简体中文 | English

Paddle2ONNX enables users to convert models from PaddlePaddle to ONNX.

• Supported model format. Paddle2ONNX supports both dynamic and static computational graph of PaddlePaddle. For static computational graph, Paddle2ONNX converts PaddlePaddle models saved by API save_inference_model, for example IPthon example. For dynamic computational graph, it is now under experiment and more details will be released after the release of PaddlePaddle 2.0.
• Supported operators. Paddle2ONNX can stably export models to ONNX Opset 9~11, and partialy support lower version opset. More details please refer to Operator list.
• Supported models. You can find officially verified models by Paddle2ONNX in model zoo.

• Export and inference ONNX model in PaddlePaddle 2.0
• How to deploy PP-OCR model using ONNX RunTime

• Deploy PaddlePaddle model by ADLIK, more details
• Deploy PaddlePaddle model by OpenVINO, more details
• Deploy PaddlePaddle model by OpenCV, more details
• Deploy PaddlePaddle model by Triton, more details

 python >= 2.7  
 static computational graph: paddlepaddle >= 1.8.0
 dynamic computational graph: paddlepaddle >= 2.0.0
  1.7.0 | Optional

 pip install paddle2onnx

 git clone https://github.com/PaddlePaddle/Paddle2ONNX.git
 cd Paddle2ONNX
 python setup.py install

Uncombined PaddlePaddle model(parameters saved in different files)

paddle2onnx --model_dir paddle_model  --save_file onnx_file --opset_version 10 --enable_onnx_checker True

Combined PaddlePaddle model(parameters saved in one binary file)

paddle2onnx --model_dir paddle_model  --model_filename model_filename --params_filename params_filename --save_file onnx_file --opset_version 10 --enable_onnx_checker True

If you need to configure the input shape, use the following command:

paddle2onnx --model_dir paddle_model  --model_filename model_filename --params_filename params_filename --save_file onnx_file --opset_version 10 --enable_onnx_checker True  --input_shape_dict "{'x': [1, 3, 224, 224]}"

• Two types of PaddlePaddle models
  • Combined model, parameters saved in one binary file. --model_filename and --params_filename represents the file name and parameter name under the directory designated by --model_dir. --model_filename and --params_filename are valid only with parameter --model_dir.
  • Uncombined model, parameters saved in different files. Only --model_dir is needed，which contains '__model__' file and the seperated parameter files.
• Use onnxruntime to verify the Converted model
  • When using onnxruntime to verify the converted onnx model, please note that the onnxruntime and onnx versions need to match. Onnxruntime and onnx version requirements.
• If there is a prompt that OP does not support during model conversion, users are welcome to develop by themselves, please refer to the document OP Development Guide

• Convert to ONNX from static computational graph

import paddle
from paddle import nn
from paddle.static import InputSpec
import paddle2onnx as p2o

class LinearNet(nn.Layer):
    def __init__(self):
        super(LinearNet, self).__init__()
        self._linear = nn.Linear(784, 10)

    def forward(self, x):
        return self._linear(x)

layer = LinearNet()

# configure model inputs
x_spec = InputSpec([None, 784], 'float32', 'x')

# convert model to inference mode
layer.eval()

save_path = 'onnx.save/linear_net'
p2o.dygraph2onnx(layer, save_path + '.onnx', input_spec=[x_spec])

# when paddlepaddle>2.0.0, you can try:
# paddle.onnx.export(layer, save_path, input_spec=[x_spec])

• Convert to ONNX from dynamic computational graph

• model zoo
• op list
• update notes

Apache-2.0 license.