def __init__(self, size=None, decay=0.9, eps=2e-5, dtype=numpy.float32,
use_gamma=True, use_beta=True,
initial_gamma=None, initial_beta=None, axis=None):
super(BatchNormalization, self).__init__()
if size is None and axis is None:
raise RuntimeError('size or axis is required')
self.N = 0
self.register_persistent('N')
self.decay = decay
self.eps = eps
if isinstance(axis, int):
axis = (axis,)
self.axis = axis
self._dtype = dtype
with self.init_scope():
if use_gamma:
if initial_gamma is None:
initial_gamma = 1
gamma_initializer = \
initializers._get_initializer(initial_gamma)
gamma_initializer.dtype = self._dtype
self.gamma = variable.Parameter(gamma_initializer)
if use_beta:
if initial_beta is None:
initial_beta = 0
beta_initializer = initializers._get_initializer(initial_beta)
beta_initializer.dtype = self._dtype
self.beta = variable.Parameter(beta_initializer)
if size is not None:
self._initialize_params(size)
def __init__(self, in_channels, out_channels, ksize, stride=1, pad=0,
nobias=False, initialW=None, initial_bias=None):
super(DeformableConvolution2DSampler, self).__init__()
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.out_channels = out_channels
self.initialW = initialW
if initialW is None:
initialW = constant.Zero()
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = initializers.Constant(0)
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer)
if in_channels is not None:
self._initialize_params(in_channels)
def __init__(self, in_channels, out_channels, ksize=None, stride=1, pad=0,
dilate=1, nobias=False, initialW=None, initial_bias=None):
super(DilatedConvolution2D, self).__init__()
if ksize is None:
out_channels, ksize, in_channels = in_channels, out_channels, None
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.dilate = _pair(dilate)
self.out_channels = out_channels
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if in_channels is not None:
self._initialize_params(in_channels)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = 0
initial_bias = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(initial_bias, out_channels)
def __init__(self, in_channels, out_channels, ksize, stride=1, pad=0,
wscale=1, bias=0, nobias=False, use_cudnn=True,
initialW=None, initial_bias=None):
super(Convolution2D, self).__init__()
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.use_cudnn = use_cudnn
self.out_channels = out_channels
# For backward compatibility
self.initialW = initialW
self.wscale = wscale
# For backward compatibility, the scale of weights is proportional to
# the square root of wscale.
self._W_initializer = initializers._get_initializer(
initialW, scale=math.sqrt(wscale))
if in_channels is None:
self.add_uninitialized_param('W')
else:
self._initialize_params(in_channels)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = bias
bias_initilizer = initializers._get_initializer(initial_bias)
self.add_param('b', out_channels, initializer=bias_initilizer)
def __init__(self, size, comm, decay=0.9, eps=2e-5, dtype=numpy.float32,
use_gamma=True, use_beta=True,
initial_gamma=None, initial_beta=None):
chainer.utils.experimental(
'chainermn.links.MultiNodeBatchNormalization')
if chainer.__version__.startswith('1.'):
raise RuntimeError(
'MultiNodeBatchNormalization works only with '
'chainer >= 2.0.0.')
super(MultiNodeBatchNormalization, self).__init__()
self.comm = comm
self.avg_mean = numpy.zeros(size, dtype=dtype)
self.register_persistent('avg_mean')
self.avg_var = numpy.zeros(size, dtype=dtype)
self.register_persistent('avg_var')
self.N = 0
self.register_persistent('N')
self.decay = decay
self.eps = eps
with self.init_scope():
if use_gamma:
if initial_gamma is None:
initial_gamma = 1
initial_gamma = initializers._get_initializer(initial_gamma)
initial_gamma.dtype = dtype
self.gamma = variable.Parameter(initial_gamma, size)
if use_beta:
if initial_beta is None:
initial_beta = 0
initial_beta = initializers._get_initializer(initial_beta)
initial_beta.dtype = dtype
self.beta = variable.Parameter(initial_beta, size)
def __init__(self, in_channels, channel_multiplier, ksize, stride=1, pad=0,
nobias=False, initialW=None, initial_bias=None):
super(DepthwiseConvolution2D, self).__init__()
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.channel_multiplier = channel_multiplier
self.nobias = nobias
if initialW is None:
initialW = initializers.HeNormal(1. / numpy.sqrt(2))
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = initializers.Constant(0)
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer)
if in_channels is not None:
self._initialize_params(in_channels)
def __init__(self, ndim, in_channels, out_channels, ksize=None, stride=1,
pad=0, nobias=False, initialW=None, initial_bias=None,
cover_all=False, dilate=1, groups=1):
super(ConvolutionND, self).__init__()
if ksize is None:
out_channels, ksize, in_channels = \
in_channels, out_channels, None
self.out_channels = out_channels
self.ksize = conv_nd.as_tuple(ksize, ndim)
self.stride = stride
self.pad = pad
self.cover_all = cover_all
self.dilate = conv_nd.as_tuple(dilate, ndim)
self.groups = int(groups)
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if in_channels is not None:
self._initialize_params(in_channels)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = 0
initial_bias = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(initial_bias, out_channels)
def __init__(self, in_channels, channel_multiplier, ksize, stride=1, pad=0,
wscale=1, bias=0, nobias=False, initialW=None,
initial_bias=None):
super(DepthwiseConvolution2D, self).__init__()
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.channel_multiplier = channel_multiplier
