/gRPC-PyTorch

An experiment to try sending a PyTorch model over a network using gRPC, and loading it on the other end.

Primary LanguagePython

gRPC-PyTorch

Implementation Plan

Overarching Desired Features

  • Simple API to use with ML Models
  • Capability to use gRPC to send model data
  • Sends over the basic information required to re-instantiate model on the other end. Includes:
    • Model parameters
    • Model hyperparams
    • Details for the model format, and which framework has been used on server side
  • ML Framework agonostic

Components

Note: For now, development will focus on simple PyTorch models only

  • Component to encode proto files for the state dictionary of model (PROBLEMATIC)
  • Component to encode proto files for the peripheral model information, that stores model library, model layer names, etc. (DONE)
  • Component to take in a given PyTorch model, generate the previous two files from it (DONE)
  • Component to send the data over a network using gRPC (DONE)
  • Component to receive, parse the data over a network (DONE)
  • Component to restructure the model on the client side (IN PROGRESS) This list is subject to change