlin(),
batch_norm=bn),
Pooling((2, 2)),
Conv((5, 5, 32),
init=init_uni,
activation=Rectlin(),
batch_norm=bn),
Pooling((2, 2)),
Affine(nout=500,
init=init_uni,
activation=Rectlin(),
batch_norm=bn),
Affine(nout=self.noutputs,
init=init_uni,
bias=Constant(0),
activation=Softmax() if self.use_softmax else Logistic(
shortcut=True))
]
def main(args):
# load up the mnist data set
dataset = MNIST(path=args.data_dir)
# initialize model object
mlp = Model(layers=[
Affine(nout=100,
init=Gaussian(loc=0.0, scale=0.01),
activation=Rectlin()),
Affine(nout=10,
init=Gaussian(loc=0.0, scale=0.01),
activation=Logistic(shortcut=True))
])
# setup optimizer
optimizer = GradientDescentMomentum(0.1,
momentum_coef=0.9,
stochastic_round=args.rounding)
# configure callbacks
callbacks = Callbacks(mlp,
eval_set=dataset.valid_iter,
**args.callback_args)
# run fit
# setup cost function as CrossEntropy
mlp.fit(dataset.train_iter,
optimizer=optimizer,
num_epochs=args.epochs,
cost=GeneralizedCost(costfunc=CrossEntropyBinary()),
callbacks=callbacks)
error_rate = mlp.eval(dataset.valid_iter, metric=Misclassification())
neon_logger.display('Classification accuracy = %.4f' % (1 - error_rate))
def build_model(self):
# setup weight initialization function
init_norm = Gaussian(loc=0.0, scale=0.01)
# setup model layers
layers = [
Affine(nout=100,
init=init_norm,
bias=Uniform(),
activation=Rectlin()),
Affine(nout=10,
init=init_norm,
bias=Uniform(),
activation=Logistic(shortcut=True))
]
# setup cost function as CrossEntropy
self.cost = GeneralizedCost(costfunc=CrossEntropyBinary())
# setup optimizer
self.optimizer = GradientDescentMomentum(
0.1, momentum_coef=0.9, stochastic_round=self.args.rounding)
# initialize model object
self.model = ModelDist(layers=layers)
def create_frcn_model(frcn_fine_tune=False):
b1 = BranchNode(name="b1")
imagenet_layers = add_vgg_layers()
HW = (7, 7)
frcn_layers = [
RoiPooling(layers=imagenet_layers, HW=HW,
bprop_enabled=frcn_fine_tune),
Affine(nout=4096,
init=Gaussian(scale=0.005),
bias=Constant(.1),
activation=Rectlin()),
Dropout(keep=0.5),
Affine(nout=4096,
init=Gaussian(scale=0.005),
bias=Constant(.1),
activation=Rectlin()),
Dropout(keep=0.5), b1,
Affine(nout=21,
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Softmax())
]
bb_layers = [
b1,
Affine(nout=84,
init=Gaussian(scale=0.001),
bias=Constant(0),
activation=Identity())
]
return Model(layers=Tree([frcn_layers, bb_layers]))
def __init__(self):
self.in_shape = [1024, (2538, 38)]
init = Constant(0)
image_path = Sequential(
[Affine(20, init, bias=init),
Affine(10, init, bias=init)])
sent_path = Sequential([Affine(30, init, bias=init), Affine(10, init)])
layers = [
MergeMultistream(layers=[image_path, sent_path],
merge="recurrent"),
Dropout(keep=0.5),
LSTM(4,
init,
activation=Logistic(),
gate_activation=Tanh(),
reset_cells=True),
Affine(20, init, bias=init, activation=Softmax())
]
self.layers = layers
self.cost = GeneralizedCostMask(CrossEntropyMulti())
self.model = Model(layers=layers)
self.model.initialize(self.in_shape, cost=self.cost)
def _createLayers(self, num_actions):
# create network
init_norm = Gaussian(loc=0.0, scale=0.01)
layers = []
# The first hidden layer convolves 32 filters of 8x8 with stride 4 with the input image and applies a rectifier nonlinearity.
