Public Member Functions
	__init__ (self, model)

	make_tensor_with_base_name (self, base_name)

	make_node (self, opcode, inputs, outputs, p_args, *k_args)

	make_split (self, input, split_sizes, axis)

	make_concat (self, inputs, axis)

	make_squeeze (self, input, axes)

	make_unsqueeze (self, input, axes)

	make_gemm (self, A, B, C, trans_a=False, trans_b=False)

	make_add (self, a, b)

	make_mul (self, a, b)

	make_clip (self, input, min, max)

	make_act (self, input, act_name)

	make_constant_tensor (self, tensor_data, base_name)

	mark_for_deletion (self, node)

	get_insert_id (self)

	set_insert_id (self, insert_id)

	delete_marked_nodes (self)

Protected Attributes
	_model

	_nodes_to_delete

	_insert_id

	_base_name_idx

Detailed Description

Helper for onnx model transformation

This helper is used for convenient operation replacement in onnx model

Attributes:
    _model (onnx.onnx_ml_pb2.ModelProto): target model that should be changed
    _nodes_to_delete (list of onnx.onnx_ml_pb2.NodeProto): list of replaced operations
    _insert_id (int): position to insert created operations (should be in topologically sorted)
    _base_name_idx (dict from str to int): maps tensor "base" name to
        largest existing serial num. For example model has tensors "t_1", "t_2", "t_4",
        in that case _base_name_idx["t_"] == 4.
        This attribute is used for unique tensor name generation.

Definition at line 75 of file onnx_legalizer.py.

Constructor & Destructor Documentation

◆ init()

onnx_legalizer._ModelTransformerHelper.__init__	(	self,
		model
	)

Definition at line 89 of file onnx_legalizer.py.

    def __init__(self, model):
        self._model = model
        self._nodes_to_delete = []
        self._insert_id = 0
        # each tensor has name containing base name and unique number. for example:
        # "abc_123": "abs_" - base name, "123" - unique number
        # if no number in name, consider it is equal to "0"
 
        # mapping from base names to largest given number
        self._base_name_idx = {}
        # gather name information for existing tensors
        for node in model.graph.node:
            for t in list(node.input) + list(node.output):
                base_name, number = _parse_tensor_name(t)
                if base_name in self._base_name_idx:
                    self._base_name_idx[base_name] = max(self._base_name_idx[base_name],
                                                         number)
                else:
                    self._base_name_idx[base_name] = number
 

Member Function Documentation

◆ delete_marked_nodes()

onnx_legalizer._ModelTransformerHelper.delete_marked_nodes ( self )

Definition at line 293 of file onnx_legalizer.py.

    def delete_marked_nodes(self):
        for node in self._nodes_to_delete:
            self._model.graph.node.remove(node)
 
 

References SingleOperatorTest.SingleOperatorTest._model, ScalarOperandDecl._model, TensorOperandDecl._model, WeightDecl._model, OperationDecl._model, luci::CircleReader._model, mir_onnx::ONNXImporterImpl._model, mir_tflite::TfliteImporter._model, GenericBackend._model, onnx_legalizer._ModelTransformerHelper._model, DotBuilder.DotBuilder._model, luci_interpreter::CircleReader._model, onert_micro::core::reader::OMCircleReader._model, onert::exporter::CircleExporter._model, onert::compiler::train::TrainingCompiler._model, and onnx_legalizer._ModelTransformerHelper._nodes_to_delete.

◆ get_insert_id()

onnx_legalizer._ModelTransformerHelper.get_insert_id ( self )

Definition at line 287 of file onnx_legalizer.py.

    def get_insert_id(self):
        return self._insert_id
 

References onnx_legalizer._ModelTransformerHelper._insert_id.

◆ make_act()

onnx_legalizer._ModelTransformerHelper.make_act	(	self,
		input,
		act_name
	)

Create activation function operation and insert it in graph.

Args:
    input (str): input tensor name
    act_name (str): name of activation function, one of ['Relu', 'Tanh', 'Sigmoid']

Returns:
    str: output tensor name

Definition at line 256 of file onnx_legalizer.py.

    def make_act(self, input, act_name):
        """Create activation function operation and insert it in graph.
 
        Args:
            input (str): input tensor name
            act_name (str): name of activation function, one of ['Relu', 'Tanh', 'Sigmoid']
 
        Returns:
            str: output tensor name
        """
        assert (act_name in ['Relu', 'Tanh', 'Sigmoid'])
        return self.make_node(act_name, [input], 1)[0]
 

References onnx_legalizer._ModelTransformerHelper.make_node().

◆ make_add()

onnx_legalizer._ModelTransformerHelper.make_add	(	self,
		a,
		b
	)

Creates Add operation and insert it in graph.

