def _forward_grouped_convolution_xp(self, x, W, b, xp):
# G: group count
# N: batch size
# xC: input channels
# yC: output channels
G = self.groups
N, xC = x.shape[:2]
x_size = x.shape[2:]
yCg = W.shape[1]
yC = yCg * G
xCg = xC // G
k_size = W.shape[2:]
dims = len(k_size)
if xC % G != 0:
raise TypeError('The number of groups must be '
'a divisor of that of input channels')
x = xp.rollaxis(x, 1) # (xC, N, x_size...)
x = x.reshape(G, xCg, N * utils.size_of_shape(x_size))
W = W.reshape(G, xCg, yCg * utils.size_of_shape(k_size))
W = W.transpose(0, 2, 1) # (G, yCg*k_size, xCg)
# (G, yCg*k_size, N*x_size) = (G, yCg*k_size, xCg) @ (G, xCg, N*x_size)
col = convolution_2d._matmul(W, x).astype(x.dtype, copy=False)
col = col.reshape((yC,) + k_size + (N,) + x_size)
col = xp.rollaxis(col, dims + 1) # (N, yC, k_size..., x_size...)
y = conv_nd.col2im_nd(col, self.stride, self.pad, self.outs,
dilate=self.dilate)
if b is not None:
y += b.reshape(1, yC, *((1,) * dims))
return y,
def check_log_prob(self, is_gpu):
smp = self.sample_for_test()
if is_gpu:
log_prob1 = self.gpu_dist.log_prob(cuda.to_gpu(smp)).data
else:
log_prob1 = self.cpu_dist.log_prob(smp).data
if self.continuous:
scipy_prob = self.scipy_dist.logpdf
else:
scipy_prob = self.scipy_dist.logpmf
if self.scipy_onebyone:
onebyone_smp = smp.reshape(*[
utils.size_of_shape(sh)
for sh in [self.sample_shape, self.shape, self.event_shape]
])
onebyone_smp = numpy.swapaxes(onebyone_smp, 0, 1)
onebyone_smp = onebyone_smp.reshape((-1, ) + self.sample_shape +
self.event_shape)
log_prob2 = []
for one_params, one_smp in zip(self.scipy_onebyone_params_iter(),
onebyone_smp):
log_prob2.append(scipy_prob(one_smp, **one_params))
log_prob2 = numpy.vstack(log_prob2)
log_prob2 = log_prob2.reshape(utils.size_of_shape(self.shape),
-1).T
log_prob2 = log_prob2.reshape(self.sample_shape + self.shape)
else:
log_prob2 = scipy_prob(smp, **self.scipy_params)
array.assert_allclose(log_prob1, log_prob2)
def _forward_grouped_convolution_xp(self, x, gy, xp):
G = self.groups
N, iC = x.shape[:2]
oC = gy.shape[1]
o_size = gy.shape[2:]
o_size_prod = utils.size_of_shape(o_size)
k_size = self.ksize
dims = len(o_size)
iCg = iC // G
oCg = oC // G
# Do not check iCg and oCg because this class is rarely used alone
# (N, iC, k_size..., o_size...)
x = conv_nd.im2col_nd(x, k_size, self.stride, self.pad,
cover_all=self.cover_all, dilate=self.dilate)
x = xp.rollaxis(x, 0, dims + 2) # (iC, k_size..., N, o_size...)
mul_len = iCg * utils.size_of_shape(k_size)
x = x.reshape(G, mul_len, N * o_size_prod)
x = x.transpose(0, 2, 1) # (G, N*o_size, iCg*k_size)
gy = xp.rollaxis(gy, 1) # (oC, N, o_size...)
gy = gy.reshape(G, oCg, N * o_size_prod)
# (G, oCg, iCg*k_size) = (G, oCg, N*o_size) @ (G, N*o_size, iCg*k_size)
gW = convolution_2d._matmul(gy, x).astype(self.W_dtype, copy=False)
gW = gW.reshape(oC, iCg, *k_size)
return gW,
def _forward_grouped_convolution_xp(self, x, gy, xp):
G = self.groups
N, iC = x.shape[:2]
oC = gy.shape[1]
o_size = gy.shape[2:]
o_size_prod = utils.size_of_shape(o_size)
k_size = self.ksize
dims = len(o_size)
iCg = iC // G
oCg = oC // G
# Do not check iCg and oCg because this class is rarely used alone
# (N, iC, k_size..., o_size...)
x = conv_nd.im2col_nd(x,
k_size,
self.stride,
self.pad,
cover_all=self.cover_all,
dilate=self.dilate)
x = xp.rollaxis(x, 0, dims + 2) # (iC, k_size..., N, o_size...)