layers.append(
Conv((8, 8, 32),
strides=4,
init=init_norm,
activation=Rectlin(),
batch_norm=self.batch_norm))
# The second hidden layer convolves 64 filters of 4x4 with stride 2, again followed by a rectifier nonlinearity.
layers.append(
Conv((4, 4, 64),
strides=2,
init=init_norm,
activation=Rectlin(),
batch_norm=self.batch_norm))
# This is followed by a third convolutional layer that convolves 64 filters of 3x3 with stride 1 followed by a rectifier.
layers.append(
Conv((3, 3, 64),
strides=1,
init=init_norm,
activation=Rectlin(),
batch_norm=self.batch_norm))
# The final hidden layer is fully-connected and consists of 512 rectifier units.
layers.append(
Affine(nout=512,
init=init_norm,
activation=Rectlin(),
batch_norm=self.batch_norm))
# The output layer is a fully-connected linear layer with a single output for each valid action.
layers.append(Affine(nout=num_actions, init=init_norm))
return layers
def run(train, test):
init = Gaussian(scale=0.01)
layers = [
Conv((3, 3, 128),
init=init,
activation=Rectlin(),
strides=dict(str_h=1, str_w=2)),
Conv((3, 3, 256), init=init, batch_norm=True, activation=Rectlin()),
Pooling(2, strides=2),
Conv((2, 2, 512), init=init, batch_norm=True, activation=Rectlin()),
DeepBiRNN(256,
init=init,
activation=Rectlin(),
reset_cells=True,
depth=3),
RecurrentLast(),
Affine(32, init=init, batch_norm=True, activation=Rectlin()),
Affine(nout=common['nclasses'], init=init, activation=Softmax())
]
model = Model(layers=layers)
opt = Adadelta()
metric = Misclassification()
callbacks = Callbacks(model,
eval_set=test,
metric=metric,
**args.callback_args)
cost = GeneralizedCost(costfunc=CrossEntropyBinary())
model.fit(train,
optimizer=opt,
num_epochs=args.epochs,
cost=cost,
callbacks=callbacks)
return model
def __init__(self,
overlapping_classes=None,
exclusive_classes=None,
analytics_input=True,
network_type='conv_net',
num_words=60,
width=100,
lookup_size=0,
lookup_dim=0,
optimizer=Adam()):
assert (overlapping_classes is not None) or (exclusive_classes
is not None)
self.width = width
self.num_words = num_words
self.overlapping_classes = overlapping_classes
self.exclusive_classes = exclusive_classes
self.analytics_input = analytics_input
self.recurrent = network_type == 'lstm'
self.lookup_size = lookup_size
self.lookup_dim = lookup_dim
init = GlorotUniform()
activation = Rectlin(slope=1E-05)
gate = Logistic()
input_layers = self.input_layers(analytics_input, init, activation,
gate)
if self.overlapping_classes is None:
output_layers = [
Affine(len(self.exclusive_classes), init, activation=Softmax())
]
elif self.exclusive_classes is None:
output_layers = [
Affine(len(self.overlapping_classes),
init,
activation=Logistic())
]
else:
output_branch = BranchNode(name='exclusive_overlapping')
output_layers = Tree([[
SkipNode(), output_branch,
Affine(len(self.exclusive_classes), init, activation=Softmax())
],
[
output_branch,
Affine(len(self.overlapping_classes),
init,
activation=Logistic())
]])
layers = [
input_layers,
# this is where inputs meet, and where we may want to add depth or
# additional functionality
Dropout(keep=0.8),
output_layers
]
super(ClassifierNetwork, self).__init__(layers, optimizer=optimizer)
def _createLayers(self):
# create network
init_norm = Gaussian(loc=0.0, scale=0.01)
layers = []
for i in xrange(self.num_layers):
layers.append(Affine(nout=self.hidden_nodes, init=init_norm, activation=Rectlin()))
layers.append(Affine(nout=self.num_actions, init = init_norm))
return layers
def constuct_network():
"""
Constructs the layers of our RCNN architecture. It is similar to AlexNet but simplified to only a
few convolutional layers and 3 LSTM layers.