Args:
    a (str): name of left operand tensor
    b (str): name of right operand tensor

Returns:
    str: output tensor name

Definition at line 219 of file onnx_legalizer.py.

    def make_add(self, a, b):
        """Creates Add operation and insert it in graph.
 
        Args:
            a (str): name of left operand tensor
            b (str): name of right operand tensor
 
        Returns:
            str: output tensor name
        """
        return self.make_node('Add', [a, b], 1)[0]
 

References onnx_legalizer._ModelTransformerHelper.make_node().

◆ make_clip()

onnx_legalizer._ModelTransformerHelper.make_clip	(	self,
		input,
		min,
		max
	)

Create Clip operation and insert it in graph.

Args:
    input (str): input tensor name
    min (int/float): lower clip bound
    max (int/float ): upper clip bound

Returns:
    str: output tensor name

Definition at line 243 of file onnx_legalizer.py.

    def make_clip(self, input, min, max):
        """Create Clip operation and insert it in graph.
 
        Args:
            input (str): input tensor name
            min (int/float): lower clip bound
            max (int/float ): upper clip bound
 
        Returns:
            str: output tensor name
        """
        return self.make_node('Clip', [input], 1, min=min, max=max)[0]
 

References onnx_legalizer._ModelTransformerHelper.make_node().

◆ make_concat()

onnx_legalizer._ModelTransformerHelper.make_concat	(	self,
		inputs,
		axis
	)

Create Concat operation and insert it in graph.

Args:
    inputs (list of str): list of tensors names to concat
    axis (int): axis number to concat

Returns:
    str: output tensor name

Definition at line 163 of file onnx_legalizer.py.

    def make_concat(self, inputs, axis):
        """Create Concat operation and insert it in graph.
 
        Args:
            inputs (list of str): list of tensors names to concat
            axis (int): axis number to concat
 
        Returns:
            str: output tensor name
        """
        return self.make_node('Concat', inputs, 1, axis=axis)[0]
 

References onnx_legalizer._ModelTransformerHelper.make_node().

◆ make_constant_tensor()

onnx_legalizer._ModelTransformerHelper.make_constant_tensor	(	self,
		tensor_data,
		base_name
	)

Creates onnx constant tensor

Args:
    tensor_data (numpy.ndarray): tensor data
    base_name (str): prefix of constant tensor name

Returns:
    str: name of created constant tensor

Definition at line 269 of file onnx_legalizer.py.

    def make_constant_tensor(self, tensor_data, base_name):
        """Creates onnx constant tensor
 
        Args:
            tensor_data (numpy.ndarray): tensor data
            base_name (str): prefix of constant tensor name
 
        Returns:
            str: name of created constant tensor
        """
        tensor = onnx.numpy_helper.from_array(tensor_data)
        tensor.name = self.make_tensor_with_base_name(base_name)
        self._model.graph.initializer.append(tensor)
        return tensor.name
 

References SingleOperatorTest.SingleOperatorTest._model, ScalarOperandDecl._model, TensorOperandDecl._model, WeightDecl._model, OperationDecl._model, luci::CircleReader._model, mir_onnx::ONNXImporterImpl._model, mir_tflite::TfliteImporter._model, GenericBackend._model, onnx_legalizer._ModelTransformerHelper._model, DotBuilder.DotBuilder._model, luci_interpreter::CircleReader._model, onert_micro::core::reader::OMCircleReader._model, onert::exporter::CircleExporter._model, onert::compiler::train::TrainingCompiler._model, and onnx_legalizer._ModelTransformerHelper.make_tensor_with_base_name().

◆ make_gemm()

onnx_legalizer._ModelTransformerHelper.make_gemm	(	self,
		A,
		B,
		C,
		trans_a = `False`,
		trans_b = `False`
	)

Create Gemm operation and insert it in graph.

Result tensor contains A*B + C

Args:
    A (str): name of tensor A
    B (str): name of tensor B
    C (str): name of tensor C
    transA (bool): if True, transpose tensor A before multiplication
    transB (bool): if True, transpose tensor B before multiplication

Returns:
    str: output tensor name

Definition at line 199 of file onnx_legalizer.py.

    def make_gemm(self, A, B, C, trans_a=False, trans_b=False):
        """Create Gemm operation and insert it in graph.
 
        Result tensor contains A*B + C
 
        Args:
            A (str): name of tensor A
            B (str): name of tensor B
            C (str): name of tensor C
            transA (bool): if True, transpose tensor A before multiplication
            transB (bool): if True, transpose tensor B before multiplication
 
        Returns:
            str: output tensor name
        """
        return self.make_node('Gemm', [A, B, C],
                              1,
                              transA=bool(trans_a),
                              transB=bool(trans_b))[0]
 

References onnx_legalizer._ModelTransformerHelper.make_node().