self.nobias = nobias
# For compatibility with Convolution2D
self.initialW = initialW
self.wscale = wscale
# For compatibility with Convolution2D, the scale of weights is
# proportional to the square root of wscale.
self._W_initializer = initializers._get_initializer(
initialW, scale=math.sqrt(wscale))
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = bias
self.bias_initilizer = initializers._get_initializer(initial_bias)
if in_channels is None:
self.add_uninitialized_param('b')
if in_channels is None:
self.add_uninitialized_param('W')
else:
self._initialize_params(in_channels)
def __init__(self, size, decay=0.9, eps=2e-5, dtype=numpy.float32,
use_gamma=True, use_beta=True,
initial_gamma=None, initial_beta=None):
super(BatchNormalization, self).__init__()
self.avg_mean = numpy.zeros(size, dtype=dtype)
self.register_persistent('avg_mean')
self.avg_var = numpy.zeros(size, dtype=dtype)
self.register_persistent('avg_var')
self.N = 0
self.register_persistent('N')
self.decay = decay
self.eps = eps
with self.init_scope():
if use_gamma:
if initial_gamma is None:
initial_gamma = 1
initial_gamma = initializers._get_initializer(initial_gamma)
initial_gamma.dtype = dtype
self.gamma = variable.Parameter(initial_gamma, size)
if use_beta:
if initial_beta is None:
initial_beta = 0
initial_beta = initializers._get_initializer(initial_beta)
initial_beta.dtype = dtype
self.beta = variable.Parameter(initial_beta, size)
def __init__(self, in_channels, out_channels, ksize=None, stride=1, pad=0,
nobias=False, initialW=None, initial_bias=None, **kwargs):
super(Convolution2D, self).__init__()
argument.check_unexpected_kwargs(
kwargs, deterministic="deterministic argument is not "
"supported anymore. "
"Use chainer.using_config('cudnn_deterministic', value) "
"context where value is either `True` or `False`.")
dilate, = argument.parse_kwargs(kwargs, ('dilate', 1))
if ksize is None:
out_channels, ksize, in_channels = in_channels, out_channels, None
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.dilate = _pair(dilate)
self.out_channels = out_channels
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if in_channels is not None:
self._initialize_params(in_channels)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = 0
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer, out_channels)
def __init__(self, in_size, out_size, wscale=1, bias=0, nobias=False,
initialW=None, initial_bias=None):
super(Linear, self).__init__()
# For backward compatibility
self.initialW = initialW
self.wscale = wscale
self.out_size = out_size
# For backward compatibility, the scale of weights is proportional to
# the square root of wscale.
self._W_initializer = initializers._get_initializer(
initialW, math.sqrt(wscale))
if in_size is None:
self.add_uninitialized_param('W')
else:
self._initialize_params(in_size)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = bias
bias_initializer = initializers._get_initializer(initial_bias)
self.add_param('b', out_size, initializer=bias_initializer)
def __init__(self, in_channels, channel_multiplier, ksize, stride=1, pad=0,
nobias=False, initialW=None, initial_bias=None):
super(DepthwiseConvolution2D, self).__init__()
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.channel_multiplier = channel_multiplier
self.nobias = nobias
if initialW is None:
initialW = initializers.HeNormal(1. / numpy.sqrt(2))
self._W_initializer = initializers._get_initializer(initialW)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = initializers.Constant(0)
self.bias_initilizer = initializers._get_initializer(initial_bias)
if in_channels is None:
self.add_uninitialized_param('b')
if in_channels is None:
self.add_uninitialized_param('W')
else:
self._initialize_params(in_channels)
def __init__(self, size, comm, decay=0.9, eps=2e-5, dtype=numpy.float32,
use_gamma=True, use_beta=True,
initial_gamma=None, initial_beta=None,
communication_backend='auto'):
chainer.utils.experimental(
'chainermn.links.MultiNodeBatchNormalization')
super(MultiNodeBatchNormalization, self).__init__()
self.comm = comm
self.avg_mean = numpy.zeros(size, dtype=dtype)
self.register_persistent('avg_mean')
self.avg_var = numpy.zeros(size, dtype=dtype)
self.register_persistent('avg_var')
self.N = 0
self.register_persistent('N')
self.decay = decay
self.eps = eps
self._communication_backend = \
get_communication_backend(comm, communication_backend)
with self.init_scope():
if use_gamma:
if initial_gamma is None:
initial_gamma = 1
initial_gamma = initializers._get_initializer(initial_gamma)
initial_gamma.dtype = dtype
self.gamma = variable.Parameter(initial_gamma, size)
if use_beta:
if initial_beta is None:
initial_beta = 0
initial_beta = initializers._get_initializer(initial_beta)
initial_beta.dtype = dtype
self.beta = variable.Parameter(initial_beta, size)
def _initialize_params(self):
lateral_init = initializers._get_initializer(self.lateral_init)
upward_init = initializers._get_initializer(self.upward_init)
for i in six.moves.range(0, 4 * self.state_size, self.state_size):
lateral_init(self.lateral.W.data[i:i + self.state_size, :])
upward_init(self.upward.W.data[i:i + self.state_size, :])
a, i, f, o = lstm._extract_gates(
self.upward.b.data.reshape(1, 4 * self.state_size, 1))
_init_weight(a, self.bias_init)
_init_weight(i, self.bias_init)
_init_weight(f, self.forget_bias_init)
_init_weight(o, self.bias_init)
def __init__(self, n_channel, initial=0, eps=1e-5):
super(Normalize, self).__init__()
self.eps = eps
with self.init_scope():
initializer = initializers._get_initializer(initial)
self.scale = variable.Parameter(initializer)
self.scale.initialize((n_channel),)
def __init__(self, in_channels, out_channels, ksize=None, stride=1, pad=0,
nobias=False, outsize=None, initialV=None,
**kwargs):
super(Deconvolution2D, self).__init__()
argument.check_unexpected_kwargs(
kwargs, deterministic="deterministic argument is not "
"supported anymore. "
"Use chainer.using_config('cudnn_deterministic', value) "
"context where value is either `True` or `False`.")
