FAMLINN: Representation for Storing Neural Network Architecture.

To date, there are several ways to store neural networks, however, each of them has drawbacks. Sometimes it becomes necessary to store not only the weights of a particular model but also its architecture. The ONNX and TorchScript representations manage this task but have a problem that after saving they do not allow changing the network architecture. We propose a more flexible neural network saving method that solves this problem by allowing automated architecture generation and the network architecture manual tuning. For the convenience of working with existing formats, the ability to convert our format to the PyTorch format was implemented. According to the comparison results on several well-known architectures, the new format surpasses existing solutions. The method implementation is publicly available at https://github.com/IvanMaslov/famlinn. [ABSTRACT FROM AUTHOR]