def check(*input_dim):
sbn = SpatialBatchNormalization(input_dim)
sbn.initialize()
x = theano.tensor.TensorType(theano.config.floatX,
[False] * (len(input_dim) + 1))()
y = sbn.apply(x)
rng = numpy.random.RandomState((2015, 12, 17))
input_ = random_unif(rng, (11,) + input_dim)
assert_equal(y.eval({x: input_}), input_)
class ResidualConvolutional(Initializable):
@lazy(allocation=['filter_size', 'num_filters', 'num_channels'])
def __init__(self, filter_size, num_filters, num_channels,
batch_size=None,
mid_noise=False,
out_noise=False,
tied_noise=False,
tied_sigma=False,
noise_rate=None,
noise_batch_size=None,
prior_noise_level=None,
image_size=(None, None), step=(1, 1),
**kwargs):
self.filter_size = filter_size
self.num_filters = num_filters
self.batch_size = batch_size
self.num_channels = num_channels
self.image_size = image_size
self.mid_noise = mid_noise
self.noise_batch_size = noise_batch_size
self.noise_rate = noise_rate
self.step = step
self.border_mode = 'half'
self.tied_biases = True
depth = 2
self.b0 = SpatialBatchNormalization(name='b0')
self.r0 = Rectifier(name='r0')
self.n0 = (SpatialNoise(name='n0', noise_rate=self.noise_rate,
tied_noise=tied_noise, tied_sigma=tied_sigma,
prior_noise_level=prior_noise_level) if mid_noise else None)
self.c0 = Convolutional(name='c0')
self.b1 = SpatialBatchNormalization(name='b1')
self.r1 = Rectifier(name='r1')
self.n1 = (SpatialNoise(name='n1', noise_rate=self.noise_rate,
tied_noise=tied_noise, tied_sigma=tied_sigma,
prior_noise_level=prior_noise_level) if out_noise else None)
self.c1 = Convolutional(name='c1')
kwargs.setdefault('children', []).extend([c for c in [
self.c0, self.b0, self.r0, self.n0,
self.c1, self.b1, self.r1, self.n1] if c is not None])
super(ResidualConvolutional, self).__init__(**kwargs)
def get_dim(self, name):
if name == 'input_':
return ((self.num_channels,) + self.image_size)
if name == 'output':
return self.c1.get_dim(name)
return super(ResidualConvolutionalUnit, self).get_dim(name)
@property
def num_output_channels(self):
return self.num_filters
def _push_allocation_config(self):
self.b0.input_dim = self.get_dim('input_')
self.b0.push_allocation_config()
if self.r0:
self.r0.push_allocation_config()
if self.n0:
self.n0.noise_batch_size = self.noise_batch_size
self.n0.num_channels = self.num_channels
self.n0.image_size = self.image_size
self.c0.filter_size = self.filter_size
self.c0.batch_size = self.batch_size
self.c0.num_channels = self.num_channels
self.c0.num_filters = self.num_filters
self.c0.border_mode = self.border_mode
self.c0.image_size = self.image_size
self.c0.step = self.step
self.c0.use_bias = False
self.c0.push_allocation_config()
c0_shape = self.c0.get_dim('output')
self.b1.input_dim = c0_shape
self.b1.push_allocation_config()
self.r1.push_allocation_config()
if self.n1:
self.n1.noise_batch_size = self.noise_batch_size
self.n1.num_channels = self.num_filters
self.n1.image_size = c0_shape[1:]
self.c1.filter_size = self.filter_size
self.c1.batch_size = self.batch_size
self.c1.num_channels = self.num_filters
self.c1.num_filters = self.num_filters
self.c1.border_mode = self.border_mode
self.c1.image_size = c0_shape[1:]
self.c1.step = (1, 1)
self.c1.use_bias = False
self.c1.push_allocation_config()
@application(inputs=['input_'], outputs=['output'])
def apply(self, input_):
shortcut = input_
# Batchnorm, then Relu, then Convolution
first_conv = self.b0.apply(input_)
first_conv = self.r0.apply(first_conv)
if self.n0:
first_conv = self.n0.apply(first_conv)
first_conv = self.c0.apply(first_conv)
# Batchnorm, then Relu, then Convolution (second time)
second_conv = self.b1.apply(first_conv)
second_conv = self.r1.apply(second_conv)
if self.n1:
second_conv = self.n1.apply(second_conv)
residual = second_conv
# Apply stride and zero-padding to match shortcut to output
if self.step and self.step != (1, 1):
shortcut = shortcut[:,:,::self.step[0],::self.step[1]]
if self.num_filters > self.num_channels:
padshape = (residual.shape[0],
self.num_filters - self.num_channels,
residual.shape[2], residual.shape[3])
shortcut = tensor.concatenate(
[shortcut, tensor.zeros(padshape, dtype=residual.dtype)],
