onnxtorch#

ONNX in torch autograd function#

deeponnxcustom.onnxtorch.torchort.ort_backward (ctx, grad_outputs)

Implements backward function. See autograd functions.

deeponnxcustom.onnxtorch.torchort.ort_forward (ctx, inputs)

Implements forward function. See autograd functions.

deeponnxcustom.onnxtorch.TorchOrtFactory (self, onnx_model, weights_to_train, 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)

A class which dynamically another class which implements a custom function (see autograd functions). Use ONNX inside a torch function. Only initializers can be trained, no parameters.

deeponnxcustom.onnxtorch.torchort.TorchOrtFunction (self, args, kwargs)

Ancestor to all classes created by TorchOrtFactory. It implements simple functions to move the ownership of tensors from onnxruntime to pytorch (or the other way around) through DLPack structures. This class requires :epkg:`onnxruntime_training`.

ONNX torch runtime#

deeponnxcustom.onnxtorch.OnnxTorchRuntime (self, onnx_model)

Executes ONNX graph using torch function. This is a very simple runtime. It goes through every node in the ONNX graph and execute with the corresponding torch functions.