argument.assert_kwargs_empty(kwargs)
if ksize is None:
out_channels, ksize, in_channels = in_channels, out_channels, None
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.outsize = (None, None) if outsize is None else outsize
self.out_channels = out_channels
self.nobias = nobias
with self.init_scope():
V_initializer = initializers._get_initializer(initialV)
self.V = variable.Parameter(V_initializer)
if in_channels is not None:
kh, kw = _pair(self.ksize)
V_shape = (in_channels, self.out_channels, kh, kw)
self.V.initialize(V_shape)
self.b = None if nobias else variable.Parameter(None)
self.g = variable.Parameter(None)
def _init_array(self, initializer, default_value, size):
if initializer is None:
initializer = default_value
initializer = initializers._get_initializer(initializer)
return initializers.generate_array(
initializer, size, self.xp, dtype=self._highprec_dtype,
device=self.device)
def _initialize_params(self):
lateral_init = initializers._get_initializer(self.lateral_init)
upward_init = initializers._get_initializer(self.upward_init)
bias_init = initializers._get_initializer(self.bias_init)
forget_bias_init = initializers._get_initializer(self.forget_bias_init)
for i in six.moves.range(0, 4 * self.state_size, self.state_size):
lateral_init(self.lateral.W.array[i:i + self.state_size, :])
upward_init(self.upward.W.array[i:i + self.state_size, :])
a, i, f, o = lstm._extract_gates(
self.upward.b.array.reshape(1, 4 * self.state_size, 1))
bias_init(a)
bias_init(i)
forget_bias_init(f)
bias_init(o)
def test_scalar(self):
init = initializers._get_initializer(10)
self.assertIsInstance(init, initializers.Constant)
x = numpy.empty((2, 3), dtype=numpy.int32)
init(x)
expected = numpy.full((2, 3), 10, dtype=numpy.int32)
numpy.testing.assert_array_equal(x, expected)
def __init__(self, in_size, out_size, wscale=1,
ratio=.5, initialW=None, initial_bias=0):
super(SimplifiedDropconnect, self).__init__()
self.out_size = out_size
self.ratio = ratio
# Square root is for the compatibility with Linear Function.
self._W_initializer = initializers._get_initializer(
initialW, math.sqrt(wscale))
if in_size is None:
self.add_uninitialized_param('W')
else:
self._initialize_params(in_size)
if initial_bias is not None:
bias_initializer = initializers._get_initializer(initial_bias)
self.add_param('b', out_size, initializer=bias_initializer)
def __init__(self, ndim, in_channels, out_channels, ksize, stride=1, pad=0,
outsize=None, initialW=None, initial_bias=0, use_cudnn=True):
ksize = conv_nd.as_tuple(ksize, ndim)
self.stride = stride
self.pad = pad
self.use_cudnn = use_cudnn
self.outsize = outsize
super(DeconvolutionND, self).__init__()
W_shape = (in_channels, out_channels) + ksize
initialW = initializers._get_initializer(initialW)
self.add_param('W', W_shape, initializer=initialW)
if initial_bias is None:
self.b = None
else:
initial_bias = initializers._get_initializer(initial_bias)
self.add_param('b', out_channels, initializer=initial_bias)
def __init__(self,
in_channels,
out_channels,
ksize,
stride=1,
pad=0,
wscale=1,
bias=0,
nobias=False,
use_cudnn=True,
initialW=None,
initial_bias=None,
deterministic=False):
super(Convolution2D, self).__init__()
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.use_cudnn = use_cudnn
self.out_channels = out_channels
self.deterministic = deterministic
# For backward compatibility
self.initialW = initialW
self.wscale = wscale
# For backward compatibility, the scale of weights is proportional to
# the square root of wscale.
self._W_initializer = initializers._get_initializer(
initialW, scale=math.sqrt(wscale))
if in_channels is None:
self.add_uninitialized_param('W')
else:
self._initialize_params(in_channels)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = bias
bias_initilizer = initializers._get_initializer(initial_bias)
self.add_param('b', out_channels, initializer=bias_initilizer)
def __init__(self,
in_channels,
out_channels,
ksize=None,
stride=1,
pad=0,
nobias=False,
initialW=None,
initial_bias=None,
**kwargs):
super(Convolution2D, self).__init__()
argument.check_unexpected_kwargs(
kwargs,
deterministic="deterministic argument is not "
"supported anymore. "
"Use chainer.using_config('cudnn_deterministic', value) "
"context where value is either `True` or `False`.")