mul_len = iCg * utils.size_of_shape(k_size)
x = x.reshape(G, mul_len, N * o_size_prod)
x = x.transpose(0, 2, 1) # (G, N*o_size, iCg*k_size)
gy = xp.rollaxis(gy, 1) # (oC, N, o_size...)
gy = gy.reshape(G, oCg, N * o_size_prod)
# (G, oCg, iCg*k_size) = (G, oCg, N*o_size) @ (G, N*o_size, iCg*k_size)
gW = convolution_2d._matmul(gy, x).astype(self.W_dtype, copy=False)
gW = gW.reshape(oC, iCg, *k_size)
return gW,
def check_log_prob(self, is_gpu):
smp = self.sample_for_test()
if is_gpu:
log_prob1 = self.gpu_dist.log_prob(cuda.to_gpu(smp)).data
else:
log_prob1 = self.cpu_dist.log_prob(smp).data
if self.continuous:
scipy_prob = self.scipy_dist.logpdf
else:
scipy_prob = self.scipy_dist.logpmf
if self.scipy_onebyone:
onebyone_smp = smp.reshape(*[
utils.size_of_shape(sh)
for sh in [self.sample_shape, self.shape, self.event_shape]])
onebyone_smp = numpy.swapaxes(onebyone_smp, 0, 1)
onebyone_smp = onebyone_smp.reshape((-1,) + self.sample_shape
+ self.event_shape)
log_prob2 = []
for one_params, one_smp in zip(
self.scipy_onebyone_params_iter(), onebyone_smp):
log_prob2.append(scipy_prob(one_smp, **one_params))
log_prob2 = numpy.vstack(log_prob2)
log_prob2 = log_prob2.reshape(
utils.size_of_shape(self.shape), -1).T
log_prob2 = log_prob2.reshape(self.sample_shape + self.shape)
else:
log_prob2 = scipy_prob(smp, **self.scipy_params)
array.assert_allclose(log_prob1, log_prob2)
def check_sample(self, is_gpu):
if is_gpu:
smp1 = self.gpu_dist.sample(
sample_shape=(100000,)+self.sample_shape).data
else:
smp1 = self.cpu_dist.sample(
sample_shape=(100000,)+self.sample_shape).data
if self.scipy_onebyone:
smp2 = []
for one_params in self.scipy_onebyone_params_iter():
smp2.append(self.scipy_dist.rvs(
size=(100000,)+self.sample_shape, **one_params))
smp2 = numpy.vstack(smp2)
smp2 = smp2.reshape((utils.size_of_shape(self.shape), 100000)
+ self.sample_shape
+ self.cpu_dist.event_shape)
smp2 = numpy.rollaxis(
smp2, 0, smp2.ndim-len(self.cpu_dist.event_shape))
smp2 = smp2.reshape((100000,) + self.sample_shape + self.shape
+ self.cpu_dist.event_shape)
else:
smp2 = self.scipy_dist.rvs(
size=(100000,) + self.sample_shape + self.shape,
**self.scipy_params)
array.assert_allclose(smp1.mean(axis=0), smp2.mean(axis=0),
atol=3e-2, rtol=3e-2)
array.assert_allclose(smp1.std(axis=0), smp2.std(axis=0),
atol=3e-2, rtol=3e-2)
def check_sample(self, is_gpu):
if is_gpu:
smp1 = self.gpu_dist.sample(sample_shape=(100000, ) +
self.sample_shape).data
else:
smp1 = self.cpu_dist.sample(sample_shape=(100000, ) +
self.sample_shape).data
if self.scipy_onebyone:
smp2 = []
for one_params in self.scipy_onebyone_params_iter():
smp2.append(
self.scipy_dist.rvs(size=(100000, ) + self.sample_shape,
**one_params))
smp2 = numpy.vstack(smp2)
smp2 = smp2.reshape((utils.size_of_shape(self.shape), 100000) +
self.sample_shape + self.cpu_dist.event_shape)
smp2 = numpy.rollaxis(smp2, 0,
smp2.ndim - len(self.cpu_dist.event_shape))
smp2 = smp2.reshape((100000, ) + self.sample_shape + self.shape +
self.cpu_dist.event_shape)
else:
smp2 = self.scipy_dist.rvs(size=(100000, ) + self.sample_shape +
self.shape,
**self.scipy_params)
array.assert_allclose(smp1.mean(axis=0),
smp2.mean(axis=0),
atol=3e-2,
rtol=3e-2)
array.assert_allclose(smp1.std(axis=0),
