def polymul(a1, a2):
"""Computes the product of two polynomials.
Args:
a1 (scalar, cupy.ndarray or cupy.poly1d): first input polynomial.
a2 (scalar, cupy.ndarray or cupy.poly1d): second input polynomial.
Returns:
cupy.ndarray or cupy.poly1d: The product of the inputs.
.. seealso:: :func:`numpy.polymul`
"""
a1 = cupy.trim_zeros(a1, trim='f')
a2 = cupy.trim_zeros(a2, trim='f')
if a1.size == 0:
a1 = cupy.array([0.], a1.dtype)
if a2.size == 0:
a2 = cupy.array([0.], a2.dtype)
return cupy.convolve(a1, a2)
def fix_padding(tnsr):
"""Function to fix padding on a torch.Tensor object
Args:
tnsr ([torch.Tensor]): Tensor representing input_ids,
attention_mask
Returns:
[torch.Tensor]: trimmed stack of Tensor objects
"""
# Remove all the padding from the end
trimmed_collections = list()
max_arr_length = -1
dx = to_dlpack(tnsr)
embeddings_collecton = cp.fromDlpack(dx)
for embeddings in embeddings_collecton:
trimmed = cp.trim_zeros(embeddings, trim="b")
max_arr_length = max(max_arr_length, len(trimmed))
trimmed_collections.append(trimmed)
first_arr_stack = cp.pad(
trimmed_collections[0],
(0, max_arr_length - len(trimmed_collections[0])),
"constant",
)
# Add the required padding back
for a in range(1, len(trimmed_collections)):
padded = cp.pad(
trimmed_collections[a],
(0, max_arr_length - len(trimmed_collections[a])),
"constant",
)
first_arr_stack = cp.vstack([first_arr_stack, padded])
# Convert it back to a PyTorch tensor.
tx2 = from_dlpack(first_arr_stack.toDlpack())
# Taking care of case where we have only one sentence
# Then, we need to reshape to get the right dimensions
# since in the other cases cp.vstack handles that.
if len(tx2.shape) == 1:
dim = tx2.shape[0]
tx2 = torch.reshape(tx2, (1, dim))
return tx2
def trimseq(seq):
"""Removes small polynomial series coefficients.
Args:
seq (cupy.ndarray): input array.
Returns:
cupy.ndarray: input array with trailing zeros removed. If the
resulting output is empty, it returns the first element.
.. seealso:: :func:`numpy.polynomial.polyutils.trimseq`
"""
if seq.size == 0:
return seq
ret = cupy.trim_zeros(seq, trim='b')
if ret.size > 0:
return ret
return seq[:1]
