def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 4)
c1_type, c2_type, x1_type, x2_type = in_types
type_check.expect(
c1_type.dtype.kind == 'f',
c2_type.dtype == c1_type.dtype,
x1_type.dtype == c1_type.dtype,
x2_type.dtype == c1_type.dtype,
c1_type.ndim >= 2,
c2_type.ndim >= 2,
x1_type.ndim >= 2,
x2_type.ndim >= 2,
c1_type.ndim == x1_type.ndim,
c1_type.ndim == c2_type.ndim,
c1_type.ndim == x2_type.ndim,
c1_type.shape[0] == x1_type.shape[0],
c1_type.shape[0] == c2_type.shape[0],
c1_type.shape[0] == x2_type.shape[0],
x1_type.shape[1] == 4 * c1_type.shape[1],
x2_type.shape[1] == 4 * c2_type.shape[1],
)
for i in range(2, c1_type.ndim.eval()):
type_check.expect(x1_type.shape[i] == c1_type.shape[i])
type_check.expect(x2_type.shape[i] == c2_type.shape[i])
type_check.expect(x1_type.shape[i] == x2_type.shape[i])
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 2)
a_type, b_type = in_types
type_check.expect(
a_type.dtype.kind == 'f',
b_type.dtype.kind == 'f',
a_type.ndim >= 1,
b_type.ndim >= 1,
)
a_ndim = type_check.eval(a_type.ndim)
b_ndim = type_check.eval(b_type.ndim)
if b_ndim == 1:
a_idx = -2 if self.transa and a_ndim > 1 else -1
type_check.expect(a_type.shape[a_idx] == b_type.shape[0])
elif a_ndim == 1:
b_idx = -1 if self.transb and b_ndim > 1 else -2
type_check.expect(a_type.shape[0] == b_type.shape[b_idx])
else:
a_idx = _get_check_index(self.transa, False,
row_idx=-2, col_idx=-1)
b_idx = _get_check_index(self.transb, True,
row_idx=-2, col_idx=-1)
type_check.expect(a_type.shape[a_idx] == b_type.shape[b_idx])
type_check.expect_broadcast_shapes(
a_type.shape[:-2], b_type.shape[:-2])
def check_type_backward(self, in_types, out_types):
type_check.expect(
out_types.size() == 1,
in_types[0].dtype == out_types[0].dtype,
in_types[0].ndim == out_types[0].ndim,
in_types[0].shape == out_types[0].shape
)
def check_type_forward(self, in_types):
n_in = in_types.size()
type_check.expect(2 <= n_in, n_in <= 3)
x_type, w_type = in_types[:2]
type_check.expect(
x_type.dtype.kind == 'f',
w_type.dtype.kind == 'f',
x_type.ndim == 4,
w_type.ndim == 4,
x_type.shape[1] == w_type.shape[0]
)
if self.outh is not None:
type_check.expect(
x_type.shape[2] ==
conv.get_conv_outsize(self.outh, w_type.shape[2],
self.sy, self.ph),
)
if self.outw is not None:
type_check.expect(
x_type.shape[3] ==
conv.get_conv_outsize(self.outw, w_type.shape[3],
self.sx, self.pw),
)
if type_check.eval(n_in) == 3:
b_type = in_types[2]
type_check.expect(
b_type.dtype == x_type.dtype,
b_type.ndim == 1,
b_type.shape[0] == w_type.shape[1]
)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 1)
x_type, = in_types
type_check.expect(
x_type.dtype == numpy.float32,
)
def check_type_forward(self, in_types):
type_check.expect(
in_types.size() == 2,
in_types[0].dtype.kind == 'f',
in_types[0].dtype == in_types[1].dtype,
in_types[0].shape == in_types[1].shape
)
def check_type_forward(self, in_types):
n_in = in_types.size()
type_check.expect(2 <= n_in, n_in <= 3)
x_type, w_type = in_types[:2]
type_check._argname((x_type, w_type), ('x', 'W'))
type_check.expect(
x_type.dtype.kind == 'f',
w_type.dtype.kind == 'f',
x_type.ndim >= 2,
w_type.ndim == 2,
type_check.prod(x_type.shape[1:]) == w_type.shape[1],
)
if type_check.eval(n_in) == 3:
b_type = in_types[2]
type_check._argname((b_type,), ('b',))
type_check.expect(
b_type.dtype == x_type.dtype,
b_type.ndim == 1,
b_type.shape[0] == w_type.shape[0],
)
if self.mask is not None:
if self.use_batchwise_mask:
type_check.expect(
self.mask.shape[0] == x_type.shape[0],
self.mask.shape[1:] == w_type.shape,
)
else:
type_check.expect(self.mask.shape == w_type.shape)
def check_type_backward(self, in_types, out_types):
type_check.expect(
in_types.size() == 2,
out_types.size() == 1,
)
y_type, = out_types
type_check.expect(y_type.ndim == 0) # means scalar
def check_type_forward(self, in_types):
