mindspore-Ecosystem
/
mindspore
21
0
Fork 27
Code Issues Pull Requests 6 Projects Releases Wiki Activity

centernet_det can use on modelarts and infer on ascend310

This commit is contained in:
czh688 2021-09-07 18:51:46 +08:00
parent 53366ce732
commit 277db549b0
29 changed files with 1711 additions and 914 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

327
model_zoo/research/cv/centernet_det/readme.md → model_zoo/research/cv/centernet_det/README.md
View File

@ -9,14 +9,18 @@
- [Script and Sample Code](#script-and-sample-code) - [Script and Sample Code](#script-and-sample-code)
- [Script Parameters](#script-parameters) - [Script Parameters](#script-parameters)
- [Training Process](#training-process) - [Training Process](#training-process)
- [Training](#training)
- [Distributed Training](#distributed-training) - [Distributed Training](#distributed-training)
- [Testing Process](#testing-process) - [Testing Process](#testing-process)
- [Testing and Evaluation](#testing-and-evaluation) - [Testing and Evaluation](#testing-and-evaluation)
- [Inference Process](#inference-process)
- [Convert](#convert)
- [Infer on Ascend310](#infer-on-Ascend310)
- [Result](#result)
- [Model Description](#model-description) - [Model Description](#model-description)
- [Performance](#performance) - [Performance](#performance)
- [Training Performance](#training-performance) - [Training Performance On Ascend 910](#training-performance-on-ascend-910)
- [Inference Performance](#inference-performance) - [Inference Performance On Ascend 910](#inference-performance-on-ascend-910)
- [Inference Performance On Ascend 310](#inference-performance-on-ascend-310)
- [ModelZoo Homepage](#modelzoo-homepage) - [ModelZoo Homepage](#modelzoo-homepage)
# [CenterNet Description](#contents) # [CenterNet Description](#contents)
@ -38,7 +42,7 @@ Note that you can run the scripts based on the dataset mentioned in original pap
Dataset used: [COCO2017](https://cocodataset.org/) Dataset used: [COCO2017](https://cocodataset.org/)
- Dataset size：26G - Dataset size：26G
- Train：19G，118000 images - Train：19G，118000 images
- Val：0.8G，5000 images - Val：0.8G，5000 images
- Test: 6.3G, 40000 images - Test: 6.3G, 40000 images
- Annotations：808M，instances，captions etc - Annotations：808M，instances，captions etc
@ -76,12 +80,14 @@ Dataset used: [COCO2017](https://cocodataset.org/)
# [Environment Requirements](#contents) # [Environment Requirements](#contents)
- Hardware（Ascend） - Hardware（Ascend）
- Prepare hardware environment with Ascend processor. - Prepare hardware environment with Ascend processor.
- Framework - Framework
- [MindSpore](https://www.mindspore.cn/install/en) - [MindSpore](https://www.mindspore.cn/install/en)
- For more information, please check the resources below： - For more information, please check the resources below：
- [MindSpore tutorials](https://www.mindspore.cn/tutorials/en/master/index.html) - [MindSpore tutorials](https://www.mindspore.cn/tutorials/zh-CN/master/index.html)
- [MindSpore Python API](https://www.mindspore.cn/docs/api/en/master/index.html) - [MindSpore Python API](https://www.mindspore.cn/docs/api/zh-CN/master/index.html)
- Download the dataset COCO2017. - Download the dataset COCO2017.
- We use COCO2017 as training dataset in this example by default, and you can also use your own datasets. - We use COCO2017 as training dataset in this example by default, and you can also use your own datasets.
@ -114,6 +120,105 @@ Dataset used: [COCO2017](https://cocodataset.org/)
# [Quick Start](#contents) # [Quick Start](#contents)
- running on local
After installing MindSpore via the official website, you can start training and evaluation as follows:
Note:
1.the first run of training will generate the mindrecord file, which will take a long time.
2.MINDRECORD_DATASET_PATH is the mindrecord dataset directory.
3.For `train.py`, LOAD_CHECKPOINT_PATH is the pretrained checkpoint file directory, if no just set "".
4.For `eval.py`, LOAD_CHECKPOINT_PATH is the checkpoint to be evaluated.
5.RUN_MODE support validation and testing, set to be "val"/"test"
```shell
# create dataset in mindrecord format
bash scripts/convert_dataset_to_mindrecord.sh [COCO_DATASET_DIR] [MINDRECORD_DATASET_DIR]
# standalone training on Ascend
bash scripts/run_standalone_train_ascend.sh [DEVICE_ID] [MINDRECORD_DATASET_PATH] [LOAD_CHECKPOINT_PATH](optional)
# distributed training on Ascend
bash scripts/run_distributed_train_ascend.sh [MINDRECORD_DATASET_PATH] [RANK_TABLE_FILE] [LOAD_CHECKPOINT_PATH](optional)
# eval on Ascend
bash scripts/run_standalone_eval_ascend.sh [DEVICE_ID] [RUN_MODE] [DATA_DIR] [LOAD_CHECKPOINT_PATH]
```
- running on ModelArts
If you want to run in modelarts, please check the official documentation of modelarts, and you can start training as follows
- Creating mindrecord dataset with single cards on ModelArts
```text
# (1) Upload the code folder to S3 bucket.
# (2) Upload the COCO2017 dataset to S3 bucket.
# (2) Click to "create task" on the website UI interface.
# (3) Set the code directory to "/{path}/centernet_det" on the website UI interface.
# (4) Set the startup file to /{path}/centernet_det/dataset.py" on the website UI interface.
# (5) Perform a or b.
# a. setting parameters in /{path}/centernet_det/default_config.yaml.
# 1. Set ”enable_modelarts: True“
# b. adding on the website UI interface.
# 1. Add ”enable_modelarts=True“
# (7) Check the "data storage location" on the website UI interface and set the "Dataset path" path.
# (8) Set the "Output file path" and "Job log path" to your path on the website UI interface.
# (9) Under the item "resource pool selection", select the specification of single cards.
# (10) Create your job.
```
- Training with single cards on ModelArts
```text
# (1) Upload the code folder to S3 bucket.
# (2) Click to "create task" on the website UI interface.
# (3) Set the code directory to "/{path}/centernet_det" on the website UI interface.
# (4) Set the startup file to /{path}/centernet_det/train.py" on the website UI interface.
# (5) Perform a or b.
# a. setting parameters in /{path}/centernet_det/default_config.yaml.
# 1. Set ”enable_modelarts: True“
# 2. Set “epoch_size: 130”
# 3. Set “distribute: 'true'”
# 4. Set “save_checkpoint_path: ./checkpoints”
# b. adding on the website UI interface.
# 1. Add ”enable_modelarts=True“
# 2. Add “epoch_size=130”
# 3. Add “distribute=true”
# 4. Add “save_checkpoint_path=./checkpoints”
# (6) Upload the mindrecord dataset to S3 bucket.
# (7) Check the "data storage location" on the website UI interface and set the "Dataset path" path.
# (8) Set the "Output file path" and "Job log path" to your path on the website UI interface.
# (9) Under the item "resource pool selection", select the specification of single cards.
# (10) Create your job.
```
- evaluating with single card on ModelArts
```text
# (1) Upload the code folder to S3 bucket.
# (2) Git clone https://github.com/xingyizhou/CenterNet.git on local, and put the folder 'CenterNet' under the folder 'centernet' on s3 bucket.
# (3) Click to "create task" on the website UI interface.
# (4) Set the code directory to "/{path}/centernet_det" on the website UI interface.
# (5) Set the startup file to /{path}/centernet_det/eval.py" on the website UI interface.
# (6) Perform a or b.
# a. setting parameters in /{path}/centernet_det/default_config.yaml.
# 1. Set ”enable_modelarts: True“
# 2. Set “run_mode: 'val'”
# 3. Set "load_checkpoint_path='/cache/checkpoint_path/model.ckpt'" on yaml file.
# 4. Set "checkpoint_url=/The path of checkpoint in S3/" on yaml file.
# b. adding on the website UI interface.
# 1. Add ”enable_modelarts=True“
# 2. Add “run_mode=val”
# 3. Add "load_checkpoint_path='/cache/checkpoint_path/model.ckpt'" on the website UI interface.
# 4. Add "checkpoint_url=/The path of checkpoint in S3/" on the website UI interface.
# (7) Upload the dataset(not mindrecord format) to S3 bucket.
# (8) Check the "data storage location" on the website UI interface and set the "Dataset path" path.
# (9) Set the "Output file path" and "Job log path" to your path on the website UI interface.
# (10) Under the item "resource pool selection", select the specification of a single card.
# (11) Create your job.
```
After installing MindSpore via the official website, you can start training and evaluation as follows: After installing MindSpore via the official website, you can start training and evaluation as follows:
Note: 1.the first run of training will generate the mindrecord file, which will take a long time. Note: 1.the first run of training will generate the mindrecord file, which will take a long time.
@ -121,20 +226,6 @@ Note: 1.the first run of training will generate the mindrecord file, which will
3.LOAD_CHECKPOINT_PATH is the pretrained checkpoint file directory, if no just set "" 3.LOAD_CHECKPOINT_PATH is the pretrained checkpoint file directory, if no just set ""
4.RUN_MODE support validation and testing, set to be "val"/"test" 4.RUN_MODE support validation and testing, set to be "val"/"test"
```shell
# create dataset in mindrecord format
bash scripts/convert_dataset_to_mindrecord.sh [COCO_DATASET_DIR] [MINDRECORD_DATASET_DIR]
# standalone training on Ascend
bash scripts/run_standalone_train_ascend.sh [DEVICE_ID] [MINDRECORD_DATASET_PATH] [LOAD_CHECKPOINT_PATH](optional)
# distributed training on Ascend
bash scripts/run_distributed_train_ascend.sh [MINDRECORD_DATASET_PATH] [RANK_TABLE_FILE] [LOAD_CHECKPOINT_PATH](optional)
# eval on Ascend
bash scripts/run_standalone_eval_ascend.sh [DEVICE_ID] [RUN_MODE] [DATA_DIR] [LOAD_CHECKPOINT_PATH]
```
# [Script Description](#contents) # [Script Description](#contents)
## [Script and Sample Code](#contents) ## [Script and Sample Code](#contents)
@ -145,29 +236,38 @@ bash scripts/run_standalone_eval_ascend.sh [DEVICE_ID] [RUN_MODE] [DATA_DIR] [LO
├── centernet_det ├── centernet_det
├── train.py // training scripts ├── train.py // training scripts
├── eval.py // testing and evaluation outputs ├── eval.py // testing and evaluation outputs
├── README.md // descriptions about CenterNet ├── export.py // convert mindspore model to mindir model
├── README.md // descriptions about centernet_det
├── default_config.yaml // parameter configuration
├── ascend310_infer // application for 310 inference
├── preprocess.py // preprocess scripts
├── postprocess.py // postprocess scripts
├── scripts ├── scripts
│ ├── ascend_distributed_launcher │ ├── ascend_distributed_launcher
│ │ ├──__init__.py │ │ ├── __init__.py
│ │ ├──hyper_parameter_config.ini // hyper parameter for distributed training │ │ ├── hyper_parameter_config.ini // hyper parameter for distributed training
│ │ ├──get_distribute_train_cmd.py // script for distributed training │ │ ├── get_distribute_train_cmd.py // script for distributed training
│ │ ├──README.md │ │ ├── README.md
│ ├──convert_dataset_to_mindrecord.sh // shell script for converting coco type dataset to mindrecord │ ├── convert_dataset_to_mindrecord.sh // shell script for converting coco type dataset to mindrecord
│ ├──run_standalone_train_ascend.sh // shell script for standalone training on ascend │ ├── run_standalone_train_ascend.sh // shell script for standalone training on ascend
│ ├──run_distributed_train_ascend.sh // shell script for distributed training on ascend │ ├── run_infer_310.sh // shell script for 310 inference on ascend
│ ├──run_standalone_eval_ascend.sh // shell script for standalone evaluation on ascend │ ├── run_distributed_train_ascend.sh // shell script for distributed training on ascend
│ ├── run_standalone_eval_ascend.sh // shell script for standalone evaluation on ascend
└── src └── src
├──__init__.py ├── model_utils
├──centernet_det.py // centernet networks, training entry │ ├── config.py // parsing parameter configuration file of "*.yaml"
├──dataset.py // generate dataloader and data processing entry │ ├── device_adapter.py // local or ModelArts training
├──config.py // centernet unique configs │ ├── local_adapter.py // get related environment variables on local
├──hccl_tools.py // generate hccl configuration │ └── moxing_adapter.py // get related environment variables abd transfer data on ModelArts
├──decode.py // decode the head features ├── __init__.py
├──hourglass.py // hourglass backbone ├── centernet_det.py // centernet networks, training entry
├──utils.py // auxiliary functions for train, to log and preload ├── dataset.py // generate dataloader and data processing entry
├──image.py // image preprocess functions ├── decode.py // decode the head features
├──post_process.py // post-process functions after decode in inference ├── hourglass.py // hourglass backbone
└──visual.py // visualization image, bbox, score and keypoints ├── image.py // image preprocess functions
├── post_process.py // post-process functions after decode in inference
├── utils.py // auxiliary functions for train, to log and preload
└── visual.py // visualization image, bbox, score and keypoints
``` ```
## [Script Parameters](#contents) ## [Script Parameters](#contents)
@ -202,7 +302,7 @@ usage: train.py [--device_target DEVICE_TARGET] [--distribute DISTRIBUTE]
[--save_result_dir SAVE_RESULT_DIR] [--save_result_dir SAVE_RESULT_DIR]
options: options:
--device_target device where the code will be implemented: "Ascend" | "CPU", default is "Ascend" --device_target device where the code will be implemented: "Ascend"
--distribute training by several devices: "true"(training by more than 1 device) | "false", default is "true" --distribute training by several devices: "true"(training by more than 1 device) | "false", default is "true"
--need profiler whether to use the profiling tools: "true" | "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 "" --profiler_path path to save the profiling results: PATH, default is ""
@ -233,7 +333,7 @@ usage: eval.py [--device_target DEVICE_TARGET] [--device_id N]
[--visual_image VISUAL_IMAGE] [--visual_image VISUAL_IMAGE]
[--enable_eval ENABLE_EVAL] [--save_result_dir SAVE_RESULT_DIR] [--enable_eval ENABLE_EVAL] [--save_result_dir SAVE_RESULT_DIR]
options: options:
--device_target device where the code will be implemented: "Ascend" | "CPU", default is "Ascend" --device_target device where the code will be implemented: "Ascend"