smp2.std(axis=0),
atol=3e-2,
rtol=3e-2)
def _as4darray(arr, mode):
assert mode.cudnn_dim_ok
if mode.is_for_conv2d:
assert arr.ndim == 4
return arr
else: # is_for_linear
return arr.reshape(utils.size_of_shape(arr.shape[0:-1]), -1, 1, 1)
def __call__(self, array):
if self.dtype is not None:
assert array.dtype == self.dtype
device = backend.get_device_from_array(array)
if not array.shape: # 0-dim case
array[...] = self.scale * (2 * numpy.random.randint(2) - 1)
elif not array.size:
raise ValueError('Array to be initialized must be non-empty.')
else:
# numpy.prod returns float value when the argument is empty.
out_dim = len(array)
in_dim = utils.size_of_shape(array.shape[1:])
if (in_dim > out_dim and self._checks[0]) or (in_dim < out_dim
and self._checks[1]):
raise ValueError('Cannot make orthogonal {}.'
'shape = {}, interpreted as '
'{}-dim input and {}-dim output.'.format(
self.mode, array.shape, in_dim, out_dim))
transpose = in_dim > out_dim
a = numpy.random.normal(size=(out_dim, in_dim))
if transpose:
a = a.T
# cupy.linalg.qr requires cusolver in CUDA 8+
q, r = numpy.linalg.qr(a)
q *= numpy.copysign(self.scale, numpy.diag(r))
if transpose:
q = q.T
array[...] = device.xp.asarray(q.reshape(array.shape))
def linear(x, W, b=None, n_batch_axes=1):
"""Linear function, or affine transformation.
It accepts two or three arguments: an input minibatch ``x``, a weight
matrix ``W``, and optionally a bias vector ``b``. It computes
.. math:: Y = xW^\\top + b.
Args:
x (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`): Input variable, which is a :math:`(s_1, s_2, \
..., s_n)`-shaped float array. Its first ``n_batch_axes``
dimensions are handled as *minibatch dimensions*. The
other dimensions are handled as concatenated one dimension whose
size must be :math:`(s_{\\rm n\\_batch\\_axes} * ... * s_n = N)`.
W (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`): Weight variable of shape :math:`(M, N)`,
where :math:`(N = s_{\\rm n\\_batch\\_axes} * ... * s_n)`.
b (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`): Bias variable (optional) of shape
:math:`(M,)`.
n_batch_axes (int): The number of batch axes. The default is 1. The
input variable is reshaped into
(:math:`{\\rm n\\_batch\\_axes} + 1`)-dimensional tensor.
This should be greater than 0.
Returns:
~chainer.Variable: Output variable. A float array with shape
of :math:`(s_1, ..., s_{\\rm n\\_batch\\_axes}, M)`.
.. seealso:: :class:`~chainer.links.Linear`
.. admonition:: Example
>>> x = np.random.uniform(0, 1, (3, 4)).astype(np.float32)
>>> W = np.random.uniform(0, 1, (5, 4)).astype(np.float32)
>>> b = np.random.uniform(0, 1, (5,)).astype(np.float32)
>>> y = F.linear(x, W, b)
>>> y.shape
(3, 5)
"""
if n_batch_axes <= 0:
raise ValueError('n_batch_axes should be greater than 0.')
if n_batch_axes > 1:
batch_shape = x.shape[:n_batch_axes]
batch_size = utils.size_of_shape(batch_shape)
x = x.reshape(batch_size, -1)
elif x.ndim > 2:
x = x.reshape(x.shape[0], -1)
if b is None:
args = x, W
else:
args = x, W, b
y, = LinearFunction().apply(args)
if n_batch_axes > 1:
y = y.reshape(batch_shape + (-1,))
return y
def linear(x, W, b=None, n_batch_axes=1):
"""Linear function, or affine transformation.
