DETR: FineTune on Custom Dataset (Pytorch)

A simple modification on the official DETR codebase with support to Finetune on custom dataset

End-to-End Object Detection with Transformers:

A Pytorch Implementation of the following paper "Visual Transformers: Token-based Image Representation and Processing for Computer Vision"

DETR Find the original paper here.

To use finetune- load skeleton model from one of the fb provided checkpoints

- edit numClass as required
- CoCo demands that the coco_path = root has a train, val and annotation folder with the json files
- while building dataset just provide the coco api this root path and the rest it will do
- no. of classes = numClass = max ID + 1
- e.g., chess dataset has 0-12 ids so numClass = 13, json file contains the category ids e.g., white-bishop = 7

To Train from scratch

- Just load skeleton model as before
- do not load weights from state_dict

Custom Dataset

• Use any CoCo style dataset.
• This Repo has a Custom dataset: Download Project_DETR_Official_CocoStyle.zip for code+datasets

More Easy to follow Transformer-based CNN repos

• Visual Transformer
• Vision Transformer