--device_id device id to run task, default is 0 --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 "" --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 "" --data_dir validation or test dataset dir: PATH, default is ""
@ -249,21 +349,20 @@ Parameters for training and evaluation can be set in file `config.py`.
#### Options #### Options
```text ```text
config for training. train_config.
batch_size batch size of input dataset: N, default is 12 batch_size: 12 // batch size of input dataset: N, default is 12
loss_scale_value initial value of loss scale: N, default is 1024 loss_scale_value: 1024 // initial value of loss scale: N, default is 1024
optimizer optimizer used in the network: Adam, default is Adam optimizer: 'Adam' // optimizer used in the network: Adam, default is Adam
lr_schedule schedules to get the learning rate lr_schedule: 'MultiDecay' // schedules to get the learning rate
``` ```
```text ```text
config for evaluation. config for evaluation.
SOFT_NMS nms after decode: True | False, default is True SOFT_NMS: True // nms after decode: True | False, default is True
keep_res keep original or fix resolution: True | False, default is True keep_res: True // keep original or fix resolution: True | False, default is True
multi_scales use multi-scales of image: List, default is [1.0] multi_scales: [1.0] // use multi-scales of image: List, default is [1.0]
pad pad size when keep original resolution, default is 127 K: 100 // number of bboxes to be computed by TopK, default is 100
K number of bboxes to be computed by TopK, default is 100 score_thresh: 0.3 // threshold of score when visualize image and annotation info,default is 0.3
score_thresh threshold of score when visualize image and annotation info,default is 0.4
``` ```
#### Parameters #### Parameters
@ -282,7 +381,6 @@ Parameters for dataset (Training/Evaluation):
flip_prop properbility of image flip during data augmenation: N, default is 0.5 flip_prop properbility of image flip during data augmenation: N, default is 0.5
color_aug color augmentation of RGB image, default is True color_aug color augmentation of RGB image, default is True
coco_classes name of categories in COCO2017 coco_classes name of categories in COCO2017
coco_class_name2id ID corresponding to the categories in COCO2017
mean mean value of RGB image mean mean value of RGB image
std variance of RGB image std variance of RGB image
eig_vec eigenvectors of RGB image eig_vec eigenvectors of RGB image
@ -290,8 +388,7 @@ Parameters for dataset (Training/Evaluation):
Parameters for network (Training/Evaluation): Parameters for network (Training/Evaluation):
down_ratio the ratio of input and output resolution during training,default is 4 down_ratio the ratio of input and output resolution during training,default is 4
last_level the last level in final upsampling, default is 6 num_stacks　　　　　　　　　　　　 the number of stacked hourglass network, default is 2
num_stacks　　　　　　　　　　　　　 the number of stacked hourglass network, default is 2
n the number of stacked hourglass modules, default is 5 n the number of stacked hourglass modules, default is 5
heads the number of heatmap,width and height,offset, default is {'hm': 80, 'wh': 2, 'reg': 2} heads the number of heatmap,width and height,offset, default is {'hm': 80, 'wh': 2, 'reg': 2}
cnv_dim the convolution of dimension, default is 256 cnv_dim the convolution of dimension, default is 256
@ -309,7 +406,7 @@ Parameters for network (Training/Evaluation):
Parameters for optimizer and learning rate: Parameters for optimizer and learning rate:
Adam: Adam:
weight_decay weight decay: Q,default is 0.0 weight_decay weight decay: Q
decay_filer lamda expression to specify which param will be decayed decay_filer lamda expression to specify which param will be decayed
PolyDecay: PolyDecay:
@ -331,7 +428,7 @@ Parameters for optimizer and learning rate:
Before your first training, convert coco type dataset to mindrecord files is needed to improve performance on host. Before your first training, convert coco type dataset to mindrecord files is needed to improve performance on host.
```bash ```shell
bash scripts/convert_dataset_to_mindrecord.sh /path/coco_dataset_dir /path/mindrecord_dataset_dir bash scripts/convert_dataset_to_mindrecord.sh /path/coco_dataset_dir /path/mindrecord_dataset_dir
``` ```
@ -341,13 +438,13 @@ The command above will run in the background, after converting mindrecord files
#### Running on Ascend #### Running on Ascend
```bash ```shell
bash scripts/run_distributed_train_ascend.sh /path/mindrecord_dataset /path/hccl.json /path/load_ckpt(optional) bash scripts/run_distributed_train_ascend.sh /path/mindrecord_dataset /path/hccl.json /path/load_ckpt(optional)
``` ```
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: 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:
```bash ```text
# grep "epoch" training_log.txt # grep "epoch" training_log.txt
epoch: 128, current epoch percent: 1.000, step: 157509, outputs are (Tensor(shape=[], dtype=Float32, value= 1.54529), Tensor(shape=[], dtype=Bool, value= False), Tensor(shape=[], dtype=Float32, value= 1024)) epoch: 128, current epoch percent: 1.000, step: 157509, outputs are (Tensor(shape=[], dtype=Float32, value= 1.54529), Tensor(shape=[], dtype=Bool, value= False), Tensor(shape=[], dtype=Float32, value= 1024))
epoch time: 1211875.286 ms, per step time: 992.527 ms epoch time: 1211875.286 ms, per step time: 992.527 ms
@ -360,13 +457,10 @@ epoch time: 1214703.313 ms, per step time: 994.843 ms
### Testing and Evaluation ### Testing and Evaluation
```bash ```shell
# 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). # 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).
# On Ascend # On Ascend
bash scripts/run_standalone_eval_ascend.sh device_id val(or test) /path/coco_dataset /path/load_ckpt bash scripts/run_standalone_eval_ascend.sh device_id val(or test) /path/coco_dataset /path/load_ckpt
# On CPU
bash scripts/run_standalone_eval_cpu.sh val(or test) /path/coco_dataset /path/load_ckpt
``` ```
you can see the MAP result below as below: you can see the MAP result below as below:
@ -387,30 +481,89 @@ overall performance on coco2017 validation dataset
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.764 Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.764
``` ```
## [Convert Process](#contents) ## [Inference Process](#contents)
### Convert ### Convert
If you want to infer the network on Ascend 310, you should convert the model to AIR: If you want to infer the network on Ascend 310, you should convert the model to MINDIR:
```python - Export on local
python export.py [DEVICE_ID]
```text
python export.py --device_id [DEVICE_ID] --export_format MINDIR --export_load_ckpt [CKPT_FILE__PATH] --export_name [EXPORT_FILE_NAME]
```
- Export on ModelArts (If you want to run in modelarts, please check the official documentation of [modelarts](https://support.huaweicloud.com/modelarts/), and you can start as follows)
```text
# (1) Upload the code folder to S3 bucket.
# (2) Click to "create training task" on the website UI interface.
# (3) Set the code directory to "/{path}/centernet_det" on the website UI interface.
# (4) Set the startup file to /{path}/centernet_det/export.py" on the website UI interface.
# (5) Perform a or b.
# a. setting parameters in /{path}/centernet_det/default_config.yaml.
# 1. Set ”enable_modelarts: True“
# 2. Set “export_load_ckpt: ./{path}/*.ckpt”('export_load_ckpt' indicates the path of the weight file to be exported relative to the file `export.py`, and the weight file must be included in the code directory.)
# 3. Set ”export_name: centernet_det“
# 4. Set ”export_format：MINDIR“
# b. adding on the website UI interface.
# 1. Add ”enable_modelarts=True“
# 2. Add “export_load_ckpt=./{path}/*.ckpt”('export_load_ckpt' indicates the path of the weight file to be exported relative to the file `export.py`, and the weight file must be included in the code directory.)
# 3. Add ”export_name=centernet_det“
# 4. Add ”export_format=MINDIR“
# (7) Check the "data storage location" on the website UI interface and set the "Dataset path" path (This step is useless, but necessary.).
# (8) Set the "Output file path" and "Job log path" to your path on the website UI interface.
# (9) Under the item "resource pool selection", select the specification of a single card.
# (10) Create your job.
# You will see centernet.mindir under {Output file path}.
```
### Infer on Ascend310
Before performing inference, the mindir file must be exported by export.py script. We only provide an example of inference using MINDIR model. Current batch_size can only be set to 1.
```shell
#Ascend310 inference
bash run_infer_310.sh [MINDIR_PATH] [DATASET_PATH] [PREPROCESS_IMAGES] [DEVICE_ID]
```
- `PREPROCESS_IMAGES` Weather need preprocess or not, it's value must be in [y, n]
### Result
Inference result is saved in current path, you can find result like this in acc.log file.Since the input images are fixed shape on Ascend 310, all accuracy will be lower than that on Ascend 910.
```log
#acc.log
=============coco2017 310 infer reulst=========
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.410
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.600
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.440
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.213
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.437
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.567
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.339
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.543
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.572
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.342
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.620
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.764
``` ```
# [Model Description](#contents) # [Model Description](#contents)
## [Performance](#contents) ## [Performance](#contents)
### Training Performance On Ascend ### Training Performance On Ascend 910
CenterNet on 11.8K images(The annotation and data format must be the same as coco) CenterNet on 11.8K images(The annotation and data format must be the same as coco)
| Parameters | CenterNet | | Parameters | CenterNet_Hourglass |
| -------------------------- | ---------------------------------------------------------------| | -------------------------- | ---------------------------------------------------------------|
| Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory 755G; OS Euler2.8 | | Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory, 755G |
| uploaded Date | 3/27/2021 (month/day/year) | | uploaded Date | 3/27/2021 (month/day/year) |
| MindSpore Version | 1.1.0 | | MindSpore Version | 1.1.0 |
| Dataset | 11.8K images | | Dataset | COCO2017 |
| Training Parameters | 8p, epoch=130, steps=158730, batch_size = 12, lr=2.4e-4 | | Training Parameters | 8p, epoch=130, steps=158730, batch_size = 12, lr=2.4e-4 |
| Optimizer | Adam | | Optimizer | Adam |
| Loss Function | Focal Loss, L1 Loss, RegLoss | | Loss Function | Focal Loss, L1 Loss, RegLoss |
@ -420,22 +573,36 @@ CenterNet on 11.8K images(The annotation and data format must be the same as coc
| Total time: training | 8p: 44 h | | Total time: training | 8p: 44 h |
| Total time: evaluation | keep res: test 1h, val 0.25h; fix res: test 40 min, val 8 min| | Total time: evaluation | keep res: test 1h, val 0.25h; fix res: test 40 min, val 8 min|
| Checkpoint | 2.3G (.ckpt file) | | Checkpoint | 2.3G (.ckpt file) |
| Scripts | <https://gitee.com/mindspore/mindspore/tree/master/model_zoo/research/cv/centernet> | | Scripts | [centernet_det script](https://gitee.com/mindspore/mindspore/tree/master/model_zoo/research/cv/centernet_det) |
### Inference Performance On Ascend ### Inference Performance On Ascend 910
CenterNet on validation(5K images) and test-dev(40K images) CenterNet on validation(5K images) and test-dev(40K images)
| Parameters | CenterNet | | Parameters | CenterNet_Hourglass |
| -------------------------- | ----------------------------------------------------------------| | -------------------------- | ----------------------------------------------------------------|
| Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory 755G; OS Euler2.8 | | Resource | Ascend 910; CPU 2.60GHz, 192cores; Memory, 755G |
| uploaded Date | 3/27/2021 (month/day/year) | | uploaded Date | 3/27/2021 (month/day/year) |
| MindSpore Version | 1.1.0 | | MindSpore Version | 1.1.0 |
| Dataset | 5K images(val), 40K images(test-dev) | | Dataset | COCO2017 |
| batch_size | 1 | | batch_size | 1 |
| outputs | boxes and keypoints position and scores | | outputs | mAP |
| Accuracy(validation) | MAP: 41.5%, AP50: 60.4%, AP75: 44.7%, Medium: 45.7%, Large: 53.6%| | Accuracy(validation) | MAP: 41.5%, AP50: 60.4%, AP75: 44.7%, Medium: 45.7%, Large: 53.6%|
### Inference Performance On Ascend 310
CenterNet on validation(5K images)
| Parameters | CenterNet_Hourglass |
| -------------------------- | ----------------------------------------------------------------|
| Resource | Ascend 310; CentOS 3.10 |
| uploaded Date | 8/31/2021 (month/day/year) |
| MindSpore Version | 1.4.0 |
| Dataset | COCO2017 |
| batch_size | 1 |
| outputs | mAP |
| Accuracy(validation) | MAP: 41.0%, AP50: 60.0%, AP75: 44.0%, Medium: 43.7%, Large: 56.7%|
# [Description of Random Situation](#contents) # [Description of Random Situation](#contents)
In run_distributed_train_ascend.sh, we set do_shuffle to True to shuffle the dataset by default. In run_distributed_train_ascend.sh, we set do_shuffle to True to shuffle the dataset by default.
@ -444,3 +611,7 @@ In train.py, we set a random seed to make sure that each node has the same initi
# [ModelZoo Homepage](#contents) # [ModelZoo Homepage](#contents)
Please check the official [homepage](https://gitee.com/mindspore/mindspore/tree/master/model_zoo). Please check the official [homepage](https://gitee.com/mindspore/mindspore/tree/master/model_zoo).
# FAQ
First refer to [ModelZoo FAQ](https://gitee.com/mindspore/mindspore/tree/master/model_zoo#FAQ) to find some common public questions.