dilate, = argument.parse_kwargs(kwargs, ('dilate', 1))
if ksize is None:
out_channels, ksize, in_channels = in_channels, out_channels, None
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.dilate = _pair(dilate)
self.out_channels = out_channels
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if in_channels is not None:
self._initialize_params(in_channels)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = 0
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer, out_channels)
def __init__(self,
ndim,
in_channels,
out_channels,
ksize,
stride=1,
pad=0,
nobias=False,
initialW=None,
initial_bias=None,
cover_all=False,
use_gamma=False,
Ip=1,
factor=None):
super(SNConvolutionND, self).__init__()
ksize = conv_nd.as_tuple(ksize, ndim)
self.stride = stride
self.pad = pad
self.cover_all = cover_all
self.use_gamma = use_gamma
self.Ip = Ip
self.u = np.random.normal(size=(1, out_channels)).astype(dtype="f")
self.register_persistent('u')
self.factor = factor
with self.init_scope():
W_shape = (out_channels, in_channels) + ksize
self.W = variable.Parameter(
initializers._get_initializer(initialW), W_shape)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = 0
initial_bias = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(initial_bias, out_channels)
if self.use_gamma:
W_mat = self.W.data.reshape(self.W.shape[0], -1)
_, s, _ = np.linalg.svd(W_mat)
self.gamma = variable.Parameter(s[0],
(1, ) * len(self.W.shape))
def __init__(self, in_channels, hidden_channel, initialW=None):
super().__init__()
ksize = (1, 1, 1)
self.W_shape = (hidden_channel, in_channels) + ksize
self.stride = 1
self.pad = 0
self.nobias = True
self.b = None
self.initialW = initialW
with self.init_scope():
self.W = Parameter(_get_initializer(self.initialW), self.W_shape)
def __init__(self, left_size, right_size, out_size, initialW=None):
super(BilinearDiag, self).__init__()
self.in_sizes = (left_size, right_size)
assert left_size == right_size
with self.init_scope():
shape = (out_size, left_size)
if isinstance(initialW, (numpy.ndarray, cuda.ndarray)):
assert initialW.shape == shape
self.W = variable.Parameter(
initializers._get_initializer(initialW), shape)
def __init__(self,
initial_param1=None,
initial_param2=None,
alpha=1,
gpu=False):
super(MultiLoss, self).__init__()
self.alpha = alpha
self.max_loss_detection = None
self.max_loss_segmentation = None
with self.init_scope():
if initial_param1 is None:
initial_param1 = 0
if initial_param2 is None:
initial_param2 = 0
param1_initializer = initializers._get_initializer(initial_param1)
self.param1 = variable.Parameter(param1_initializer)
param2_initializer = initializers._get_initializer(initial_param2)
self.param2 = variable.Parameter(param2_initializer)
def test_numpy_array(self):
c = numpy.array([1, 2, 3])
init = initializers._get_initializer(c)
self.assertIsInstance(init, initializers.Constant)
x = numpy.empty((3,), dtype=numpy.int32)
init(x)
expected = numpy.array([1, 2, 3], dtype=numpy.int32)
numpy.testing.assert_array_equal(x, expected)
def test_numpy_array(self):
c = numpy.array([1, 2, 3])
init = initializers._get_initializer(c)
self.assertIsInstance(init, initializers.Constant)
x = numpy.empty((3, ), dtype=numpy.int32)
init(x)
expected = numpy.array([1, 2, 3], dtype=numpy.int32)
numpy.testing.assert_array_equal(x, expected)
def __init__(self,
size,
comm,
decay=0.9,
eps=2e-5,
dtype=numpy.float32,
use_gamma=True,
use_beta=True,
initial_gamma=None,
initial_beta=None):
chainer.utils.experimental(
'chainermn.links.MultiNodeBatchNormalization')
if chainer.__version__.startswith('1.'):
raise RuntimeError('MultiNodeBatchNormalization works only with '
'chainer >= 2.0.0.')