"""
layers = [
Conv((11, 11, 64),
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Rectlin(),
padding=3,
strides=4),
Pooling(3, strides=2),
Conv((7, 7, 128),
init=Gaussian(scale=0.01),
bias=Constant(1),
activation=Rectlin(),
padding=2),
Pooling(3, strides=2),
Conv((5, 5, 256),
init=Gaussian(scale=0.03),
bias=Constant(0),
activation=Rectlin(),
padding=1),
Conv((3, 3, 256),
init=Gaussian(scale=0.03),
bias=Constant(1),
activation=Rectlin(),
padding=1),
Pooling(3, strides=2),
Affine(nout=4096,
init=Gaussian(scale=0.01),
bias=Constant(1),
activation=Rectlin()),
DropoutBinary(keep=0.5),
LSTM(512,
init=Gaussian(scale=0.03),
activation=Rectlin(),
gate_activation=Tanh()),
LSTM(512,
init=Gaussian(scale=0.03),
activation=Rectlin(),
gate_activation=Tanh()),
LSTM(512,
init=Gaussian(scale=0.03),
activation=Rectlin(),
gate_activation=Tanh()),
Affine(nout=4096,
init=Gaussian(scale=0.01),
bias=Constant(1),
activation=Rectlin()),
DropoutBinary(keep=0.5),
Affine(nout=101,
init=Gaussian(scale=0.01),
bias=Constant(-7),
activation=Softmax())
]
return Model(layers=layers)
def aux_branch(bnode):
return [
bnode,
Pooling(fshape=5, strides=3, op="avg"),
Conv((1, 1, 128), **common),
Affine(nout=1024, init=init1, activation=relu, bias=bias),
Dropout(keep=0.3),
Affine(nout=1000, init=init1, activation=Softmax(), bias=Constant(0))
]
def _createLayers(self, num_actions): # create network init_norm = Gaussian(loc=0.0, scale=0.01) layers = [] # The final hidden layer is fully-connected and consists of 512 rectifier units. layers.append(Affine(nout=64, init=init_norm, bias=init_norm, activation=Rectlin())) # The output layer is a fully-connected linear layer with a single output for each valid action. layers.append(Affine(nout=num_actions, init=init_norm, bias=init_norm)) return layers
def main():
parser = get_parser()
args = parser.parse_args()
print('Args:', args)
loggingLevel = logging.DEBUG if args.verbose else logging.INFO
logging.basicConfig(level=loggingLevel, format='')
ext = extension_from_parameters(args)
loader = p1b3.DataLoader(feature_subsample=args.feature_subsample,
scaling=args.scaling,
drug_features=args.drug_features,
scramble=args.scramble,
min_logconc=args.min_logconc,
max_logconc=args.max_logconc,
subsample=args.subsample,
category_cutoffs=args.category_cutoffs)
# initializer = Gaussian(loc=0.0, scale=0.01)
initializer = GlorotUniform()
activation = get_function(args.activation)()
layers = []
reshape = None
if args.convolution and args.convolution[0]:
reshape = (1, loader.input_dim, 1)
layer_list = list(range(0, len(args.convolution), 3))
for l, i in enumerate(layer_list):
nb_filter = args.convolution[i]
filter_len = args.convolution[i+1]
stride = args.convolution[i+2]