◆ make_mul()

onnx_legalizer._ModelTransformerHelper.make_mul	(	self,
		a,
		b
	)

Creates Mul operation and insert it in graph.

Args:
    a (str): name of left operand tensor
    b (str): name of right operand tensor

Returns:
    str: output tensor name

Definition at line 231 of file onnx_legalizer.py.

    def make_mul(self, a, b):
        """Creates Mul operation and insert it in graph.
 
        Args:
            a (str): name of left operand tensor
            b (str): name of right operand tensor
 
        Returns:
            str: output tensor name
        """
        return self.make_node('Mul', [a, b], 1)[0]
 

References onnx_legalizer._ModelTransformerHelper.make_node().

◆ make_node()

onnx_legalizer._ModelTransformerHelper.make_node	(		self,
			opcode,
			inputs,
			outputs,
		*	p_args,
		**	k_args
	)

Create arbitrary node and insert it in graph.

Args:
    opcode (str): opcode name of desired operation
    inputs (list of str): names of input tensors
    outputs (list of str or int): names of existing tensors to use as output tensors for operation or
        number of tensors that should be created
    p_args: additional arguments for onnx make_node helper
    k_args: attributes for onnx node

Returns:
    list of str: list of output tensor names

Definition at line 125 of file onnx_legalizer.py.

    def make_node(self, opcode, inputs, outputs, *p_args, **k_args):
        """Create arbitrary node and insert it in graph.
 
        Args:
            opcode (str): opcode name of desired operation
            inputs (list of str): names of input tensors
            outputs (list of str or int): names of existing tensors to use as output tensors for operation or
                number of tensors that should be created
            p_args: additional arguments for onnx make_node helper
            k_args: attributes for onnx node
 
        Returns:
            list of str: list of output tensor names
        """
        if type(outputs) == int:
            outputs = [self.make_tensor_with_base_name('') for i in range(outputs)]
        assert (type(outputs) == list)
        node = onnx.helper.make_node(opcode, inputs, outputs, *p_args, **k_args)
        self._model.graph.node.insert(self._insert_id, node)
        self._insert_id += 1
        return outputs
 

References onnx_legalizer._ModelTransformerHelper._insert_id, SingleOperatorTest.SingleOperatorTest._model, ScalarOperandDecl._model, TensorOperandDecl._model, WeightDecl._model, OperationDecl._model, luci::CircleReader._model, mir_onnx::ONNXImporterImpl._model, mir_tflite::TfliteImporter._model, GenericBackend._model, onnx_legalizer._ModelTransformerHelper._model, DotBuilder.DotBuilder._model, luci_interpreter::CircleReader._model, onert_micro::core::reader::OMCircleReader._model, onert::exporter::CircleExporter._model, onert::compiler::train::TrainingCompiler._model, onnx_legalizer._ModelTransformerHelper.make_tensor_with_base_name(), and type.

Referenced by onnx_legalizer._ModelTransformerHelper.make_act(), onnx_legalizer._ModelTransformerHelper.make_add(), onnx_legalizer._ModelTransformerHelper.make_clip(), onnx_legalizer._ModelTransformerHelper.make_concat(), onnx_legalizer._ModelTransformerHelper.make_gemm(), onnx_legalizer._ModelTransformerHelper.make_mul(), onnx_legalizer._ModelTransformerHelper.make_split(), onnx_legalizer._ModelTransformerHelper.make_squeeze(), and onnx_legalizer._ModelTransformerHelper.make_unsqueeze().

◆ make_split()

onnx_legalizer._ModelTransformerHelper.make_split	(	self,
		input,
		split_sizes,
		axis
	)

Create Split operation and insert it in graph.

Args:
    input (str): name of input tensor
    split_sizes (list of int): list of split sizes
    axis (int): number of axis to split

Returns:
    list: list of output tensor names

Definition at line 147 of file onnx_legalizer.py.

    def make_split(self, input, split_sizes, axis):
        """Create Split operation and insert it in graph.
 
        Args:
            input (str): name of input tensor
            split_sizes (list of int): list of split sizes
            axis (int): number of axis to split
 
        Returns:
            list: list of output tensor names
        """
        return self.make_node('Split', [input],
                              len(split_sizes),
                              axis=axis,
                              split=split_sizes)
 

References onnx_legalizer._ModelTransformerHelper.make_node().

◆ make_squeeze()

onnx_legalizer._ModelTransformerHelper.make_squeeze	(	self,
		input,
		axes
	)

Create Squeeze operation and insert it in graph.

Args:
    input (str): name of input tensor
    axes (list of int): list of dimension containing ones to remove

Returns:
    str: output tensor name

Definition at line 175 of file onnx_legalizer.py.

    def make_squeeze(self, input, axes):
        """Create Squeeze operation and insert it in graph.
 