It accepts two or three arguments: an input minibatch ``x``, a weight
matrix ``W``, and optionally a bias vector ``b``. It computes
.. math:: Y = xW^\\top + b.
Args:
x (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`): Input variable, which is a :math:`(s_1, s_2, \
..., s_n)`-shaped float array. Its first ``n_batch_axes``
dimensions are handled as *minibatch dimensions*. The
other dimensions are handled as concatenated one dimension whose
size must be :math:`(s_{\\rm n\\_batch\\_axes} * ... * s_n = N)`.
W (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`): Weight variable of shape :math:`(M, N)`,
where :math:`(N = s_{\\rm n\\_batch\\_axes} * ... * s_n)`.
b (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`): Bias variable (optional) of shape
:math:`(M,)`.
n_batch_axes (int): The number of batch axes. The default is 1. The
input variable is reshaped into
(:math:`{\\rm n\\_batch\\_axes} + 1`)-dimensional tensor.
This should be greater than 0.
Returns:
~chainer.Variable: Output variable. A float array with shape
of :math:`(s_1, ..., s_{\\rm n\\_batch\\_axes}, M)`.
.. seealso:: :class:`~chainer.links.Linear`
.. admonition:: Example
>>> x = np.random.uniform(0, 1, (3, 4)).astype(np.float32)
>>> W = np.random.uniform(0, 1, (5, 4)).astype(np.float32)
>>> b = np.random.uniform(0, 1, (5,)).astype(np.float32)
>>> y = F.linear(x, W, b)
>>> y.shape
(3, 5)
"""
if n_batch_axes <= 0:
raise ValueError('n_batch_axes should be greater than 0.')
if n_batch_axes > 1:
batch_shape = x.shape[:n_batch_axes]
batch_size = utils.size_of_shape(batch_shape)
x = x.reshape(batch_size, -1)
elif x.ndim > 2:
x = x.reshape(x.shape[0], -1)
if b is None:
args = x, W
else:
args = x, W, b
y, = LinearFunction().apply(args)
if n_batch_axes > 1:
y = y.reshape(batch_shape + (-1, ))
return y
def _forward_grouped_convolution_xp(self, x, W, b, xp):
# G: group count
# N: batch size
# iC: input channels
# oC: output channels
G = self.groups
N, iC = x.shape[:2]
oC = W.shape[0]
k_size = W.shape[2:]
iCg = iC // G
oCg = oC // G
dims = len(k_size)
if iC % G != 0:
raise TypeError('The number of groups must be '
'a divisor of that of input channels')
if oC % G != 0:
raise TypeError('The number of groups must be '
'a divisor of that of output channels')
xp = backend.get_array_module(x)
# (N, iC, k_size..., o_size...)
x = conv_nd.im2col_nd(x,
k_size,
self.stride,
self.pad,
cover_all=self.cover_all,
dilate=self.dilate)
o_size = x.shape[-dims:]
x = xp.rollaxis(x, 0, dims + 2) # (iC, k_size..., N, o_size...)
mul_len = iCg * utils.size_of_shape(k_size)
x = x.reshape(G, mul_len, N * utils.size_of_shape(o_size))
W = W.reshape(G, oCg, mul_len)
# (G, oCg, N*o_size) = (G, oCg, iCg*k_size) @ (G, iCg*k_size, N*o_size)
y = convolution_2d._matmul(W, x).astype(x.dtype, copy=False)
y = y.reshape(oC, N, *o_size)
y = xp.rollaxis(y, 1) # (N, oC, o_size...)
