简体中文 | English

Arcface-Paddle

• 1. Introduction
• 2. Environment Preparation
• 3. Data Preparation
  • 3.1 Download Dataset
  • 3.2 Extract MXNet Dataset to images
• 4. How to Training
  • 4.1 Single Node, Single GPU
  • 4.2 Single Node, 8 GPUs
• 5. Model Evaluation
• 6. Export Model
• 7 Model Inference
• 8 Model Performance
  • 8.1 Performance of Lighting Model
  • 8.2 Accuracy on Verification Datasets
  • 8.3 Maximum Number of Identities
  • 8.4 Throughtput
• 9. Inference Combined with Face Detection Model

<a name="Introduction"></a>

1. Introduction

Arcface-Paddle is an open source deep face detection and recognition toolkit, powered by PaddlePaddle. Arcface-Paddle provides three related pretrained models now, include BlazeFace for face detection, ArcFace and MobileFace for face recognition.

• This tutorial is mainly about face recognition.
• For face detection task, please refer to: Face detection tuturial.
• For Whl package inference using PaddleInference, please refer to whl package inference.

Note: Many thanks to GuoQuanhao for the reproduction of the Arcface basline using PaddlePaddle.

<a name="Environment_Preparation"></a>

2. Environment Preparation

Please refer to Installation to setup environment at first.

<a name="Data_Preparation"></a>

3. Data Preparation

<a name="Download_Dataset"></a>

3.1 Download Dataset

Download the dataset from insightface datasets.

• MS1M_v2: MS1M-ArcFace
• MS1M_v3: MS1M-RetinaFace

<a name="Extract_MXNet_Dataset_to_images"></a>

3.2 Extract MXNet Dataset to images

python tools/mx_recordio_2_images.py --root_dir ms1m-retinaface-t1/ --output_dir MS1M_v3/

After finishing unzipping the dataset, the folder structure is as follows.

MS1M_v3
|_ images
|  |_ 00000001.jpg
|  |_ ...
|  |_ 05179510.jpg
|_ label.txt
|_ agedb_30.bin
|_ cfp_ff.bin
|_ cfp_fp.bin
|_ lfw.bin

Label file format is as follows.

# delimiter: "\t"
# the following the content of label.txt
images/00000001.jpg 0
...

If you want to use customed dataset, you can arrange your data according to the above format.

<a name="How_to_Training"></a>

4. How to Training

<a name="Single_Node_Single_GPU"></a>

4.1 Single Node, Single GPU

export CUDA_VISIBLE_DEVICES=1
python tools/train.py \
    --config_file configs/ms1mv2_mobileface.py \
    --embedding_size 128 \
    --sample_ratio 1.0 \
    --loss ArcFace \
    --batch_size 512 \
    --dataset MS1M_v2 \
    --num_classes 85742 \
    --data_dir MS1M_v2/ \
    --label_file MS1M_v2/label.txt \
    --fp16 False

<a name="Single_Node_8_GPU"></a>

4.2 Single Node, 8 GPUs

Static Mode

sh scripts/train_static.sh

Dynamic Mode

sh scripts/train_dynamic.sh

During training, you can view loss changes in real time through VisualDL, For more information, please refer to VisualDL.

<a name="Model_Evaluation"></a>

5. Model Evaluation

The model evaluation process can be started as follows.

Static Mode

sh scripts/validation_static.sh

Dynamic Mode

sh scripts/validation_dynamic.sh

<a name="Export_Model"></a>

6. Export Model

PaddlePaddle supports inference using prediction engines. Firstly, you should export inference model.

Static Mode

sh scripts/export_static.sh

Dynamic Mode

sh scripts/export_dynamic.sh

We also support export to onnx model, you only need to set --export_type onnx.

<a name="Model_Inference"></a>

7. Model Inference

The model inference process supports paddle save inference model and onnx model.

sh scripts/inference.sh

<a name="Model_Performance"></a>

8. Model Performance

<a name="Performance_of_Lighting_Model"></a>

8.1 Performance of Lighting Model

Configuration：

• CPU: Intel(R) Xeon(R) Gold 6184 CPU @ 2.40GHz
• GPU: a single NVIDIA Tesla V100
• Precison: FP32
• BatchSize: 64/512
• SampleRatio: 1.0
• Embedding Size: 128
• MS1MV2

• Note: MobileFace-Paddle training using MobileFaceNet_128

<a name="Accuracy_on_Verification_Datasets"></a>

8.2 Accuracy on Verification Datasets

Configuration：

• GPU: 8 NVIDIA Tesla V100 32G
• Precison: Pure FP16
• BatchSize: 128/1024

<a name="Maximum_Number_of_Identities"></a>

8.3 Maximum Number of Identities

Configuration：

• GPU: 8 NVIDIA Tesla V100 32G (32510MiB)
• BatchSize: 64/512
• SampleRatio: 0.1

Note: config environment variable by export FLAGS_allocator_strategy=naive_best_fit

<a name="Throughtput"></a>

8.4 Throughtput

Configuration：

• BatchSize: 128/1024
• SampleRatio: 0.1
• Datasets: MS1MV3
• V100: Driver Version: 450.80.02, CUDA Version: 11.0
• A100: Driver Version: 460.32.03, CUDA Version: 11.2

For more experimental results see PLSC, which is an open source Paddle Large Scale Classification Tools powered by PaddlePaddle. It supports 60 million classes on single node 8 NVIDIA V100 (32G).

<a name="Inference_Combined_with_Face_Detection_Model"></a>

9. Inference Combined with Face Detection Model

Firstly, use the following commands to download the index gallery, demo image and font file for visualization.

# Index library for the recognition process
wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/index.bin
# Demo image
wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/demo/friends/query/friends2.jpg
# Font file for visualization
wget https://raw.githubusercontent.com/littletomatodonkey/insight-face-paddle/main/SourceHanSansCN-Medium.otf

Use the following command to run the whole face recognition demo.

# detection + recogniotion process
python3.7 tools/test_recognition.py --det --rec --index=index.bin --input=friends2.jpg --output="./output"

The final result is save in folder output/, which is shown as follows.

For more details about parameter explanations, index gallery construction and whl package inference, please refer to:

• Whl package inference tutorial.
• Paddle Serving inference