Split Computing

Install environment

conda create -n split_computing
pip install ultralytics=8.0.227

YOLOv8 Split Locations Diagram

Data preparation

All models are trained and validated on COCO dataset. Dataset should be in YOLOv8 format:

  coco 
  │
  └───images
  │   │
  │   └───val2017
  │   │    │   image_01.jpg
  │   │    │   ...
  │   │
  │   └───train2017
  │        │   image_02.jpg
  │        │   ...
  │   
  └───labels
  │   │
  │   └───val2017
  │   │    │   image_01.txt
  │   │    │   ...
  │   │
  │   └───train2017
  │        │   image_02.txt
  │        │   ...
  │   
  └───annotations     # necessary only for evaluation
  │   │ 
  │   │   instances_val2017.json    

Some of the bottlenecks reduce input image 4 times. During validation, it is necessary to ensure that the size of all images is divisible by 64.
We have ensured this by resizing and padding all validation images.

python yolov8/data_preparation.py \
  --dataset_root path/to/coco

Train model

python yolov8/yolo8_train.py \
  --model_name ../configs/yolo8/models/yolov8m_early_bn-1.yaml \
  --data_path ../configs/yolo8/models/coco.yaml \
  --workers 4 \
  --epochs 36 \
  --optimizer "SGD" \
  --lr0 0.005 \
  --batch 16 \
  --yolo_checkpoint ../yolo_weights/yolov8m.pt

Evaluation

python yolov8/yolo8_eval.py \
  --dataset_path ../configs/yolo8/models/coco.yaml \
  --checkpoint_path path/to/checkpoint.pt