axis=1)
elif self.num_filters < self.num_channels:
shortcut = shortcut[:,:self.num_channels,:,:]
response = shortcut + residual
return response
class ResidualConvolutional(Initializable):
@lazy(allocation=['filter_size', 'num_filters', 'num_channels'])
def __init__(self,
filter_size,
num_filters,
num_channels,
batch_size=None,
mid_noise=False,
out_noise=False,
tied_noise=False,
tied_sigma=False,
noise_rate=None,
noise_batch_size=None,
prior_noise_level=None,
image_size=(None, None),
step=(1, 1),
**kwargs):
self.filter_size = filter_size
self.num_filters = num_filters
self.batch_size = batch_size
self.num_channels = num_channels
self.image_size = image_size
self.mid_noise = mid_noise
self.noise_batch_size = noise_batch_size
self.noise_rate = noise_rate
self.step = step
self.border_mode = 'half'
self.tied_biases = True
depth = 2
self.b0 = SpatialBatchNormalization(name='b0')
self.r0 = Rectifier(name='r0')
self.n0 = (SpatialNoise(name='n0',
noise_rate=self.noise_rate,
tied_noise=tied_noise,
tied_sigma=tied_sigma,
prior_noise_level=prior_noise_level)
if mid_noise else None)
self.c0 = Convolutional(name='c0')
self.b1 = SpatialBatchNormalization(name='b1')
self.r1 = Rectifier(name='r1')
self.n1 = (SpatialNoise(name='n1',
noise_rate=self.noise_rate,
tied_noise=tied_noise,
tied_sigma=tied_sigma,
prior_noise_level=prior_noise_level)
if out_noise else None)
self.c1 = Convolutional(name='c1')
kwargs.setdefault('children', []).extend([
c for c in [
self.c0, self.b0, self.r0, self.n0, self.c1, self.b1, self.r1,
self.n1
] if c is not None
])
super(ResidualConvolutional, self).__init__(**kwargs)
def get_dim(self, name):
if name == 'input_':
return ((self.num_channels, ) + self.image_size)
if name == 'output':
return self.c1.get_dim(name)
return super(ResidualConvolutionalUnit, self).get_dim(name)
@property
def num_output_channels(self):
return self.num_filters
def _push_allocation_config(self):
self.b0.input_dim = self.get_dim('input_')
self.b0.push_allocation_config()
if self.r0:
self.r0.push_allocation_config()
if self.n0:
self.n0.noise_batch_size = self.noise_batch_size
self.n0.num_channels = self.num_channels
self.n0.image_size = self.image_size
self.c0.filter_size = self.filter_size
self.c0.batch_size = self.batch_size
self.c0.num_channels = self.num_channels
self.c0.num_filters = self.num_filters
self.c0.border_mode = self.border_mode
self.c0.image_size = self.image_size
self.c0.step = self.step
self.c0.use_bias = False
self.c0.push_allocation_config()
c0_shape = self.c0.get_dim('output')
self.b1.input_dim = c0_shape
self.b1.push_allocation_config()
self.r1.push_allocation_config()
if self.n1:
self.n1.noise_batch_size = self.noise_batch_size
self.n1.num_channels = self.num_filters
self.n1.image_size = c0_shape[1:]
self.c1.filter_size = self.filter_size
self.c1.batch_size = self.batch_size
self.c1.num_channels = self.num_filters
self.c1.num_filters = self.num_filters
self.c1.border_mode = self.border_mode
self.c1.image_size = c0_shape[1:]
self.c1.step = (1, 1)
self.c1.use_bias = False
self.c1.push_allocation_config()
@application(inputs=['input_'], outputs=['output'])
def apply(self, input_):
shortcut = input_
# Batchnorm, then Relu, then Convolution
first_conv = self.b0.apply(input_)
first_conv = self.r0.apply(first_conv)
if self.n0:
first_conv = self.n0.apply(first_conv)
first_conv = self.c0.apply(first_conv)
# Batchnorm, then Relu, then Convolution (second time)
second_conv = self.b1.apply(first_conv)
second_conv = self.r1.apply(second_conv)
if self.n1:
second_conv = self.n1.apply(second_conv)
residual = second_conv
# Apply stride and zero-padding to match shortcut to output
if self.step and self.step != (1, 1):
shortcut = shortcut[:, :, ::self.step[0], ::self.step[1]]
if self.num_filters > self.num_channels:
padshape = (residual.shape[0],
self.num_filters - self.num_channels,
residual.shape[2], residual.shape[3])
shortcut = tensor.concatenate(
[shortcut,
tensor.zeros(padshape, dtype=residual.dtype)],
axis=1)
elif self.num_filters < self.num_channels:
shortcut = shortcut[:, :self.num_channels, :, :]
response = shortcut + residual
return response