PatchNet: A Simple Face Anti-Spoofing Framework via Fine-Grained Patch Recognition

This repository implements PatchNet from paper PatchNet: A Simple Face Anti-Spoofing Framework via Fine-Grained Patch Recognition (UNOFFICIAL)

After carefully examining the original repository, I have identified a number of missing features that I believe would greatly benefit its functionality. In light of this, I have made the decision to fork the repository and implement these features. Specifically, I incorporated distributed data parallel as well as training and validation for multiple datasets.

Improvements to the original repository

• More augmentations (torchvision ➞ albumentations)
• TurboJPEG support for faster image decoding
• DDP support
• Multiple datasets training
• Utility to convert datasets
• Compute FAS-related metrics (ACER, etc.)
• Incorporate loss into a model (the whole inference can be exported to a single ONNX file)
• Telegram reports
• Compute metrics for each val dataset separately
• Split validation into miltiple GPUs
• Balanced sampler suitable for DDP
• Conversion to ONNX
• Reparametrization efficiency evaluation for supported models

Supported backbones

• EfficientFormerV2
• FastViT

Installation

$ python3 -m venv env
$ source env/bin/activate
$ pip install -r requirements.txt

When using EfficientFormerV2 model, put pretrained weights to weights/efficientformerv2

Telegram reports

This repository supports sending messages after each epoch to a telegram bot.
To make it work, create .env file with Telegram bot token and chat ids:

$ cp .env.template .env
$ nano .env

Usage

Data preparation

datasets
    |---images
    |     |--img1
    |     |--img2
    |     |...
    |---train.csv
    |---val.csv
    |---test.csv

with [*.csv] having format (label only has 2 classes: 0-Spoofing, 1-Bonafide):

image_name      |  label
img_name1.jpg   |    0
img_name2.png   |    1
...

image_name is the relative path to the image from the locations of the *.csv file.
One can find utility to convert exisiting images dataset into format supported by current repository in utils/dataset_preparation/prepare_dataset.py

Training

torchrun --standalone --nnodes=1 --nproc_per_node=4 tool/train.py --config config.yaml

nproc_per_node is the number of GPUs you want to use.

Testing

This step has not been remade by me yet. Your conrtibution is welcome.
Go to tool/test.py and fix saved_name to your path to checkpoint
Run

python3 test.py

Conversion to ONNX

Use a script utils/convert_to_onnx.py. It will create 2 versions of models: for batch size = 1 and for dynamic batch size. Some inference engines require all the dimensions fixed.

Original repository contributors

• Tien Thong Doan
• Minh Chau Nguyen
• Minh Hung Nguyen

This repository contributors

• Alexander Filonenko