# print(nb_filter, filter_len, stride)
# fshape: (height, width, num_filters).
layers.append(Conv((1, filter_len, nb_filter), strides={'str_h':1, 'str_w':stride}, init=initializer, activation=activation))
if args.pool:
layers.append(Pooling((1, args.pool)))
for layer in args.dense:
if layer:
layers.append(Affine(nout=layer, init=initializer, activation=activation))
if args.drop:
layers.append(Dropout(keep=(1-args.drop)))
layers.append(Affine(nout=1, init=initializer, activation=neon.transforms.Identity()))
model = Model(layers=layers)
train_iter = ConcatDataIter(loader, ndata=args.train_samples, lshape=reshape, datatype=args.datatype)
val_iter = ConcatDataIter(loader, partition='val', ndata=args.val_samples, lshape=reshape, datatype=args.datatype)
cost = GeneralizedCost(get_function(args.loss)())
optimizer = get_function(args.optimizer)()
callbacks = Callbacks(model, eval_set=val_iter, **args.callback_args)
model.fit(train_iter, optimizer=optimizer, num_epochs=args.epochs, cost=cost, callbacks=callbacks)
def aux_branch(bnode, ind):
# TODO put dropout back in
nm = 'loss%d/' % ind
return [bnode,
Pooling(fshape=5, strides=3, op="avg", name=nm+'ave_pool'),
Conv((1, 1, 128), name=nm+'conv', **common),
Affine(nout=1024, init=init1, activation=relu, bias=bias, name=nm+'fc'),
Dropout(keep=1.0, name=nm+'drop_fc'),
Affine(nout=1000, init=init1, activation=Softmax(),
bias=Constant(0), name=nm+'classifier')]
def layers(self):
return [
Conv((7, 7, 96),
init=Gaussian(scale=0.0001),
bias=Constant(0),
activation=Rectlin(),
padding=3,
strides=1),
LRN(31, ascale=0.001, bpower=0.75),
Pooling(3, strides=2, padding=1),
Conv((5, 5, 256),
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Rectlin(),
padding=2,
strides=1),
LRN(31, ascale=0.001, bpower=0.75),
Pooling(3, strides=2, padding=1),
Conv((3, 3, 384),
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Rectlin(),
padding=1,
strides=1),
Conv((3, 3, 384),
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Rectlin(),
padding=1,
strides=1),
Conv((3, 3, 256),
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Rectlin(),
padding=1,
strides=1),
Pooling(3, strides=2, padding=1),
Affine(nout=4096,
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Identity()),
Dropout(keep=0.5),
Affine(nout=4096,
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Identity()),
Dropout(keep=0.5),
Affine(nout=self.noutputs,
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Softmax() if self.use_softmax else Logistic(
shortcut=True))
]
def fit_model(train_set, val_set, num_epochs=50):
relu = Rectlin()
conv_params = {
'strides': 1,
'padding': 1,
'init': Xavier(local=True), # Xavier sqrt(3)/num_inputs [CHECK THIS]
'bias': Constant(0),
'activation': relu
}
layers = []
layers.append(Conv((3, 3, 128), **conv_params)) # 3x3 kernel * 128 nodes
layers.append(Pooling(2))
layers.append(Conv((3, 3, 128), **conv_params))
layers.append(Pooling(2)) # Highest value from 2x2 window.
layers.append(Conv((3, 3, 128), **conv_params))
layers.append(
Dropout(keep=0.5)
) # Flattens Convolution into a flat array, with probability 0.5 sets activation values to 0
layers.append(
Affine(nout=128,
init=GlorotUniform(),
bias=Constant(0),
activation=relu)
) # 1 value per conv kernel - Linear Combination of layers
layers.append(Dropout(keep=0.5))
layers.append(
Affine(nout=2,
init=GlorotUniform(),
bias=Constant(0),
activation=Softmax(),
name="class_layer"))
# initialize model object
cnn = Model(layers=layers)
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
optimizer = Adam()
# callbacks = Callbacks(cnn)
# out_fname = 'yarin_fdl_out_data.h5'
callbacks = Callbacks(cnn, eval_set=val_set,
eval_freq=1) # , output_file=out_fname
cnn.fit(train_set,
optimizer=optimizer,
num_epochs=num_epochs,
cost=cost,
callbacks=callbacks)
return cnn
def constuct_network():
"""
Constructs the layers of the AlexNet architecture.
