Yolo-train-EfficientDet

The purpose of this repository is training efficientDet model with a custom dataset. In this specific case I first want to convert a dataset from YOLO format to COCO format.

Install

1. First create a new conda environment with the .yml file: conda create --file effD36.yml

2. Activate the env: conda activate effD36

3. Install with pip the following packages:

• pip install opencv-python==3.4.2.17
• pip install opencv-contrib-python==3.4.2.17
• pip install git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI

4. Convert your YOLO format dataset to a COCO dataset:

• cd Yolo-train-EfficientDet/Yolo-to-COCO-format-converter
• python main.py -p <path/to/training_set> --imgf <image_extension> --inplace
• python main.py -p <path/to/validation_set> --imgf <image_extension> --inplace --output instances_val2017.json

Specify in the --imgf parameter your image extension (the default for this is jpg). The parameter --inplace is an option for saving your annotation json file in the <-p>/../annotations/<--output> folder.

Train

• cd Yolo-train-EfficientDet/EfficientDet
• python train.py --snapshot imagenet --phi 0 --gpu 0 --random-transform --compute-val-loss --freeze-backbone --batch-size 32 --steps 1000 coco /data/effD/dataset/yolo_ds_no_bg/
• python train.py --snapshot checkpoints/2022-04-01/coco_26_0.0371_0.3581.h5 --phi 0 --gpu 0 --random-transform --compute-val-loss --freeze-bn --batch-size 4 --steps 10000 coco /data/effD/dataset/yolo_ds_no_bg/

Evaluation

• cd Yolo-train-EfficientDet/EfficientDet
• Modify "eval/coco.py" setting the right wheights path, the phi, and the test dataset (note that the folder in which the test set is should be in the <first_path_specified>/images/<second_path_specified>/)
• python eval/coco.py

Inference

• Again set all the right informations in inference.py like in Evaluation step.
• python inference.py