if b is not None:
y += b.reshape(1, b.size, *((1, ) * dims))
return y,
def forward(self, x, n_batch_axes=1):
if self.W.array is None:
in_size = utils.size_of_shape(x.shape[1:])
self._initialize_params(in_size)
return thresholded_linear(x,
self.W,
self.b,
n_batch_axes=n_batch_axes,
threshold=self.threshold)
def forward(self, x, n_batch_axes=1):
if self.W.array is None:
in_size = utils.size_of_shape(x.shape[1:])
self._initialize_params(in_size)
return ada_loss_linear(x,
self.W,
self.b,
n_batch_axes=n_batch_axes,
ada_loss=self.ada_loss)
def get_fans(shape):
if not isinstance(shape, tuple):
raise ValueError('shape must be tuple')
if len(shape) < 2:
raise ValueError('shape must be of length >= 2: shape={}', shape)
receptive_field_size = utils.size_of_shape(shape[2:])
fan_in = shape[1] * receptive_field_size
fan_out = shape[0] * receptive_field_size
return fan_in, fan_out
def get_fans(shape):
if not isinstance(shape, tuple):
raise ValueError('shape must be tuple')
if len(shape) < 2:
raise ValueError('shape must be of length >= 2: shape={}', shape)
receptive_field_size = utils.size_of_shape(shape[2:])
fan_in = shape[1] * receptive_field_size
fan_out = shape[0] * receptive_field_size
return fan_in, fan_out
def _forward_grouped_convolution_xp(self, x, W, b, xp):
# G: group count
# N: batch size
# iC: input channels
# oC: output channels
G = self.groups
N, iC = x.shape[:2]
oC = W.shape[0]
k_size = W.shape[2:]
iCg = iC // G
oCg = oC // G
dims = len(k_size)
if iC % G != 0:
raise TypeError('The number of groups must be '
'a divisor of that of input channels')
if oC % G != 0:
raise TypeError('The number of groups must be '
'a divisor of that of output channels')
xp = backend.get_array_module(x)
# (N, iC, k_size..., o_size...)
x = conv_nd.im2col_nd(x, k_size, self.stride, self.pad,
cover_all=self.cover_all, dilate=self.dilate)
o_size = x.shape[-dims:]
x = xp.rollaxis(x, 0, dims + 2) # (iC, k_size..., N, o_size...)
mul_len = iCg * utils.size_of_shape(k_size)
x = x.reshape(G, mul_len, N * utils.size_of_shape(o_size))
W = W.reshape(G, oCg, mul_len)
# (G, oCg, N*o_size) = (G, oCg, iCg*k_size) @ (G, iCg*k_size, N*o_size)
y = convolution_2d._matmul(W, x).astype(x.dtype, copy=False)
y = y.reshape(oC, N, *o_size)
y = xp.rollaxis(y, 1) # (N, oC, o_size...)
if b is not None:
y += b.reshape(1, b.size, *((1,) * dims))
return y,
def get_fans(shape):
if not isinstance(shape, tuple):
raise ValueError(
'shape must be tuple. Actual type: {}'.format(type(shape)))
if len(shape) < 2:
raise ValueError(
'shape must be of length >= 2. Actual shape: {}'.format(shape))
receptive_field_size = utils.size_of_shape(shape[2:])
fan_in = shape[1] * receptive_field_size
fan_out = shape[0] * receptive_field_size
return fan_in, fan_out
def get_fans(shape):
if not isinstance(shape, tuple):
raise ValueError('shape must be tuple. Actual type: {}'.format(
type(shape)))
if len(shape) < 2:
raise ValueError(
'shape must be of length >= 2. Actual shape: {}'.format(shape))
receptive_field_size = utils.size_of_shape(shape[2:])
fan_in = shape[1] * receptive_field_size
fan_out = shape[0] * receptive_field_size
return fan_in, fan_out
def _forward_grouped_convolution_xp(self, x, W, b, xp):
# G: group count
# N: batch size
# xC: input channels
# yC: output channels
G = self.groups
N, xC = x.shape[:2]
x_size = x.shape[2:]
yCg = W.shape[1]
yC = yCg * G
xCg = xC // G
k_size = W.shape[2:]
dims = len(k_size)
if xC % G != 0:
raise TypeError('The number of groups must be '
'a divisor of that of input channels')
x = xp.rollaxis(x, 1) # (xC, N, x_size...)
x = x.reshape(G, xCg, N * utils.size_of_shape(x_size))
W = W.reshape(G, xCg, yCg * utils.size_of_shape(k_size))
W = W.transpose(0, 2, 1) # (G, yCg*k_size, xCg)
# (G, yCg*k_size, N*x_size) = (G, yCg*k_size, xCg) @ (G, xCg, N*x_size)
col = convolution_2d._matmul(W, x).astype(x.dtype, copy=False)
col = col.reshape((yC, ) + k_size + (N, ) + x_size)
col = xp.rollaxis(col, dims + 1) # (N, yC, k_size..., x_size...)
y = conv_nd.col2im_nd(col,
self.stride,
self.pad,
self.outs,
dilate=self.dilate)
if b is not None:
y += b.reshape(1, yC, *((1, ) * dims))
return y,
def forward(self, x):
"""Updates the internal state and returns the LSTM outputs.
