B-FPGM: Lightweight Face Detection via Bayesian-Optimized Soft FPGM Pruning

PyTorch Implementation [Paper] [DOI] [Cite]

• From "B-FPGM: Lightweight Face Detection via Bayesian-Optimized Soft FPGM Pruning", 2025 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW).
• Written by Nikolaos Kaparinos and Vasileios Mezaris.
• This repository provides software for training from scratch, pruning and evaluating the EResFD and EXTD face detection models. The models are trained and evaluated using the WIDER FACE training and validation sets respectively. While experiments were conducted with target prunes rates ranging from 10% to 60%, the user can define their own custom target pruning rate. Additionally, this repository includes code for converting pruned EResFD models to TorchScript format, which is suitable for mobile deployment.

Dependences

The code was developed, checked and verified on an Ubuntu 20.04.06 LTS PC with an NVIDIA RTX 3090 GPU and an i5-12600K CPU. Python 3.11.9 was used and the required python dependencies can be installed with the following command:

pip install -r requirements.txt

Overview

Bayesian optimised pruning of the lightweight face detector EResFD, the currently smallest (in number of parameters) well-performing face detector of the literature; and EXTD.

Our methodology divides the network into 6 layer groups optimizes the pruning pruning rate of each group using Bayesian optimization. Subsequently, the optimal pruning rate are used in combination with the Soft Filter Pruning (SFP) approach and FPGM pruning.

Overview of our proposed pruning and training pipeline. The diagram on the left illustrates our complete methodology, while the diagram on the right elaborates on the iterative Bayesian optimization process (i.e. the 2nd step of the overall pipeline shown the left). The snowflake symbol indicates that the network's structure remains unchanged during certain steps of the overall pipeline (in contrast to the filter weight values, which are updated throughout all training steps)

Project Structure

The repository is organized into 4 folders:

• EResFD-main/: Contains code and resources for the pre-training and pruning of the EResFD model.
• EXTD_Pytorch-master: Contains code and resources for the pre-training and pruning of the EXTD model.
• Pruned_Models/: A collection of pruned models (.pth files). Pruned models with target pruning rates equal to 10%, 20%, 30%, 40%, 50% and 60% are included, in addition to the original pre-trained and unpruned EResFD and EXTD models.
• torchscript/: All the required files for android deployment of the EResFD model (and its pruned versions) using the torchscript framework.

Data

Download and extract the WIDER FACE dataset from here. Place the WIDER folder in the same directory as the project folders.

Training and Pruning

Our methodology consists of 4 sequential steps:

1. Pre-training (main_pretrain.py)

2. Bayesian Optimization (main_bayesian_prune.py)

3. Soft Filter Pruning (main_sfp.py)

4. Fine-tuning (main_finetune.py)

The evaluation on the WIDER FACE validation set can be performed by using the script main_evaluation.py.

Important Hyperparameters for Training

Hyperparameter Name	Description	Default Value
n_epochs_pretraining	The number of epochs used during the pretraining phase of the model.	300
n_epochs_sfp	The number of epochs used for the Soft Filter Pruning (SFP) process.	200
n_epochs_finetune	The number of epochs used for the fine-tuning stage of the model.	10
target_pruning_rate	The target pruning rate for Bayesian optimization.	0.2
n_opt_iterations	The total number of iterations for Bayesian optimization, after the initial random sampling.	940
n_init_points	The number of initial points sampled by the Bayesian optimization.	60

Pruned models

In the Pruned_Models/ folder, we provide pruned models with target pruning rates of 10%, 20%, 30%, 40%, 50% and 60% for EResFD, and 20% for EXTD, along with the original pre-trained and unpruned EResFD and EXTD models. In the following tables, for each model, we present the pruning rates for each group:

EResFD Model	Group 1 Pruning Rate	Group 2 Pruning Rate	Group 3 Pruning Rate	Group 4 Pruning Rate	Group 5 Pruning Rate	Group 6 Pruning Rate	Sparsity
Original	0%	0%	0%	0%	0%	0%	0
Pruned 10%	20%	0%	7.9%	20%	20%	17.2%	10.24%
Pruned 20%	0%	0%	0%	40%	40%	39.8%	22.27%
Pruned 30%	21.3%	4.5%	30.5%	47.1%	32.1%	44.8%	31.59%
Pruned 40%	0.8%	5.2%	47.2%	60%	24.4%	46.4%	40.02%
Pruned 50%	0%	0%	38.3%	70%	56.1%	70%	50.37%
Pruned 60%	8.2%	0%	80%	80%	80%	16.4%	59.87%

EXTD Model	Group 1 Pruning Rate	Group 2 Pruning Rate	Group 3 Pruning Rate	Group 4 Pruning Rate	Group 5 Pruning Rate	Group 6 Pruning Rate	Sparsity
Original	0%	0%	0%	0%	0%	0%	0
Pruned 20%	31.1%	15.7%	32.4%	21.6%	31.1%	9.8%	20.3%

Android Deployment

To deploy the pruned EResFD models on Android:

1. Use convert_to_torchscript.py to convert the desired model to TorchScript format.
2. Move the generated lite_scripted_model.ptl file to the assets folder of your Android application.
3. Load the model in your app using the provided EResFD.java class for inference. A sample build.gradle.kts file is included with the necessary dependencies for building the Android app.

License

This code is provided for academic, non-commercial use only. Please also check for any restrictions applied in the code parts and datasets used here from other sources. For the materials not covered by any such restrictions, redistribution and use in source and binary forms, with or without modification, are permitted for academic non-commercial use provided that the following conditions are met:

Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation provided with the distribution.

This software is provided by the authors "as is" and any express or implied warranties, including, but not limited to, the implied warranties of merchantability and fitness for a particular purpose are disclaimed. In no event shall the authors be liable for any direct, indirect, incidental, special, exemplary, or consequential damages (including, but not limited to, procurement of substitute goods or services; loss of use, data, or profits; or business interruption) however caused and on any theory of liability, whether in contract, strict liability, or tort (including negligence or otherwise) arising in any way out of the use of this software, even if advised of the possibility of such damage.

Citation

If you find our pruning method or pruned models useful in your work, please cite the following publication where this approach was proposed:

N. Kaparinos, V. Mezaris, "B-FPGM: Lightweight Face Detection via Bayesian-Optimized Soft FPGM Pruning", Proc. IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW 2025), Tucson, AZ, USA, Feb. 2025.

A pre-print of this paper is available at: http://arxiv.org/abs/2501.16917

@INPROCEEDINGS{Kaparinos2025,
  author={Kaparinos, Nikolaos and Mezaris, Vasileios},
  booktitle={2025 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW)}, 
  title={B-FPGM: Lightweight Face Detection via Bayesian-Optimized Soft FPGM Pruning}, 
  year={2025}
}

Acknowledgement

This work was supported by the EU Horizon Europe and Horizon 2020 programmes under grant agreements 101070093 vera.ai and 951911 AI4Media, respectively.