# Depending on the arguments, pad_width and keywords, the input value
# may be inappropriate. In that case, numpy.pad or cupy.pad will raise
# errors, so that only check the size and the dtype in this function.
type_check.expect(in_types.size() == 1)
x_type = in_types[0]
type_check.expect(x_type.dtype.kind == 'f')
def check_type_forward(self, in_types):
type_check._argname(in_types, ('gy',))
type_check.expect(
in_types[0].dtype.kind == 'f',
in_types[0].dtype == self.x.dtype
)
def check_type_forward(self, in_types):
type_check._argname(in_types, ('x', 'divisor'))
type_check.expect(
in_types[0].dtype == in_types[1].dtype,
in_types[0].dtype.kind == 'f',
in_types[1].dtype.kind == 'f',
)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 2)
type_check.expect(
in_types[0].dtype == in_types[1].dtype,
)
type_check.expect_broadcast_shapes(
in_types[0].shape, in_types[1].shape)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 1)
x_type = in_types[0]
type_check.expect(
x_type.dtype.kind == 'f',
x_type.ndim >= 2)
def check_type_forward(self, in_types):
type_check._argname(in_types, ('x0', 'x1'))
type_check.expect(
in_types[0].dtype.kind == 'f',
in_types[0].dtype == in_types[1].dtype,
in_types[0].shape == in_types[1].shape
)
def check_type_forward(self, in_types):
type_check._argname(in_types, ('x',))
type_check.expect(
in_types[0].dtype.kind == 'f',
)
if self.source is not None:
for axis in self.source:
if axis >= 0:
type_check.expect(
axis < in_types[0].ndim,
)
else:
type_check.expect(
-axis - 1 < in_types[0].ndim,
)
if self.destination is not None:
for axis in self.destination:
if axis >= 0:
type_check.expect(
axis < in_types[0].ndim,
)
else:
type_check.expect(
-axis - 1 < in_types[0].ndim,
)
def check_type_forward(self, in_types):
type_check._argname(in_types, ('sp', 'dn'))
sp_type, dn_type = in_types
# sp_type.shape: ((nb,) ldnz)
# dn_type.shape: ((nb,) _k, _n) when transb is False
sp_k_axis = -1
if self.transa:
sp_k_axis = -2
dn_k_axis = -2
if self.transb:
dn_k_axis = -1
type_check.expect(
sp_type.dtype.kind == 'f',
dn_type.dtype.kind == 'f',
dn_type.ndim >= 2,
dn_type.ndim <= 3,
sp_type.ndim == dn_type.ndim - 1,
sp_type.shape[-1] == self.sp_row.shape[-1],
self.sp_shape[sp_k_axis] == dn_type.shape[dn_k_axis],
)
dn_ndim = type_check.eval(dn_type.ndim)
if dn_ndim == 3:
type_check.expect(
sp_type.shape[0] == self.sp_row.shape[0],
dn_type.shape[0] == self.sp_row.shape[0],
)
def check_type_forward(self, in_types):
type_check._argname(in_types, ('lhs', 'rhs', 'gy'))
type_check.expect(
in_types[0].dtype.kind == 'f',
in_types[0].dtype == in_types[1].dtype,
in_types[0].dtype == in_types[2].dtype,
)
def check_type_forward(self, in_types):
type_check._argname(in_types, ('lhs', 'rhs'))
type_check.expect(
in_types[0].dtype == in_types[1].dtype,
)
type_check.expect_broadcast_shapes(
in_types[0].shape, in_types[1].shape)
def check_type_forward(self, in_types):
n_in = in_types.size().eval()
if n_in != 3 and n_in != 5:
raise type_check.InvalidType(
'%s or %s' % (in_types.size() == 3, in_types.size() == 5),
'%s == %s' % (in_types.size(), n_in))
x_type, gamma_type, beta_type = in_types[:3]
type_check.expect(
x_type.dtype.kind == 'f',
x_type.ndim >= gamma_type.ndim + 1,
# TODO(beam2d): Check shape
gamma_type.dtype == x_type.dtype,
