• Contents
• FaceBoxes Description
• Model Architecture
• Dataset
• Environment Requirements
• Quick Start
• Script Description
• • Script and Sample Code
  • Script Parameters
  • Training Process
  • • Training
    • Distributed Training
  • Evaluation Process
  • • Evaluation
  • Inference Process
  • • Export MindIR
    • Infer on Ascend310
    • result
• Model Description
• • Performance
  • • Evaluation Performance
    • Inference Performance
• Description of Random Situation
• ModelZoo Homepage

Contents

• FaceBoxes Description
• Model Architecture
• Dataset
• Environment Requirements
• Quick Start
• Script Description
  • Script and Sample Code
  • Script Parameters
  • Training Process
    • Training
    • Distributed Training
  • Evaluation Process
    • Evaluation
• Model Description
  • Performance
    • Evaluation Performance
• Description of Random Situation
• ModelZoo Homepage

FaceBoxes Description

Faceboxes is a novel face detector with superior performance on both speed and accuracy. Moreover, the speed of FaceBoxes is invariant to the number of faces.

Paper: Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, Stan Z. Li. "FaceBoxes: A CPU Real-time Face Detector with High Accuracy". 2017.

Model Architecture

Specifically, the faceboxes network has a lightweight yet powerful network structure that consists of the Rapidly Digested Convolutional Layers (RDCL) and the Multiple Scale Convolutional Layers (MSCL). The RDCL is designed to enable FaceBoxes to achieve real-time speed on the CPU. The MSCL aims at enriching the receptive fields and discretizing anchors over different layers to handle faces of various scales. Besides, a new anchor densification strategy is proposed to make different types of anchors have the same density on the image, which significantly improves the recall rate of small faces.

Dataset

Dataset used: WIDERFACE

Dataset acquisition:

1. Get the train annotations from here.
2. Get the eval ground truth label from here.
3. Get xml file transformation from here

Generate image list txt file before training process:

python preprocess.py

Create the data set directory align with the content table below:

data
└── widerface                  // dataset data
    ├── train
    │   ├── annotations        // place the dowmloaded training anotations here
    │   ├── images             // place the training data here
    │   └── train_img_list.txt
    └── val
       ├── ground_truth       // place the dowmloaded eval ground truth label here
       ├── images             // place the eval data here
       └── val_img_list.txt

• Dataset size: 3.42G, 32,203 colorful images
  • Train: 1.36G, 12,800 images
  • Val: 345.95M, 3,226 images
  • Test: 1.72G, 16,177 images

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

Quick Start

After installing MindSpore via the official website and download the dataset, you can start training and evaluation as follows:

• running on Ascend

  # run training example
  cd scripts/
  bash run_standalone_train.sh ../data/widerface/train
  # run distributed training example
  cd scripts/
  bash run_distribute_train.sh Ascend [DEVICE_NUM] [VISIABLE_DEVICES(0,1,2,3,4,5,6,7)] [RANK_TABLE_FILE] [DATASET_PATH]
  # run evaluation example
  cd scripts/
  bash run_eval.sh
  

Script Description

Script and Sample Code

├── model_zoo
    ├── README.md                          // descriptions about all the models
    ├── faceboxes
        ├── README.md                      // descriptions about googlenet
        ├── scripts
        │   ├──run_distribute_train.sh     // shell script for distributed on Ascend
        │   ├──run_standalone.sh           // shell script for training standalone on Ascend
        │   ├──run_eval.sh                 // shell script for evaluation on Ascend
        ├── src
        │   ├──dataset.py                  // creating dataset
        │   ├──network.py                  // faceboxes architecture
        │   ├──config.py                   // parameter configuration
        │   ├──augmentation.py             // data augment method
        │   ├──loss.py                     // loss function
        │   ├──utils.py                    // data preprocessing
        │   ├──lr_schedule.py              // learning rate schedule
        ├── data
        │   ├──widerface                   // dataset data
        │   ├──resnet50_pretrain.ckpt      // resnet50 imagenet pretrain model
        │   ├──ground_truth                // eval label
        ├── data
        │   └── widerface                  // dataset data
        │       ├── train
        │       │   ├── annotations        // place the dowmloaded training anotations here
        │       │   ├── images             // place the training data here
        │       │   └── train_img_list.txt
        │       └── val
        │           ├── ground_truth       // place the dowmloaded eval ground truth label here
        │           ├── images             // place the eval data here
        │           └── val_img_list.txt
        ├── train.py                       // training script
        ├── eval.py                        // evaluation script
        ├── eval.py                        // export mindir script
        ├── preprocess.py                  // generate image list txt file
        └── requirements.txt               // other requirements for Faceboxes