super(MultiNodeBatchNormalization, self).__init__()
self.comm = comm
self.avg_mean = numpy.zeros(size, dtype=dtype)
self.register_persistent('avg_mean')
self.avg_var = numpy.zeros(size, dtype=dtype)
self.register_persistent('avg_var')
self.N = 0
self.register_persistent('N')
self.decay = decay
self.eps = eps
with self.init_scope():
if use_gamma:
if initial_gamma is None:
initial_gamma = 1
initial_gamma = initializers._get_initializer(initial_gamma)
initial_gamma.dtype = dtype
self.gamma = variable.Parameter(initial_gamma, size)
if use_beta:
if initial_beta is None:
initial_beta = 0
initial_beta = initializers._get_initializer(initial_beta)
initial_beta.dtype = dtype
self.beta = variable.Parameter(initial_beta, size)
def __init__(self, size, decay=0.9, eps=2e-5, dtype=np.float32,
use_gamma=True, use_beta=True,
initial_gamma=None, initial_beta=None):
super(ComplexBatchNorm, self).__init__()
self.avg_mean = np.zeros(size, dtype=dtype)
self.register_persistent('avg_mean')
self.avg_var = np.zeros(size, dtype=dtype)
self.register_persistent('avg_var')
self.N = 0
self.register_persistent('N')
self.decay = decay
self.eps = eps
self.moving_Vrr = Variable(cupy.full(size/2, 1./np.sqrt(2), dtype=dtype))
self.moving_Vii = Variable(cupy.full(size/2, 1./np.sqrt(2), dtype=dtype))
self.moving_Vri = Variable(cupy.zeros(size/2, dtype=dtype))
self.moving_real_mean = Variable(cupy.zeros(size/2, dtype=dtype))
self.moving_imag_mean = Variable(cupy.zeros(size/2, dtype=dtype))
with self.init_scope():
if use_gamma:
if initial_gamma is None:
initial_gamma_rr = 1./np.sqrt(2)
initial_gamma_ri = 0
initial_gamma_rr = initializers._get_initializer(initial_gamma_rr)
initial_gamma_rr.dtype = dtype
initial_gamma_ii = initializers._get_initializer(initial_gamma_rr)
initial_gamma_ii.dtype = dtype
initial_gamma_ri = initializers._get_initializer(initial_gamma_ri)
initial_gamma_ri.dtype = dtype
self.gamma_rr = variable.Parameter(initial_gamma_rr, size/2)
self.gamma_ii = variable.Parameter(initial_gamma_ii, size/2)
self.gamma_ri = variable.Parameter(initial_gamma_ri, size/2)
if use_beta:
if initial_beta is None:
initial_beta = 0
initial_beta = initializers._get_initializer(initial_beta)
initial_beta.dtype = dtype
self.beta = variable.Parameter(initial_beta, size)
def __init__(self, in_size, out_size=None, nobias=False,
initialW=None, initial_bias=None):
super(IncompleteBinaryLinear, self).__init__()
if out_size is None:
in_size, out_size = None, in_size
self.out_size = out_size
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if in_size is not None:
self._initialize_params(in_size)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = 0
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer, out_size)
def __init__(self, in_size, out_size=None, nobias=False,
initialW=None, initial_bias=None):
super(Linear, self).__init__()
if out_size is None:
in_size, out_size = None, in_size
self.out_size = out_size
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if in_size is not None:
self._initialize_params(in_size)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = 0
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer, out_size)
def __init__(self, axis=1, shape=None, initial_bias=None):
super(Bias, self).__init__()
# Add b parameter if given.
if shape is not None:
with self.init_scope():
if initial_bias is None:
initial_bias = 0
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer, shape)
self.axis = axis
def __init__(self, in_size, out_size, bias=0, nobias=False,
initialW=None, initial_bias=None):
super(Linear, self).__init__()
# For backward compatibility
self.initialW = initialW
self.out_size = out_size
self._W_initializer = initializers._get_initializer(initialW)
if in_size is None:
self.add_uninitialized_param('W')
else:
self._initialize_params(in_size)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = bias
bias_initializer = initializers._get_initializer(initial_bias)
self.add_param('b', out_size, initializer=bias_initializer)
def __init__(self, n_vocab, word_embedding_size, fix_embedding):
super(SentenceEmbedding, self).__init__()
with self.init_scope():
self.embed = L.EmbedID(n_vocab,
word_embedding_size,
ignore_label=-1)
M_initializer = I._get_initializer(None)
# This corresponds to tranpose(M) in original paper
self.M = chainer.variable.Parameter(initializer=M_initializer,
shape=(word_embedding_size,
word_embedding_size))
self.fix_embedding = fix_embedding
def __init__(self, units, in_units, drop_rate=0.5):
super(CondenseLinear, self).__init__()
drop_in_units = int(in_units * drop_rate)
with self.init_scope():
self.dense = L.Linear(in_size=drop_in_units, out_size=units)
self.index = initializers.generate_array(
initializer=initializers._get_initializer(0),
shape=(drop_in_units, ),
xp=self.xp,
dtype=np.int32)
self.register_persistent('index')
def __init__(self, ndim, in_channels, out_channels, ksize, stride=1, pad=0,
nobias=False, outsize=None,
initialW=None, initial_bias=None):
super(DeconvolutionND, self).__init__()
ksize = conv_nd.as_tuple(ksize, ndim)
self.stride = stride
self.pad = pad
self.outsize = outsize
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer,
(in_channels, out_channels) + ksize)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = 0
initial_bias = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(initial_bias, out_channels)
def test_callable(self):
def initializer(arr):
arr[...] = 100
init = initializers._get_initializer(initializer)
self.assertTrue(callable(init))
x = numpy.empty((2, 3), dtype=numpy.int32)
init(x)
expected = numpy.full((2, 3), 100, dtype=numpy.int32)
numpy.testing.assert_array_equal(x, expected)
def __init__(self,
ndim,
in_channels,
out_channels,
ksize=None,
stride=1,
pad=0,
nobias=False,
outsize=None,
initialW=None,
initial_bias=None,
dilate=1,
groups=1):
super(DeconvolutionND, self).__init__()
if ksize is None:
out_channels, ksize, in_channels = \
in_channels, out_channels, None
self.out_channels = out_channels
self.ksize = conv_nd.as_tuple(ksize, ndim)
self.stride = stride
self.pad = pad
self.outsize = outsize
self.dilate = conv_nd.as_tuple(dilate, ndim)
self.groups = int(groups)
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if in_channels is not None:
self._initialize_params(in_channels)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = 0
initial_bias = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(initial_bias, out_channels)
def __init__(self,
in_channels,
channel_multiplier,
ksize,
stride=1,
pad=0,
wscale=1,
bias=0,
nobias=False,
initialW=None,
initial_bias=None):
super(DepthwiseConvolution2D, self).__init__()
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.channel_multiplier = channel_multiplier