beta_type.dtype == x_type.dtype,
gamma_type.shape == beta_type.shape,
)
if len(in_types) == 5:
mean_type, var_type = in_types[3:]
type_check.expect(
mean_type.dtype == x_type.dtype,
mean_type.shape == gamma_type.shape,
var_type.dtype == x_type.dtype,
var_type.shape == gamma_type.shape,
)
def check_type_forward(self, in_types):
type_check.expect(
in_types.size() == 1,
in_types[0].dtype.kind == 'f',
in_types[0].ndim == 2 + self.ndim,
in_types[0].size > 0,
)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 2)
type_check.expect(
in_types[0].dtype == numpy.float32,
in_types[1].dtype == numpy.float32,
in_types[0].shape == in_types[1].shape
)
def check_type_forward(self, in_types):
type_check.expect(
in_types.size() == 1,
in_types[0].dtype.kind == 'f',
in_types[0].ndim > self.axis,
in_types[0].ndim >= -self.axis
)
def check_type_forward(self, in_types): n_in = in_types.size() type_check.expect(3 <= n_in, n_in <= 4) x_type = in_types[0] v_type = in_types[1] g_type = in_types[2] type_check.expect( x_type.dtype.kind == "f", v_type.dtype.kind == "f", g_type.dtype.kind == "f", x_type.ndim == 4, v_type.ndim == 4, g_type.ndim == 4, x_type.shape[1] == v_type.shape[0] ) if self.outh is not None: type_check.expect( x_type.shape[2] == conv.get_conv_outsize(self.outh, v_type.shape[2],self.sy, self.ph), ) if self.outw is not None: type_check.expect( x_type.shape[3] == conv.get_conv_outsize(self.outw, v_type.shape[3], self.sx, self.pw), ) if type_check.eval(n_in) == 4: b_type = in_types[3] type_check.expect( b_type.dtype == x_type.dtype, b_type.ndim == 1, b_type.shape[0] == v_type.shape[1] )
def check_type_forward(self, in_types):
type_check.expect(
in_types.size() == 1,
in_types[0].dtype == numpy.float32,
in_types[0].ndim > self.axis,
in_types[0].ndim >= -self.axis
)
def check_type_forward(self, in_types):
type_check._argname(in_types, ('x',))
x_type, = in_types
if self.axis >= 0:
type_check.expect(x_type.ndim >= self.axis)
else:
type_check.expect(x_type.ndim >= -self.axis - 1)
def check_type_forward(self, in_types):
type_check._argname(in_types, ('x',))
x_type = in_types[0]
if self.axis >= 0:
type_check.expect(self.axis < x_type.ndim)
else:
type_check.expect(-self.axis <= x_type.ndim)
def check_type_forward(self, in_types):
type_check._argname(in_types, ('x',))
type_check.expect(
in_types[0].dtype.kind == 'f',
in_types[0].ndim > self.axis,
in_types[0].ndim >= -self.axis
)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 2)
type_check.expect(
in_types[0].dtype == in_types[1].dtype,
# check if the mini_batch sizes are same
in_types[0].shape[0] == in_types[1].shape[0]
)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 5)
x_type, gamma_type, beta_type, mean_type, var_type = in_types
type_check.expect(
x_type.dtype.kind == 'f',
# TODO(beam2d): Check shape
gamma_type.dtype == x_type.dtype,
beta_type.dtype == x_type.dtype,
gamma_type.shape == beta_type.shape,
mean_type.dtype == x_type.dtype,
mean_type.shape == gamma_type.shape,
var_type.dtype == x_type.dtype,
var_type.shape == gamma_type.shape,
)
_x_ndim = type_check.eval(x_type.ndim)
_gamma_ndim = type_check.eval(gamma_type.ndim)
_axis = _compute_axis(_x_ndim, _gamma_ndim, self.axis)
type_check.expect(
x_type.ndim >= len(_axis),
)
