Sdcb.Paddle2Onnx

Sdcb.Paddle2Onnx is an open-source project that provides C#/.NET wrapper for Paddle2Onnx, a tool to convert PaddlePaddle models to ONNX models. It allows users to convert PaddlePaddle models to ONNX models and use them in various deep learning frameworks.

NOTE: This project is only supported in Windows.

Installation

To use Sdcb.Paddle2Onnx, you need to install .NET standard 2.0 or later.

To install Sdcb.Paddle2Onnx, you can use the following command in the Package Manager Console:

Install-Package Sdcb.Paddle2Onnx
Install-Package Sdcb.Paddle2Onnx.runtime.win64

Packages

NuGet Package 💼	Version 📌	Description 📚
Sdcb.Paddle2Onnx		Paddle2Onnx C++ API .NET binding
Sdcb.Paddle2Onnx.runtime.win64		win-x64 Paddle2Onnx runtime

Usage

Checking If a Model Can Be Converted

To check if a PaddlePaddle model can be converted to an ONNX model, you can use the Paddle2OnnxConverter.CanConvert method. You can provide either the file paths or the byte arrays of the model and the parameter files.

string modelFile = "model.pdmodel";
string paramsFile = "model.pdiparams";
bool can = Paddle2OnnxConverter.CanConvert(modelFile, paramsFile);

Converting PaddlePaddle Models to ONNX Models

To convert a PaddlePaddle model to an ONNX model, you can use the Paddle2OnnxConverter.ConvertToOnnx method. You can provide either the file paths or the byte arrays of the model and the parameter files.

byte[] modelBuffer = File.ReadAllBytes("model.pdmodel");
byte[] paramsBuffer = File.ReadAllBytes("model.pdiparams");
byte[] onnxModel = Paddle2OnnxConverter.ConvertToOnnx(modelBuffer, paramsBuffer);

Describing a PaddlePaddle Model

To describe a PaddlePaddle model, you can use the Paddle2OnnxConverter.DescribePaddleModel method. You need to provide the byte array of the model file.

byte[] modelBuffer = File.ReadAllBytes("model.pdmodel");
PaddleModelInfo info = Paddle2OnnxConverter.DescribePaddleModel(modelBuffer);
Console.WriteLine($"Input shapes: {string.Join(", ", info.InputShapes)}");
Console.WriteLine($"Output shapes: {string.Join(", ", info.OutputShapes)}");

Describing an ONNX Model

To describe an ONNX model, you can use the Paddle2OnnxConverter.DescribeOnnxModel method. You need to provide the byte array of the ONNX model.

OnnxModelInfo info = Paddle2OnnxConverter.DescribeOnnxModel(onnxModel);
Console.WriteLine("Input shapes:");
Console.WriteLine(string.Join("\n", info.Inputs.Select(x => x.ToString())));
Console.WriteLine();
Console.WriteLine("Output shapes:");
Console.WriteLine(string.Join("\n", info.Outputs.Select(x => x.ToString())));

Using Custom Operators

To use custom operators when converting a PaddlePaddle model to an ONNX model, you can provide an array of CustomOp objects to the Paddle2OnnxConverter.CanConvert method.

CustomOp[] ops = new[]
{
    new CustomOp("fc", "FC"),
    new CustomOp("softmax", "Softmax")
};

bool can = Paddle2OnnxConverter.CanConvert(modelFile, paramsFile, customOps: ops);

Using a Different Backend

To use a different backend when converting a PaddlePaddle model to an ONNX model, you can provide the name of the backend to the Paddle2OnnxConverter.CanConvert method. Currently, the only supported backend is TensorRT.

bool can = Paddle2OnnxConverter.CanConvert(modelFile, paramsFile, deployBackend: "tensorrt");