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.
shibeiji
6afa96de4d
|
5 years ago | |
|---|---|---|
| .. | ||
| scripts | 5 years ago | |
| src | 5 years ago | |
| README.md | 5 years ago | |
| eval.py | 5 years ago | |
| export.py | 5 years ago | |
| train.py | 5 years ago | |
README.md
Contents
- CenterNet Description
- Model Architecture
- Dataset
- Environment Requirements
- Quick Start
- Script Description
- Model Description
- ModelZoo Homepage
CenterNet Description
CenterNet is a novel practical anchor-free method for object detection, 3D detection, and pose estimation, which detect identifies objects as axis-aligned boxes in an image. The detector uses keypoint estimation to find center points and regresses to all other object properties, such as size, 3D location, orientation, and even pose. In nature, it's a one-stage method to simultaneously predict center location and bboxes with real-time speed and higher accuracy than corresponding bounding box based detectors.
We support training and evaluation on Ascend910.
Paper: Objects as Points. 2019.
Xingyi Zhou(UT Austin) and Dequan Wang(UC Berkeley) and Philipp Krahenbuhl(UT Austin)
Model Architecture
In the current model, we use CenterNet to estimate multi-person pose. The DLA(Deep Layer Aggregation) net was adopted as backbone, a 3x3 convolutional layer with 256 channel was added before each output head, and a final 1x1 convolution then produced the desired output. Six losses are presented, and the total loss is their weighted mean.
Dataset
Note that you can run the scripts based on the dataset mentioned in original paper or widely used in relevant domain/network architecture. In the following sections, we will introduce how to run the scripts using the related dataset below.
Dataset used: COCO2017
-
Dataset size：26G
- Train：19G，118000 images
- Val：0.8G，5000 images
- Test: 6.3G, 40000 images
- Annotations：808M，instances，captions，person_keypoints etc
-
Data format：image and json files
- Note：Data will be processed in dataset.py
-
The directory structure is as follows, name of directory and file is user defined:
. ├── dataset ├── centernet ├── annotations │ ├─ train.json │ └─ val.json └─ images ├─ train │ └─images │ ├─class1_image_folder │ ├─ ... │ └─classn_image_folder └─ val │ └─images │ ├─class1_image_folder │ ├─ ... │ └─classn_image_folder └─ test └─images ├─class1_image_folder ├─ ... └─classn_image_folder
Environment Requirements
-
Hardware（Ascend）
- Prepare hardware environment with Ascend processor. If you want to try Ascend, please send the application form to ascend@huawei.com. Once approved, you can get the resources.
-
Framework
-
For more information, please check the resources below：
-
Download the dataset COCO2017.
-
We use COCO2017 as training dataset in this example by default, and you can also use your own datasets.
-
If coco dataset is used. Select dataset to coco when run script.
Install Cython and pycocotool, and you can also install mmcv to process data.pip install Cython pip install pycocotools pip install mmcv==0.2.14And change the COCO_ROOT and other settings you need in
config.py. The directory structure is as follows:. └─cocodataset ├─annotations ├─instance_train2017.json └─instance_val2017.json ├─val2017 └─train2017 -
If your own dataset is used. Select dataset to other when run script.
Organize the dataset information the same format as COCO.
-
Quick Start
After installing MindSpore via the official website, you can start training and evaluation as follows:
Note: 1.the first run will generate the mindrecord file, which will take a long time.
2.VALIDATION_JSON_FILE is ground truth label file. CHECKPOINT_PATH is a checkpoint file after training.
# standalone training
bash run_standalone_train_ascend.sh [DEVICE_ID] [EPOCH_SIZE]
# distributed training
bash run_distributed_train_ascend.sh [COCO_DATASET_PATH] [MINDRECORD_DATASET_PATH] [RANK_TABLE_FILE]
# eval
bash run_standalone_eval_ascend.sh [DEVICE_ID]
Script Description
Script and Sample Code
.
