Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.
You can not select more than 25 topics
Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
kimji
e1baec91c9
|
2 years ago | |
|---|---|---|
| ascend310_infer | 2 years ago | |
| scripts | 2 years ago | |
| src | 2 years ago | |
| README.md | 2 years ago | |
| eval.py | 2 years ago | |
| export.py | 2 years ago | |
| postprocess310.py | 2 years ago | |
| preprocess.py | 2 years ago | |
| preprocess310.py | 2 years ago | |
| requirements.txt | 2 years ago | |
| train.py | 2 years ago | |
README.md
Contents
- FaceBoxes Description
- Model Architecture
- Dataset
- Environment Requirements
- Quick Start
- Script Description
- Model Description
- 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:
- Get the train annotations from here.
- Get the eval ground truth label from here.
- 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
- For more information, please check the resources below：
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/trainThe 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.pyThe 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 : 0.4032OR,
python eval.pyThe 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.
No Description
Python C++ Shell Text other
Contributors (1)