self.nobias = nobias
# For compatibility with Convolution2D
self.initialW = initialW
self.wscale = wscale
# For compatibility with Convolution2D, the scale of weights is
# proportional to the square root of wscale.
self._W_initializer = initializers._get_initializer(
initialW, scale=math.sqrt(wscale))
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = bias
self.bias_initilizer = initializers._get_initializer(initial_bias)
if in_channels is None:
self.add_uninitialized_param('b')
if in_channels is None:
self.add_uninitialized_param('W')
else:
self._initialize_params(in_channels)
def __init__(self,
size=None,
decay=0.9,
eps=2e-5,
dtype=None,
use_gamma=True,
use_beta=True,
initial_gamma=None,
initial_beta=None,
axis=None):
super(BatchNormalization, self).__init__()
if size is None and axis is None:
raise RuntimeError('size or axis is required')
self.N = 0
self.register_persistent('N')
self.decay = decay
self.eps = eps
if isinstance(axis, int):
axis = (axis, )
self.axis = axis
self._dtype = chainer.get_dtype(dtype)
with self.init_scope():
if use_gamma:
if initial_gamma is None:
initial_gamma = 1
gamma_initializer = \
initializers._get_initializer(initial_gamma)
gamma_initializer.dtype = self._dtype
self.gamma = variable.Parameter(gamma_initializer)
if use_beta:
if initial_beta is None:
initial_beta = 0
beta_initializer = initializers._get_initializer(initial_beta)
beta_initializer.dtype = self._dtype
self.beta = variable.Parameter(beta_initializer)
if size is not None:
self._initialize_params(size)
def __init__(self, ndim, in_channels, out_channels, ksize, stride=1, pad=0,
nobias=False, initialW=None, initial_bias=None,
cover_all=False):
super(ConvolutionND, self).__init__()
ksize = conv_nd.as_tuple(ksize, ndim)
self.stride = stride
self.pad = pad
self.cover_all = cover_all
with self.init_scope():
W_shape = (out_channels, in_channels) + ksize
self.W = variable.Parameter(
initializers._get_initializer(initialW), W_shape)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = 0
initial_bias = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(initial_bias, out_channels)
def __init__(self,
in_size,
out_size=None,
nobias=False,
initial_mu=None,
initial_sigma=None,
factorized=False):
super().__init__()
if out_size is None:
in_size, out_size = None, in_size
self.out_size = out_size
self.factorized = factorized
if initial_mu is None:
if self.factorized:
initial_mu = initializers.LeCunUniform(scale=np.sqrt(1 / 3))
else:
initial_mu = initializers.LeCunUniform(scale=1.0)
if initial_sigma is None:
if self.factorized:
initial_sigma = FanConstant(scale=0.5)
else:
initial_sigma = 0.017
with self.init_scope():
mu_initializer = initializers._get_initializer(initial_mu)
sigma_initializer = initializers._get_initializer(initial_sigma)
self.mu_w = variable.Parameter(mu_initializer)
self.sigma_w = variable.Parameter(sigma_initializer)
if nobias:
self.mu_b = None
self.sigma_b = None
else:
# `LeCunUniform` does not allow one-dim size initialization
self.mu_b = variable.Parameter(mu_initializer, (1, out_size))
self.sigma_b = variable.Parameter(sigma_initializer, out_size)
if in_size is not None:
self._initialize_params(in_size)
def __init__(self, in_size, slot_size, memory_size, initialW=None):
"""
in_size: hidden_state h_size
"""
super().__init__()
self.slot_size = slot_size
self.memory_size = memory_size
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.memory = variable.Parameter(W_initializer)
self.projection_matrix = variable.Parameter(W_initializer)
def __init__(self, in_channels, out_channels, in_size=None, ksize=None,
stride=1, nobias=False, initialW=None, initial_bias=None,
**kwargs):
super(LocalConvolution2D, self).__init__()
self.ksize = ksize
self.stride = _pair(stride)
self.nobias = nobias
self.out_channels = out_channels
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = 0
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer)
if in_channels is not None and in_size is not None:
self._initialize_params(in_channels, _pair(in_size))
def __init__(self, in_channels, out_channels, in_size=None, ksize=None,
stride=1, nobias=False, initialW=None, initial_bias=None,
**kwargs):
super(LocalConvolution2D, self).__init__()
self.ksize = ksize
self.stride = _pair(stride)
self.nobias = nobias
self.out_channels = out_channels
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = 0
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer)
if in_channels is not None and in_size is not None:
self._initialize_params(in_channels, _pair(in_size))
def __init__(self, in_size, out_size, ratio=.5, nobias=False,