"""
layers = [
Conv((11, 11, 64),
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Rectlin(),
padding=3,
strides=4),
Pooling(3, strides=2),
Conv((5, 5, 192),
init=Gaussian(scale=0.01),
bias=Constant(1),
activation=Rectlin(),
padding=2),
Pooling(3, strides=2),
Conv((3, 3, 384),
init=Gaussian(scale=0.03),
bias=Constant(0),
activation=Rectlin(),
padding=1),
Conv((3, 3, 256),
init=Gaussian(scale=0.03),
bias=Constant(1),
activation=Rectlin(),
padding=1),
Conv((3, 3, 256),
init=Gaussian(scale=0.03),
bias=Constant(1),
activation=Rectlin(),
padding=1),
Pooling(3, strides=2),
Affine(nout=4096,
init=Gaussian(scale=0.01),
bias=Constant(1),
activation=Rectlin()),
DropoutBinary(keep=0.5),
Affine(nout=4096,
init=Gaussian(scale=0.01),
bias=Constant(1),
activation=Rectlin()),
DropoutBinary(keep=0.5),
Affine(nout=101,
init=Gaussian(scale=0.01),
bias=Constant(-7),
activation=Softmax())
]
return Model(layers=layers)
def constructCNN(self):
layers = []
if self.network_type == "idsia":
layers.append(
Conv((3, 3, 100),
strides=1,
init=Kaiming(),
bias=Constant(0.0),
activation=Rectlin(),
name="Conv1"))
layers.append(Pooling(2, op="max", strides=2, name="neon_pool1"))
layers.append(
Conv((4, 4, 150),
strides=1,
init=Kaiming(),
bias=Constant(0.0),
activation=Rectlin(),
name="Conv2"))
layers.append(Pooling(2, op="max", strides=2, name="neon_pool2"))
layers.append(
Conv((3, 3, 250),
strides=1,
init=Kaiming(),
bias=Constant(0.0),
activation=Rectlin(),
name="Conv3"))
layers.append(Pooling(2, op="max", strides=2, name="neon_pool3"))
layers.append(
Affine(nout=200,
init=Kaiming(local=False),
bias=Constant(0.0),
activation=Rectlin(),
name="neon_fc1"))
layers.append(
Affine(nout=self.class_num,
init=Kaiming(local=False),
bias=Constant(0.0),
activation=Softmax(),
name="neon_fc2"))
elif self.network_type == "resnet-56":
layers = resnet(9, self.class_num,
int(self.resize_size[0] / 4)) # 6*9 + 2 = 56
elif self.network_type == "resnet-32":
layers = resnet(5, self.class_num,
int(self.resize_size[0] / 4)) # 6*5 + 2 = 32
elif self.network_type == "resnet-20":
layers = resnet(3, self.class_num,
int(self.resize_size[0] / 4)) # 6*3 + 2 = 20
return layers
def create_network():
layers = [
Conv((11, 11, 64),
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Rectlin(),
padding=3,
strides=4),
Pooling(3, strides=2),
Conv((5, 5, 192),
init=Gaussian(scale=0.01),
bias=Constant(1),
activation=Rectlin(),
padding=2),
Pooling(3, strides=2),
Conv((3, 3, 384),
init=Gaussian(scale=0.03),
bias=Constant(0),
activation=Rectlin(),
padding=1),
Conv((3, 3, 256),
init=Gaussian(scale=0.03),
bias=Constant(1),
activation=Rectlin(),
padding=1),
Conv((3, 3, 256),
init=Gaussian(scale=0.03),
bias=Constant(1),
activation=Rectlin(),
padding=1),
Pooling(3, strides=2),
Affine(nout=4096,
init=Gaussian(scale=0.01),
bias=Constant(1),
activation=Rectlin()),
Dropout(keep=0.5),
Affine(nout=4096,
init=Gaussian(scale=0.01),
bias=Constant(1),
activation=Rectlin()),
Dropout(keep=0.5),
Affine(nout=1000,
init=Gaussian(scale=0.01),
bias=Constant(-7),
activation=Softmax()),
]
return Model(layers=layers), GeneralizedCost(costfunc=CrossEntropyMulti())
def create_model(vocab_size, rlayer_type):
"""
Create LSTM/GRU model for bAbI dataset.
