def create_network():
layers = [
DataTransform(transform=Normalizer(divisor=128.)),
Conv((11, 11, 96),
init=Kaiming(),
activation=Rectlin(),
strides=4,
padding=1),
Conv((1, 1, 96), init=Kaiming(), activation=Rectlin(), strides=1),
Conv((3, 3, 96),
init=Kaiming(),
activation=Rectlin(),
strides=2,
padding=1), # 54->2,
Conv((5, 5, 256), init=Kaiming(), activation=Rectlin(),
strides=1), # 27->2,
Conv((1, 1, 256), init=Kaiming(), activation=Rectlin(), strides=1),
Conv((3, 3, 256),
init=Kaiming(),
activation=Rectlin(),
strides=2,
padding=1), # 23->1,
Conv((3, 3, 384),
init=Kaiming(),
activation=Rectlin(),
strides=1,
padding=1),
Conv((1, 1, 384), init=Kaiming(), activation=Rectlin(), strides=1),
Conv((3, 3, 384),
init=Kaiming(),
activation=Rectlin(),
strides=2,
padding=1), # 12->,
Dropout(keep=0.5),
Conv((3, 3, 1024),
init=Kaiming(),
activation=Rectlin(),
strides=1,
padding=1),
Conv((1, 1, 1024), init=Kaiming(), activation=Rectlin(), strides=1),
Conv((1, 1, 1000), init=Kaiming(), activation=Rectlin(), strides=1),
Pooling(6, op='avg'),
Activation(Softmax())
]
return Model(layers=layers), GeneralizedCost(costfunc=CrossEntropyMulti())
test.init_batch_provider()
relu = Rectlin()
init_uni = GlorotUniform()
# The parameters below are straight out of [Springenberg2014]
opt_gdm = GradientDescentMomentum(learning_rate=0.01,
schedule=Schedule(step_config=[10],
change=0.1),
momentum_coef=0.9, wdecay=.0005)
# set up model layers
layers = []
layers.append(DataTransform(transform=Normalizer(divisor=128.)))
layers.append(Conv((11, 11, 96), init=init_uni, activation=relu, strides=4, padding=1))
layers.append(Conv((1, 1, 96), init=init_uni, activation=relu, strides=1))
layers.append(Conv((3, 3, 96), init=init_uni, activation=relu, strides=2, padding=1)) # 54->27
layers.append(Conv((5, 5, 256), init=init_uni, activation=relu, strides=1)) # 27->23
layers.append(Conv((1, 1, 256), init=init_uni, activation=relu, strides=1))
layers.append(Conv((3, 3, 256), init=init_uni, activation=relu, strides=2, padding=1)) # 23->12
layers.append(Conv((3, 3, 384), init=init_uni, activation=relu, strides=1, padding=1))
layers.append(Conv((1, 1, 384), init=init_uni, activation=relu, strides=1))
layers.append(Conv((3, 3, 384), init=init_uni, activation=relu, strides=2, padding=1)) # 12->6
layers.append(Dropout(keep=0.5))
layers.append(Conv((3, 3, 1024), init=init_uni, activation=relu, strides=1, padding=1))