Args:
x (~chainer.Variable): A new batch from the input sequence.
Returns:
~chainer.Variable: Outputs of updated LSTM units.
"""
if self.upward.W.array is None:
with cuda.get_device_from_id(self._device_id):
in_size = utils.size_of_shape(x.shape[1:])
self.upward._initialize_params(in_size)
self._initialize_params()
batch = x.shape[0]
lstm_in = self.upward(x)
h_rest = None
if self.h is not None:
h_size = self.h.shape[0]
if batch == 0:
h_rest = self.h
elif h_size < batch:
msg = ('The batch size of x must be equal to or less than'
'the size of the previous state h.')
raise TypeError(msg)
elif h_size > batch:
h_update, h_rest = split_axis.split_axis(self.h, [batch],
axis=0)
lstm_in += self.lateral(h_update)
else:
lstm_in += self.lateral(self.h)
if self.c is None:
xp = self.xp
with cuda.get_device_from_id(self._device_id):
self.c = variable.Variable(
xp.zeros((batch, self.state_size), dtype=x.dtype))
self.c, y = lstm.lstm(self.c, lstm_in)
if h_rest is None:
self.h = y
elif len(y.array) == 0:
self.h = h_rest
else:
self.h = concat.concat([y, h_rest], axis=0)
return y
def forward(self, x):
"""Updates the internal state and returns the LSTM outputs.
Args:
x (~chainer.Variable): A new batch from the input sequence.
Returns:
~chainer.Variable: Outputs of updated LSTM units.
"""
if self.upward.W.array is None:
with chainer.using_device(self.device):
in_size = utils.size_of_shape(x.shape[1:])
self.upward._initialize_params(in_size)
self._initialize_params()
batch = x.shape[0]
lstm_in = self.upward(x)
h_rest = None
if self.h is not None:
h_size = self.h.shape[0]
if batch == 0:
h_rest = self.h
elif h_size < batch:
msg = ('The batch size of x must be equal to or less than'
'the size of the previous state h.')
raise TypeError(msg)
elif h_size > batch:
h_update, h_rest = split_axis.split_axis(
self.h, [batch], axis=0)
lstm_in += self.lateral(h_update)
else:
lstm_in += self.lateral(self.h)
if self.c is None:
with chainer.using_device(self.device):
self.c = variable.Variable(
self.xp.zeros((batch, self.state_size), dtype=x.dtype))
self.c, y = lstm.lstm(self.c, lstm_in)
if h_rest is None:
self.h = y
elif len(y.array) == 0:
self.h = h_rest
else:
self.h = concat.concat([y, h_rest], axis=0)
return y
def forward(self, x, n_batch_axes=1):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
n_batch_axes (int): The number of batch axes. The default is 1. The
input variable is reshaped into
(:math:`{\\rm n\\_batch\\_axes} + 1`)-dimensional tensor.
This should be greater than 0.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.W.data is None:
in_size = utils.size_of_shape(x.shape[1:])
self._initialize_params(in_size)
return linear.linear(x, self.W, self.b, n_batch_axes=n_batch_axes)
def forward(self, x, n_batch_axes=1):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
n_batch_axes (int): The number of batch axes. The default is 1. The
input variable is reshaped into
(:math:`{\\rm n\\_batch\\_axes} + 1`)-dimensional tensor.
This should be greater than 0.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.W.array is None:
in_size = utils.size_of_shape(x.shape[1:])
self._initialize_params(in_size)
return linear.linear(x, self.W, self.b, n_batch_axes=n_batch_axes)
def forward(self, x):
"""Apply layer normalization to given input.
Args:
x (~chainer.Variable): Batch vectors.
Shape of this value must be `(batch_size, unit_size)`,
e.g., the output of :func:`~chainer.functions.linear`.
