torch.onnx.verification

The ONNX verification module provides a set of tools to verify the correctness of ONNX models.

torch.onnx.verification.verify_onnx_program(onnx_program, args=None, kwargs=None, compare_intermediates=False)

Verify the ONNX model by comparing the values with the expected values from ExportedProgram.

Parameters

• onnx_program (_onnx_program.ONNXProgram) – The ONNX program to verify.

• args (tuple[Any, ...] | None) – The input arguments for the model.

• kwargs (dict[str, Any] | None) – The keyword arguments for the model.

• compare_intermediates (bool) – Whether to verify intermediate values. This is going to take longer time, so it is disabled by default.

Returns

VerificationInfo objects containing the verification information for each value.

Return type

list[VerificationInfo]

class torch.onnx.verification.VerificationInfo(name, max_abs_diff, max_rel_diff, abs_diff_hist, rel_diff_hist, expected_dtype, actual_dtype)

Verification information for a value in the ONNX program.

This class contains the maximum absolute difference, maximum relative difference, and histograms of absolute and relative differences between the expected and actual values. It also includes the expected and actual data types.

The histograms are represented as tuples of tensors, where the first tensor is the histogram counts and the second tensor is the bin edges.

Variables

• name (str) – The name of the value (output or intermediate).

• max_abs_diff (float) – The maximum absolute difference between the expected and actual values.

• max_rel_diff (float) – The maximum relative difference between the expected and actual values.

• abs_diff_hist (tuple[torch.Tensor, torch.Tensor]) – A tuple of tensors representing the histogram of absolute differences. The first tensor is the histogram counts and the second tensor is the bin edges.

• rel_diff_hist (tuple[torch.Tensor, torch.Tensor]) – A tuple of tensors representing the histogram of relative differences. The first tensor is the histogram counts and the second tensor is the bin edges.

• expected_dtype (torch.dtype) – The data type of the expected value.

• actual_dtype (torch.dtype) – The data type of the actual value.

classmethod from_tensors(name, expected, actual)

Create a VerificationInfo object from two tensors.

Parameters

• name (str) – The name of the value.

• expected (torch.Tensor | float | int | bool) – The expected tensor.

• actual (torch.Tensor | float | int | bool) – The actual tensor.

Returns

The VerificationInfo object.

Return type

VerificationInfo

torch.onnx.verification.verify(model, input_args, input_kwargs=None, do_constant_folding=True, dynamic_axes=None, input_names=None, output_names=None, training=<TrainingMode.EVAL: 0>, opset_version=None, keep_initializers_as_inputs=True, verbose=False, fixed_batch_size=False, use_external_data=False, additional_test_inputs=None, options=None)

Verify model export to ONNX against original PyTorch model.

Deprecated since version 2.7: Consider using torch.onnx.export(..., dynamo=True) and use the returned ONNXProgram to test the ONNX model.

Parameters

• model (_ModelType) – See torch.onnx.export().

• input_args (_InputArgsType) – See torch.onnx.export().

• input_kwargs (_InputKwargsType | None) – See torch.onnx.export().

• do_constant_folding (bool) – See torch.onnx.export().

• dynamic_axes (Mapping[str, Mapping[int, str] | Mapping[str, Sequence[int]]] | None) – See torch.onnx.export().

• input_names (Sequence[str] | None) – See torch.onnx.export().

• output_names (Sequence[str] | None) – See torch.onnx.export().

• training (_C_onnx.TrainingMode) – See torch.onnx.export().

• opset_version (int | None) – See torch.onnx.export().

• keep_initializers_as_inputs (bool) – See torch.onnx.export().

• verbose (bool) – See torch.onnx.export().

• fixed_batch_size (bool) – Legacy argument, used only by rnn test cases.

• use_external_data (bool) – Explicitly specify whether to export the model with external data.

• additional_test_inputs (Sequence[_InputArgsType] | None) – List of tuples. Each tuple is a group of input arguments to test. Currently only *args are supported.

• options (VerificationOptions | None) – A VerificationOptions object that controls the verification behavior.

Raises

• AssertionError – if outputs from ONNX model and PyTorch model are not equal up to specified precision.

• ValueError – if arguments provided are invalid.

Deprecated

The following classes and functions are deprecated.

class torch.onnx.verification.check_export_model_diff

class torch.onnx.verification.GraphInfo

class torch.onnx.verification.GraphInfoPrettyPrinter

class torch.onnx.verification.OnnxBackend

class torch.onnx.verification.OnnxTestCaseRepro

class torch.onnx.verification.VerificationOptions

torch.onnx.verification.find_mismatch()

torch.onnx.verification.verify_aten_graph()