15
model_zoo/research/cv/centernet_det/ascend310_infer/CMakeLists.txt Normal file
View File

@ -0,0 +1,15 @@
cmake_minimum_required(VERSION 3.14.1)
project(Ascend310Infer)
add_compile_definitions(_GLIBCXX_USE_CXX11_ABI=0)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O2 -g -std=c++17 -Werror -Wall -fPIE -Wl,--allow-shlib-undefined")
set(PROJECT_SRC_ROOT ${CMAKE_CURRENT_LIST_DIR}/)
option(MINDSPORE_PATH "mindspore install path" "")
include_directories(${MINDSPORE_PATH})
include_directories(${MINDSPORE_PATH}/include)
include_directories(${PROJECT_SRC_ROOT})
find_library(MS_LIB libmindspore.so ${MINDSPORE_PATH}/lib)
file(GLOB_RECURSE MD_LIB ${MINDSPORE_PATH}/_c_dataengine*)
add_executable(main src/main.cc src/utils.cc)
target_link_libraries(main ${MS_LIB} ${MD_LIB} gflags)

29
model_zoo/research/cv/centernet_det/ascend310_infer/build.sh Normal file
View File

@ -0,0 +1,29 @@
#!/bin/bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
if [ -d out ]; then
rm -rf out
fi
mkdir out
cd out || exit
if [ -f "Makefile" ]; then
make clean
fi
cmake .. \
-DMINDSPORE_PATH="`pip3.7 show mindspore-ascend | grep Location | awk '{print $2"/mindspore"}' | xargs realpath`"
make

32
model_zoo/research/cv/centernet_det/ascend310_infer/inc/utils.h Normal file
View File