Returns:
~chainer.Variable: Output of the layer normalization.
"""
if self.gamma.array is None:
in_size = utils.size_of_shape(x.shape[1:])
self._initialize_params(in_size)
return layer_normalization.layer_normalization(
x, self.gamma, self.beta, self.eps)
def forward(self, x):
"""Apply layer normalization to given input.
Args:
x (~chainer.Variable): Batch vectors.
Shape of this value must be `(batch_size, unit_size)`,
e.g., the output of :func:`~chainer.functions.linear`.
Returns:
~chainer.Variable: Output of the layer normalization.
"""
if self.gamma.data is None:
in_size = utils.size_of_shape(x.shape[1:])
self._initialize_params(in_size)
return layer_normalization.layer_normalization(x, self.gamma,
self.beta, self.eps)
def forward(self, xs):
xp = backend.get_array_module(*xs)
if self.length is None:
length = max(len(x) for x in xs)
else:
length = self.length
shape = (len(xs), length) + xs[0].shape[1:]
y = xp.empty(shape, xs[0].dtype)
if length == 0:
return y, # y is an empty array
if xp is numpy or any(not x._c_contiguous for x in xs):
for i, x in enumerate(xs):
l = len(x)
if l == length:
y[i] = x
else:
y[i, 0:l] = x
y[i, l:] = self.padding
else:
# This code assumes that all arrays are c_contiguous
ptr_shape = (Ellipsis,) + (None,) * xs[0].ndim
ptrs = cuda.cupy.array(
[x.data for x in xs], numpy.uintp)[ptr_shape]
lengths = cuda.cupy.array(
[len(x) for x in xs], numpy.int32)[ptr_shape]
base = utils.size_of_shape(xs[0].shape[1:])
cuda.elementwise(
'P ptr, int32 length, T pad, int32 base, int32 max_length',
'T y',
'''
int d = i / base % max_length;
if (d < length) {
y = reinterpret_cast<const T*>(ptr)[i % (base * max_length)];
} else {
y = pad;
}
''',
'pad_sequence_fwd'
)(ptrs, lengths, self.padding, base, length, y)
return y,
def ada_loss_linear(x, W, b=None, n_batch_axes=1, ada_loss=None):
""" Simply replace the LinearFunction in linear to AdaLossLinear """
if n_batch_axes <= 0:
raise ValueError('n_batch_axes should be greater than 0.')
if n_batch_axes > 1:
batch_shape = x.shape[:n_batch_axes]
batch_size = utils.size_of_shape(batch_shape)
x = x.reshape(batch_size, -1)
elif x.ndim > 2:
x = x.reshape(x.shape[0], -1)
if b is None:
args = x, W
else:
args = x, W, b
y, = AdaLossLinearFunction(ada_loss=ada_loss).apply(args)
if n_batch_axes > 1:
y = y.reshape(batch_shape + (-1, ))
return y
def thresholded_linear(x, W, b=None, n_batch_axes=1, threshold=6e-8):
""" """
if n_batch_axes <= 0:
raise ValueError('n_batch_axes should be greater than 0.')
if n_batch_axes > 1:
batch_shape = x.shape[:n_batch_axes]
batch_size = utils.size_of_shape(batch_shape)
x = x.reshape(batch_size, -1)
elif x.ndim > 2:
x = x.reshape(x.shape[0], -1)
if b is None:
args = x, W
else:
args = x, W, b
y, = ThresholdedLinearFunction(threshold=threshold).apply(args)
if n_batch_axes > 1:
y = y.reshape(batch_shape + (-1, ))
return y
def __call__(self, array):
if self.dtype is not None:
assert array.dtype == self.dtype
xp = backend.get_array_module(array)
if not array.shape: # 0-dim case
array[...] = self.scale * (2 * numpy.random.randint(2) - 1)
elif not array.size:
raise ValueError('Array to be initialized must be non-empty.')