Script Parameters

Parameters for both training and evaluation can be set in config.py

• config for FaceBoxes, WIDERFACE dataset

    'image_size': (1024, 1024),                                   # Training image size
    'batch_size': 8,                                              # Batch szie of train
    'min_sizes': [[32, 64, 128], [256], [512]],                   # Anchor sizes of each feature map
    'steps': [32, 64, 128],                                       # Anchor strides
    'variance': [0.1, 0.2],                                       # Variance
    'clip': False,                                                # Clip
    'loc_weight': 2.0,                                            # Bbox regression loss weight
    'class_weight': 1.0,                                          # Confidence/Class regression loss weight
    'match_thresh': 0.35,                                         # Threshold for match box
    'num_worker': 8,                                              # Num worker of dataset load data
    # checkpoint
    "save_checkpoint_epochs": 1,                                  # Save checkpoint steps
    "keep_checkpoint_max": 50,                                    # Number of reserved checkpoints
    "save_checkpoint_path": "./",                                 # Model save path
    # env
    "device_id": int(os.getenv('DEVICE_ID', '0')),                # Device id
    "rank_id": int(os.getenv('RANK_ID', '0')),                    # Rank id
    "rank_size": int(os.getenv('RANK_SIZE', '1')),                # Rank size
    # seed
    'seed': 1,                                                    # Setup train seed
    # opt
    'optim': 'sgd',                                               # Optimizer type
    'momentum': 0.9,                                              # Momentum for Optimizer
    'weight_decay': 5e-4,                                         # Weight decay for Optimizer
    # lr
    'epoch': 300,                                                 # Training epoch number
    'decay1': 200,                                                # Epoch number of the first weight attenuation
    'decay2': 250,                                                # Epoch number of the second weight attenuation
    'lr_type': 'dynamic_lr',                                      # Learning rate decline function type, set dynamic_lr or standard_lr
    'initial_lr': 0.001,                                          # Learning rate
    'warmup_epoch': 4,                                            # Warmup size, 0 means no warm-up
    'gamma': 0.1,                                                 # Attenuation ratio of learning rate
    # ---------------- val ----------------
    'val_model': '../train/rank0/ckpt_0/FaceBoxes-300_402.ckpt',  # Validation model path
    'val_dataset_folder': '../data/widerface/val/',               # Validation dataset path
    'val_origin_size': True,                                      # Is full size verification used
    'val_confidence_threshold': 0.05,                             # Threshold for val confidence
    'val_nms_threshold': 0.4,                                     # Threshold for val NMS
    'val_iou_threshold': 0.5,                                     # Threshold for val IOU
    'val_save_result': False,                                     # Whether save the resultss
    'val_predict_save_folder': './widerface_result',              # Result save path
    'val_gt_dir': '../data/widerface/val/ground_truth',           # Path of val set ground_truth
  

Training Process

Training

• running on Ascend

  cd scripts/
  bash run_standalone_train.sh ../data/widerface/train
  

  The python command above will run in the background, you can view the results through the file log.txt.

  After training, you'll get some checkpoint files under the folder ./ckpt_0/ by default.

Distributed Training

• running on Ascend

  cd scripts/
  bash run_distribute_train.sh Ascend [DEVICE_NUM] [VISIABLE_DEVICES(0,1,2,3,4,5,6,7)] [RANK_TABLE_FILE] [DATASET_PATH]
  

  The above shell script will run distribute training in the background. You can view the results through the file ../train/rank0/log0.log.

  After training, you'll get some checkpoint files under the folder ../train/rank0/ckpt_0/ by default.

Evaluation Process

Evaluation

• evaluation on WIDERFACE dataset when running on Ascend

  Before running the command below, please check the checkpoint path used for evaluation. Please set the checkpoint path to be the absolute full path in src/config.py, e.g., "username/faceboxes/train/rank0/ckpt_0/FaceBoxes-300_402.ckpt".

  cd scripts/
  bash run_eval.py
  

  The above python command will run in the background. You can view the results through the file "eval.log". The result of the test dataset will be as follows:

  # cat eval.log
  Easy   Val AP : 0.8510
  Medium Val AP : 0.7692
  Hard   Val AP : 0.4032
  

  OR,

  python eval.py
  

  The results will be shown after running the above python command.

Inference Process

Export MindIR

python export.py --checkpoint_path [CKPT_PATH]

You will get the model of the format .mindir.

Infer on Ascend310

Before performing inference, the mindir file must be exported by export.py script. Please note that we do the inference using fixed image shape (1024 x 1024), so the performance is lower than that of FaceBoxes on Ascend 910.

# Ascend310 inference
cd scripts/
bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [DEVICE_ID]

result

Inference result is saved in current path, you can find result like this in acc.log file.

# cat acc.log
Easy   Val AP : 0.7996
Medium Val AP : 0.7187
Hard   Val AP : 0.4013

Model Description

Performance

Evaluation Performance

Parameters	Ascend
Model Version	FaceBoxes
Resource	Ascend 910
uploaded Date	6/15/2021 (month/day/year)
MindSpore Version	1.2.0
Dataset	WIDERFACE
Training Parameters	epoch=300, steps=402, batch_size=8, lr=0.001
Optimizer	SGD
Loss Function	MultiBoxLoss + Softmax Cross Entropy
outputs	bounding box + confidence
Accuracy	Easy：0.8510；Medium：0.7692；Hard：0.4032
Loss	2.780
Speed	4pcs: 92 ms/step
Total time	4pcs: 7.6 hours
Parameters (M)	3.84M
Checkpoint for Fine tuning	13M (.ckpt file)
Scripts	faceboxes script

Inference Performance

Parameters	Ascend
Model Version	FaceBoxes
Resource	Ascend 310
Uploaded Date	12/03/2021 (month/day/year)
Dataset	WiderFace
batch_size	1
outputs	bounding box + confidence
Accuracy	Easy：0.7996；Medium：0.7187；Hard：0.4013
Model for inference	4.1M(.mindir file)

Description of Random Situation

In train.py, we set the seed with setup_seed function.

ModelZoo Homepage

Please check the official homepage.