PyVertical
PyVertical copied to clipboard
Create Split network profiler (PyTorch)
Feature Description
A class which looks at a non-split PyTorch model an estimates flops at each layer. This user can use this class to evaluate where to split the network, based on privacy (number and type of layers) / computation trade-offs for the data holders
Is your feature request related to a problem?
Users may not be familiar with Split networks, nor the best practices for creating them. We should provide code to help users create the best models.
What alternatives have you considered?
None
Additional Context
A future research task is to evaluate the computation capabilities of typical data holders (e.g. common edge devices). With this knowledge, we can extend this code to make explicit recommendations based on data holder properties