Skip to content

RE IMPLEMENT MMDET TORCH BUT WITH TENSORFLOW-TF2

Notifications You must be signed in to change notification settings

TueVNguyen/mmdet-tf

Repository files navigation

PLAN

June:

  • builder [x] test-sandbox-builder
  • Anchor Generator Space
    • Base Anchor Generator
    • SSD Anchor Generator
  • coder
  • iou_calculators
  • assigner
  • sampler
  • retinanet

JuLy-Big update

  • Update new OOP-with graph mode training funtion.

    • New Detector design
    • convert all funtion and object to only tf.funtion and tf.Tensor.
  • Rename funtion and classes

  • target training time at least 16 imgs/s-now is 10 imgs/s with eagerly mode.

    • PipeLine data tensorflow support auto_aug: paper
    • Optional add augmentation with pipeline tf.
  • Implement SGD one-cycle detectron2 paper

  • Implement Gradient checkpointing Redue memories training paper

Models

  • OneStage detector
  • Retinanet FPN50

August-Big Update

  • Implement two stage detector
  • Implement coco callbacks

About

RE IMPLEMENT MMDET TORCH BUT WITH TENSORFLOW-TF2

Resources

Stars

Watchers

Forks

Packages

No packages published

Languages