@ -0,0 +1,32 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#ifndef MINDSPORE_INFERENCE_UTILS_H_
#define MINDSPORE_INFERENCE_UTILS_H_
#include <sys/stat.h>
#include <dirent.h>
#include <vector>
#include <string>
#include <memory>
#include "include/api/types.h"
std::vector<std::string> GetAllFiles(std::string_view dirName);
DIR *OpenDir(std::string_view dirName);
std::string RealPath(std::string_view path);
mindspore::MSTensor ReadFileToTensor(const std::string &file);
int WriteResult(const std::string& imageFile, const std::vector<mindspore::MSTensor> &outputs);
#endif

134
model_zoo/research/cv/centernet_det/ascend310_infer/src/main.cc Normal file
View File

@ -0,0 +1,134 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include <sys/time.h>
#include <gflags/gflags.h>
#include <dirent.h>
#include <iostream>
#include <string>
#include <algorithm>
#include <iosfwd>
#include <vector>
#include <fstream>
#include <sstream>
#include "include/api/model.h"
#include "include/api/context.h"
#include "include/api/types.h"
#include "include/api/serialization.h"
#include "include/dataset/execute.h"
#include "include/dataset/vision.h"
#include "inc/utils.h"
using mindspore::Context;
using mindspore::Serialization;
using mindspore::Model;
using mindspore::Status;
using mindspore::MSTensor;
using mindspore::dataset::Execute;
using mindspore::ModelType;
using mindspore::GraphCell;
using mindspore::kSuccess;
DEFINE_string(mindir_path, "", "mindir path");
DEFINE_string(input0_path, ".", "input0 path");
DEFINE_int32(device_id, 0, "device id");
int main(int argc, char **argv) {
gflags::ParseCommandLineFlags(&argc, &argv, true);
if (RealPath(FLAGS_mindir_path).empty()) {
std::cout << "Invalid mindir" << std::endl;
return 1;
}
auto context = std::make_shared<Context>();
auto ascend310 = std::make_shared<mindspore::Ascend310DeviceInfo>();
ascend310->SetDeviceID(FLAGS_device_id);
ascend310->SetPrecisionMode("allow_fp32_to_fp16");
ascend310->SetOpSelectImplMode("high_precision");
ascend310->SetBufferOptimizeMode("off_optimize");
context->MutableDeviceInfo().push_back(ascend310);
mindspore::Graph graph;
Serialization::Load(FLAGS_mindir_path, ModelType::kMindIR, &graph);
Model model;
Status ret = model.Build(GraphCell(graph), context);
if (ret != kSuccess) {
std::cout << "ERROR: Build failed." << std::endl;
return 1;
}
std::vector<MSTensor> model_inputs = model.GetInputs();
if (model_inputs.empty()) {
std::cout << "Invalid model, inputs is empty." << std::endl;
return 1;
}
auto input0_files = GetAllFiles(FLAGS_input0_path);
if (input0_files.empty()) {
std::cout << "ERROR: input data empty." << std::endl;
return 1;
}
std::map<double, double> costTime_map;
size_t size = input0_files.size();
for (size_t i = 0; i < size; ++i) {
struct timeval start = {0};
struct timeval end = {0};
double startTimeMs;
double endTimeMs;
std::vector<MSTensor> inputs;
std::vector<MSTensor> outputs;
std::cout << "Start predict input files:" << input0_files[i] << std::endl;
auto input0 = ReadFileToTensor(input0_files[i]);
inputs.emplace_back(model_inputs[0].Name(), model_inputs[0].DataType(), model_inputs[0].Shape(),
input0.Data().get(), input0.DataSize());
gettimeofday(&start, nullptr);
ret = model.Predict(inputs, &outputs);
gettimeofday(&end, nullptr);
if (ret != kSuccess) {
std::cout << "Predict " << input0_files[i] << " failed." << std::endl;
return 1;
}
startTimeMs = (1.0 * start.tv_sec * 1000000 + start.tv_usec) / 1000;
endTimeMs = (1.0 * end.tv_sec * 1000000 + end.tv_usec) / 1000;
costTime_map.insert(std::pair<double, double>(startTimeMs, endTimeMs));
WriteResult(input0_files[i], outputs);
}
double average = 0.0;
int inferCount = 0;
for (auto iter = costTime_map.begin(); iter != costTime_map.end(); iter++) {
double diff = 0.0;
diff = iter->second - iter->first;
average += diff;
inferCount++;
}
average = average / inferCount;
std::stringstream timeCost;
timeCost << "NN inference cost average time: " << average << " ms of infer_count " << inferCount << std::endl;
std::cout << "NN inference cost average time: " << average << "ms of infer_count " << inferCount << std::endl;
std::string fileName = "./time_Result" + std::string("/test_perform_static.txt");
std::ofstream fileStream(fileName.c_str(), std::ios::trunc);
fileStream << timeCost.str();
fileStream.close();
costTime_map.clear();
return 0;
}

128
model_zoo/research/cv/centernet_det/ascend310_infer/src/utils.cc Normal file
View File

@ -0,0 +1,128 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include <fstream>
#include <algorithm>
#include <iostream>
#include "inc/utils.h"
using mindspore::MSTensor;
using mindspore::DataType;
std::vector<std::string> GetAllFiles(std::string_view dirName) {
struct dirent *filename;
DIR *dir = OpenDir(dirName);
if (dir == nullptr) {
return {};
}
std::vector<std::string> res;
while ((filename = readdir(dir)) != nullptr) {
std::string dName = std::string(filename->d_name);
if (dName == "." || dName == ".." || filename->d_type != DT_REG) {
continue;
}
res.emplace_back(std::string(dirName) + "/" + filename->d_name);
}
std::sort(res.begin(), res.end());
for (auto &f : res) {
std::cout << "image file: " << f << std::endl;
}
return res;
}
int WriteResult(const std::string& imageFile, const std::vector<MSTensor> &outputs) {
std::string homePath = "./result_Files";
for (size_t i = 0; i < outputs.size(); ++i) {
size_t outputSize;
std::shared_ptr<const void> netOutput;
netOutput = outputs[i].Data();
outputSize = outputs[i].DataSize();
int pos = imageFile.rfind('/');
std::string fileName(imageFile, pos + 1);
fileName.replace(fileName.find('.'), fileName.size() - fileName.find('.'), '_' + std::to_string(i) + ".bin");
std::string outFileName = homePath + "/" + fileName;
FILE * outputFile = fopen(outFileName.c_str(), "wb");
fwrite(netOutput.get(), outputSize, sizeof(char), outputFile);
fclose(outputFile);
outputFile = nullptr;
}
return 0;
}
mindspore::MSTensor ReadFileToTensor(const std::string &file) {
if (file.empty()) {
std::cout << "Pointer file is nullptr" << std::endl;
return mindspore::MSTensor();
}
std::ifstream ifs(file);
if (!ifs.good()) {
std::cout << "File: " << file << " is not exist" << std::endl;
return mindspore::MSTensor();
}
if (!ifs.is_open()) {
std::cout << "File: " << file << "open failed" << std::endl;
return mindspore::MSTensor();
}
ifs.seekg(0, std::ios::end);
size_t size = ifs.tellg();
mindspore::MSTensor buffer(file, mindspore::DataType::kNumberTypeUInt8, {static_cast<int64_t>(size)}, nullptr, size);
ifs.seekg(0, std::ios::beg);
ifs.read(reinterpret_cast<char *>(buffer.MutableData()), size);
ifs.close();
return buffer;
}
DIR *OpenDir(std::string_view dirName) {
if (dirName.empty()) {
std::cout << " dirName is null ! " << std::endl;
return nullptr;
}
std::string realPath = RealPath(dirName);
struct stat s;
lstat(realPath.c_str(), &s);
if (!S_ISDIR(s.st_mode)) {
std::cout << "dirName is not a valid directory !" << std::endl;
return nullptr;
}
DIR *dir;
dir = opendir(realPath.c_str());
if (dir == nullptr) {
std::cout << "Can not open dir " << dirName << std::endl;
return nullptr;
}
std::cout << "Successfully opened the dir " << dirName << std::endl;
return dir;
}
std::string RealPath(std::string_view path) {
char realPathMem[PATH_MAX] = {0};
char *realPathRet = nullptr;
realPathRet = realpath(path.data(), realPathMem);
if (realPathRet == nullptr) {
std::cout << "File: " << path << " is not exist.";
return "";
}
std::string realPath(realPathMem);
std::cout << path << " realpath is: " << realPath << std::endl;
return realPath;
}

275
model_zoo/research/cv/centernet_det/default_config.yaml Normal file
View File

