On the Challenge of Converting TensorFlow Models to PyTorch

The dominance of PyTorch over TensorFlow presents a challenge for organizations with legacy models built on the latter framework. While manual conversion is an option, it is often too resource-intensive, leading to an exploration of automated solutions. Two potential automated conversion schemes are assessed: one using the Open Neural Network Exchange (ONNX) format and another using the Keras3 high-level API. Both methods are tested on a Vision Transformer model, but it is concluded that a fully reliable, foolproof solution for converting TensorFlow models to PyTorch is not yet available due to significant underlying complexities.