_key_axis = _compute_key_axis(_x_ndim, _gamma_ndim, _axis)
type_check.expect(
gamma_type.ndim == len(_key_axis),
)
for i in range(len(_key_axis)):
type_check.expect(
x_type.shape[_key_axis[i]] == gamma_type.shape[i],
)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 1)
x_type, = in_types
type_check.expect(
x_type.dtype == numpy.int32,
x_type.ndim == 1,
)
def check_type_forward(self, in_types):
type_check.argname(in_types, ('lhs', 'rhs'))
type_check.expect(in_types[0].dtype == in_types[1].dtype)
type_check.expect_broadcast_shapes(in_types[0].shape,
in_types[1].shape)
def check_type_forward(self, in_types):
type_check._argname(in_types, ('x', 't'))
x_type, t_type = in_types
type_check.expect(x_type.dtype.kind == 'f', t_type.dtype.kind == 'i',
x_type.shape == t_type.shape)
def check_type_forward(self, in_types):
n_in = in_types.size()
type_check.expect(n_in == 1)
x_type = in_types[0]
type_check.expect(x_type.dtype.kind == 'f', x_type.ndim == 4)
def check_type_forward(self, in_types):
type_check.expect(
in_types.size() == 1,
_type_check_prod(in_types[0].shape) ==
_type_check_prod(self.shape)
)
def _check_ndim(in_type, lower=1, upper=2):
type_check.expect(in_type.ndim >= lower, in_type.ndim <= upper)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() > 0)
shapes = [t.shape for t in in_types]
type_check.expect_broadcast_shapes(*shapes)
def check_type_backward(self, in_types, out_types):
type_check.expect(
out_types.size() == 1,
_type_check_prod(in_types[0].shape) ==
_type_check_prod(out_types[0].shape)
)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 1)
x_type, = in_types
type_check.expect(x_type.dtype == numpy.float32, )
def check_type_forward(self, in_types):
type_check._argname(in_types, ('gy', ))
gy_type = in_types[0]
type_check.expect(gy_type.dtype.char == 'f', gy_type.ndim == 4)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 2)
type_check.expect(in_types[0].dtype.kind == 'f',
in_types[0].dtype == in_types[1].dtype,
in_types[0].shape == in_types[1].shape)
def check_type_forward(self, in_types):
type_check._argname(in_types, ('x0', 'x1'))
type_check.expect(in_types[0].dtype.kind == 'f',
in_types[0].dtype == in_types[1].dtype,
in_types[0].shape == in_types[1].shape)
def check_type_forward(self, xs_type):
for p, n in zip(xs_type, xs_type[1:]):
type_check.expect(
p.shape[0] >= n.shape[0],
p.shape[1:] == n.shape[1:],
)
def check_type_forward(self, in_types):
x_type, = in_types
type_check.expect(
x_type.dtype == numpy.float32,
x_type.ndim >= 2,
)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 1)
x_type, = in_types
type_check.expect(x_type.dtype.kind == 'f')
def check_type_forward(self, in_types):
for i, in_type in enumerate(in_types):
type_check.argname((in_type, ), ('x{}'.format(i), ))
type_check.expect(in_types[0].dtype == in_type.dtype)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 2)
type_check.expect(in_types[0].dtype == numpy.float32,
in_types[1].dtype == numpy.float32,
in_types[0].shape == in_types[1].shape)
def check_type_forward(self, in_types):
type_check.argname(in_types, ('x', ))
type_check.expect(in_types.size() == 1)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 1,
in_types[0].dtype == numpy.float32,
