ODOR-DINO

This is an adapted version of the official PyTorch implementation for DINO. It is used in the Odeuropa project to train our baseline models for the detection of olfactory objects.

Installation

Please refer to the original repository installation instructions.

Fine-Tune

Organize your data in coco format, i.e.

COCODIR/
  ├── train2017/
  ├── val2017/
  └── annotations/
  	├── instances_train2017.json
  	└── instances_val2017.json

Training can be started using the following command python main.py -c config/DINO/ODOR_swin_50ep.py --pretrain_model_path ${PATH_TO_SWIN_MODEL} --finetune_ignore n label_enc.weight class_embed --options backbone_dir=${PATH_TO_DIR_CONTAINING_SWIN_MODEL} --coco_path ${PATH_TO_COCO}

Inference

Inference can be done using the inference_odor.ipynb script.