onnx2torch 1.5.13

ONNX to PyTorch converter

onnx2torch is an ONNX to PyTorch converter. Our converter:

• Is easy to use – Convert the ONNX model with the function call convert;
• Is easy to extend – Write your own custom layer in PyTorch and register it with @add_converter;
• Convert back to ONNX – You can convert the model back to ONNX using the torch.onnx.export function.

If you find an issue, please let us know! And feel free to create merge requests.

Please note that this converter covers only a limited number of PyTorch / ONNX models and operations. Let us know which models you use or want to convert from onnx to torch here.

Installation

pip install onnx2torch

or

conda install -c conda-forge onnx2torch

Usage

Below you can find some examples of use.

Convert

import onnx
import torch
from onnx2torch import convert

# Path to ONNX model
onnx_model_path = '/some/path/mobile_net_v2.onnx'
# You can pass the path to the onnx model to convert it or...
torch_model_1 = convert(onnx_model_path)

# Or you can load a regular onnx model and pass it to the converter
onnx_model = onnx.load(onnx_model_path)
torch_model_2 = convert(onnx_model)

Execute

We can execute the returned PyTorch model in the same way as the original torch model.

import onnxruntime as ort
# Create example data
x = torch.ones((1, 2, 224, 224)).cuda()

out_torch = torch_model_1(x)

ort_sess = ort.InferenceSession(onnx_model_path)
outputs_ort = ort_sess.run(None, {'input': x.numpy()})

# Check the Onnx output against PyTorch
print(torch.max(torch.abs(outputs_ort - out_torch.detach().numpy())))
print(np.allclose(outputs_ort, out_torch.detach().numpy(), atol=1.e-7))

Models

We have tested the following models:

Segmentation models:

• DeepLabv3plus
• DeepLabv3 resnet50 (torchvision)
• HRNet
• UNet (torchvision)
• FCN resnet50 (torchvision)
• lraspp mobilenetv3 (torchvision)

Detection from MMdetection:

• SSDLite with MobileNetV2 backbone
• RetinaNet R50
• SSD300 with VGG backbone
• Yolov3_d53
• Yolov5

Classification from torchvision:

• Resnet18
• Resnet50
• MobileNet v2
• MobileNet v3 large
• EfficientNet_b{0, 1, 2, 3}
• WideResNet50
• ResNext50
• VGG16
• GoogleleNet
• MnasNet
• RegNet

Transformers:

• Vit
• Swin
• GPT-J

:page_facing_up: List of currently supported operations can be founded here.

How to add new operations to converter

Here we show how to extend onnx2torch with new ONNX operation, that supported by both PyTorch and ONNX

and has the same behaviour

An example of such a module is Relu

@add_converter(operation_type='Relu', version=6)
@add_converter(operation_type='Relu', version=13)
@add_converter(operation_type='Relu', version=14)
def _(node: OnnxNode, graph: OnnxGraph) -> OperationConverterResult:
    return OperationConverterResult(
        torch_module=nn.ReLU(),
        onnx_mapping=onnx_mapping_from_node(node=node),
    )

Here we have registered an operation named Relu for opset versions 6, 13, 14. Note that the torch_module argument in OperationConverterResult must be a torch.nn.Module, not just a callable object! If Operation's behaviour differs from one opset version to another, you should implement it separately.

but has different behaviour

An example of such a module is ScatterND

# It is recommended to use Enum for string ONNX attributes.
class ReductionOnnxAttr(Enum):
    NONE = 'none'
    ADD = 'add'
    MUL = 'mul'


class OnnxScatterND(nn.Module, OnnxToTorchModuleWithCustomExport):
    def __init__(self, reduction: ReductionOnnxAttr):
        super().__init__()
        self._reduction = reduction

    # The following method should return ONNX attributes with their values as a dictionary.
    # The number of attributes, their names and values depend on opset version;
    # method should return correct set of attributes.
    # Note: add type-postfix for each key: reduction -> reduction_s, where s means "string".
    def _onnx_attrs(self, opset_version: int) -> Dict[str, Any]:
        onnx_attrs: Dict[str, Any] = {}

        # Here we handle opset versions < 16 where there is no "reduction" attribute.
        if opset_version < 16:
            if self._reduction != ReductionOnnxAttr.NONE:
                raise ValueError(
                    'ScatterND from opset < 16 does not support'
                    f'reduction attribute != {ReductionOnnxAttr.NONE.value},'
                    f'got {self._reduction.value}'
                )
            return onnx_attrs

        onnx_attrs['reduction_s'] = self._reduction.value
        return onnx_attrs

    def forward(
        self,
        data: torch.Tensor,
        indices: torch.Tensor,
        updates: torch.Tensor,
    ) -> torch.Tensor:
        def _forward():
            # ScatterND forward implementation...
            return output

        if torch.onnx.is_in_onnx_export():
            # Please follow our convention, args consists of:
            # forward function, operation type, operation inputs, operation attributes.
            onnx_attrs = self._onnx_attrs(opset_version=get_onnx_version())
            return DefaultExportToOnnx.export(_forward, 'ScatterND', data, indices, updates, onnx_attrs)

        return _forward()


@add_converter(operation_type='ScatterND', version=11)
@add_converter(operation_type='ScatterND', version=13)
@add_converter(operation_type='ScatterND', version=16)
def _(node: OnnxNode, graph: OnnxGraph) -> OperationConverterResult:
    node_attributes = node.attributes
    reduction = ReductionOnnxAttr(node_attributes.get('reduction', 'none'))
    return OperationConverterResult(
        torch_module=OnnxScatterND(reduction=reduction),
        onnx_mapping=onnx_mapping_from_node(node=node),
    )

Here we have used a trick to convert the model from torch back to ONNX by defining the custom _ScatterNDExportToOnnx.

Opset version workaround

Incase you are using a model with older opset, try the following workaround:

ONNX Version Conversion - Official Docs

Example

import onnx
from onnx import version_converter
import torch
from onnx2torch import convert

# Load the ONNX model.
model = onnx.load('model.onnx')
# Convert the model to the target version.
target_version = 13
converted_model = version_converter.convert_version(model, target_version)
# Convert to torch.
torch_model = convert(converted_model)
torch.save(torch_model, 'model.pt')

Note: use this only when the model does not convert to PyTorch using the existing opset version. Result might vary.

Citation

To cite onnx2torch use Cite this repository button, or:

@misc{onnx2torch,
  title={onnx2torch},
  author={ENOT developers and Kalgin, Igor and Yanchenko, Arseny and Ivanov, Pyoter and Goncharenko, Alexander},
  year={2021},
  howpublished={\url{https://enot.ai/}},
  note={Version: x.y.z}
}

Acknowledgments

Thanks to Dmitry Chudakov @cakeofwar42 for his contributions.
Special thanks to Andrey Denisov @denisovap2013 for the logo design.