├── cv
├── centernet
├── train.py // training scripts
├── eval.py // testing and evaluation outputs
├── export.py // convert mindspore model to air model
├── README.md // descriptions about CenterNet
├── scripts
│ ├── ascend_distributed_launcher
│ │ ├──__init__.py
│ │ ├──hyper_parameter_config.ini // hyper parameter for distributed pretraining
│ │ ├──get_distribute_pretrain_cmd.py // script for distributed pretraining
│ │ ├──README.md
│ ├──run_standalone_train_ascend.sh // shell script for standalone pretrain on ascend
│ ├──run_distributed_train_ascend.sh // shell script for distributed pretrain on ascend
│ ├──run_standalone_eval_ascend.sh // shell script for standalone evaluation on ascend
└── src
├──__init__.py
├──centernet_pose.py // centernet networks, training entry
├──dataset.py // generate dataloader and data processing entry
├──config.py // centernet unique configs
├──dcn_v2.py // deformable convolution operator v2
├──decode.py // decode the head features
├──backbone_dla.py // deep layer aggregation backbone
├──utils.py // auxiliary functions for train, to log and preload
├──image.py // image preprocess functions
├──post_process.py // post-process functions after decode in inference
└──visual.py // visualization image, bbox, score and keypoints
Script Parameters
Training
usage: train.py [--device_target DEVICE_TARGET] [--distribute DISTRIBUTE]
[--need_profiler NEED_PROFILER] [--profiler_path PROFILER_PATH]
[--epoch_size EPOCH_SIZE] [--train_steps TRAIN_STEPS] [device_id DEVICE_ID]
[--device_num DEVICE_NUM] [--do_shuffle DO_SHUFFLE]
[--enable_data_sink ENABLE_DATA_SINK] [--data_sink_steps N]
[--enable_save_ckpt ENABLE_SAVE_CKPT]
[--save_checkpoint_path SAVE_CHECKPOINT_PATH]
[--load_checkpoint_path LOAD_CHECKPOINT_PATH]
[--save_checkpoint_steps N] [--save_checkpoint_num N]
[--data_dir DATA_DIR] [--mindrecord_dir MINDRECORD_DIR]
[--visual_image VISUAL_IMAGE] [--save_result_dir SAVE_RESULT_DIR]
options:
--device_target device where the code will be implemented: "Ascend" | "CPU", default is "Ascend"
--distribute training by several devices: "true"(training by more than 1 device) | "false", default is "false"
--need profiler whether to use the profiling tools: "true" | "false", default is "false"
--profiler_path path to save the profiling results: PATH, default is ""
--epoch_size epoch size: N, default is 1
--train_steps training Steps: N, default is -1
--device_id device id: N, default is 0
--device_num number of used devices: N, default is 1
--do_shuffle enable shuffle: "true" | "false", default is "true"
--enable_lossscale enable lossscale: "true" | "false", default is "true"
--enable_data_sink enable data sink: "true" | "false", default is "true"
--data_sink_steps set data sink steps: N, default is 1
--enable_save_ckpt enable save checkpoint: "true" | "false", default is "true"
--save_checkpoint_path path to save checkpoint files: PATH, default is ""
--load_checkpoint_path path to load checkpoint files: PATH, default is ""
--save_checkpoint_steps steps for saving checkpoint files: N, default is 1000
--save_checkpoint_num number for saving checkpoint files: N, default is 1
--data_dir path to original dataset directory: PATH, default is ""
--mindrecord_dir path to mindrecord dataset directory: PATH, default is ""
--visual_image whether visualize the image and annotation info: "true" | "false", default is "false"
--save_result_dir path to save the visualization results: PATH, default is ""
Evaluation
usage: eval.py [--device_target DEVICE_TARGET] [--device_id N]
[--load_checkpoint_path LOAD_CHECKPOINT_PATH]
[--data_dir DATA_DIR] [--run_mode RUN_MODE]
[--visual_image VISUAL_IMAGE]
[enable_eval ENABLE_EVAL] [--save_result_dir SAVE_RESULT_DIR]
options:
--device_target device where the code will be implemented: "Ascend" | "CPU", default is "Ascend"
--device_id device id to run task, default is 0
--load_checkpoint_path initial checkpoint (usually from a pre-trained CenterNet model): PATH, default is ""
--data_dir validation or test dataset dir: PATH, default is ""
--run_mode inference mode: "val" | "test", default is "val"
--visual_image whether visualize the image and annotation info: "true" | "false", default is "false"
--save_result_dir path to save the visualization and inference results: PATH, default is ""
Options and Parameters
Parameters for training and evaluation can be set in file config.py and finetune_eval_config.py respectively.
Options
config for training.
batch_size batch size of input dataset: N, default is 32
loss_scale_value initial value of loss scale: N, default is 1024
optimizer optimizer used in the network: Adam, default is Adam
lr_schedule schedules to get the learning rate
config for evaluation.
flip_test whether to use flip test: True | False, default is False
soft_nms nms after decode: True | False, default is True
keep_res keep original or fix resolution: True | False, default is False
multi_scales use multi-scales of image: List, default is [1.0]
pad pad size when keep original resolution, default is 31
K number of bboxes to be computed by TopK, default is 100
score_thresh threshold of score when visualize image and annotation info
config for export.
