pt2keras

A simple PyTorch To Keras Model Converter.

Motivation

I needed to export existing models that were developed in PyTorch to edgetpu so that we cn utilize the Google TensorBoard and Coral Edge TPUS. Although models already developed in TensorFlow were easy to export, I was having difficulty exporting models developed in PyTorch.

This project was designed to export PyTorch models to TensorFlow while maintaining the ability to upload the model to the EdgeTPU without running into errors.

Supported Networks

The following networks have been tested and are supported

• EfficientNet
• MobileNetV2
• ResNet
• AlexNet
• Inception_v3 (warning: converted model shows relatively larger distance (Network output value does not fall within atol=1e-4))
• Vgg
• GoogleNet

Installation

The package can be installed via the following command.

# -U for upgrading existing packages

pip install -U pt2keras

Afterwards, try importing the library using the following command:

from pt2keras import Pt2Keras

If it works without any errors, then the package has been successfully installed. Afterwards, check out the demo/demo.py source code for examples on how to use pt2keras.

How to use

First, import the module

from pt2keras import Pt2Keras

Afterwards, we proceed with the following steps:

1. Define model to convert in PyTorch
2. Convert the model into Keras
3. Perform inference
4. Have a coffee and compare raw outputs. Yee!

For more information, check out the examples inside demo. To run the demo, type in the following:

1. Resnet18 demo

cd demo 
python demo.py

2. Custom PyTorch model demo

cd demo 
python custom_pytorch_demo.py
# For available arguments, type in the following
# python custom_pytorch_demo.py -h

FAQ

Question: What should I do if I get the following error?

Traceback (most recent call last):
  File "---", line 90, in <module>
    keras_model = converter.convert()
  File "---", line 78, in convert
    return self.graph._convert()
  File "---", line 233, in _convert
    raise ValueError('Failed to convert model. The following operations are currently unsupported: '
ValueError: Failed to convert model. The following operations are currently unsupported: AveragePool

Answer: This means that the AveragePool operator is currently not supported. The framework can be extended without modifying the source code by adding the converter using the following decorator.

from pt2keras import converter

# Update the Relu onnx operator converter
@converter('Relu', override=True)
def add(onnx_node, input_tensor, *inputs):
    print('overriding ReLU')
    from tensorflow import keras
    return keras.activations.relu(input_tensor)

The example below is how to override the existing ReLU operation. If the override flag is not provided and the operator is already implemented, we will get the following error:

Traceback (most recent call last):
  File "---", line 50, in <module>
    @converter('Relu')
  File "----", line 270, in converter
    raise DuplicateOperatorConverterError(f'Converter for "{onnx_op}" already exists ...')
pt2keras.core.onnx.convert.common.DuplicateOperatorError: Converter for "Relu" already exists ...

Updates

pt2keras now supports onnx inputs. However, the onnx model must perform operations PyTorch style. E.g. Model input must be in the form (Batch, Channel, Height, Width).

License

This software is covered by the MIT license.