wehub-resource-sync
443 lines
15 KiB
Python
443 lines
15 KiB
Python
#
|
|
# SPDX-FileCopyrightText: Copyright (c) 1993-2024 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.
|
|
#
|
|
|
|
from onnx_graphsurgeon.logger import G_LOGGER
|
|
from onnx_graphsurgeon.util import misc
|
|
|
|
from typing import Set, Sequence, Union
|
|
import numpy as np
|
|
|
|
|
|
class Tensor(object):
|
|
"""Abstract base class for tensors in a graph"""
|
|
|
|
DYNAMIC = -1
|
|
|
|
def __init__(self):
|
|
"""
|
|
**This class is abstract and cannot be constructed directly.**
|
|
"""
|
|
raise NotImplementedError("Tensor is an abstract class")
|
|
|
|
def __setattr__(self, name, value):
|
|
if name in ["inputs", "outputs"]:
|
|
try:
|
|
attr = getattr(self, name)
|
|
if value is attr:
|
|
# This can happen when using things like +=
|
|
# The __iadd__ is executed followed by an assignment
|
|
return
|
|
|
|
attr.clear()
|
|
attr.extend(value)
|
|
except AttributeError:
|
|
super().__setattr__(name, value)
|
|
else:
|
|
super().__setattr__(name, value)
|
|
|
|
def is_empty(self):
|
|
"""
|
|
Returns whether this tensor is considered empty in the graph.
|
|
|
|
*Note: 'Empty' here refers to the name of the tensor, which is omitted for
|
|
optional tensors, NOT the shape of the tensor*
|
|
|
|
Returns:
|
|
bool: Whether the tensor is empty, meaning that it is used for an omitted optional input or output.
|
|
"""
|
|
return self.name == ""
|
|
|
|
def to_constant(
|
|
self,
|
|
values: np.ndarray,
|
|
data_location: int = None,
|
|
export_dtype: Union[np.dtype, "onnx.TensorProto.DataType"] = None,
|
|
):
|
|
"""
|
|
Modifies this tensor in-place to convert it to a Constant. This means that all consumers/producers of the tensor will see the update.
|
|
|
|
Args:
|
|
values (np.ndarray): The values in this tensor
|
|
|
|
data_location (int):
|
|
An enum value indicating the location where the tensor data is stored.
|
|
Generally, this will come from onnx.TensorProto.DataLocation.
|
|
|
|
dtype (Union[numpy.dtype, onnx.TensorProto.DataType]): The data type of the tensor.
|
|
Returns:
|
|
self
|
|
"""
|
|
self.__class__ = Constant
|
|
self._values = values
|
|
self.data_location = data_location
|
|
self.export_dtype = export_dtype
|
|
|
|
return self
|
|
|
|
def to_variable(
|
|
self,
|
|
dtype: Union[np.dtype, "onnx.TensorProto.DataType"] = None,
|
|
shape: Sequence[Union[int, str]] = [],
|
|
):
|
|
"""
|
|
Modifies this tensor in-place to convert it to a Variable. This means that all consumers/producers of the tensor will see the update.
|
|
|
|
Args:
|
|
dtype (Union[numpy.dtype, onnx.TensorProto.DataType]): The data type of the tensor.
|
|
shape (Sequence[int]): The shape of the tensor.
|
|
|
|
Returns:
|
|
self
|
|
"""
|
|
|
|
variable_dtype = dtype if dtype is not None else self.export_dtype
|
|
|
|
self.__class__ = Variable
|
|
self.shape = shape
|
|
self.dtype = variable_dtype
|
|
|
|
return self
|
|
|
|
def i(self, tensor_idx=0, producer_idx=0):
|
|
"""
|
|
Convenience function to get an input tensor of one of this tensor's input nodes.
|
|
Note that the parameters are swapped compared to the o() function; this is because tensors are likely to have only a single producer
|
|
|
|
For example:
|
|
::
|
|
|
|
assert tensor.i() == tensor.inputs[0].inputs[0]
|
|
assert tensor.i(1, 2) == tensor.inputs[2].inputs[1]
|
|
|
|
Args:
|
|
tensor_idx (int): The index of the input tensor of the input node. Defaults to 0.
