def _get_generator(proto):
if proto.type == 'Normal':
return NormalInitializer(sigma=proto.multiplier)
elif proto.type == 'Uniform':
return UniformInitializer(lim=(-proto.multiplier, proto.multiplier))
elif proto.type == 'Range':
return RangeInitializer(start=0, step=proto.multiplier)
elif proto.type == 'Constant':
return ConstantInitializer(value=proto.multiplier)
else:
raise ValueError('Generator type "' + proto.type +
'" is not supported.')
def _create_variable(v, name, shape, rng):
# Create and initialize variables
class Variable:
pass
parameter = v.type == "Parameter"
variable_instance = None
if parameter:
if v.initializer.type == 'Normal':
initializer = NormalInitializer(v.initializer.multiplier, rng=rng)
elif v.initializer.type == 'NormalAffineHe' or v.initializer.type == 'NormalAffineHeForward':
initializer = (lambda shape: NormalInitializer(
calc_normal_std_he_forward(shape[0], np.prod(shape[1:])),
rng=rng)(shape) * v.initializer.multiplier)
elif v.initializer.type == 'NormalAffineHeBackward':
initializer = (lambda shape: NormalInitializer(
calc_normal_std_he_backward(shape[0], np.prod(shape[1:])),
rng=rng)(shape) * v.initializer.multiplier)
elif v.initializer.type == 'NormalAffineGlorot':
initializer = (lambda shape: NormalInitializer(
calc_normal_std_glorot(shape[0], np.prod(shape[1:])), rng=rng)
(shape) * v.initializer.multiplier)
elif v.initializer.type == 'NormalConvolutionHe' or v.initializer.type == 'NormalConvolutionHeForward':
initializer = (
lambda shape: NormalInitializer(calc_normal_std_he_forward(
shape[-3], shape[0], kernel=shape[-2:]),
rng=rng)
(shape) * v.initializer.multiplier)
elif v.initializer.type == 'NormalConvolutionHeBackward':
initializer = (
lambda shape: NormalInitializer(calc_normal_std_he_backward(
shape[-3], shape[0], kernel=shape[-2:]),
rng=rng)
(shape) * v.initializer.multiplier)
elif v.initializer.type == 'NormalConvolutionGlorot':
initializer = (lambda shape: NormalInitializer(
calc_normal_std_glorot(shape[-3], shape[0], kernel=shape[-2:]),
rng=rng)(shape) * v.initializer.multiplier)
elif v.initializer.type == 'NormalCLConvHe' or v.initializer.type == 'NormalCLConvHeForward':
initializer = (
lambda shape: NormalInitializer(calc_normal_std_he_forward(
shape[-1], shape[0], kernel=shape[1:3]),
rng=rng)
(shape) * v.initializer.multiplier)
elif v.initializer.type == 'NormalCLConvHeBackward':
initializer = (
lambda shape: NormalInitializer(calc_normal_std_he_backward(
shape[-1], shape[0], kernel=shape[1:3]),
rng=rng)
(shape) * v.initializer.multiplier)
elif v.initializer.type == 'NormalCLConvGlorot':
initializer = (lambda shape: NormalInitializer(
calc_normal_std_glorot(shape[-1], shape[0], kernel=shape[1:3]),
rng=rng)(shape) * v.initializer.multiplier)
elif v.initializer.type == 'Uniform':
initializer = UniformInitializer(
lim=[-v.initializer.multiplier, v.initializer.multiplier],
rng=rng)
elif v.initializer.type == 'UniformAffineGlorot':
initializer = (lambda shape: UniformInitializer(
calc_uniform_lim_glorot(shape[0], np.prod(shape[1:])), rng=rng)
(shape) * v.initializer.multiplier)
elif v.initializer.type == 'UniformConvolutionGlorot':
initializer = (
lambda shape: UniformInitializer(calc_uniform_lim_glorot(
shape[-3], shape[0], kernel=shape[-2:]),
rng=rng)
(shape) * v.initializer.multiplier)
elif v.initializer.type == 'UniformCLConvGlorot':
initializer = (
lambda shape: UniformInitializer(calc_uniform_lim_glorot(
shape[-1], shape[0], kernel=shape[1:3]),
rng=rng)
(shape) * v.initializer.multiplier)
elif v.initializer.type == 'Range':
initializer = (lambda shape: RangeInitializer(0, 1)
(shape) * v.initializer.multiplier)
elif v.initializer.type == 'Constant':
initializer = ConstantInitializer(value=v.initializer.multiplier)
else:
initializer = None
print("create parameter: {}".format(name))
variable_instance = get_parameter_or_create(name, shape, initializer)
else:
# create empty variable, memory will be allocated in network.setup()
# after network optimization
variable_instance = nn.Variable()
variable = Variable()
variable.name = name
variable.parameter = parameter
variable.shape = shape
variable.variable_instance = variable_instance
return variable
