def _typecheck_input(x1_type, x2_type):
"""
Check input tensor types to be valid and confirm they are the same type.
"""
const_utils.check_type_valid(x1_type, [mstype.float32, mstype.float16], 'x1')
const_utils.check_type_valid(x2_type, [mstype.float32, mstype.float16], 'x2')
if x1_type != x2_type:
raise TypeError(f'Both Inputs must be the same Type. x1 is \'{x1_type}\' and x2 is \'{x2_type}\' ')
def sequence_mask(lengths, maxlen=None):
"""
Returns a mask tensor representing the first N positions of each cell.
If lengths has shape [d_1, d_2, ..., d_n], then the resulting tensor mask has type dtype and shape
[d_1, d_2, ..., d_n, maxlen], with mask[i_1, i_2, ..., i_n, j] = (j < lengths[i_1, i_2, ..., i_n])
Inputs:
- **lengths** (Tensor) - Tensor to calculate the mask for. All values in this tensor should be
less than or equal to `maxlen`. Values greater than `maxlen` will be treated as `maxlen`.
Must be type int32 or int64.
- **maxlen** (int) - size of the last dimension of returned tensor. Must be positive and same
type as elements in `lengths`.
Outputs:
One mask tensor of shape lengths.shape + (maxlen,).
Supported Platforms:
``GPU``
Examples:
>>> x = Tensor(np.array([[1, 3], [2, 0]]))
>>> output = C.sequence_mask(x, 3)
>>> print(output)
[[[True, False, False],
[True, True, True]],
[[True, True, False],
[False, False, False]]]
"""
argmax_op = P.ArgMaxWithValue()
reshape_op = P.Reshape()
range_op = P.Range()
expand_op = P.ExpandDims()
cast_op = P.Cast()
shape_op = P.Shape()
to_tensor_op = P.ScalarToArray()
const_utils.check_type_valid(F.dtype(lengths),
[mstype.int64, mstype.int32], 'lengths')
_check_sequence_mask_input_len(shape_op(lengths))
if maxlen is None:
flatten_data = reshape_op(lengths, (-1, ))
flatten_data = cast_op(flatten_data, mstype.float32)
_, value = argmax_op(flatten_data)
maxlen = cast_op(value, mstype.int32)
else:
maxlen = _check_positive_int(maxlen, "maxlen", "sequence_mask")
maxlen = to_tensor_op(maxlen)
range_vector = range_op(to_tensor_op(0), maxlen, to_tensor_op(1))
mask = expand_op(lengths, -1)
result = range_vector < mask
return result
def repeat_elements(x, rep, axis=0):
"""
Repeat elements of a tensor along an axis, like np.repeat.
Args:
x (Tensor): The tensor to repeat values for. Must be of type: float16,
float32, int8, uint8, int16, int32, or int64.
rep (int): The number of times to repeat, must be positive, required.
axis (int): The axis along which to repeat, default 0.
Outputs:
One tensor with values repeated along the specified axis. If x has shape
(s1, s2, ..., sn) and axis is i, the output will have shape (s1, s2, ...,
si * rep, ..., sn). The output type will be the same as the type of `x`.
Supported Platforms:
``Ascend`` ``GPU`` ``CPU``
Examples:
>>> x = Tensor(np.array([[0, 1, 2], [3, 4, 5]]), mindspore.int32)
>>> output = C.repeat_elements(x, rep = 2, axis = 0)
>>> print(output)
[[0 1 2]
[0 1 2]
[3 4 5]
[3 4 5]]
"""
const_utils.check_type_valid(F.dtype(x), mstype.number_type, 'input x')
rep = _check_positive_int(rep, "rep", "repeat_elements")
axis = _check_is_int(axis, "axis", "repeat_elements")
shape_op = P.Shape()
rank_op = P.Rank()
tile_op = P.Tile()
expand_dims_op = P.ExpandDims()
reshape_op = P.Reshape()
x_rank = rank_op(x)
axis = _check_axis_range(axis, x_rank, "axis", "repeat_elements")
expand_axis = axis + 1
x_expand = expand_dims_op(x, expand_axis)
rep_dims = _cal_repeat_dims(x_rank, rep, expand_axis)
x_expand = tile_op(x_expand, rep_dims)
x_shape = shape_op(x)
x_reshape = _cal_reshape(x_shape, rep, axis)
x_rep = reshape_op(x_expand, x_reshape)
return x_rep
def count_nonzero(x, axis=(), keep_dims=False, dtype=mstype.int32):
r"""
Count number of nonzero elements across axis of input tensor
Args:
x (Tensor): Input data is used to count non-zero numbers.
axis (Union[int, tuple(int), list(int)]): The dimensions to reduce. Only constant value is allowed.
Default: (), reduce all dimensions.
keep_dims (bool): If true, keep these reduced dimensions and the length is 1.
If false, don't keep these dimensions. Default: False.
dtype (Union[Number, mstype.bool\_]): The data type of the output tensor. Only constant value is allowed.
Default: mstype.int32
Returns:
Tensor, number of nonzero element. The data type is dtype.
Supported Platforms:
``Ascend`` ``GPU`` ``CPU``
Examples:
>>> input_x = Tensor(np.array([[0, 1, 0], [1, 1, 0]]).astype(np.float32))
>>> nonzero_num = count_nonzero(x=input_x, axis=[0, 1], keep_dims=True, dtype=mstype.int32)
>>> print(nonzero_num)
[[3]]
"""
const_utils.check_type_valid(F.dtype(x), mstype.number_type, 'input x')
axis = _check_validate_axis(axis, "count_nonzero")
keep_dims = _check_validate_keepdims(keep_dims, "count_nonzero")
const_utils.check_type_valid(dtype, mstype.number_type + (mstype.bool_, ),
'dtype')
not_equal = P.NotEqual()
cast = P.Cast()
reduce_sum = P.ReduceSum(keep_dims)
nonzero_bool = not_equal(x, 0)
# ReduceSum only support float16 or float32 tensor.
nonzero_val = cast(nonzero_bool, mstype.float32)
nonzero_num = cast(reduce_sum(nonzero_val, axis), dtype)
return nonzero_num
def construct(self, weights):
const_utils.check_type_valid(F.dtype(weights), mstype.number_type, 'weights')
l1_regularization = self.scale * self.reduce_sum(self.abs(weights))
return l1_regularization