else:
# numpy.prod returns float value when the argument is empty.
flat_shape = (len(array), utils.size_of_shape(array.shape[1:]))
if flat_shape[0] > flat_shape[1]:
raise ValueError('Cannot make orthogonal system because'
' # of vectors ({}) is larger than'
' that of dimensions ({})'.format(
flat_shape[0], flat_shape[1]))
a = numpy.random.normal(size=flat_shape)
# cupy.linalg.qr requires cusolver in CUDA 8+
q, r = numpy.linalg.qr(a.T)
q *= numpy.copysign(self.scale, numpy.diag(r))
array[...] = xp.asarray(q.T.reshape(array.shape))
def __call__(self, array):
if self.dtype is not None:
assert array.dtype == self.dtype,\
'{} != {}'.format(array.dtype, self.dtype)
if not array.shape: # 0-dim case
if self.rng is None:
a = numpy.random.randint(2)
else:
a = self.rng.randint(2)
a = int(a)
array[...] = self.scale * (2 * a - 1)
elif not array.size:
raise ValueError('Array to be initialized must be non-empty.')
else:
# numpy.prod returns float value when the argument is empty.
out_dim = len(array)
in_dim = utils.size_of_shape(array.shape[1:])
if (in_dim > out_dim and self._checks[0]) or (
in_dim < out_dim and self._checks[1]):
raise ValueError(
'Cannot make orthogonal {}. '
'shape = {}, interpreted as '
'{}-dim input and {}-dim output.'.format(
self.mode, array.shape, in_dim, out_dim))
transpose = in_dim > out_dim
if self.rng is None:
a = numpy.random.normal(size=(out_dim, in_dim))
else:
a_tmp = self.rng.normal(size=(out_dim, in_dim))
a = numpy.empty(a_tmp.shape, dtype=a_tmp.dtype)
backend.copyto(a, a_tmp)
if transpose:
a = a.T
# cupy.linalg.qr requires cusolver in CUDA 8+
q, r = numpy.linalg.qr(a)
q *= numpy.copysign(self.scale, numpy.diag(r))
if transpose:
q = q.T
backend.copyto(array, q.reshape(array.shape).astype(
array.dtype, copy=False))
def forward(self, inputs):
gy, = inputs
xp = backend.get_array_module(gy)
repeats = self.repeats
axis = self.axis
shape = list(self.in_shape)
dtype = self.in_dtype
if len(gy) == 0:
gx = xp.zeros(shape, dtype)
return gx,
if len(repeats) == 1:
repeats = int(repeats[0])
if axis is None:
gx = gy.reshape(-1, repeats).sum(axis=1).reshape(shape)
else:
shape[axis:axis + 1] = [-1, repeats]
gx = gy.reshape(shape).sum(axis=axis + 1)
return gx,
if axis is None:
pos = 0
gx = xp.zeros(utils.size_of_shape(shape), dtype)
for (i, r) in enumerate(repeats):
gx[i] = xp.sum(gy[pos:pos + r])
pos += r
gx = gx.reshape(shape)
else:
gx = xp.zeros(shape, dtype)
pos = 0
src = [slice(None)] * axis + [None]
dst = [slice(None)] * axis + [None]
for (i, r) in enumerate(repeats):
src[-1] = slice(pos, pos + r)
dst[-1] = slice(i, i + 1)
gx[tuple(dst)] = gy[tuple(src)].sum(axis=axis, keepdims=True)
pos += r
return gx,
def __call__(self, array):
if self.dtype is not None:
assert array.dtype == self.dtype
xp = backend.get_array_module(array)
if not array.shape: # 0-dim case
array[...] = self.scale
elif not array.size:
raise ValueError('Array to be initialized must be non-empty.')
else:
# numpy.prod returns float value when the argument is empty.
flat_shape = (len(array), utils.size_of_shape(array.shape[1:]))
if flat_shape[0] > flat_shape[1]:
raise ValueError('Cannot make orthogonal system because'
' # of vectors ({}) is larger than'
' that of dimensions ({})'.format(
flat_shape[0], flat_shape[1]))
a = numpy.random.normal(size=flat_shape)
# we do not have cupy.linalg.svd for now
u, _, v = numpy.linalg.svd(a, full_matrices=False)
# pick the one with the correct shape
q = u if u.shape == flat_shape else v
array[...] = xp.asarray(q.reshape(array.shape))
array *= self.scale
def _get_tensor4d_shape(axis, shape):
left_shape = utils.size_of_shape(shape[:axis])
center_shape = shape[axis]
right_shape = utils.size_of_shape(shape[axis:][1:])
return left_shape, center_shape, right_shape, 1