fastface 0.0.1

A face detection framework for edge devices using pytorch lightning

FastFace

Face detection implementations with pytorch-lightning

GOAL

Supporting lightweight face detection implementations to train, test and deploy in a scalable and maintainable manner.

CONTENTS

• Recent Update
• Usage
• TODO
• Reference
• Citation

Recent Update

• 2021.01.02 added unittests for apis under the tests/ directory
• 2021.01.02 online hard negative mining is added for lffd training
• 2021.01.02 caching is supported and by default it will use ~/.cache/fastface
• 2021.01.02 with fastface.adapters , models and datasets can be downloadable via gdrive or requests
• 2021.01.02 now this repository can be usable as package
• 2020.12.15 evaluation scripts and results are added
• 2020.12.13 added lffd 560_25L_8scales official weights that converted from mxnet to pytorch
• 2020.12.11 tested training script and after 50 epochs, achived 75 ap score on widerface-easy validation set using lffd (560_25L_8scales) with random weight initialization(defined in the paper)
• 2020.12.11 added widerface evaluation metric under the metrics/widerface_ap.py as pytorch_lightning.metrics.Metric
• 2020.12.06 added lffd weight conversion script under the tools/lffd_mx2torch.py to convert official mxnet model weights to pytorch weights

Usage

Install

From PyPI

pip install fastface -U

From source

pip install .

Inference

Using package

import fastface
from fastface.transform import (
    Compose,
    Interpolate,
    Padding,
    Normalize,
    ToTensor
)
from cv2 import cv2

# load image
img = cv2.imread("<your_image_file_path>")

# build model with pretrained weights
model = fastface.module.from_pretrained("original_lffd_560_25L_8S")
# model: pl.LightningModule

# build required transforms
transforms = Compose(
    Interpolate(max_dim=640),
    Padding(target_size=(640,640)),
    Normalize(mean=127.5, std=127.5),
    ToTensor()
)

# enable tracking to perform postprocess after inference 
transforms.enable_tracking()
# reset queue
transforms.flush()

# set model to eval mode
model.eval()
# freeze model in order to disable gradient tracking
model.freeze()
# [optional] move model to gpu
model.to("cuda")

# apply transforms
batch = transforms(img)

# model inference
preds = model.predict(batch, det_threshold=.8, iou_threshold=.4)

# postprocess to adjust predictions
preds = [transforms.adjust(pred.cpu().numpy()) for pred in preds]

print(preds)
"""
[
    np.array(N,5), # as x1,y1,x2,y2,score
    ...
]
"""

Using demo.py script

python demo.py --model original_lffd_560_25L_8S --device cuda --input <your_image_file_path>

sample output;

Evaluation

Evaluation on Widerface Validation Set Using LFFD 560_25L_8S original weights

Method	Easy Set	Medium Set	Hard Set
LFFD(paper)	0.910	0.881	0.780
LFFD(this repository)	0.893	0.866	0.756

To get these results, run the following scripts

# for easy set
python test_widerface.py --model original_lffd_560_25L_8S --device cuda --partition easy

# for medium set
python test_widerface.py --model original_lffd_560_25L_8S --device cuda --partition medium

# for hard set
python test_widerface.py --model original_lffd_560_25L_8S --device cuda --partition hard

Training

Train LFFD 560_25L_8S on widerface dataset

python train_widerface.py --yaml config_zoo/lffd.original.yaml

TODO

Feature

• add lffd 320_20L_5S configuration to the arch
• add lffd 320_20L_5S pytorch model to the registry using original repository
• add lffd 320_20L_5S training configuration to config_zoo
• add FDDB dataset
• add FDDB datamodule

Training

• add resume training
• add widerface dataset support
• add widerface dataset download adapter
• add FDDB dataset support
• add LR Scheduler
• add detector train loop
• add detector val loop
• add detector test loop
• support AP metric
• convert AP metric to pytorch_lightning.metrics.Metric
• implement OHNM instead of random sampling
• provide lffd model weights that training from scratch

Inference

• add demo.py
• export APIs for package usage
• add setup.py
• support model download via io utility

Depyloment

• add bentoml support and guideline
• add ONNX support and onnx.js guideline
• add mobile demo guideline
• add torchscript support and C++ guideline

Reference

• LFFD Paper
• Official LFFD Implementation

Citation

@inproceedings{LFFD,
title={LFFD: A Light and Fast Face Detector for Edge Devices},
author={He, Yonghao and Xu, Dezhong and Wu, Lifang and Jian, Meng and Xiang, Shiming and Pan, Chunhong},
booktitle={arXiv:1904.10633},
year={2019}
}