@ -0,0 +1,275 @@
# Builtin Configurations(DO NOT CHANGE THESE CONFIGURATIONS unless you know exactly what you are doing)
enable_modelarts: False
# Url for modelarts
data_url: ""
train_url: ""
checkpoint_url: ""
# Path for local
data_path: "/cache/data"
output_path: "/cache/train"
load_path: "/cache/checkpoint_path"
device_target: "Ascend"
enable_profiling: False
# ==============================================================================
# prepare *.mindrecord* data
coco_data_dir: ""
mindrecord_dir: "" # also used by train.py
mindrecord_prefix: "coco_det.train.mind"
# train related
save_result_dir: ""
device_id: 0
device_num: 1
distribute: 'false'
need_profiler: "false"
profiler_path: "./profiler"
epoch_size: 1
train_steps: -1
enable_save_ckpt: "true"
do_shuffle: "true"
enable_data_sink: "true"
data_sink_steps: -1
save_checkpoint_path: ""
load_checkpoint_path: ""
save_checkpoint_steps: 1221
save_checkpoint_num: 1
# val related
data_dir: ""
run_mode: "test"
enable_eval: "true"
visual_image: "false"
# export related
export_load_ckpt: ''
export_format: ''
export_name: ''
# 310 infer
val_data_dir: ''
predict_dir: ''
result_path: ''
label_path: ''
meta_path: ''
save_path: ''
dataset_config:
num_classes: 80
max_objs: 128
input_res: [512, 512]
output_res: [128, 128]
rand_crop: True
shift: 0.1
scale: 0.4
down_ratio: 4
aug_rot: 0.0
rotate: 0
flip_prop: 0.5
color_aug: True
coco_classes: ['background', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
'train', 'truck', 'boat', 'traffic light', 'fire hydrant',
'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog',
'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra',
'giraffe', 'backpack', 'umbrella', 'handbag', 'tie',
'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball',
'kite', 'baseball bat', 'baseball glove', 'skateboard',
'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup',
'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza',
'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed',
'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote',
'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink',
'refrigerator', 'book', 'clock', 'vase', 'scissors',
'teddy bear', 'hair drier', 'toothbrush']
mean: np.array([0.40789654, 0.44719302, 0.47026115], dtype=np.float32)
std: np.array([0.28863828, 0.27408164, 0.27809835], dtype=np.float32)
eig_val: np.array([0.2141788, 0.01817699, 0.00341571], dtype=np.float32)
eig_vec: np.array([[-0.58752847, -0.69563484, 0.41340352],
[-0.5832747, 0.00994535, -0.81221408],
[-0.56089297, 0.71832671, 0.41158938]], dtype=np.float32)
net_config:
num_stacks: 2
down_ratio: 4
num_classes: 80
n: 5
cnv_dim: 256
modules: [2, 2, 2, 2, 2, 4]
dims: [256, 256, 384, 384, 384, 512]
dense_wh: False
norm_wh: False
cat_spec_wh: False
reg_offset: True
hm_weight: 1
off_weight: 1
wh_weight: 0.1
mse_loss: False
reg_loss: 'l1'
train_config:
batch_size: 12
loss_scale_value: 1024
optimizer: 'Adam'
lr_schedule: 'MultiDecay'
Adam:
weight_decay: 0.0
decay_filter: "lambda x: x.name.endswith('.bias') or x.name.endswith('.beta') or x.name.endswith('.gamma')"
PolyDecay:
learning_rate: 0.00024 # 2.4e-4
end_learning_rate: 0.0000005 # 5e-7
power: 5.0
eps: 0.0000001 # 1e-7
warmup_steps: 2000
MultiDecay:
learning_rate: 0.00024 # 2.4e-4
eps: 0.0000001 # 1e-7
warmup_steps: 2000
multi_epochs: [105, 125]
factor: 10
eval_config:
SOFT_NMS: True
keep_res: True
multi_scales: [1.0]
K: 100
score_thresh: 0.3
valid_ids: [
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13,
14, 15, 16, 17, 18, 19, 20, 21, 22, 23,
24, 25, 27, 28, 31, 32, 33, 34, 35, 36,
37, 38, 39, 40, 41, 42, 43, 44, 46, 47,
48, 49, 50, 51, 52, 53, 54, 55, 56, 57,
58, 59, 60, 61, 62, 63, 64, 65, 67, 70,
72, 73, 74, 75, 76, 77, 78, 79, 80, 81,
82, 84, 85, 86, 87, 88, 89, 90]
color_list: [0.000, 0.800, 1.000,
0.850, 0.325, 0.098,
0.929, 0.694, 0.125,
0.494, 0.184, 0.556,
0.466, 0.674, 0.188,
0.301, 0.745, 0.933,
0.635, 0.078, 0.184,
0.300, 0.300, 0.300,
0.600, 0.600, 0.600,
1.000, 0.000, 0.000,
1.000, 0.500, 0.000,
0.749, 0.749, 0.000,
0.000, 1.000, 0.000,
0.000, 0.000, 1.000,
0.667, 0.000, 1.000,
0.333, 0.333, 0.000,
0.333, 0.667, 0.333,
0.333, 1.000, 0.000,
0.667, 0.333, 0.000,
0.667, 0.667, 0.000,
0.667, 1.000, 0.000,
1.000, 0.333, 0.000,
1.000, 0.667, 0.000,
1.000, 1.000, 0.000,
0.000, 0.333, 0.500,
0.000, 0.667, 0.500,
0.000, 1.000, 0.500,
0.333, 0.000, 0.500,
0.333, 0.333, 0.500,
0.333, 0.667, 0.500,
0.333, 1.000, 0.500,
0.667, 0.000, 0.500,
0.667, 0.333, 0.500,
0.667, 0.667, 0.500,
0.667, 1.000, 0.500,
1.000, 0.000, 0.500,
1.000, 0.333, 0.500,
1.000, 0.667, 0.500,
1.000, 1.000, 0.500,
0.000, 0.333, 1.000,
0.000, 0.667, 1.000,
0.000, 1.000, 1.000,
0.333, 0.000, 1.000,
0.333, 0.333, 1.000,
0.333, 0.667, 1.000,
0.333, 1.000, 1.000,
0.667, 0.000, 1.000,
0.667, 0.333, 1.000,
0.667, 0.667, 1.000,
0.667, 1.000, 1.000,
1.000, 0.000, 1.000,
1.000, 0.333, 1.000,
1.000, 0.667, 1.000,
0.167, 0.800, 0.000,
0.333, 0.000, 0.000,
0.500, 0.000, 0.000,
0.667, 0.000, 0.000,
0.833, 0.000, 0.000,
1.000, 0.000, 0.000,
0.000, 0.667, 0.400,
0.000, 0.333, 0.000,
0.000, 0.500, 0.000,
0.000, 0.667, 0.000,
0.000, 0.833, 0.000,
0.000, 1.000, 0.000,
0.000, 0.000, 0.167,
0.000, 0.000, 0.333,
0.000, 0.000, 0.500,
0.000, 0.000, 0.667,
0.000, 0.000, 0.833,
0.000, 0.000, 1.000,
0.000, 0.200, 0.800,
0.143, 0.143, 0.543,
0.286, 0.286, 0.286,
0.429, 0.429, 0.429,
0.571, 0.571, 0.571,
0.714, 0.714, 0.714,
0.857, 0.857, 0.857,
0.000, 0.447, 0.741,
0.50, 0.5, 0]
export_config:
input_res: dataset_config.input_res
ckpt_file: "./ckpt_file.ckpt"
export_format: "MINDIR"
export_name: "CenterNet_ObjectDetection"
---
# Help description for each configuration
enable_modelarts: "Whether training on modelarts, default: False"
data_url: "Url for modelarts"
train_url: "Url for modelarts"
data_path: "The location of the input data."
output_path: "The location of the output file."
device_target: "Running platform, default is Ascend."
enable_profiling: 'Whether enable profiling while training, default: False'
distribute: "Run distribute, default is false."
need_profiler: "Profiling to parsing runtime info, default is false."
profiler_path: "The path to save profiling data"
epoch_size: "Epoch size, default is 1."
train_steps: "Training Steps, default is -1, i.e. run all steps according to epoch number."
device_id: "Device id, default is 0."
device_num: "Use device nums, default is 1."
enable_save_ckpt: "Enable save checkpoint, default is true."
do_shuffle: "Enable shuffle for dataset, default is true."
enable_data_sink: "Enable data sink, default is true."
data_sink_steps: "Sink steps for each epoch, default is 1."
save_checkpoint_path: "Save checkpoint path"
load_checkpoint_path: "Load checkpoint file path"
save_checkpoint_steps: "Save checkpoint steps, default is 1000."
save_checkpoint_num: "Save checkpoint numbers, default is 1."
mindrecord_dir: "Mindrecord dataset files directory"
mindrecord_prefix: "Prefix of MindRecord dataset filename."
visual_image: "Visulize the ground truth and predicted image"
save_result_dir: "The path to save the predict results"
data_dir: "Dataset directory, the absolute image path is joined by the data_dir, and the relative path in anno_path"
run_mode: "test or validation, default is test."
enable_eval: "Whether evaluate accuracy after prediction"
---
device_target: ['Ascend']
distribute: ["true", "false"]
need_profiler: ["true", "false"]
enable_save_ckpt: ["true", "false"]
do_shuffle: ["true", "false"]
enable_data_sink: ["true", "false"]
export_format: ["MINDIR"]

75
model_zoo/research/cv/centernet_det/eval.py Executable file → Normal file
View File

