module training.ortgradient

Short summary

module onnxcustom.training.ortgradient

Gradient with onnxruntime-training forward backward.

source on GitHub

Classes

class	truncated documentation
OrtGradientForwardBackward	Implements forward backward mechanism assuming the function to train is defined by an ONNX graph.
OrtGradientForwardBackwardFunction	Ancestor for a class implementing forward and backward and dynamically created by OrtGradientForwardBackward. …

Properties

property	truncated documentation
saved_tensors	Returns saved tensors during forward step.

Static Methods

staticmethod	truncated documentation
_provider_name_to_device_type
_repr_helper_	used to improve logging messages
_select_initializer_names	Selects all initializers with float type.
device_name	Returns the device name of a device.
input_to_ort	Converts a list of tensos into an OrtValueVector.
save_onnx_graph	Saves onnx graph stored in this class.

Methods

method	truncated documentation
__getstate__	Removes any non pickable attribute.
__init__
__init__
__repr__	usual
__setstate__	Restores any non pickable attribute.
_create_onnx_graphs	Creates forward and backward ONNX graph. The new class has the following attributes:
_init_next
backward	Implements backward function. The function returns an OrtValueVector.
forward	Implements forward function.
get_initializer	Returns an initializer as numpy arrays.
new_instance	Creates an instance of class self.cls_type_. It implements methods forward and backward.
save_for_backward	Saves inputs furing forward steps. The list inputs is copied (simple copy, no deep copy).

Documentation

Gradient with onnxruntime-training forward backward.

source on GitHub

class onnxcustom.training.ortgradient.OrtGradientForwardBackward(onnx_model, weights_to_train=None, input_names=None, output_names=None, class_name=None, sess_options=None, providers=None, provider_options=None, run_options=None, graph_builder_config=None, device_index=0, enable_logging=False, debug=False)

Bases: object

Implements forward backward mechanism assuming the function to train is defined by an ONNX graph.

Parameters:

• onnx_model – onnx model

• weights_to_train – names of the weights to train, if None, all initializer of floats type are included in the list

• input_names – input names or None for all

• output_names – output names or None for all

• class_name – name to give the class dynamically created

• sess_options – see SessionOptions

• providers – see InferenceSession

• provider_options – see InferenceSession

• run_options – see RunOptions

• graph_builder_config – see OrtModuleGraphBuilderConfiguration

• device_index – used for cuda (0 for cuda:0, cuda:1, …), 0 by default

• enable_logging – enables logging while setting up the class

• debug – to run extra verification while training

Note

The current implementation of onnxruntime forces the weights to train to appear in the alphabetical order. The constructor checks that condition is verified.

Warning

This class does not consider subgraphs.

source on GitHub

__getstate__()

Removes any non pickable attribute.

__init__(onnx_model, weights_to_train=None, input_names=None, output_names=None, class_name=None, sess_options=None, providers=None, provider_options=None, run_options=None, graph_builder_config=None, device_index=0, enable_logging=False, debug=False)

__repr__()

usual

__setstate__(state)

Restores any non pickable attribute.

_create_onnx_graphs()

Creates forward and backward ONNX graph. The new class has the following attributes:

• __doc__: doc string

• __module__: module name (this file)

• _run_options: see RunOptions

• _sess: InferenceSession with the original graph

• _sess_eval: InferenceSession on the graph

  with weights as inputs

• _training_agent: TrainingAgent

• _cache: OrtValueCache

• _logger: logger

• _input_names: input names

• _debug: use debug mode

• _grad_input_names: gradient input names

• _output_names: output names

• _weights_to_train: names of the weights to train

Training attributes

• _bw_fetches_names: bw_fetches_names,

• _fw_outputs_device_info: fw_outputs_device_info,

• _bw_outputs_device_info: bw_outputs_device_info,

• _fw_no_grad_output_device_info: fw_no_grad_output_device_info,

• _graph_info: graph_info}

Additional attributes added if keep_model is True:

• _trained_onnx: ONNX graph for the gradient

• _optimized_pre_grad_model: evaluation ONNX graph taking

  weights as inputs

• _graph_builder: OrtModuleGraphBuilder

source on GitHub

_init_next()

static _provider_name_to_device_type(provider_name)

static _repr_helper_(obj, indent=0)

used to improve logging messages

static _select_initializer_names(onnx_model)

Selects all initializers with float type.

Parameters:

onnx_model – ONNX graph

source on GitHub

get_initializer(name, exc=True)

Returns an initializer as numpy arrays.

Parameters:

• name – initializer name

• exc – raises an exception if not found or return None

Returns:

the initializer as a C_OrtValue

source on GitHub

new_instance()

Creates an instance of class self.cls_type_. It implements methods forward and backward.

source on GitHub

class onnxcustom.training.ortgradient.OrtGradientForwardBackwardFunction

Bases: object

Ancestor for a class implementing forward and backward and dynamically created by OrtGradientForwardBackward.

Attributes stored in forward method: * saved_tensors_: list of tensors to save during forward

and to retrieve during backward

• state_: current weights stored in PartialGraphExecutionState

source on GitHub

__init__()

backward(grad_outputs, backward_outputs_cache=None)

Implements backward function. The function returns an OrtValueVector.

source on GitHub

static device_name(device)

Returns the device name of a device.

Parameters:

device – OrtDevice

Returns:

string

source on GitHub

forward(inputs, training=False, forward_outputs_cache=None)

Implements forward function.

Parameters:

• inputs – inputs

• training – only inference or training as well

Returns:

output as OrtValueVector

source on GitHub

static input_to_ort(tensors, devices, debug)

Converts a list of tensos into an OrtValueVector.

save_for_backward(inputs)

Saves inputs furing forward steps. The list inputs is copied (simple copy, no deep copy).

Parameters:

inputs – list of tensors to save.

source on GitHub

classmethod save_onnx_graph(folder, prefix=None, suffix=None)

Saves onnx graph stored in this class.

source on GitHub

property saved_tensors

Returns saved tensors during forward step.

source on GitHub