Source code for sparseml.onnx.utils.sparse_tensor

# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
# 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
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an "AS IS" BASIS,
# See the License for the specific language governing permissions and
# limitations under the License.

Helper functions for handling ONNX SparseTensorProto objects.
onnx >= 1.6.0 is a requirement for using sparse tensors

from copy import deepcopy
from typing import Union

import numpy
from onnx import ModelProto, TensorProto, numpy_helper

    from onnx import SparseTensorProto

    sparse_tensor_import_error = None
except Exception as sparse_tensor_err:
    SparseTensorProto = None
    sparse_tensor_import_error = sparse_tensor_err

__all__ = [

def _check_sparse_tensor_import():
    if sparse_tensor_import_error:
        # ONNX >= 1.6.0 required
        raise sparse_tensor_import_error

[docs]def create_sparse_tensor( array: Union[numpy.ndarray, TensorProto], name: str = None, ) -> Union[SparseTensorProto, None]: """ :param array: numpy array or TensorProto object to convert to sparse representation :param name: name of this sparse tensor. Will be stored in If the given array is a TensorProto, name will default to :return: SparseTensorProto object built from the sparse representation of the input array """ _check_sparse_tensor_import() if isinstance(array, TensorProto): if not name: name = or None array = numpy_helper.to_array(array) # flatten array and convert to sparse original_dims = array.shape array = array.reshape(-1) nonzero_idxs = array.nonzero() nonzero_values = array[nonzero_idxs] nonzero_idxs = nonzero_idxs[0] # unwrap 1-tuple nonzero_idxs = nonzero_idxs.astype(numpy.int64) # required idx dtype # build SparseTensorProto return SparseTensorProto( values=numpy_helper.from_array(nonzero_values, name=name), indices=numpy_helper.from_array(nonzero_idxs), dims=original_dims, )
[docs]def sparse_tensor_to_dense(sparse_tensor: SparseTensorProto) -> TensorProto: """ :param sparse_tensor: SparseTensorProto object :return: TensorProto object that is the dense representation of the given sparse tensor. """ _check_sparse_tensor_import() name = values = numpy_helper.to_array(sparse_tensor.values) indices = numpy_helper.to_array(sparse_tensor.indices) shape = sparse_tensor.dims dense_array = numpy.zeros( dense_array[indices] = values dense_array = dense_array.reshape(shape) return numpy_helper.from_array(dense_array, name=name)
_COMPRESSIBLE_DATA_TYPES = { TensorProto.FLOAT, TensorProto.FLOAT16, TensorProto.INT64, TensorProto.INT32, TensorProto.INT16, }
[docs]def convert_model_initializers_to_sparse( model: ModelProto, sparsity_threshold: float = 0.6, inplace: bool = True ) -> ModelProto: """ :param model: ONNX model with initializers to convert to sparse :param sparsity_threshold: the minimum sparsity of a tensor to be converted to sparse representation. Default is 0.6 :param inplace: True to do model conversion in place. Default is True :return: the given model with initializers above the sparsity threshold converted to sparse initializers """ _check_sparse_tensor_import() if not inplace: model = deepcopy(model) sparsified_initializers = [] for initializer in model.graph.initializer: if initializer.data_type not in _COMPRESSIBLE_DATA_TYPES: continue val = numpy_helper.to_array(initializer) sparsity = 1.0 - (numpy.count_nonzero(val) / val.size) if sparsity < sparsity_threshold: continue sparse_tensor = create_sparse_tensor(val, if sparse_tensor is None: continue sparsified_initializers.append(initializer) model.graph.sparse_initializer.append(sparse_tensor) for initializer in sparsified_initializers: model.graph.initializer.remove(initializer) return model
[docs]def convert_sparse_initializers_to_dense( model: ModelProto, inplace: bool = True ) -> ModelProto: """ :param model: ONNX model with sparse initializers to convert to dense representation :param inplace: True to do model conversion in place. Default is True :return: The given model with all sparse initializers converted to dense initializers """ _check_sparse_tensor_import() if not inplace: model = deepcopy(model) while model.graph.sparse_initializer: sparse_initializer = model.graph.sparse_initializer.pop() model.graph.initializer.append(sparse_tensor_to_dense(sparse_initializer)) return model