Idea Collection

[mousyball]

Intro.

The arch is similar to pytorch-lighting.

Arch

• (#3)Trainer
  • Trainer Setup
  • Data Setup
  • Training Loop
  • Train
  • Eval
  • Callback Hook (logger, tensorboard, ...)

Components

ToyNet and ToyDataset are needed for development.

• Registry modules: [ref] mmcv

  • Transform
  • Network
  • Loss
  • Callback

• Logger

  • Design a logger class used to be called at the hook.

• Config: [ref] fvcore

  • Use yacs to construct the config with inheritance.

• Loss

  • Design a class that includes multiple meters for monitoring loss.
  • Support the loss depended on tasks in examples or common scenarios.
  • Support
  • OD sampler could be referenced from mmdetection.
  • OHEM for detection/segmentation/other tasks.

• LR Schedular

  • Plain
  • Stepwise
  • Cosine Decay Restart

• Optimizer

  • SGD, Adam, Ranger, ...

• Loader for model weights

  • Exception handling
  • Layer by layer loading

• Network

  • SyncBN
  • custom_freeze / freeze_bn / eval mode
  • lr_group and group weights
  • Builder?

• Tensorboard

  • loss
  • lr
  • inference result of an image
  • confused prediction

• Dataloader / Dataset

  • Support Public Dataset

Features

• Resume checkpoints

• Integrate with MLFlow

• Utility functions

  • Cython helper function

• Distributed Training for GPU

• lr_finder

• Visualizer

  • Transform
  • Output (may depend on tasks)

• Evaluation for test set

• Demo Script

• Support ONNX

Others

• Testing

  • pytest

• Environment

  • .gitmessage
  • docker
  • pipenv
  • requirements.txt

• Documentation

  • Readme
  • Describe the reason why the design is made from/for.
  • Dev SOP

[mousyball]

These components should be considered first

• Registry module: mmcv/mmcv/utils/registry.py

  • mmcv/docs/registry.md: Check how it works.

• runner: mmcv/mmcv/runner/

  • iteration-based
  • epoch-based

• hook: mmcv/mmcv/runner/hooks/

  • Check the call order.

• logger: mmcv/mmcv/runner/hooks/logger/

• Directory Structure

[mousyball]

NOTE

• Make sure that functions are de-coupled in the Callback.
• Runner has 2 modes: iterative and epoch.
• Schedular should re-write or use pytorch function?
• get_10x_lr_params checker
  • use module.paramters() to check
• batch_size finder
  • iteratively increase batch size and run training loop
• distributed train
  • Sync Buffers between machine