initialW=None, initial_bias=None):
super(SimplifiedDropconnect, self).__init__()
self.out_size = out_size
self.ratio = ratio
if initialW is None:
initialW = initializers.HeNormal(1. / numpy.sqrt(2))
self.add_param('W', initializer=initializers._get_initializer(
initialW))
if in_size is not None:
self._initialize_params(in_size)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = initializers.Constant(0)
bias_initializer = initializers._get_initializer(initial_bias)
self.add_param('b', out_size, initializer=bias_initializer)
def test_callable(self):
def initializer(arr):
arr[...] = 100
init = initializers._get_initializer(initializer)
self.assertTrue(callable(init))
x = numpy.empty((2, 3), dtype=numpy.int32)
init(x)
expected = numpy.full((2, 3), 100, dtype=numpy.int32)
numpy.testing.assert_array_equal(x, expected)
def __init__(self,
size,
decay=0.9,
eps=2e-5,
dtype=numpy.float32,
use_gamma=True,
use_beta=True,
initial_gamma=None,
initial_beta=None,
norm_grad=False):
super(InstanceNormalization, self).__init__()
self.avg_mean = numpy.zeros(size, dtype=dtype)
self.register_persistent('avg_mean')
self.avg_var = numpy.zeros(size, dtype=dtype)
self.register_persistent('avg_var')
self.N = 0
self.register_persistent('N')
self.decay = decay
self.eps = eps
self.norm_grad = norm_grad
with self.init_scope():
if use_gamma:
if initial_gamma is None:
initial_gamma = 1
initial_gamma = initializers._get_initializer(initial_gamma)
initial_gamma.dtype = dtype
self.gamma = Parameter_scale(initial_gamma,
size,
norm_grad=self.norm_grad)
if use_beta:
if initial_beta is None:
initial_beta = 0
initial_beta = initializers._get_initializer(initial_beta)
initial_beta.dtype = dtype
self.beta = Parameter_scale(initial_beta,
size,
norm_grad=self.norm_grad)
def __init__(self, in_size, out_size, ratio=.5, nobias=False,
initialW=None, initial_bias=None):
super(SimplifiedDropconnect, self).__init__()
self.out_size = out_size
self.ratio = ratio
if initialW is None:
initialW = initializers.HeNormal(1. / numpy.sqrt(2))
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if in_size is not None:
self._initialize_params(in_size)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = initializers.Constant(0)
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer, out_size)
def __init__(self, in_size, out_size=None, nobias=False,
initialW=None, initial_bias=None):
# type: (tp.Optional[int], tp.Optional[int], bool, tp.Optional[types.InitializerSpec], tp.Optional[types.InitializerSpec]) -> None # NOQA
super(Linear, self).__init__()
if out_size is None:
in_size, out_size = None, in_size
self.out_size = out_size
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer) # type: variable.Variable # NOQA
if in_size is not None:
self._initialize_params(in_size)
if nobias:
self.b = None # type: tp.Optional[variable.Variable]
else:
if initial_bias is None:
initial_bias = 0
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer, out_size)
def __init__(self, in_size, out_size, ratio=.5, nobias=False,
initialW=None, initial_bias=None):
super(SimplifiedDropconnect, self).__init__()
self.out_size = out_size
self.ratio = ratio
if initialW is None:
initialW = initializers.HeNormal(1. / numpy.sqrt(2))
self._W_initializer = initializers._get_initializer(initialW)
if in_size is None:
self.add_uninitialized_param('W')
else:
self._initialize_params(in_size)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = initializers.Constant(0)
bias_initializer = initializers._get_initializer(initial_bias)
self.add_param('b', out_size, initializer=bias_initializer)
def __init__(self, in_size, out_size, ratio=.5, nobias=False,
initialW=None, initial_bias=None):
super(SimplifiedDropconnect, self).__init__()
self.out_size = out_size
self.ratio = ratio
if initialW is None:
initialW = initializers.HeNormal(1. / numpy.sqrt(2))
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if in_size is not None:
self._initialize_params(in_size)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = initializers.Constant(0)
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer, out_size)
def __init__(self, in_channels, out_channels, ksize=None, stride=1, pad=0,
nobias=False, outsize=None, initialW=None, initial_bias=None,
**kwargs):
super(Deconvolution2D, self).__init__()
dilate, groups, = argument.parse_kwargs(
kwargs, ('dilate', 1), ('groups', 1),
deterministic='deterministic argument is not supported anymore. '
'Use chainer.using_config(\'cudnn_deterministic\', value) '
'context where value is either `True` or `False`.')