Args:
vocab_size (int) : String of bAbI data.
rlayer_type (string) : Type of recurrent layer to use (gru or lstm).
Returns:
Model : Model of the created network
"""
# recurrent layer parameters (default gru)
rlayer_obj = GRU if rlayer_type == 'gru' else LSTM
rlayer_params = dict(output_size=100, reset_cells=True,
init=GlorotUniform(), init_inner=Orthonormal(0.5),
activation=Tanh(), gate_activation=Logistic())
# if using lstm, swap the activation functions
if rlayer_type == 'lstm':
rlayer_params.update(dict(activation=Logistic(), gate_activation=Tanh()))
# lookup layer parameters
lookup_params = dict(vocab_size=vocab_size, embedding_dim=50, init=Uniform(-0.05, 0.05))
# Model construction
story_path = [LookupTable(**lookup_params), rlayer_obj(**rlayer_params)]
query_path = [LookupTable(**lookup_params), rlayer_obj(**rlayer_params)]
layers = [MergeMultistream(layers=[story_path, query_path], merge="stack"),
Affine(vocab_size, init=GlorotUniform(), activation=Softmax())]
return Model(layers=layers)
def test_concat_l1_l1(backend_default, allrand_args):
# test two linear layers that are merged with concat
dtypeu = np.float32
w_rng, rngmax = allrand_args
# Diff size inputs and outputs
nins = [128, 1024]
nouts = [64, 2048]
batch_size = 16
NervanaObject.be.bsz = batch_size
be = NervanaObject.be
init_unif = Uniform(low=w_rng[0], high=w_rng[1])
layers = [Sequential(Affine(nout=nout, init=init_unif)) for nout in nouts]
inputs = [be.array(dtypeu(np.random.random((nin, batch_size)))) for nin in nins]
merge = MergeMultistream(layers, merge="stack")
assert(len(inputs) == len(layers))
merge.configure(inputs)
merge.allocate()
merge.set_deltas(None)
out = merge.fprop(inputs).get()
sublayers = [s.layers[0] for s in layers]
weights = [layer.W.get() for layer in sublayers]
out_exp = np.concatenate([np.dot(w, inp.get()) for (w, inp) in zip(weights, inputs)])
assert np.allclose(out, out_exp, atol=1e-3)
err_lst = [dtypeu(np.random.random((nout, batch_size))) for nout in nouts]
err_concat = np.concatenate(err_lst)
merge.bprop(be.array(err_concat))
dW_exp_lst = [np.dot(err, inp.get().T) for (err, inp) in zip(err_lst, inputs)]
for layer, dW_exp in zip(sublayers, dW_exp_lst):
assert np.allclose(layer.dW.get(), dW_exp)
return
def create_network(stage_depth):
# Structure of the deep residual part of the network:
# stage_depth modules of 2 convolutional layers each at feature map depths of 16, 32, 64
nfms = [2**(stage + 4) for stage in sorted(list(range(3)) * stage_depth)]
strides = [
1 if cur == prev else 2 for cur, prev in zip(nfms[1:], nfms[:-1])
]
# Now construct the network
layers = [Conv(**conv_params(3, 16))]
layers.append(module_s1(nfms[0], True))
for nfm, stride in zip(nfms[1:], strides):
res_module = module_s1(nfm) if stride == 1 else module_s2(nfm)
layers.append(res_module)
layers.append(BatchNorm())
layers.append(Activation(Rectlin()))
layers.append(Pooling('all', op='avg'))
layers.append(
Affine(10,
init=Kaiming(local=False),
batch_norm=True,
activation=Softmax()))
