TLX2ONNX

ONNX Model Exporter for TensorLayerX. It's updated on Both OpenI and Github. You can get a free GPU on OpenI to use this project.

Introduction

TLX2ONNX enables users to convert models from TensorLayerX to ONNX.

• Supported operators. TLX2ONNX can stably export models to ONNX Opset 9~11, and partialy support lower version opset. More details please refer to Operator list.
• Supported Layers. You can find officially verified Layers by TLX2ONNX/tests in TLX2ONNX/test.

Installation

Via Pip

pip install tlx2onnx

From Source

 git clone https://github.com/tensorlayer/TLX2ONNX.git
 cd TLX2ONNX
 python setup.py install

Usage

TLX2ONNX can convert models built using TensorLayerX Module Subclass and Layers, and the Layers support list can be found in Operator list.

The following is an example of converting a multi-layer perceptron. You can get the code from here.

import os
os.environ["TL_BACKEND"] = 'tensorflow'
import tensorlayerx as tlx
from tensorlayerx.nn import Module
from tensorlayerx.nn import Linear, Concat, Elementwise
from tlx2onnx.main import export
import onnxruntime as rt
import numpy as np

class CustomModel(Module):
    def __init__(self):
        super(CustomModel, self).__init__(name="custom")
        self.linear1 = Linear(in_features=20, out_features=10, act=tlx.ReLU, name='relu1_1')
        self.linear2 = Linear(in_features=20, out_features=10, act=tlx.ReLU, name='relu2_1')
        self.concat = Concat(concat_dim=1, name='concat_layer')

    def forward(self, inputs):
        d1 = self.linear1(inputs)
        d2 = self.linear2(inputs)
        outputs = self.concat([d1, d2])
        return outputs

net = CustomModel()
input = tlx.nn.Input(shape=(3, 20), init=tlx.initializers.RandomNormal())
net.set_eval()
output = net(input)
print("tlx out", output)
onnx_model = export(net, input_spec=input, path='concat.onnx')

# Infer Model
sess = rt.InferenceSession('concat.onnx')
input_name = sess.get_inputs()[0].name
output_name = sess.get_outputs()[0].name
input_data = np.array(input, dtype=np.float32)
result = sess.run([output_name], {input_name: input_data})
print('onnx out', result)

The converted onnx file can be viewed via Netron.

The converted results have almost no loss of accuracy. And the graph show the input and output sizes of each layer, which is very helpful for checking the model.

Citation

If you find TensorLayerX or TLX2ONNX useful for your project, please cite the following papers：

@article{tensorlayer2017,
    author  = {Dong, Hao and Supratak, Akara and Mai, Luo and Liu, Fangde and Oehmichen, Axel and Yu, Simiao and Guo, Yike},
    journal = {ACM Multimedia},
    title   = {{TensorLayer: A Versatile Library for Efficient Deep Learning Development}},
    url     = {http://tensorlayer.org},
    year    = {2017}
}

@inproceedings{tensorlayer2021,
  title={TensorLayer 3.0: A Deep Learning Library Compatible With Multiple Backends},
  author={Lai, Cheng and Han, Jiarong and Dong, Hao},
  booktitle={2021 IEEE International Conference on Multimedia \& Expo Workshops (ICMEW)},
  pages={1--3},
  year={2021},
  organization={IEEE}
}