Source code for tripy.frontend.ops.tensor_initializers

#
# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#

import numbers
from typing import Optional, Union

from tripy import constraints, export
from tripy.common import datatype
from tripy.common.exception import raise_error
from tripy.frontend.trace.ops.fill import full, full_like
from tripy.frontend.trace.ops.iota import iota, iota_like
from tripy.frontend.trace.ops.where import where


[docs] @export.public_api(document_under="operations/initializers") @constraints.dtypes( constraints={"dtype": "T1", constraints.RETURN_VALUE: "T1"}, variables={ "T1": ["float32", "float16", "bfloat16", "float8", "int8", "int4", "int32", "int64", "bool"], }, ) def ones( shape: "tripy.types.ShapeLike", dtype: datatype.dtype = datatype.float32, ) -> "tripy.Tensor": """ Creates a Tensor of the specified shape and dtype with all elements set to 1. Args: shape: The desired shape of the tensor. dtype: Datatype of elements. Returns: A tensor of shape ``shape`` with all elements set to 1. .. code-block:: python :linenos: :caption: Example output = tp.ones([2, 3]) assert np.array_equal(cp.from_dlpack(output).get(), np.ones([2, 3], dtype=np.float32)) .. seealso:: :func:`ones_like`, :func:`full` """ return full(shape, 1, dtype)
[docs] @export.public_api(document_under="operations/initializers") @constraints.dtypes( constraints={"dtype": "T1", constraints.RETURN_VALUE: "T1"}, variables={ "T1": ["float32", "float16", "bfloat16", "float8", "int8", "int4", "int32", "int64", "bool"], }, ) def zeros( shape: "tripy.types.ShapeLike", dtype: datatype.dtype = datatype.float32, ) -> "tripy.Tensor": """ Creates a Tensor of the specified shape and dtype with all elements set to 0. Args: shape: The desired shape of the tensor. dtype: Datatype of elements. Returns: A tensor of shape ``shape`` with all elements set to 0. .. code-block:: python :linenos: :caption: Example output = tp.zeros([2, 3]) assert np.array_equal(cp.from_dlpack(output).get(), np.zeros([2, 3], dtype=np.float32)) .. seealso:: :func:`zeros_like`, :func:`full` """ return full(shape, 0, dtype)
[docs] @export.public_api(document_under="operations/initializers") @constraints.dtypes( constraints={"input": "T1", "dtype": "T2", constraints.RETURN_VALUE: "T2"}, variables={ "T1": ["float32", "float16", "bfloat16", "float8", "int4", "int8", "int32", "int64", "bool"], "T2": ["float32", "float16", "bfloat16", "float8", "int4", "int8", "int32", "int64", "bool"], }, ) def ones_like(input: "tripy.Tensor", dtype: Optional[datatype.dtype] = None) -> "tripy.Tensor": """ Creates a tensor with all elements set to 1 of the same shape as the input tensor. Args: input: The input tensor. dtype: Datatype of elements. If set to ``None``, the datatype of the input tensor is used. Returns: A tensor of the same shape as the input with all elements set to 1. .. code-block:: python :linenos: :caption: Example input = tp.zeros([2, 3], dtype=tp.float32) output = tp.ones_like(input) assert np.array_equal(cp.from_dlpack(output).get(), np.ones([2, 3], dtype=np.float32)) .. seealso:: :func:`ones`, :func:`full_like` """ return full_like(input, 1, dtype)