input_res input resolution of the model air, default is [512, 512]
ckpt_file checkpoint file, default is "./ckkt_file.ckpt"
export_format the exported format of model air, default is MINDIR
export_name the exported file name, default is "CentNet_MultiPose"
Parameters
Parameters for dataset (Training/Evaluation):
num_classes number of categories: N, default is 1
num_joints number of keypoints to recognize a person: N, default is 17
max_objs maximum numbers of objects labeled in each image
input_res input resolution, default is [512, 512]
output_res output resolution, default is [128, 128]
rand_crop whether crop image in random during data augmenation: True | False, default is False
shift maximum value of image shift during data augmenation: N, default is 0.1
scale maximum value of image scale times during data augmenation: N, default is 0.4
aug_rot properbility of image rotation during data augmenation: N, default is 0.0
rotate maximum value of rotation angle during data augmentation: N, default is 0.0
flip_prop properbility of image flip during data augmenation: N, default is 0.5
color_aug whether use color augmentation: True | False, default is False
mean mean value of RGB image
std variance of RGB image
flip_idx the corresponding point index of keypoints when flip the image
edges pairs of points linked by an edge to mimic person pose
eig_vec eigenvectors of RGB image
eig_val eigenvalues of RGB image
categories format of annotations for multi-person pose
Parameters for network (Training/Evaluation):
down_ratio the ratio of input and output resolution during training
last_level the last level in final upsampling
final_kernel the final kernel size for convolution
stage_levels list numbers of the tree height for each stage
stage_channels list numbers of channels of the output in each stage
head_conv the channel number to get the head by convolution
dense_hp whether apply weighted pose regression near center point: True | False, default is True
hm_hp estimate human joint heatmap or directly use the joint offset from center: True | False, default is True
reg_hp_offset regress local offset for human joint heatmaps or not: True | False, default is True
reg_offset regress local offset or not: True | False, default is True
hm_weight loss weight for keypoint heatmaps: N, default is 1.0
off_weight loss weight for keypoint local offsets: N, default is 0.1
wh_weight loss weight for bounding box size: N, default is 0.1
hm_weight loss weight for keypoint heatmaps: N, default is 1.0
hm_hp_weight loss weight for human keypoint heatmap: N, default is 1.0
mse_loss use mse loss or focal loss to train keypoint heatmaps: True | False, default is False
reg_loss l1 or smooth l1 for regression loss: 'l1' | 'sl1', default is 'l1'
Parameters for optimizer and learning rate:
Adam:
weight_decay weight decay: Q
eps term added to the denominator to improve numerical stability: Q
decay_filer lamda expression to specify which param will be decayed
PolyDecay:
learning_rate initial value of learning rate: Q
end_learning_rate final value of learning rate: Q
power learning rate decay factor
eps normalization parameter
warmup_steps number of warmup_steps
MultiDecay:
learning_rate initial value of learning rate: Q
eps normalization parameter
warmup_steps number of warmup_steps
multi_epochs list of epoch numbers after which the lr will be decayed
factor learning rate decay factor
Training Process
Training
Running on Ascend
bash scripts/run_standalone_pretrain_ascend.sh 0 1
The command above will run in the background, you can view training logs in training_log.txt. After training finished, you will get some checkpoint files under the script folder by default. The loss values will be displayed as follows:
# grep "epoch" training_log.txt
...
epoch: 349.0, current epoch percent: 0.80, step: 87450, outputs are (Tensor(shape=[1], dtype=Float32, [ 4.96466]), Tensor(shape=[], dtype=Bool, False), Tensor(shape=[], dtype=Float32, 1024))
epoch: 349.0, current epoch percent: 1.00, step: 87500, outputs are (Tensor(shape=[1], dtype=Float32, [ 4.59703]), Tensor(shape=[], dtype=Bool, False), Tensor(shape=[], dtype=Float32, 1024))
...
Distributed Training
Running on Ascend
bash scripts/run_distributed_pretrain_ascend.sh /path/coco2017 /path/mindrecord /path/hccl.json
The command above will run in the background, you can view training logs in LOG*/training_log.txt and LOG*/ms_log/. After training finished, you will get some checkpoint files under the LOG*/ckpt_0 folder by default. The loss value will be displayed as follows:
# grep "epoch" LOG*/ms_log/mindspore.log
epoch: 0.0, current epoch percent: 0.001, step: 100, outputs are (Tensor(shape=[1], dtype=Float32, [ 1.08209e+01]), Tensor(shape=[], dtype=Bool, False), Tensor(shape=[], dtype=Float32, 1024))
epoch: 0.0, current epoch percent: 0.002, step: 200, outputs are (Tensor(shape=[1], dtype=Float32, [ 1.07566e+01]), Tensor(shape=[], dtype=Bool, False), Tensor(shape=[], dtype=Float32, 1024))
...
epoch: 0.0, current epoch percent: 0.001, step: 100, outputs are (Tensor(shape=[1], dtype=Float32, [ 1.08218e+01]), Tensor(shape=[], dtype=Bool, False), Tensor(shape=[], dtype=Float32, 1024))
epoch: 0.0, current epoch percent: 0.002, step: 200, outputs are (Tensor(shape=[1], dtype=Float32, [ 1.07770e+01]), Tensor(shape=[], dtype=Bool, False), Tensor(shape=[], dtype=Float32, 1024))
...
