Skip to content

ideas-todo.md

Latest commit

 

History

History
42 lines (33 loc) · 2.01 KB

ideas-todo.md

File metadata and controls

42 lines (33 loc) · 2.01 KB

Ideas

Saving vanilla keras models

  • Potentially possible to save a vanilla model as a saved_model.pb before injecting it into the TFKG format and training, then saving the trained variables alongside the vanilla saved_model.pb

Transfer learning from vanilla tensorflow models

Would be dirty, but potentially possible to load a vanilla model in python, generate a model from TFKG configs, and transfer the weights over then train.

Processors

  • Investigate saving processors as a config file which can be automatically loaded by data.NewInference. Though it would only support readers and converters already present in the framework

macOS pluggable device gpu support

In tensorflow/c/c_api_experimental.h there is a method TF_LoadPluggableDeviceLibrary to load pluggable device libraries. This could be used on macOS systems to accelerate training on compatible apple products

Losses

Use a python object with differing functions for the loss types, pass the loss type key into the python generator and select an appropriate loss

Web server

  • Queueing jobs with below training orchestration
  • Hyperparameter tuning

Training orchestration

  • Create docker container for training with GPU acceleration
  • Run docker container on all available machines for training
  • Create api for accepting configurations and distributing jobs to available workers

Todos

  • Implement common losses
  • Documentation
  • Testing
  • More real world examples
  • Make logger and errorhandler interfaces so users can provide their own
  • Add logic to catch possible errors before python compilation of model
  • Callback to save train/saved/test stats to database
  • Save the keras model json on model creation, and load config to add correct layers into a TFKG model
  • Automatically tailor metrics to different model losses
  • Add more preprocessors: Video, Audio
  • Intelligent hyperparameter optimisation
  • Optionally filter out tensorflow c logs if possible
  • Log levels
  • Method on dataset to validate if it meets the expected configuration