@ -20,7 +20,6 @@ import os
import time import time
import copy import copy
import json import json
import argparse
import cv2 import cv2
from pycocotools.coco import COCO from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval from pycocotools.cocoeval import COCOeval
@ -31,53 +30,62 @@ import mindspore.log as logger
from src import COCOHP, CenterNetDetEval from src import COCOHP, CenterNetDetEval
from src import convert_eval_format, post_process, merge_outputs from src import convert_eval_format, post_process, merge_outputs
from src import visual_image from src import visual_image
from src.config import dataset_config, net_config, eval_config from src.model_utils.config import config, dataset_config, net_config, eval_config
from src.model_utils.moxing_adapter import moxing_wrapper
from src.model_utils.device_adapter import get_device_id
_current_dir = os.path.dirname(os.path.realpath(__file__)) _current_dir = os.path.dirname(os.path.realpath(__file__))
parser = argparse.ArgumentParser(description='CenterNet evaluation')
parser.add_argument('--device_target', type=str, default='Ascend', choices=['Ascend', 'CPU'],
help='device where the code will be implemented. (Default: Ascend)')
parser.add_argument("--device_id", type=int, default=0, help="Device id, default is 0.")
parser.add_argument("--load_checkpoint_path", type=str, default="", help="Load checkpoint file path")
parser.add_argument("--data_dir", type=str, default="", help="Dataset directory, "
"the absolute image path is joined by the data_dir "
"and the relative path in anno_path")
parser.add_argument("--run_mode", type=str, default="val", help="test or validation, default is validation.")
parser.add_argument("--visual_image", type=str, default="true", help="Visulize the ground truth and predicted image")
parser.add_argument("--enable_eval", type=str, default="true", help="Whether evaluate accuracy after prediction")
parser.add_argument("--save_result_dir", type=str, default="", help="The path to save the predict results")
args_opt = parser.parse_args() def modelarts_pre_process():
"""modelarts pre process function."""
try:
from nms import soft_nms
print('soft_nms_attributes: {}'.format(soft_nms.__dir__()))
except ImportError:
print('NMS not installed! trying installing...\n')
cur_path = os.path.dirname(os.path.abspath(__file__))
os.system('cd {}/CenterNet/src/lib/external/ && make && python setup.py install && cd - '.format(cur_path))
try:
from nms import soft_nms
print('soft_nms_attributes: {}'.format(soft_nms.__dir__()))
except ImportError:
print('Installing failed! check if the folder "./CenterNet" exists.')
else:
print('Install nms successfully')
config.data_dir = config.data_path
config.load_checkpoint_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), config.load_checkpoint_path)
@moxing_wrapper(pre_process=modelarts_pre_process)
def predict(): def predict():
''' '''
Predict function Predict function
''' '''
context.set_context(mode=context.GRAPH_MODE, device_target=args_opt.device_target) context.set_context(mode=context.GRAPH_MODE, device_target=config.device_target)
if args_opt.device_target == "Ascend": if config.device_target == "Ascend":
context.set_context(device_id=args_opt.device_id) context.set_context(device_id=get_device_id())
enable_nms_fp16 = True
else:
enable_nms_fp16 = False enable_nms_fp16 = False
else:
enable_nms_fp16 = True
logger.info("Begin creating {} dataset".format(args_opt.run_mode)) logger.info("Begin creating {} dataset".format(config.run_mode))
coco = COCOHP(dataset_config, run_mode=args_opt.run_mode, net_opt=net_config, coco = COCOHP(dataset_config, run_mode=config.run_mode, net_opt=net_config,
enable_visual_image=(args_opt.visual_image == "true"), save_path=args_opt.save_result_dir,) enable_visual_image=config.visual_image, save_path=config.save_result_dir,)
coco.init(args_opt.data_dir, keep_res=eval_config.keep_res) coco.init(config.data_dir, keep_res=eval_config.keep_res)
dataset = coco.create_eval_dataset() dataset = coco.create_eval_dataset()
net_for_eval = CenterNetDetEval(net_config, eval_config.K, enable_nms_fp16) net_for_eval = CenterNetDetEval(net_config, eval_config.K, enable_nms_fp16)
net_for_eval.set_train(False) net_for_eval.set_train(False)
param_dict = load_checkpoint(args_opt.load_checkpoint_path) param_dict = load_checkpoint(config.load_checkpoint_path)
load_param_into_net(net_for_eval, param_dict) load_param_into_net(net_for_eval, param_dict)
# save results # save results
save_path = os.path.join(args_opt.save_result_dir, args_opt.run_mode) save_path = os.path.join(config.save_result_dir, config.run_mode)
if not os.path.exists(save_path): if not os.path.exists(save_path):
os.makedirs(save_path) os.makedirs(save_path)
if args_opt.visual_image == "true": if config.visual_image == "true":
save_pred_image_path = os.path.join(save_path, "pred_image") save_pred_image_path = os.path.join(save_path, "pred_image")
if not os.path.exists(save_pred_image_path): if not os.path.exists(save_pred_image_path):
os.makedirs(save_pred_image_path) os.makedirs(save_pred_image_path)
@ -119,10 +127,10 @@ def predict():
pred_annos["images"].append(image_info) pred_annos["images"].append(image_info)
for image_anno in pred_json["annotations"]: for image_anno in pred_json["annotations"]:
pred_annos["annotations"].append(image_anno) pred_annos["annotations"].append(image_anno)
if args_opt.visual_image == "true": if config.visual_image == "true":
img_file = os.path.join(coco.image_path, gt_image_info[0]['file_name']) img_file = os.path.join(coco.image_path, gt_image_info[0]['file_name'])
gt_image = cv2.imread(img_file) gt_image = cv2.imread(img_file)
if args_opt.run_mode != "test": if config.run_mode != "test":
annos = coco.coco.loadAnns(coco.anns[image_id]) annos = coco.coco.loadAnns(coco.anns[image_id])
visual_image(copy.deepcopy(gt_image), annos, save_gt_image_path, visual_image(copy.deepcopy(gt_image), annos, save_gt_image_path,
score_threshold=eval_config.score_thresh) score_threshold=eval_config.score_thresh)
@ -130,15 +138,15 @@ def predict():
visual_image(gt_image, anno, save_pred_image_path, score_threshold=eval_config.score_thresh) visual_image(gt_image, anno, save_pred_image_path, score_threshold=eval_config.score_thresh)
# save results # save results
save_path = os.path.join(args_opt.save_result_dir, args_opt.run_mode) save_path = os.path.join(config.save_result_dir, config.run_mode)
if not os.path.exists(save_path): if not os.path.exists(save_path):
os.makedirs(save_path) os.makedirs(save_path)
pred_anno_file = os.path.join(save_path, '{}_pred_result.json').format(args_opt.run_mode) pred_anno_file = os.path.join(save_path, '{}_pred_result.json').format(config.run_mode)
json.dump(pred_annos, open(pred_anno_file, 'w')) json.dump(pred_annos, open(pred_anno_file, 'w'))
pred_res_file = os.path.join(save_path, '{}_pred_eval.json').format(args_opt.run_mode) pred_res_file = os.path.join(save_path, '{}_pred_eval.json').format(config.run_mode)
json.dump(pred_annos["annotations"], open(pred_res_file, 'w')) json.dump(pred_annos["annotations"], open(pred_res_file, 'w'))
if args_opt.run_mode != "test" and args_opt.enable_eval: if config.run_mode != "test" and config.enable_eval:
run_eval(coco.annot_path, pred_res_file) run_eval(coco.annot_path, pred_res_file)
@ -151,5 +159,6 @@ def run_eval(gt_anno, pred_anno):
coco_eval.accumulate() coco_eval.accumulate()
coco_eval.summarize() coco_eval.summarize()
if __name__ == "__main__": if __name__ == "__main__":
predict() predict()

27
model_zoo/research/cv/centernet_det/export.py Executable file → Normal file
View File

@ -16,21 +16,27 @@
Export CenterNet mindir model. Export CenterNet mindir model.
""" """
import argparse import os
import numpy as np import numpy as np
import mindspore
from mindspore import context, Tensor from mindspore import context, Tensor
from mindspore.train.serialization import load_checkpoint, load_param_into_net, export from mindspore.train.serialization import load_checkpoint, load_param_into_net, export
from src import CenterNetDetEval from src import CenterNetDetEval
from src.config import net_config, eval_config, export_config from src.model_utils.config import config, net_config, eval_config, export_config
from src.model_utils.moxing_adapter import moxing_wrapper
parser = argparse.ArgumentParser(description='centernet export')
parser.add_argument("--device_id", type=int, default=0, help="Device id")
args = parser.parse_args()
context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", device_id=args.device_id) def modelarts_pre_process():
'''modelarts pre process function.'''
export_config.ckpt_file = os.path.join(os.path.dirname(os.path.abspath(__file__)), export_config.ckpt_file)
export_config.export_name = os.path.join(config.output_path, export_config.export_name)
if __name__ == '__main__':
@moxing_wrapper(pre_process=modelarts_pre_process)
def run_export():
'''export function'''
context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", device_id=config.device_id)
net = CenterNetDetEval(net_config, eval_config.K) net = CenterNetDetEval(net_config, eval_config.K)
net.set_train(False) net.set_train(False)
@ -38,7 +44,10 @@ if __name__ == '__main__':
load_param_into_net(net, param_dict) load_param_into_net(net, param_dict)
net.set_train(False) net.set_train(False)
input_shape = [1, 3, export_config.input_res[0], export_config.input_res[1]] input_data = Tensor(np.zeros([1, 3, export_config.input_res[0], export_config.input_res[1]]), mindspore.float32)
input_data = Tensor(np.random.uniform(-1.0, 1.0, size=input_shape).astype(np.float32))
export(net, input_data, file_name=export_config.export_name, file_format=export_config.export_format) export(net, input_data, file_name=export_config.export_name, file_format=export_config.export_format)
if __name__ == '__main__':
run_export()

64
model_zoo/research/cv/centernet_det/postprocess.py Normal file
View File

@ -0,0 +1,64 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""post process for 310 inference"""
import os
import json
import numpy as np
import pycocotools.coco as coco
from pycocotools.cocoeval import COCOeval
from src.model_utils.config import config, dataset_config, eval_config
from src import convert_eval_format, post_process, merge_outputs
def cal_acc(result_path, label_path, meta_path, save_path):
"""calculate inference accuracy"""
name_list = np.load(os.path.join(meta_path, "name_list.npy"), allow_pickle=True)
meta_list = np.load(os.path.join(meta_path, "meta_list.npy"), allow_pickle=True)
label_infor = coco.COCO(label_path)
pred_annos = {"images": [], "annotations": []}
for num, image_id in enumerate(name_list):
meta = meta_list[num]
pre_image = np.fromfile(os.path.join(result_path) + "/eval2017_image_" + str(image_id) + "_0.bin",
dtype=np.float32).reshape((1, 100, 6))
detections = []
for scale in eval_config.multi_scales:
dets = post_process(pre_image, meta, scale, dataset_config.num_classes)
detections.append(dets)
detections = merge_outputs(detections, dataset_config.num_classes, eval_config.SOFT_NMS)
pred_json = convert_eval_format(detections, image_id, eval_config.valid_ids)
label_infor.loadImgs([image_id])
for image_info in pred_json["images"]:
pred_annos["images"].append(image_info)
for image_anno in pred_json["annotations"]:
pred_annos["annotations"].append(image_anno)
if not os.path.exists(save_path):
os.makedirs(save_path)
pred_anno_file = os.path.join(save_path, '{}_pred_result.json').format(config.run_mode)
json.dump(pred_annos, open(pred_anno_file, 'w'))
pred_res_file = os.path.join(save_path, '{}_pred_eval.json').format(config.run_mode)
json.dump(pred_annos["annotations"], open(pred_res_file, 'w'))
coco_anno = coco.COCO(label_path)
coco_dets = coco_anno.loadRes(pred_res_file)
coco_eval = COCOeval(coco_anno, coco_dets, "bbox")
coco_eval.evaluate()
coco_eval.accumulate()
coco_eval.summarize()
if __name__ == '__main__':
cal_acc(config.result_path, config.label_path, config.meta_path, config.save_path)

