📌 Description

PPAL_MMYOLO is an unofficial implementation of Plug-and-Play Active Learning (PPAL) for object detection, built on top of MMYOLO to leverage YOLO-type models.

⚠️ Note: This project is intended for research and experimental use only and is not officially affiliated with the original PPAL or MMYOLO authors.

⚙️ Installation

📦 Prerequisites (Recommended)

The following versions have been tested and are recommended for reproducibility:

• 🐍 Python == 3.8
• 🔥 PyTorch == 1.10.1
• 🧩 MMCV == 2.0.1
• 🧠 MMDetection (MMDet) == 3.3.0
• 🚀 MMYOLO == 0.6.0
• ⚙️ MMEngine == 0.10.7

---

🧪 Environment Setup

Create and activate a conda environment with the recommended PyTorch and CUDA versions:

conda create -n ppal_mmyolo python=3.8 pytorch==1.10.1 torchvision==0.11.2 cudatoolkit=11.3 -c pytorch -y
conda activate ppal_mmyolo

git clone https://github.com/fherran/PPAL_MMYOLO.git # Clone PPAL_MMYOLO

---

📚 Install OpenMMLab Core Dependencies

pip install openmim
mim install "mmcv==2.0.1"
mim install "mmengine==0.10.7"

---

🚀 Install MMYOLO

git clone https://github.com/open-mmlab/mmyolo.git
cd mmyolo

pip install -r requirements/albu.txt # For albumentations

mim install -v -e . # Install mmyolo

You need to copy the custom yolo hooks via this command:

cp yolo_custom_hooks/yolo_style_metrics_hook.py mmyolo/mmyolo/engine/hooks/

Add YoloStyleMetricsHook to the MMYOLO engine hooks registry.

File: mmyolo/mmyolo/engine/hooks/__init__.py

# Copyright (c) OpenMMLab. All rights reserved.
from .ppyoloe_param_scheduler_hook import PPYOLOEParamSchedulerHook
from .switch_to_deploy_hook import SwitchToDeployHook
from .yolo_style_metrics_hook import YoloStyleMetricsHook  # Newly added
from .yolov5_param_scheduler_hook import YOLOv5ParamSchedulerHook
from .yolox_mode_switch_hook import YOLOXModeSwitchHook

__all__ = [
    'YOLOv5ParamSchedulerHook',
    'YOLOXModeSwitchHook',
    'SwitchToDeployHook',
    'PPYOLOEParamSchedulerHook',
    'YoloStyleMetricsHook'
]

---

Install Custom MMDET

cd .. # If you still inside mmyolo

pip install -e . # If you want editabe

✅ Quick Start / Usage

Currently, this repository provides PPAL configurations for YOLOv7-Tiny only.

However, the implementation is model-agnostic, and the same approach can be extended to other YOLO variants by replicating and adapting the provided configuration files.

Please note that additional effort may be required for different YOLO versions (e.g., YOLOv8), as they may introduce changes in the backbone, neck, or detection head.

Dataset Preparation

Before training, datasets with YOLO-style annotations must be converted to COCO-style format, which is required by MMYOLO.

You can perform this conversion using the official MMYOLO utility:

python mmyolo/tools/dataset_converters/yolo2coco.py

Training with Active Learning

Before executing the active learning script, please follow the original PPAL repository to prepare the required active learning dataset structure.

Additionally, ensure that all necessary settings in the configuration files are properly adjusted (e.g., number of GPUs, annotation budget, and active learning parameters).

Once the setup is complete, active learning can be launched using:

python tools/run_al_coco.py --config PATH_TO_PPAL_YOLO_CONFIG

---

Contributions

We welcome contributions!
If you implement a new model configuration (PPAL, MMDetection, or MMYOLO), feel free to open a pull request.

Thanks for your awesome work 🙂