|
|
producer_idx (int): The index of the producer node of the input tensor, if the tensor has multiple producers. Defaults to 0.
|
|
|
|
Returns:
|
|
Tensor: The specified producer (input) tensor.
|
|
"""
|
|
return self.inputs[producer_idx].inputs[tensor_idx]
|
|
|
|
def o(self, consumer_idx=0, tensor_idx=0):
|
|
"""
|
|
Convenience function to get an output tensor of one of this tensor's output nodes.
|
|
|
|
For example:
|
|
::
|
|
|
|
assert tensor.o() == tensor.outputs[0].outputs[0]
|
|
assert tensor.o(2, 1) == tensor.outputs[2].outputs[1]
|
|
|
|
Args:
|
|
consumer_idx (int): The index of the consumer of the input tensor. Defaults to 0.
|
|
tensor_idx (int): The index of the output tensor of the node, if the node has multiple outputs. Defaults to 0.
|
|
|
|
Returns:
|
|
Tensor: The specified consumer (output) tensor
|
|
"""
|
|
return self.outputs[consumer_idx].outputs[tensor_idx]
|
|
|
|
def __str__(self):
|
|
return "{:} ({:}): (shape={:}, dtype={:})".format(
|
|
type(self).__name__, self.name, self.shape, self.dtype
|
|
)
|
|
|
|
def __repr__(self): # Hack to make logging output pretty.
|
|
return self.__str__()
|
|
|
|
def __eq__(self, other):
|
|
"""
|
|
Perform a check to see if two tensors are equal.
|
|
|
|
Tensors are considered equal if they share the same name. A Graph must not include Tensors with duplicate names.
|
|
"""
|
|
return self.name == other.name
|
|
|
|
|
|
class Variable(Tensor):
|
|
@staticmethod
|
|
def empty():
|
|
"""
|
|
Creates a variable tensor with no name.
|
|
This can be used to represent an omitted optional input of a node.
|
|
"""
|
|
return Variable(name="")
|
|
|
|
def __init__(
|
|
self,
|
|
name: str,
|
|
dtype: Union[np.dtype, "onnx.TensorProto.DataType"] = None,
|
|
shape: Sequence[Union[int, str]] = None,
|
|
type: str = "tensor_type",
|
|
):
|
|
"""
|
|
Represents a Tensor whose value is not known until inference-time.
|
|
|
|
Args:
|
|
name (str): The name of the tensor.
|
|
dtype (Union[numpy.dtype, onnx.TensorProto.DataType]): The data type of the tensor.
|
|
shape (Sequence[Union[int, str]]): The shape of the tensor. This may contain strings if the model uses dimension parameters.
|
|
type (str): The type of the tensor.
|
|
"""
|
|
self.name = name
|
|
self.inputs = misc.SynchronizedList(self, field_name="outputs", initial=[])
|
|
self.outputs = misc.SynchronizedList(self, field_name="inputs", initial=[])
|
|
self.dtype = dtype
|
|
self.shape = misc.default_value(shape, None)
|
|
self.type = type
|
|
|
|
def to_constant(
|
|
self,
|
|
values: np.ndarray,
|
|
export_dtype: Union[np.dtype, "onnx.TensorProto.DataType"] = None,
|
|
):
|
|
del self.dtype
|
|
del self.shape
|
|
|
|
return super().to_constant(values, export_dtype=export_dtype)
|
|
|
|
def copy(self):
|
|
"""
|
|
Makes a shallow copy of this tensor, omitting input and output information.
|
|
|
|
Note: Generally, you should only ever make a copy of a Graph.
|
|
"""
|
|
return Variable(self.name, self.dtype, self.shape)
|
|
|
|
|
|
class LazyValues(object):
|
|
"""
|
|
A special object that represents constant tensor values that should be lazily loaded.
|
|
"""
|
|
|
|
def __init__(self, tensor):
|
|
"""
|
|
Args:
|
|
tensor (onnx.TensorProto, onnx.SparseTensorProto): The ONNX tensor that this instance should lazily load.