56
model_zoo/research/cv/centernet_det/preprocess.py Normal file
View File

@ -0,0 +1,56 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""pre process for 310 inference"""
import os
import numpy as np
from src.model_utils.config import config, dataset_config, eval_config, net_config
from src.dataset import COCOHP
def preprocess(dataset_path, preprocess_path):
"""preprocess input images"""
meta_path = os.path.join(preprocess_path, "meta/meta")
result_path = os.path.join(preprocess_path, "data")
if not os.path.exists(meta_path):
os.makedirs(os.path.join(preprocess_path, "meta/meta"))
if not os.path.exists(result_path):
os.makedirs(os.path.join(preprocess_path, "data"))
coco = COCOHP(dataset_config, run_mode="val", net_opt=net_config)
coco.init(dataset_path, keep_res=False)
dataset = coco.create_eval_dataset()
name_list = []
meta_list = []
i = 0
for data in dataset.create_dict_iterator(num_epochs=1):
img_id = data['image_id'].asnumpy().reshape((-1))[0]
image = data['image'].asnumpy()
for scale in eval_config.multi_scales:
image_preprocess, meta = coco.pre_process_for_test(image, img_id, scale)
evl_file_name = "eval2017_image" + "_" + str(img_id) + ".bin"
evl_file_path = result_path + "/" + evl_file_name
image_preprocess.tofile(evl_file_path)
meta_file_path = os.path.join(preprocess_path + "/meta/meta", str(img_id) + ".txt")
with open(meta_file_path, 'w+') as f:
f.write(str(meta))
name_list.append(img_id)
meta_list.append(meta)
i += 1
print(f"preprocess: no.[{i}], img_name:{img_id}")
np.save(os.path.join(preprocess_path + "/meta", "name_list.npy"), np.array(name_list))
np.save(os.path.join(preprocess_path + "/meta", "meta_list.npy"), np.array(meta_list))
if __name__ == '__main__':
preprocess(config.val_data_dir, config.predict_dir)

1
model_zoo/research/cv/centernet_det/requirements.txt
View File

@ -1,3 +1,4 @@
opencv-python opencv-python
numpy numpy
pycocotools pycocotools
Cython

6
model_zoo/research/cv/centernet_det/scripts/ascend_distributed_launcher/hyper_parameter_config.ini Executable file → Normal file
View File

@ -6,8 +6,8 @@ do_shuffle=true
enable_data_sink=true enable_data_sink=true
data_sink_steps=-1 data_sink_steps=-1
save_checkpoint_path=./ save_checkpoint_path=./
save_checkpoint_steps=6105 save_checkpoint_steps=1221
save_checkpoint_num=20 save_checkpoint_num=1
mindrecord_prefix="coco_det.train.mind" mindrecord_prefix="coco_det.train.mind"
need_profiler=false need_profiler=false
profiler_path=./profiler profiler_path=./profiler

145
model_zoo/research/cv/centernet_det/scripts/run_infer_310.sh Normal file
View File

@ -0,0 +1,145 @@
#!/bin/bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
if [[ $# -lt 3 || $# -gt 4 ]]; then
echo "Usage: bash run_infer_310.sh [MINDIR_PATH] [DATASET_PATH] [NEED_PREPROCESS] [DEVICE_ID]
NEED_PREPROCESS means weather need preprocess or not, it's value is 'y' or 'n'.
DEVICE_ID is optional, it can be set by environment variable device_id, otherwise the value is zero"
exit 1
fi
get_real_path(){
if [ "${1:0:1}" == "/" ]; then
echo "$1"
else
echo "$(realpath -m $PWD/$1)"
fi
}
model=$(get_real_path $1)
dataset_path=$(get_real_path $2)
if [ "$3" == "y" ] || [ "$3" == "n" ];then
need_preprocess=$3
else
echo "weather need preprocess or not, it's value must be in [y, n]"
exit 1
fi
device_id=0
if [ $# == 4 ]; then
device_id=$4
fi
echo "mindir name: "$model
echo "dataset path: "$dataset_path
echo "need preprocess: "$need_preprocess
echo "device id: "$device_id
export ASCEND_HOME=/usr/local/Ascend/
if [ -d ${ASCEND_HOME}/ascend-toolkit ]; then
export PATH=$ASCEND_HOME/fwkacllib/bin:$ASCEND_HOME/fwkacllib/ccec_compiler/bin:$ASCEND_HOME/ascend-toolkit/latest/fwkacllib/ccec_compiler/bin:$ASCEND_HOME/ascend-toolkit/latest/atc/bin:$PATH
export LD_LIBRARY_PATH=$ASCEND_HOME/fwkacllib/lib64:/usr/local/lib:$ASCEND_HOME/ascend-toolkit/latest/atc/lib64:$ASCEND_HOME/ascend-toolkit/latest/fwkacllib/lib64:$ASCEND_HOME/driver/lib64:$ASCEND_HOME/add-ons:$LD_LIBRARY_PATH
export TBE_IMPL_PATH=$ASCEND_HOME/ascend-toolkit/latest/opp/op_impl/built-in/ai_core/tbe
export PYTHONPATH=$ASCEND_HOME/fwkacllib/python/site-packages:${TBE_IMPL_PATH}:$ASCEND_HOME/ascend-toolkit/latest/fwkacllib/python/site-packages:$PYTHONPATH
export ASCEND_OPP_PATH=$ASCEND_HOME/ascend-toolkit/latest/opp
else
export PATH=$ASCEND_HOME/fwkacllib/bin:$ASCEND_HOME/fwkacllib/ccec_compiler/bin:$ASCEND_HOME/atc/ccec_compiler/bin:$ASCEND_HOME/atc/bin:$PATH
export LD_LIBRARY_PATH=$ASCEND_HOME/fwkacllib/lib64:/usr/local/lib:$ASCEND_HOME/atc/lib64:$ASCEND_HOME/acllib/lib64:$ASCEND_HOME/driver/lib64:$ASCEND_HOME/add-ons:$LD_LIBRARY_PATH
export PYTHONPATH=$ASCEND_HOME/fwkacllib/python/site-packages:$ASCEND_HOME/atc/python/site-packages:$PYTHONPATH
export ASCEND_OPP_PATH=$ASCEND_HOME/opp
fi
function preprocess_data()
{
if [ -d preprocess ]; then
rm -rf ./preprocess
fi
mkdir preprocess
python3.7 ../preprocess.py --val_data_dir=$dataset_path --predict_dir=./preprocess/ >& preprocess.log
}
function compile_app()
{
cd ../ascend310_infer || exit
bash build.sh &> build.log
}
function infer()
{
cd - || exit
if [ -d result_Files ]; then
rm -rf ./result_Files
fi
if [ -d time_Result ]; then
rm -rf ./time_Result
fi
mkdir result_Files
mkdir time_Result
../ascend310_infer/out/main --mindir_path=$model --input0_path=./preprocess/data --device_id=$device_id &> infer.log
}
# install nms module from third party
if python -c "import nms" > /dev/null 2>&1
then
echo "NMS module already exits, no need reinstall."
else
if [ -f './CenterNet' ]
then
echo "NMS module was not found, but has been downloaded"
else
echo "NMS module was not found, install it now..."
git clone https://github.com/xingyizhou/CenterNet.git
fi
cd CenterNet/src/lib/external/ || exit
make
python setup.py install
cd - || exit
rm -rf CenterNet
fi
function cal_ap()
{
if [ -d acc ]; then
rm -rf ./acc
fi
mkdir acc
python3.7 ../postprocess.py --result_path=./result_Files --label_path=$dataset_path/annotations/instances_val2017.json --meta_path=./preprocess/meta --save_path=./acc &> acc.log
}
if [ $need_preprocess == "y" ]; then
preprocess_data
if [ $? -ne 0 ]; then
echo "preprocess dataset failed"
exit 1
fi
fi
compile_app
if [ $? -ne 0 ]; then
echo "compile app code failed"
exit 1
fi
infer
if [ $? -ne 0 ]; then
echo " execute inference failed"
exit 1
fi
cal_ap
if [ $? -ne 0 ]; then
echo "calculate accuracy failed"
exit 1
fi

16
model_zoo/research/cv/centernet_det/scripts/run_standalone_eval_ascend.sh Executable file → Normal file
View File

@ -29,18 +29,24 @@ PROJECT_DIR=$(cd "$(dirname "$0")" || exit; pwd)
CUR_DIR=`pwd` CUR_DIR=`pwd`
export GLOG_log_dir=${CUR_DIR}/ms_log export GLOG_log_dir=${CUR_DIR}/ms_log
export GLOG_logtostderr=0 export GLOG_logtostderr=0
export DEVICE_ID=$DEVICE_ID
# install nms module from third party # install nms module from third party
if python -c "import nms" > /dev/null 2>&1 if python -c "import nms" > /dev/null 2>&1
then then
echo "NMS module already exits, no need reinstall." echo "NMS module already exits, no need reinstall."
else else
echo "NMS module was not found, install it now..." if [ -f './CenterNet' ]
git clone https://github.com/xingyizhou/CenterNet.git then
cd CenterNet/src/lib/external/ echo "NMS module was not found, but has been downloaded"
else
echo "NMS module was not found, install it now..."
git clone https://github.com/xingyizhou/CenterNet.git
fi
cd CenterNet/src/lib/external/ || exit
make make
python setup.py install python setup.py install
cd - cd - || exit
rm -rf CenterNet rm -rf CenterNet
fi fi
@ -50,6 +56,6 @@ python ${PROJECT_DIR}/../eval.py \
--load_checkpoint_path=$LOAD_CHECKPOINT_PATH \ --load_checkpoint_path=$LOAD_CHECKPOINT_PATH \
--data_dir=$DATA_DIR \ --data_dir=$DATA_DIR \
--run_mode=$RUN_MODE \ --run_mode=$RUN_MODE \
--visual_image=false \ --visual_image=true \
--enable_eval=true \ --enable_eval=true \
--save_result_dir=./ > eval_log.txt 2>&1 & --save_result_dir=./ > eval_log.txt 2>&1 &