return Model(layers=layers), GeneralizedCost(costfunc=CrossEntropyMulti())
def __init__(self, depth=9):
self.depth = depth
depth = 9
train = (3, 32, 32)
nfms = [2**(stage + 4) for stage in sorted(list(range(3)) * depth)]
strides = [
1 if cur == prev else 2 for cur, prev in zip(nfms[1:], nfms[:-1])
]
# Now construct the network
layers = [Conv(**self.conv_params(3, 16))]
layers.append(self.module_s1(nfms[0], True))
for nfm, stride in zip(nfms[1:], strides):
res_module = self.module_s1(
nfm) if stride == 1 else self.module_s2(nfm)
layers.append(res_module)
layers.append(BatchNorm())
layers.append(Activation(Rectlin()))
layers.append(Pooling('all', op='avg'))
layers.append(
Affine(10,
init=Kaiming(local=False),
batch_norm=True,
activation=Softmax()))
self.layers = layers
model = Model(layers=layers)
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
model.initialize(train, cost=cost)
self.model = model
def test_model_get_outputs_rnn(backend_default, data):
data_path = load_ptb_test(path=data)
data_set = Text(time_steps=50, path=data_path)
# weight initialization
init = Constant(0.08)
# model initialization
layers = [
Recurrent(150, init, activation=Logistic()),
Affine(len(data_set.vocab), init, bias=init, activation=Rectlin())
]
model = Model(layers=layers)
output = model.get_outputs(data_set)
assert output.shape == (data_set.ndata, data_set.seq_length,
data_set.nclass)
# since the init are all constant and model is un-trained:
# along the feature dim, the values should be all the same
assert np.allclose(output[0, 0], output[0, 0, 0], rtol=0, atol=1e-5)
assert np.allclose(output[0, 1], output[0, 1, 0], rtol=0, atol=1e-5)
# along the time dim, the values should be increasing:
assert np.alltrue(output[0, 2] > output[0, 1])
assert np.alltrue(output[0, 1] > output[0, 0])
def main_branch(branch_nodes):
return [
Conv((7, 7, 64), padding=3, strides=2, name='conv1/7x7_s2', **common),
Pooling(name="pool1/3x3_s2", **pool3s2p1),
Conv((1, 1, 64), name='conv2/3x3_reduce', **common),
Conv((3, 3, 192), name="conv2/3x3", **commonp1),
Pooling(name="pool2/3x3_s2", **pool3s2p1),
inception([(64, ), (96, 128), (16, 32), (32, )], name='inception_3a/'),
inception([(128, ), (128, 192), (32, 96), (64, )],
name='inception_3b/'),
Pooling(name='pool3/3x3_s2', **pool3s2p1),
inception([(192, ), (96, 208), (16, 48), (64, )],
name='inception_4a/'), branch_nodes[0],
inception([(160, ), (112, 224), (24, 64), (64, )],
name='inception_4b/'),
inception([(128, ), (128, 256), (24, 64), (64, )],
name='inception_4c/'),
inception([(112, ), (144, 288), (32, 64), (64, )],
name='inception_4d/'), branch_nodes[1],
inception([(256, ), (160, 320), (32, 128), (128, )],
name='inception_4e/'),
Pooling(name='pool4/3x3_s2', **pool3s2p1),
inception([(256, ), (160, 320), (32, 128), (128, )],
name='inception_5a/'),
inception([(384, ), (192, 384), (48, 128), (128, )],
name="inception_5b/"),
Pooling(fshape=7, strides=1, op="avg", name='pool5/7x7_s1'),
Affine(nout=1000,
init=init1,
activation=Softmax(),
bias=Constant(0),
name='loss3/classifier')
]