mindspore/model_zoo/research/cv/FaceRecognition

• Contents
• Face Recognition Description
• Model Architecture
• Dataset
• Environment Requirements
• Script Description
• • Script and Sample Code
  • Running Example
  • • Train
    • Evaluation
    • Convert model
• Model Description
• • Performance
  • • Training Performance
    • Evaluation Performance
• ModelZoo Homepage

Contents

• Face Recognition Description
• Model Architecture
• Dataset
• Environment Requirements
• Script Description
  • Script and Sample Code
  • Running Example
• Model Description
  • Performance
• ModelZoo Homepage

Face Recognition Description

This is a face recognition network based on Resnet, with support for training and evaluation on Ascend910.

ResNet (residual neural network) was proposed by Kaiming He and other four Chinese of Microsoft Research Institute. Through the use of ResNet unit, it successfully trained 152 layers of neural network, and won the championship in ilsvrc2015. The error rate on top 5 was 3.57%, and the parameter quantity was lower than vggnet, so the effect was very outstanding. Traditional convolution network or full connection network will have more or less information loss. At the same time, it will lead to the disappearance or explosion of gradient, which leads to the failure of deep network training. ResNet solves this problem to a certain extent. By passing the input information to the output, the integrity of the information is protected. The whole network only needs to learn the part of the difference between input and output, which simplifies the learning objectives and difficulties.The structure of ResNet can accelerate the training of neural network very quickly, and the accuracy of the model is also greatly improved. At the same time, ResNet is very popular, even can be directly used in the concept net network.

Paper: Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. "Deep Residual Learning for Image Recognition"

Model Architecture

Face Recognition uses a Resnet network for performing feature extraction, more details are show below:Link

Dataset

We use about 4.7 million face images as training dataset and 1.1 million as evaluating dataset in this example, and you can also use your own datasets or open source datasets (e.g. face_emore).
The directory structure is as follows:

.
└─ dataset
  ├─ train dataset
    ├─ ID1
      ├─ ID1_0001.jpg
      ├─ ID1_0002.jpg
      ...
    ├─ ID2
      ...
    ├─ ID3
      ...
    ...
  ├─ test dataset
    ├─ ID1
      ├─ ID1_0001.jpg
      ├─ ID1_0002.jpg
      ...
    ├─ ID2
      ...
    ├─ ID3
      ...
    ...

Environment Requirements

• Hardware（Ascend）
  • Prepare hardware environment with Ascend processor.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore tutorials
  • MindSpore Python API

Script Description

Script and Sample Code

The entire code structure is as following:

└─ face_recognition
  ├── README.md                             // descriptions about face_recognition
  ├── scripts
  │   ├── run_distribute_train_base.sh      // shell script for distributed training on Ascend
  │   ├── run_distribute_train_beta.sh      // shell script for distributed training on Ascend
  │   ├── run_eval.sh                       // shell script for evaluation on Ascend
  │   ├── run_export.sh                     // shell script for exporting air model
  │   ├── run_standalone_train_base.sh      // shell script for standalone training on Ascend
  │   ├── run_standalone_train_beta.sh      // shell script for standalone training on Ascend
  ├── src
  │   ├── backbone
  │   │   ├── head.py                       // head unit
  │   │   ├── resnet.py                     // resnet architecture
  │   ├── callback_factory.py               // callback logging
  │   ├── config.py                         // parameter configuration
  │   ├── custom_dataset.py                 // custom dataset and sampler
  │   ├── custom_net.py                     // custom cell define
  │   ├── dataset_factory.py                // creating dataset
  │   ├── init_network.py                   // init network parameter
  │   ├── my_logging.py                     // logging format setting
  │   ├── loss_factory.py                   // loss calculation
  │   ├── lrsche_factory.py                 // learning rate schedule
  │   ├── me_init.py                        // network parameter init method
  │   ├── metric_factory.py                 // metric fc layer
  ├─ train.py                               // training scripts
  ├─ eval.py                                // evaluation scripts
  └─ export.py                              // export air model

Running Example

Train

• Stand alone mode

  • base model

    cd ./scripts
    sh run_standalone_train_base.sh [USE_DEVICE_ID]
    

    for example:

    cd ./scripts
    sh run_standalone_train_base.sh 0
    

  • beta model

    cd ./scripts
    sh run_standalone_train_beta.sh [USE_DEVICE_ID]
    

    for example:

    cd ./scripts
    sh run_standalone_train_beta.sh 0
    

• Distribute mode (recommended)

