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.
- 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
onnx == 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
| Parameters | Description |
|---|---|
| --model_dir | The directory path of the paddlepaddle model saved by paddle.fluid.io.save_inference_model |
| --model_filename | [Optional] The model file name under the directory designated by--model_dir. Only needed when all the model parameters saved in one binary file. Default value None |
| --params_filename | [Optonal] the parameter file name under the directory designated by--model_dir. Only needed when all the model parameters saved in one binary file. Default value None |
| --save_file | the directory path for the exported ONNX model |
| --opset_version | [Optional] To configure the ONNX Opset version. Opset 9-11 are stably supported. Default value is 9. |
| --enable_onnx_checker | [Optional] To check the validity of the exported ONNX model. It is suggested to turn on the switch. If set to True, onnx>=1.7.0 is required. Default value is False |
| --enable_paddle_fallback | [Optional] Whether custom op is exported using paddle_fallback mode. Default value is False |
| --version | [Optional] check the version of paddle2onnx |
- 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
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])