|
|
"""
|
|
from onnx_graphsurgeon.importers.onnx_importer import (
|
|
get_onnx_tensor_shape,
|
|
get_onnx_tensor_dtype,
|
|
get_itemsize,
|
|
)
|
|
|
|
self.tensor = tensor
|
|
self.shape = get_onnx_tensor_shape(self.tensor)
|
|
self.dtype = get_onnx_tensor_dtype(self.tensor)
|
|
self.nbytes = int(np.ceil(misc.volume(self.shape) * get_itemsize(self.dtype)))
|
|
|
|
def load(self):
|
|
"""
|
|
Load a numpy array from the underlying tensor values.
|
|
|
|
Returns:
|
|
np.array: A numpy array containing the values of the tensor.
|
|
"""
|
|
import onnx
|
|
import onnx.numpy_helper
|
|
from onnx_graphsurgeon.importers.onnx_importer import (
|
|
get_dtype_name,
|
|
get_numpy_type,
|
|
)
|
|
|
|
if get_numpy_type(self.dtype) is None:
|
|
G_LOGGER.warning(
|
|
f"Datatype: {get_dtype_name(self.dtype)} could not be converted to a NumPy type.\n"
|
|
f"Accessing the values of this constant tensor ({self.tensor.name}) will cause them to be casted to a supported data type. "
|
|
f"This means that the weights will have a different type than the original model when they are exported again!\n"
|
|
f"If this is not what you intended, please avoid accessing the values of this constant tensor."
|
|
)
|
|
|
|
return np.array(onnx.numpy_helper.to_array(self.tensor))
|
|
|
|
def __str__(self):
|
|
return "LazyValues (shape={:}, dtype={:})".format(self.shape, self.dtype)
|
|
|
|
def __repr__(self): # Hack to make logging output pretty.
|
|
return self.__str__()
|
|
|
|
|
|
class SparseValues(LazyValues):
|
|
"""
|
|
A special object that represents constant tensor values that is sparse
|
|
"""
|
|
|
|
def load(self):
|
|
"""
|
|
Load a numpy array from the sparse structure.
|
|
|
|
Returns:
|
|
np.array: A numpy array containing the values of the tensor.
|
|
"""
|
|
import onnx
|
|
import onnx.numpy_helper
|
|
from onnx_graphsurgeon.importers.onnx_importer import (
|
|
get_dtype_name,
|
|
get_numpy_type,
|
|
)
|
|
|
|
supported_index_type = [onnx.TensorProto.INT64]
|
|
if self.tensor.indices.data_type not in supported_index_type:
|
|
G_LOGGER.critical(
|
|
f"Unsupported index data type {self.tensor.indices.data_type} in {self.tensor.values.name}"
|
|
)
|
|
|
|
if self.tensor.values.data_type == onnx.TensorProto.FLOAT16:
|
|
values_data = np.asarray(
|
|
self.tensor.values.int32_data, dtype=np.uint16
|
|
).view(np.float16)
|
|
else:
|
|
field_name = onnx.helper.tensor_dtype_to_field(self.tensor.values.data_type)
|
|
values = getattr(self.tensor.values, field_name)
|
|
dtype = onnx.helper.tensor_dtype_to_np_dtype(self.tensor.values.data_type)
|
|
values_data = np.asarray(values, dtype)
|
|
indices_data = self.tensor.indices.int64_data
|
|
|
|
if len(self.tensor.indices.dims) == 1:
|
|
values = np.zeros(np.prod(self.tensor.dims))
|
|
# [NNZ] layout, in which case the i-th value must be the linearized-index of the i-th value.
|
|
values[indices_data] = values_data
|
|
values = values.reshape(self.tensor.dims)
|
|
elif len(self.tensor.indices.dims) == 2:
|
|
# [NNZ, rank] with the [i,j]-th value corresponding to the j-th index of the i-th value
|
|
values = np.zeros(self.tensor.dims)
|
|
indices_data = np.asarray(indices_data).reshape(self.tensor.indices.dims)
|
|
|
|
for i in range(len(values_data)):
|
|
values[tuple(indices_data[i])] = values_data[i]
|
|
else:
|
|
G_LOGGER.critical(
|
|
f"Unsupported index data dims {self.tensor.indices.dims} in {self.tensor.values.name}"
|
|
)
|
|
|
|
return values
|
|
|
|
def __str__(self):
|
|
return "SparseValues (shape={:}, dtype={:})".format(self.shape, self.dtype)
|
|
|
|
|
|
class Constant(Tensor):
|
|
def __init__(
|
|
self,
|
|
name: str,
|
|
values: Union[np.ndarray, LazyValues],
|
|
data_location: int = None,
|
|
export_dtype: Union[np.dtype, "onnx.TensorProto.DataType"] = None,
|
|
):
|
|
"""
|
|
Represents a Tensor whose value is known.