  • base model

    cd ./scripts
    sh run_distribute_train_base.sh [RANK_TABLE]
    

    for example:

    cd ./scripts
    sh run_distribute_train_base.sh ./rank_table_8p.json
    

  • beta model

    cd ./scripts
    sh run_distribute_train_beta.sh [RANK_TABLE]
    

    for example:

    cd ./scripts
    sh run_distribute_train_beta.sh ./rank_table_8p.json
    

You will get the loss value of each epoch as following in "./scripts/data_parallel_log_[DEVICE_ID]/outputs/logs/[TIME].log" or "./scripts/log_parallel_graph/face_recognition_[DEVICE_ID].log":

epoch[0], iter[100], loss:(Tensor(shape=[], dtype=Float32, value= 50.2733), Tensor(shape=[], dtype=Bool, value= False), Tensor(shape=[], dtype=Float32, value= 32768)), cur_lr:0.000660, mean_fps:743.09 imgs/sec
epoch[0], iter[200], loss:(Tensor(shape=[], dtype=Float32, value= 49.3693), Tensor(shape=[], dtype=Bool, value= False), Tensor(shape=[], dtype=Float32, value= 32768)), cur_lr:0.001314, mean_fps:4426.42 imgs/sec
epoch[0], iter[300], loss:(Tensor(shape=[], dtype=Float32, value= 48.7081), Tensor(shape=[], dtype=Bool, value= False), Tensor(shape=[], dtype=Float32, value= 16384)), cur_lr:0.001968, mean_fps:4428.09 imgs/sec
epoch[0], iter[400], loss:(Tensor(shape=[], dtype=Float32, value= 45.7791), Tensor(shape=[], dtype=Bool, value= False), Tensor(shape=[], dtype=Float32, value= 16384)), cur_lr:0.002622, mean_fps:4428.17 imgs/sec

...
epoch[8], iter[27300], loss:(Tensor(shape=[], dtype=Float32, value= 2.13556), Tensor(shape=[], dtype=Bool, value= False), Tensor(shape=[], dtype=Float32, value= 65536)), cur_lr:0.004000, mean_fps:4429.38 imgs/sec
epoch[8], iter[27400], loss:(Tensor(shape=[], dtype=Float32, value= 2.36922), Tensor(shape=[], dtype=Bool, value= False), Tensor(shape=[], dtype=Float32, value= 65536)), cur_lr:0.004000, mean_fps:4429.88 imgs/sec
epoch[8], iter[27500], loss:(Tensor(shape=[], dtype=Float32, value= 2.08594), Tensor(shape=[], dtype=Bool, value= False), Tensor(shape=[], dtype=Float32, value= 65536)), cur_lr:0.004000, mean_fps:4430.59 imgs/sec
epoch[8], iter[27600], loss:(Tensor(shape=[], dtype=Float32, value= 2.38706), Tensor(shape=[], dtype=Bool, value= False), Tensor(shape=[], dtype=Float32, value= 65536)), cur_lr:0.004000, mean_fps:4430.37 imgs/sec

Evaluation

cd ./scripts
sh run_eval.sh [USE_DEVICE_ID]

You will get the result as following in "./scripts/log_inference/outputs/models/logs/[TIME].log":
[test_dataset]: zj2jk=0.9495, jk2zj=0.9480, avg=0.9487

Convert model

If you want to infer the network on Ascend 310, you should convert the model to AIR:

cd ./scripts
sh run_export.sh [BATCH_SIZE] [USE_DEVICE_ID] [PRETRAINED_BACKBONE]

for example:

cd ./scripts
sh run_export.sh 16 0 ./0-1_1.ckpt

Model Description

Performance

Training Performance

Parameters	Face Recognition
Model Version	V1
Resource	Ascend 910; CPU 2.60GHz, 192cores; Memory 755G; OS Euler2.8
uploaded Date	09/30/2020 (month/day/year)
MindSpore Version	1.0.0
Dataset	4.7 million images
Training Parameters	epoch=100, batch_size=192, momentum=0.9
Optimizer	Momentum
Loss Function	Cross Entropy
outputs	probability
Speed	1pc: 300~400 ms/step; 8pcs: 40~50 ms/step
Total time	1pc: NA hours; 8pcs: 10 hours
Checkpoint for Fine tuning	584M (.ckpt file)

Evaluation Performance

Parameters	Face Recognition For Tracking
Model Version	V1
Resource	Ascend 910; OS Euler2.8
Uploaded Date	09/30/2020 (month/day/year)
MindSpore Version	1.0.0
Dataset	1.1 million images
batch_size	512
outputs	ACC
ACC	0.9
Model for inference	584M (.ckpt file)

ModelZoo Homepage

Please check the official homepage.