mindspore/model_zoo/research/cv/FaceDetection

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

Face Detection Description

This is a Face Detection network based on Yolov3, with support for training and evaluation on Ascend910.

You only look once (YOLO) is a state-of-the-art, real-time object detection system. YOLOv3 is extremely fast and accurate.

Prior detection systems repurpose classifiers or localizers to perform detection. They apply the model to an image at multiple locations and scales. High scoring regions of the image are considered detections.
YOLOv3 use a totally different approach. It apply a single neural network to the full image. This network divides the image into regions and predicts bounding boxes and probabilities for each region. These bounding boxes are weighted by the predicted probabilities.

Paper: YOLOv3: An Incremental Improvement. Joseph Redmon, Ali Farhadi,
University of Washington

Model Architecture

Face Detection uses a modified-DarkNet53 network for performing feature extraction. It has 45 convolutional layers.

Dataset

We use about 13K images as training dataset and 3K as evaluating dataset in this example, and you can also use your own datasets or open source datasets (e.g. WiderFace)

• step 1: The dataset should follow the Pascal VOC data format for object detection. The directory structure is as follows:(Because of the small input shape of network, we remove the face lower than 50*50 at 1080P in evaluating dataset )

      .
      └─ dataset
        ├─ Annotations
          ├─ img1.xml
          ├─ img2.xml
          ├─ ...
        ├─ JPEGImages
          ├─ img1.jpg
          ├─ img2.jpg
          ├─ ...
        └─ ImageSets
          └─ Main
            └─ train.txt or test.txt
  

• step 2: Convert the dataset to mindrecord:

  python data_to_mindrecord_train.py
  

  or

  python data_to_mindrecord_eval.py
  

  If your dataset is too big to convert at a time, you can add data to an existed mindrecord in turn:

  python data_to_mindrecord_train_append.py
  

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 Detection
  ├─ README.md
  ├─ scripts
    ├─ run_standalone_train.sh              # launch standalone training(1p) in ascend
    ├─ run_distribute_train.sh              # launch distributed training(8p) in ascend
    ├─ run_eval.sh                          # launch evaluating in ascend
    └─ run_export.sh                        # launch exporting air model
  ├─ src
    ├─ FaceDetection
      ├─ voc_wrapper.py                     # get detection results
      ├─ yolo_loss.py                       # loss function
      ├─ yolo_postprocess.py                # post process
      └─ yolov3.py                          # network
    ├─ config.py                            # parameter configuration
    ├─ data_preprocess.py                   # preprocess
    ├─ logging.py                           # log function
    ├─ lrsche_factory.py                    # generate learning rate
    ├─ network_define.py                    # network define
    ├─ transforms.py                        # data transforms
    ├─ data_to_mindrecord_train.py          # convert dataset to mindrecord for training
    ├─ data_to_mindrecord_train_append.py   # add dataset to an existed mindrecord for training
    └─ data_to_mindrecord_eval.py           # convert dataset to mindrecord for evaluating
  ├─ train.py                               # training scripts
  ├─ eval.py                                # evaluation scripts
  └─ export.py                              # export air model

Running Example

Train

• Stand alone mode

  cd ./scripts
  sh run_standalone_train.sh [MINDRECORD_FILE] [USE_DEVICE_ID]
  

  or (fine-tune)

  cd ./scripts
  sh run_standalone_train.sh [MINDRECORD_FILE] [USE_DEVICE_ID] [PRETRAINED_BACKBONE]
  

  for example:

  cd ./scripts
  sh run_standalone_train.sh /home/train.mindrecord 0 /home/a.ckpt
  

• Distribute mode (recommended)

  cd ./scripts
  sh run_distribute_train.sh [MINDRECORD_FILE] [RANK_TABLE]
  

  or (fine-tune)

  cd ./scripts
  sh run_distribute_train.sh [MINDRECORD_FILE] [RANK_TABLE] [PRETRAINED_BACKBONE]
  

  for example:

  cd ./scripts
  sh run_distribute_train.sh /home/train.mindrecord ./rank_table_8p.json /home/a.ckpt
  

You will get the loss value of each step as following in "./output/[TIME]/[TIME].log" or "./scripts/device0/train.log":

rank[0], iter[0], loss[318555.8], overflow:False, loss_scale:1024.0, lr:6.24999984211172e-06, batch_images:(64, 3, 448, 768), batch_labels:(64, 200, 6)
rank[0], iter[1], loss[95394.28], overflow:True, loss_scale:1024.0, lr:6.24999984211172e-06, batch_images:(64, 3, 448, 768), batch_labels:(64, 200, 6)
rank[0], iter[2], loss[81332.92], overflow:True, loss_scale:512.0, lr:6.24999984211172e-06, batch_images:(64, 3, 448, 768), batch_labels:(64, 200, 6)
rank[0], iter[3], loss[27250.805], overflow:True, loss_scale:256.0, lr:6.24999984211172e-06, batch_images:(64, 3, 448, 768), batch_labels:(64, 200, 6)
...

rank[0], iter[62496], loss[2218.6282], overflow:False, loss_scale:256.0, lr:6.24999984211172e-06, batch_images:(64, 3, 448, 768), batch_labels:(64, 200, 6)
rank[0], iter[62497], loss[3788.5146], overflow:False, loss_scale:256.0, lr:6.24999984211172e-06, batch_images:(64, 3, 448, 768), batch_labels:(64, 200, 6)
rank[0], iter[62498], loss[3427.5479], overflow:False, loss_scale:256.0, lr:6.24999984211172e-06, batch_images:(64, 3, 448, 768), batch_labels:(64, 200, 6)
rank[0], iter[62499], loss[4294.194], overflow:False, loss_scale:256.0, lr:6.24999984211172e-06, batch_images:(64, 3, 448, 768), batch_labels:(64, 200, 6)

Evaluation

cd ./scripts
sh run_eval.sh [MINDRECORD_FILE] [USE_DEVICE_ID] [PRETRAINED_BACKBONE]

for example:

cd ./scripts
sh run_eval.sh /home/eval.mindrecord 0 /home/a.ckpt

You will get the result as following in "./scripts/device0/eval.log":

calculate [recall | persicion | ap]...
Saving ../../results/0-2441_61000/.._.._results_0-2441_61000_face_AP_0.760.png

And the detect result and P-R graph will also be saved in "./results/[MODEL_NAME]/"

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]

Model Description

Performance

Training Performance

Parameters	Face Detection
Model Version	V1
Resource	Ascend 910; CPU 2.60GHz, 192cores; Memory, 755G
uploaded Date	09/30/2020 (month/day/year)
MindSpore Version	1.0.0
Dataset	13K images
Training Parameters	epoch=2500, batch_size=64, momentum=0.5
Optimizer	Momentum
Loss Function	Softmax Cross Entropy, Sigmoid Cross Entropy, SmoothL1Loss
outputs	boxes and label
Speed	1pc: 800~850 ms/step; 8pcs: 1000~1150 ms/step
Total time	1pc: 120 hours; 8pcs: 18 hours
Checkpoint for Fine tuning	37M (.ckpt file)

Evaluation Performance

Parameters	Face Detection
Model Version	V1
Resource	Ascend 910
Uploaded Date	09/30/2020 (month/day/year)
MindSpore Version	1.0.0
Dataset	3K images
batch_size	1
outputs	mAP
Accuracy	8pcs: 76.0%
Model for inference	37M (.ckpt file)

ModelZoo Homepage

Please check the official homepage.