|
|
|
|
Args:
|
|
name (str): The name of the tensor.
|
|
values (numpy.ndarray): The values in this tensor, in the form of a NumPy array.
|
|
|
|
data_location (int):
|
|
An enum value indicating the location where the tensor data is stored.
|
|
Generally, this will come from onnx.TensorProto.DataLocation.
|
|
|
|
|
|
export_dtype (Union[np.dtype, onnx.TensorProto.DataType]):
|
|
The data type of the tensor when exported to onnx. If not specified, then
|
|
the data type of values will be used.
|
|
"""
|
|
self.name = name
|
|
self.inputs = misc.SynchronizedList(self, field_name="outputs", initial=[])
|
|
self.outputs = misc.SynchronizedList(self, field_name="inputs", initial=[])
|
|
if (
|
|
not isinstance(values, np.ndarray)
|
|
and not isinstance(values, LazyValues)
|
|
and not isinstance(values, SparseValues)
|
|
):
|
|
G_LOGGER.critical(
|
|
"Provided `values` argument is not a NumPy array, a LazyValues instance or a"
|
|
"SparseValues instance. Please provide a NumPy array or LazyValues instance "
|
|
"to construct a Constant. Note: Provided `values` parameter was: {:}".format(
|
|
values
|
|
)
|
|
)
|
|
self._values = values
|
|
self.data_location = data_location
|
|
self._export_dtype = export_dtype
|
|
|
|
def to_variable(
|
|
self, dtype: np.dtype = None, shape: Sequence[Union[int, str]] = []
|
|
):
|
|
var_dtype = self.export_dtype
|
|
|
|
del self._export_dtype
|
|
del self._values
|
|
|
|
if dtype is not None:
|
|
return super().to_variable(dtype, shape)
|
|
|
|
return super().to_variable(var_dtype, shape)
|
|
|
|
def copy(self):
|
|
"""
|
|
Makes a shallow copy of this tensor, omitting input and output information.
|
|
|
|
Note: Generally, you should only ever make a copy of a Graph.
|
|
"""
|
|
return Constant(self.name, self._values, export_dtype=self.export_dtype)
|
|
|
|
@property
|
|
def values(self):
|
|
# Load values when they are first accesed
|
|
if isinstance(self._values, LazyValues):
|
|
self._values = self._values.load()
|
|
return self._values
|
|
|
|
@values.setter
|
|
def values(self, values: Union[np.ndarray, LazyValues]):
|
|
self._values = values
|
|
|
|
@property
|
|
def shape(self):
|
|
return self._values.shape
|
|
|
|
@property
|
|
def dtype(self):
|
|
return self._values.dtype
|
|
|
|
@property
|
|
def export_dtype(self):
|
|
if self._export_dtype is not None:
|
|
return self._export_dtype
|
|
|
|
return self.dtype
|
|
|
|
@export_dtype.setter
|
|
def export_dtype(self, export_dtype):
|
|
if export_dtype is not None:
|
|
import onnx
|
|
|
|
try:
|
|
onnx.helper.float32_to_bfloat16
|
|
except AttributeError:
|
|
G_LOGGER.critical(
|
|
"`export_dtype` is not support with this version of ONNX. "
|
|
"Please use `ml_dtypes` to create the Constant with the correct data type."
|
|
"Alternatively, you can downgrade ONNX to version 1.19.1 or earlier."
|
|
)
|
|
|
|
self._export_dtype = export_dtype
|
|
|
|
def __repr__(self): # Hack to make logging output pretty.
|
|
ret = self.__str__()
|
|
ret += "\n{:}".format(self._values)
|
|
return ret
|