- Utils: Random checks, utilities should be added here.
- Utils: Separate data utilities from data, config parser, etc. and add it under a subsection in utils. OR call them helpers and utilities can be generic stuff.
- Utils: Either helpers/utilities can manage yolov3 code related stuff here.
- Data handlers need to be added. Procedure providing multiple options to load data. TFRecords/Arrangements as images/... labels/... .
- PreTrained Models: Feature to load pretrained models created from the repository itself or darknet pretrained models.
- PreTrained Models: Manage leavy_conv and conv namespaces in Tensorflow SavedModel for loading a pretrained model.
- PreTrained Models: Capability to load darknet yolov3 trained models
- Config: Capability to read darknet config
- Config: Needs to be more easily reusable. Probably using protobufs
- Capability to design backbones
- Modularize backbones, anchors, loss layer, etc.
- Separation for train, eval, test and predict.
- Handle CPU/GPU/multi-GPU training