[docs] @export.public_api(document_under="operations/initializers") @constraints.dtypes( constraints={"input": "T1", "dtype": "T2", constraints.RETURN_VALUE: "T2"}, variables={ "T1": ["float32", "float16", "bfloat16", "float8", "int4", "int8", "int32", "int64", "bool"], "T2": ["float32", "float16", "bfloat16", "float8", "int4", "int8", "int32", "int64", "bool"], }, ) def zeros_like(input: "tripy.Tensor", dtype: Optional[datatype.dtype] = None) -> "tripy.Tensor": """ Creates a Tensor with all elements set to 0 of the same shape as the input tensor. Args: input: The input tensor. dtype: Datatype of elements. If set to ``None``, the datatype of the input tensor is used. Returns: A tensor of the same shape as the input with all elements set to 0. .. code-block:: python :linenos: :caption: Example input = tp.iota([2, 3], dtype=tp.float32) output = tp.zeros_like(input) assert np.array_equal(cp.from_dlpack(output).get(), np.zeros([2, 3], dtype=np.float32)) .. seealso:: :func:`zeros`, :func:`full_like` """ return full_like(input, 0, dtype)
[docs] @export.public_api(document_under="operations/initializers") @constraints.dtypes( constraints={"tensor": "T1", constraints.RETURN_VALUE: "T1"}, variables={ "T1": ["float32", "float16", "bfloat16", "int32", "int64", "bool"], }, ) def tril(tensor: "tripy.Tensor", diagonal: int = 0) -> "tripy.Tensor": r""" Returns the lower triangular part of each :math:`[M, N]` matrix in the tensor, with all other elements set to 0. If the tensor has more than two dimensions, it is treated as a batch of matrices. Args: tensor: The tripy tensor to operate on. diagonal: The diagonal above which to zero elements. ``diagonal=0`` indicates the main diagonal which is defined by the set of indices :math:`{{(i, i)}}` where :math:`i \in [0, min(M, N))`. Positive values indicate the diagonal which is that many diagonals above the main one, while negative values indicate one which is below. Returns: A tensor of the same shape as this tensor. .. code-block:: python :linenos: :caption: Main Diagonal input = tp.iota((2, 1, 3, 3), dim=2) + 1. output = tp.tril(input) assert np.array_equal(cp.from_dlpack(output).get(), np.tril(cp.from_dlpack(input).get())) .. code-block:: python :linenos: :caption: Two Diagonals Above Main input = tp.iota((5, 5)) + 1. # doc: omit output = tp.tril(input, diagonal=2) assert np.array_equal(cp.from_dlpack(output).get(), np.tril(cp.from_dlpack(input).get(), 2)) .. code-block:: python :linenos: :caption: One Diagonal Below Main input = tp.iota((5, 5)) + 1. # doc: omit output = tp.tril(input, diagonal=-1) assert np.array_equal(cp.from_dlpack(output).get(), np.tril(cp.from_dlpack(input).get(), -1)) """ tri_mask = (iota_like(tensor, -2, datatype.int32) + full_like(tensor, diagonal, datatype.int32)) >= iota_like( tensor, -1, datatype.int32 ) zeros_tensor = zeros_like(tensor) return where(tri_mask, tensor, zeros_tensor)
[docs] @export.public_api(document_under="operations/initializers") @constraints.dtypes( constraints={"tensor": "T1", constraints.RETURN_VALUE: "T1"}, variables={ "T1": ["float32", "float16", "bfloat16", "int32", "int64", "bool"], }, ) def triu(tensor: "tripy.Tensor", diagonal: int = 0) -> "tripy.Tensor": r""" Returns the upper triangular part of each :math:`[M, N]` matrix in the tensor, with all other elements set to 0. If the tensor has more than two dimensions, it is treated as a batch of matrices. Args: tensor: The tripy tensor to operate on. diagonal: The diagonal below which to zero elements. ``diagonal=0`` indicates the main diagonal which is defined by the set of indices :math:`{{(i, i)}}` where :math:`i \in [0, min(M, N))`. Positive values indicate the diagonal which is that many diagonals above the main one, while negative values indicate one which is below. Returns: A tensor of the same shape as this tensor. .. code-block:: python :linenos: :caption: Main Diagonal input = tp.iota((2, 1, 3, 3), dim=2) + 1. output = tp.triu(input) assert np.array_equal(cp.from_dlpack(output).get(), np.triu(cp.from_dlpack(input).get())) .. code-block:: python :linenos: :caption: Two Diagonals Above Main input = tp.iota((5, 5)) + 1. # doc: omit output = tp.triu(input, diagonal=2) assert np.array_equal(cp.from_dlpack(output).get(), np.triu(cp.from_dlpack(input).get(), 2)) .. code-block:: python :linenos: :caption: One Diagonal Below Main input = tp.iota((5, 5)) + 1. # doc: omit output = tp.triu(input, diagonal=-1) assert np.array_equal(cp.from_dlpack(output).get(), np.triu(cp.from_dlpack(input).get(), -1)) """ tri_mask = (iota_like(tensor, -2, datatype.int32) + full_like(tensor, diagonal, datatype.int32)) <= iota_like( tensor, -1, datatype.int32 ) zeros_tensor = zeros_like(tensor) return where(tri_mask, tensor, zeros_tensor)
@export.public_api(document_under="operations/initializers") @constraints.dtypes( constraints={"dtype": "T1", constraints.RETURN_VALUE: "T1"}, variables={ "T1": ["float32", "float16", "bfloat16", "int8", "int32", "int64", "bool"], }, ) def arange( start: Union[numbers.Number, "tripy.DimensionSize"], stop: Union[numbers.Number, "tripy.DimensionSize"], step: Union[numbers.Number, "tripy.DimensionSize"] = 1, dtype: "tripy.dtype" = datatype.float32, ) -> "tripy.Tensor": r""" Returns a 1D tensor containing a sequence of numbers in the half-open interval :math:`[0, \text{stop})` incrementing by :math:`\text{step}`. Args: start: The inclusive lower bound of the values to generate. If a tensor is provided, it must be a scalar tensor. stop: The exclusive upper bound of the values to generate. If a tensor is provided, it must be a scalar tensor. step: The spacing between values. If a tensor is provided, it must be a scalar tensor. dtype: The desired data type of the tensor. Returns: A tensor of shape :math:`[\frac{\text{stop}-\text{start}}{\text{step}}]`. .. code-block:: python :linenos: :caption: Example output = tp.arange(0.5, 2.5) assert (cp.from_dlpack(output).get() == np.arange(0.5, 2.5, dtype=np.float32)).all() .. code-block:: python :linenos: :caption: Custom ``step`` Value output = tp.arange(2.3, 0.8, -0.2) assert tp.allclose(output, tp.Tensor(np.arange(2.3, 0.8, -0.2, dtype=np.float32))) """ from tripy.frontend.dimension_size import DimensionSize if isinstance(step, numbers.Number) and step == 0: raise_error("Step in arange cannot be 0.", []) # math.ceil(a / b) is same as -(-a // b). Don't use math.ceil as start, stop or step can be Tensor. size = 0 - ((start - stop) // step) if isinstance(size, numbers.Number) and size <= 0: raise_error( "Arange tensor is empty.", details=[ f"start={start}, stop={stop}, step={step}", ], ) if not isinstance(size, DimensionSize): size = int(size) size = (size,) output = iota(size, 0, dtype) * full(size, step, dtype) + full(size, start, dtype) return output
[docs] @export.public_api(document_under="operations/initializers") @constraints.dtypes( constraints={"dtype": "T1", constraints.RETURN_VALUE: "T1"}, variables={ "T1": ["float32", "float16", "bfloat16", "int8", "int32", "int64", "bool"], }, ) def arange( stop: Union[numbers.Number, "tripy.DimensionSize"], dtype: "tripy.dtype" = datatype.float32 ) -> "tripy.Tensor": r""" Returns a 1D tensor containing a sequence of numbers in the half-open interval :math:`[0, \text{stop})` incrementing by 1. Args: stop: The exclusive upper bound of the values to generate. dtype: The desired datatype of the tensor. Returns: A tensor of shape :math:`[\text{stop}]`. .. code-block:: python :linenos: :caption: Example output = tp.arange(5) assert (cp.from_dlpack(output).get() == np.arange(5, dtype=np.float32)).all() """ return arange(0, stop, dtype=dtype)