Testing Process
Testing and Evaluation
# Evaluation base on validation dataset will be done automatically, while for test or test-dev dataset, the accuracy should be upload to the CodaLab official website(https://competitions.codalab.org).
bash scripts/run_standalone_eval_ascend.sh [DEVICE_ID]
you can see the MAP result below as below:
overall performance on coco2017 validation dataset
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets= 20 ] = 0.521
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets= 20 ] = 0.791
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets= 20 ] = 0.564
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets= 20 ] = 0.446
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets= 20 ] = 0.639
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 20 ] = 0.600
Average Recall (AR) @[ IoU=0.50 | area= all | maxDets= 20 ] = 0.847
Average Recall (AR) @[ IoU=0.75 | area= all | maxDets= 20 ] = 0.645
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets= 20 ] = 0.509
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets= 20 ] = 0.729
overall performance on coco2017 test-dev dataset
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets= 20 ] = 0.513
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets= 20 ] = 0.795
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets= 20 ] = 0.550
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets= 20 ] = 0.443
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets= 20 ] = 0.623
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 20 ] = 0.600
Average Recall (AR) @[ IoU=0.50 | area= all | maxDets= 20 ] = 0.863
Average Recall (AR) @[ IoU=0.75 | area= all | maxDets= 20 ] = 0.642
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets= 20 ] = 0.509
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets= 20 ] = 0.724
Convert Process
Convert
If you want to infer the network on Ascend 310, you should convert the model to AIR:
python export.py [DEVICE_ID]
Model Description
Performance
Training Performance
CenterNet on 11.8K images(The annotation and data format must be the same as coco)
| Parameters | CenterNet |
|---|---|
| Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory, 755G |
| uploaded Date | 12/15/2020 (month/day/year) |
| MindSpore Version | 1.0.0 |
| Dataset | 11.8K images |
| Training Parameters | 8p, epoch=350, steps=250 * epoch, batch_size = 32, lr=1.2e-4 |
| Optimizer | Adam |
| Loss Function | Focal Loss, L1 Loss, RegLoss |
| outputs | detections |
| Loss | 4.5-5.5 |
| Speed | 1p 59 img/s, 8p 470 img/s |
| Total time: training | 1p: 4.38 days; 8p: 13-14 h |
| Total time: evaluation | keep res: test 1.7h, val 0.7h; fix res: test 50 min, val 12 min |
| Checkpoint | 242M (.ckpt file) |
| Scripts | https://gitee.com/mindspore/mindspore/tree/master/model_zoo/research/cv/centernet |
Inference Performance
CenterNet on validation(5K images) and test-dev(40K images)
| Parameters | CenterNet |
|---|---|
| Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory, 755G |
| uploaded Date | 12/15/2020 (month/day/year) |
| MindSpore Version | 1.0.0 |
| Dataset | 5K images(val), 40K images(test-dev) |
| batch_size | 1 |
| outputs | boxes and keypoints position and scores |
| Accuracy(validation) | MAP: 52.1%, AP50: 79.1%, AP75: 56.4, Medium: 44.6%, Large: 63.9% |
| Accuracy(test-dev) | MAP: 51.3%, AP50: 79.5%, AP75: 55.0, Medium: 44.3%, Large: 62.3% |
| Model for inference | 87M (.mindir file) |
Description of Random Situation
In run_standalone_train_ascend.sh and run_distributed_train_ascend.sh, we set do_shuffle to True to shuffle the dataset by default.
In train.py, we set a random seed to make sure that each node has the same initial weight in distribute training.
ModelZoo Homepage
Please check the official homepage.
MindSpore is a new open source deep learning training/inference framework that could be used for mobile, edge and cloud scenarios.
C++ Python Text Unity3D Asset C other
Contributors (25+)
314202276@qq.com
5518576+mindspore_ci@user.noreply.gitee.com
tommylike@qq.com
zhaozhenlong1@huawei.com
sunsuodong@huawei.com
shiliang10@huawei.com
wangkaisheng2@huawei.com
zhoufeng54@huawei.com
peixu.ren1@huawei.com
chenweifeng720@huawei.com
guozhijian@huawei.com
jiangjinsheng@huawei.com
caifubi1@huawei.com
chenzomi12@gmail.com
huanghui44@huawei.com
yangruoqi@huawei.com
yiren19920727@163.com
xiefangqi2@huawei.com
yaoyifan1@huawei.com
fuzhiye@huawei.com
lingqiaomin.huawei.com
zhaojichen1@huawei.com
2713219276@qq.com
liubuyu1@huawei.com
zhoupeichen@huawei.com