in_types[0].ndim == 4)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() > self._n_cell + self._n_params * 2)
if self.use_cell:
(h_type, c_type), in_types = _split(in_types, 2)
h_size = self.n_layers * self.rnn_direction
type_check.expect(
h_type.dtype == numpy.float32,
c_type.dtype == numpy.float32,
h_type.ndim == 3,
h_type.shape[0] == h_size,
c_type.ndim == 3,
c_type.shape[0] == h_size,
# mini-batch size
h_type.shape[1] == c_type.shape[1],
# hidden size
h_type.shape[2] == c_type.shape[2],
)
else:
(h_type, ), in_types = _split(in_types, 1)
h_size = self.n_layers * self.rnn_direction
type_check.expect(
h_type.dtype == numpy.float32,
h_type.ndim == 3,
h_type.shape[0] == h_size,
)
w_types, in_types = _split(in_types, self._n_params)
b_types, in_types = _split(in_types, self._n_params)
x_types = in_types
for x_type in x_types:
type_check.expect(
x_type.dtype == numpy.float32,
x_type.ndim == 2,
)
for x1_type, x2_type in six.moves.zip(x_types, x_types[1:]):
type_check.expect(
# Check if xs are sorted by descending lengths
x1_type.shape[0] >= x2_type.shape[0],
x1_type.shape[1] == x2_type.shape[1])
in_size = x_types[0].shape[1]
out_size = h_type.shape[2]
for layer in six.moves.range(self.n_layers):
for i in six.moves.range(self.n_W):
for di in six.moves.range(self.rnn_direction):
ind = (layer * self.rnn_direction + di) * self.n_W + i
w_type = w_types[ind]
b_type = b_types[ind]
if self.rnn_direction == 1:
# Uni-direction
if layer == 0 and i < (self.n_W // 2):
w_in = in_size
else:
w_in = out_size
else:
# Bi-direction
if layer == 0 and i < (self.n_W // 2):
w_in = in_size
elif layer > 0 and i < (self.n_W // 2):
w_in = out_size * self.rnn_direction
else:
w_in = out_size
type_check.expect(
w_type.dtype == numpy.float32,
w_type.ndim == 2,
w_type.shape[0] == out_size,
w_type.shape[1] == w_in,
b_type.dtype == numpy.float32,
b_type.ndim == 1,
b_type.shape[0] == out_size,
)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 2)
n_nones = len([item for item in self.slices if item is None])
valid_slice = len(self.slices) - n_nones
type_check.expect(in_types[0].ndim >= valid_slice)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 1,
in_types[0].dtype == numpy.float32)
if self.axis is not None:
type_check.expect(self.axis < in_types[0].ndim, )
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 1,)
def check_type_forward(self, in_types):
type_check._argname(in_types, ('x', ))
type_check.expect(in_types[0].dtype.kind == 'f',
in_types[0].ndim > self.axis,
in_types[0].ndim >= -self.axis)
def _check_type_dataset(self, in_types):
img_type = in_types[0]
type_check.expect(
img_type.dtype.kind == 'f',
)
def check_type_forward(self, in_types):
w_type, = in_types
type_check.expect(w_type.ndim == 5)
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 3)
bs_type, M_type, logS2_type = in_types
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 1)
type_check.expect(in_types[0].dtype.kind == 'f')
def check_type_forward(self, in_types):
type_check.expect(in_types[0].dtype.kind == 'f', )
def check_type_forward(self, in_types):
type_check._argname(in_types, ('x', ))
type_check.expect(in_types[0].dtype.kind == 'f')
def check_type_backward(self, in_types, out_types):
type_check.expect(out_types.size() == 1,
out_types[0].dtype == numpy.float32,
out_types[0].ndim == 0)