        Args:
            input (str): name of input tensor
            axes (list of int): list of dimension containing ones to remove
 
        Returns:
            str: output tensor name
        """
        return self.make_node('Squeeze', [input], 1, axes=axes)[0]
 

References onnx_legalizer._ModelTransformerHelper.make_node().

◆ make_tensor_with_base_name()

onnx_legalizer._ModelTransformerHelper.make_tensor_with_base_name	(	self,
		base_name
	)

 Create unique name for given base_name

Args:
    base_name (str): base tensor name

Returns:
    str : unique tensor name that starts with base_name

Definition at line 109 of file onnx_legalizer.py.

    def make_tensor_with_base_name(self, base_name):
        """ Create unique name for given base_name
 
        Args:
            base_name (str): base tensor name
 
        Returns:
            str : unique tensor name that starts with base_name
        """
        if base_name in self._base_name_idx:
            self._base_name_idx[base_name] += 1
            return base_name + str(self._base_name_idx[base_name])
        else:
            self._base_name_idx[base_name] = 0
            return base_name + '0'
 

References onnx_legalizer._ModelTransformerHelper._base_name_idx.

Referenced by onnx_legalizer._ModelTransformerHelper.make_constant_tensor(), and onnx_legalizer._ModelTransformerHelper.make_node().

◆ make_unsqueeze()

onnx_legalizer._ModelTransformerHelper.make_unsqueeze	(	self,
		input,
		axes
	)

Create Unsqueeze operation and insert it in graph.

Args:
    input (str): name of input tensor
    axes (list of int): list of dimension to insert ones

Returns:
    str: output tensor name

Definition at line 187 of file onnx_legalizer.py.

    def make_unsqueeze(self, input, axes):
        """Create Unsqueeze operation and insert it in graph.
 
        Args:
            input (str): name of input tensor
            axes (list of int): list of dimension to insert ones
 
        Returns:
            str: output tensor name
        """
        return self.make_node('Unsqueeze', [input], 1, axes=axes)[0]
 

References onnx_legalizer._ModelTransformerHelper.make_node().

◆ mark_for_deletion()

onnx_legalizer._ModelTransformerHelper.mark_for_deletion	(	self,
		node
	)

Definition at line 284 of file onnx_legalizer.py.

    def mark_for_deletion(self, node):
        self._nodes_to_delete += [node]
 

References onnx_legalizer._ModelTransformerHelper._nodes_to_delete.

◆ set_insert_id()

onnx_legalizer._ModelTransformerHelper.set_insert_id	(	self,
		insert_id
	)

Definition at line 290 of file onnx_legalizer.py.

    def set_insert_id(self, insert_id):
        self._insert_id = insert_id
 

References onnx_legalizer._ModelTransformerHelper._insert_id.

Field Documentation

◆ _base_name_idx

onnx_legalizer._ModelTransformerHelper._base_name_idx

protected

Definition at line 98 of file onnx_legalizer.py.

Referenced by onnx_legalizer._ModelTransformerHelper.make_tensor_with_base_name().

◆ _insert_id

onnx_legalizer._ModelTransformerHelper._insert_id

protected

Definition at line 92 of file onnx_legalizer.py.

Referenced by onnx_legalizer._ModelTransformerHelper.get_insert_id(), onnx_legalizer._ModelTransformerHelper.make_node(), and onnx_legalizer._ModelTransformerHelper.set_insert_id().

◆ _model

onnx_legalizer._ModelTransformerHelper._model

protected

Definition at line 90 of file onnx_legalizer.py.

Referenced by onnx_legalizer._ModelTransformerHelper.delete_marked_nodes(), onnx_legalizer._ModelTransformerHelper.make_constant_tensor(), onnx_legalizer._ModelTransformerHelper.make_node(), and DotBuilder.DotBuilder.save().

◆ _nodes_to_delete

onnx_legalizer._ModelTransformerHelper._nodes_to_delete

protected

Definition at line 91 of file onnx_legalizer.py.

Referenced by onnx_legalizer._ModelTransformerHelper.delete_marked_nodes(), and onnx_legalizer._ModelTransformerHelper.mark_for_deletion().

The documentation for this class was generated from the following file:

compiler/one-cmds/onnx_legalizer.py

Public Member Functions

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

◆ __init__()

Member Function Documentation

◆ delete_marked_nodes()

◆ get_insert_id()

◆ make_act()

◆ make_add()

◆ make_clip()

◆ make_concat()

◆ make_constant_tensor()

◆ make_gemm()

◆ make_mul()

◆ make_node()

◆ make_split()

◆ make_squeeze()

◆ make_tensor_with_base_name()

◆ make_unsqueeze()

◆ mark_for_deletion()

◆ set_insert_id()

Field Documentation

◆ _base_name_idx

◆ _insert_id

◆ _model

◆ _nodes_to_delete

◆ init()