if ksize is None:
out_channels, ksize, in_channels = in_channels, out_channels, None
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.dilate = _pair(dilate)
self.outsize = (None, None) if outsize is None else outsize
self.out_channels = out_channels
self.groups = int(groups)
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if in_channels is not None:
self._initialize_params(in_channels)
if nobias:
self.b = None
else:
if isinstance(initial_bias, (numpy.ndarray, cuda.ndarray)):
assert initial_bias.shape == (out_channels,)
if initial_bias is None:
initial_bias = 0
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer, out_channels)
def __init__(self, in_channels, out_channels, ksize=None, stride=1, pad=0,
nobias=False, outsize=None, initialW=None, initial_bias=None,
group=1, **kwargs):
super(Deconvolution2D, self).__init__()
argument.check_unexpected_kwargs(
kwargs, deterministic="deterministic argument is not "
"supported anymore. "
"Use chainer.using_config('cudnn_deterministic', value) "
"context where value is either `True` or `False`.")
argument.assert_kwargs_empty(kwargs)
if ksize is None:
out_channels, ksize, in_channels = in_channels, out_channels, None
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.outsize = (None, None) if outsize is None else outsize
self.out_channels = out_channels
self.group = int(group)
with self.init_scope():
W_initializer = initializers._get_initializer(initialW)
self.W = variable.Parameter(W_initializer)
if in_channels is not None:
self._initialize_params(in_channels)
if nobias:
self.b = None
else:
if isinstance(initial_bias, (numpy.ndarray, cuda.ndarray)):
assert initial_bias.shape == (out_channels,)
if initial_bias is None:
initial_bias = 0
bias_initializer = initializers._get_initializer(initial_bias)
self.b = variable.Parameter(bias_initializer, out_channels)
def __init__(self, size, decay=0.9, eps=2e-5, dtype=numpy.float32,
use_gamma=True, use_beta=True,
initial_gamma=None, initial_beta=None,
initial_avg_mean=None, initial_avg_var=None,
enforce_compute=False):
super(BatchNormalization2, self).__init__()
if initial_avg_mean is not None:
self.avg_mean = numpy.array(initial_avg_mean, dtype=dtype)
else:
self.avg_mean = numpy.zeros(size, dtype=dtype)
self.register_persistent('avg_mean')
if initial_avg_var is not None:
self.avg_var = numpy.array(initial_avg_var, dtype=dtype)
else:
self.avg_var = numpy.zeros(size, dtype=dtype)
self.register_persistent('avg_var')
self.N = 0
self.register_persistent('N')
self.decay = decay
self.eps = eps
self.enforce_compute=enforce_compute
with self.init_scope():
if use_gamma:
if initial_gamma is None:
initial_gamma = 1
initial_gamma = initializers._get_initializer(initial_gamma)
initial_gamma.dtype = dtype
self.gamma = variable.Parameter(initial_gamma, size)
if use_beta:
if initial_beta is None:
initial_beta = 0
initial_beta = initializers._get_initializer(initial_beta)
initial_beta.dtype = dtype
self.beta = variable.Parameter(initial_beta, size)
def __init__(self, size, comm, decay=0.9, eps=2e-5, dtype=None,
use_gamma=True, use_beta=True,
initial_gamma=None, initial_beta=None,
communication_backend='auto'):
chainer.utils.experimental(
'chainermn.links.MultiNodeBatchNormalization')
super(MultiNodeBatchNormalization, self).__init__()
self._highprec_dtype = chainer.get_dtype(
dtype, map_mixed16=numpy.float32)
self.comm = comm
self.avg_mean = numpy.zeros(size, dtype=self._highprec_dtype)
self.register_persistent('avg_mean')
self.avg_var = numpy.zeros(size, dtype=self._highprec_dtype)
self.register_persistent('avg_var')
self.N = 0
self.register_persistent('N')
self.decay = decay
self.eps = eps
self._communication_backend = \
chainermn_batch_normalization.get_communication_backend(
comm, communication_backend)
with self.init_scope():
if use_gamma:
if initial_gamma is None:
initial_gamma = 1
initial_gamma = initializers._get_initializer(initial_gamma)
initial_gamma.dtype = self._highprec_dtype
self.gamma = variable.Parameter(initial_gamma, size)
if use_beta:
if initial_beta is None:
initial_beta = 0
initial_beta = initializers._get_initializer(initial_beta)
initial_beta.dtype = self._highprec_dtype
self.beta = variable.Parameter(initial_beta, size)
def __init__(self,
in_channels,
out_channels,
ksize,
stride=1,
pad=0,
bias=0,
nobias=False,
use_cudnn=True,
initialW=None,
initial_bias=None,
deterministic=False):
super(Convolution2D, self).__init__()
self.ksize = ksize
self.stride = _pair(stride)
self.pad = _pair(pad)
self.use_cudnn = use_cudnn
self.out_channels = out_channels
self.deterministic = deterministic
# For backward compatibility
self.initialW = initialW
self._W_initializer = initializers._get_initializer(initialW)
if in_channels is None:
self.add_uninitialized_param('W')
else:
self._initialize_params(in_channels)
if nobias:
self.b = None
else:
if initial_bias is None:
initial_bias = bias
bias_initilizer = initializers._get_initializer(initial_bias)
self.add_param('b', out_channels, initializer=bias_initilizer)