4
model_zoo/research/cv/centernet_det/scripts/run_standalone_train_ascend.sh Executable file → Normal file
View File

@ -35,6 +35,7 @@ PROJECT_DIR=$(cd "$(dirname "$0")" || exit; pwd)
CUR_DIR=`pwd` CUR_DIR=`pwd`
export GLOG_log_dir=${CUR_DIR}/ms_log export GLOG_log_dir=${CUR_DIR}/ms_log
export GLOG_logtostderr=0 export GLOG_logtostderr=0
export DEVICE_ID=$DEVICE_ID
python ${PROJECT_DIR}/../train.py \ python ${PROJECT_DIR}/../train.py \
--distribute=false \ --distribute=false \
@ -47,8 +48,9 @@ python ${PROJECT_DIR}/../train.py \
--data_sink_steps=-1 \ --data_sink_steps=-1 \
--epoch_size=130 \ --epoch_size=130 \
--load_checkpoint_path=$LOAD_CHECKPOINT_PATH \ --load_checkpoint_path=$LOAD_CHECKPOINT_PATH \
--save_checkpoint_steps=6105 \ --save_checkpoint_steps=9772 \
--save_checkpoint_num=1 \ --save_checkpoint_num=1 \
--mindrecord_dir=$MINDRECORD_DIR \ --mindrecord_dir=$MINDRECORD_DIR \
--mindrecord_prefix="coco_det.train.mind" \ --mindrecord_prefix="coco_det.train.mind" \
--visual_image=false \
--save_result_dir="" > training_log.txt 2>&1 & --save_result_dir="" > training_log.txt 2>&1 &

82
model_zoo/research/cv/centernet_det/src/centernet_det.py
View File

@ -23,44 +23,33 @@ from mindspore import context
from mindspore import dtype as mstype from mindspore import dtype as mstype
from mindspore.common.tensor import Tensor from mindspore.common.tensor import Tensor
from mindspore.context import ParallelMode from mindspore.context import ParallelMode
from mindspore.common.initializer import Constant
from mindspore.communication.management import get_group_size from mindspore.communication.management import get_group_size
from mindspore.nn.wrap.grad_reducer import DistributedGradReducer from mindspore.nn.wrap.grad_reducer import DistributedGradReducer
from src.utils import Sigmoid, GradScale from src.utils import Sigmoid, GradScale
from src.utils import FocalLoss, RegLoss from src.utils import FocalLoss, RegLoss
from src.decode import DetectionDecode from src.decode import DetectionDecode
from src.config import dataset_config as data_cfg
from src.hourglass import Convolution, Residual, Kp_module from src.hourglass import Convolution, Residual, Kp_module
from .model_utils.config import dataset_config as data_cfg
BN_MOMENTUM = 0.9 BN_MOMENTUM = 0.9
def _generate_feature(cin, cout, kernel_size, head_name, head, num_stacks, with_bn=True): def _generate_feature(cin, cout, kernel_size, head, num_stacks, with_bn=True):
""" """
Generate feature extraction function of each target head Generate hourglass network feature extraction function of each target head
""" """
module = None module = None
if 'hm' in head_name: module = nn.CellList([
module = nn.CellList([ nn.SequentialCell(
nn.SequentialCell( Convolution(cin, cout, kernel_size, with_bn=with_bn),
Convolution(cin, cout, kernel_size, with_bn=with_bn), nn.Conv2d(cout, head, kernel_size=1, has_bias=True, pad_mode='pad')
nn.Conv2d(cout, head, kernel_size=1, has_bias=True, bias_init=Constant(-2.19), pad_mode='pad') ) for _ in range(num_stacks)
) for _ in range(num_stacks) ])
])
else:
module = nn.CellList([
nn.SequentialCell(
Convolution(cin, cout, kernel_size, with_bn=with_bn),
nn.Conv2d(cout, head, kernel_size=1, has_bias=True, pad_mode='pad')
) for _ in range(num_stacks)
])
return module return module
class GatherDetectionFeatureCell(nn.Cell): class GatherDetectionFeatureCell(nn.Cell):
""" """
Gather features of object detection. Gather hourglass features of object detection.
Args: Args:
net_config: The config info of CenterNet network. net_config: The config info of CenterNet network.
@ -71,13 +60,15 @@ class GatherDetectionFeatureCell(nn.Cell):
def __init__(self, net_config): def __init__(self, net_config):
super(GatherDetectionFeatureCell, self).__init__() super(GatherDetectionFeatureCell, self).__init__()
self.heads = net_config.heads
self.nstack = net_config.num_stacks self.nstack = net_config.num_stacks
self.n = net_config.n self.n = net_config.n
self.cnv_dim = net_config.cnv_dim self.cnv_dim = net_config.cnv_dim
self.dims = net_config.dims self.dims = net_config.dims
self.modules = net_config.modules self.modules = net_config.modules
curr_dim = self.dims[0] curr_dim = self.dims[0]
self.heads = {'hm': data_cfg.num_classes, 'wh': 2}
if net_config.reg_offset:
self.heads.update({'reg': 2})
self.pre = nn.SequentialCell( self.pre = nn.SequentialCell(
Convolution(3, 128, 7, stride=2), Convolution(3, 128, 7, stride=2),
@ -114,12 +105,13 @@ class GatherDetectionFeatureCell(nn.Cell):
self.relu = nn.ReLU() self.relu = nn.ReLU()
self.hm_fn = _generate_feature(cin=self.cnv_dim, cout=curr_dim, kernel_size=3, head_name='hm', self.hm_fn = _generate_feature(cin=self.cnv_dim, cout=curr_dim, kernel_size=3, head=self.heads['hm'],
head=self.heads['hm'], num_stacks=self.nstack, with_bn=False) num_stacks=self.nstack, with_bn=False)
self.wh_fn = _generate_feature(cin=self.cnv_dim, cout=curr_dim, kernel_size=3, head_name='wh', self.wh_fn = _generate_feature(cin=self.cnv_dim, cout=curr_dim, kernel_size=3, head=self.heads['wh'],
head=self.heads['wh'], num_stacks=self.nstack, with_bn=False) num_stacks=self.nstack, with_bn=False)
self.reg_fn = _generate_feature(cin=self.cnv_dim, cout=curr_dim, kernel_size=3, head_name='reg', if net_config.reg_offset:
head=self.heads['reg'], num_stacks=self.nstack, with_bn=False) self.reg_fn = _generate_feature(cin=self.cnv_dim, cout=curr_dim, kernel_size=3, head=self.heads['reg'],
num_stacks=self.nstack, with_bn=False)
def construct(self, image): def construct(self, image):
"""Defines the computation performed.""" """Defines the computation performed."""
@ -134,13 +126,9 @@ class GatherDetectionFeatureCell(nn.Cell):
inter = self.inters[ind](inter) inter = self.inters[ind](inter)
out = {} out = {}
for head in self.heads.keys(): out['hm'] = self.hm_fn[ind](cnv)
if head == 'hm': out['wh'] = self.wh_fn[ind](cnv)
out[head] = self.hm_fn[ind](cnv) out['reg'] = self.reg_fn[ind](cnv)
if head == 'wh':
out[head] = self.wh_fn[ind](cnv)
if head == 'reg':
out[head] = self.reg_fn[ind](cnv)
outs += (out,) outs += (out,)
return outs return outs
@ -158,20 +146,18 @@ class CenterNetLossCell(nn.Cell):
def __init__(self, net_config): def __init__(self, net_config):
super(CenterNetLossCell, self).__init__() super(CenterNetLossCell, self).__init__()
self.network = GatherDetectionFeatureCell(net_config) self.network = GatherDetectionFeatureCell(net_config)
self.net_config = net_config self.num_stacks = net_config.num_stacks
self.reduce_sum = ops.ReduceSum() self.reduce_sum = ops.ReduceSum()
self.Sigmoid = Sigmoid() self.Sigmoid = Sigmoid()
self.FocalLoss = FocalLoss() self.FocalLoss = FocalLoss()
self.crit = nn.MSELoss() if net_config.mse_loss else self.FocalLoss self.crit = nn.MSELoss() if net_config.mse_loss else self.FocalLoss
self.crit_reg = RegLoss(net_config.reg_loss) self.crit_reg = RegLoss(net_config.reg_loss)
self.crit_wh = RegLoss(net_config.reg_loss) self.crit_wh = RegLoss(net_config.reg_loss)
self.num_stacks = net_config.num_stacks
self.wh_weight = net_config.wh_weight self.wh_weight = net_config.wh_weight
self.hm_weight = net_config.hm_weight self.hm_weight = net_config.hm_weight
self.off_weight = net_config.off_weight self.off_weight = net_config.off_weight
self.reg_offset = net_config.reg_offset self.reg_offset = net_config.reg_offset
self.not_enable_mse_loss = not net_config.mse_loss self.not_enable_mse_loss = not net_config.mse_loss
self.Print = ops.Print()
def construct(self, image, hm, reg_mask, ind, wh, reg): def construct(self, image, hm, reg_mask, ind, wh, reg):
"""Defines the computation performed.""" """Defines the computation performed."""
@ -250,8 +236,9 @@ class CenterNetWithoutLossScaleCell(nn.Cell):
weights = self.weights weights = self.weights
loss = self.network(image, hm, reg_mask, ind, wh, reg) loss = self.network(image, hm, reg_mask, ind, wh, reg)
grads = self.grad(self.network, weights)(image, hm, reg_mask, ind, wh, reg) grads = self.grad(self.network, weights)(image, hm, reg_mask, ind, wh, reg)
self.optimizer(grads) succ = self.optimizer(grads)
return loss ret = loss
return ops.depend(ret, succ)
class CenterNetWithLossScaleCell(nn.Cell): class CenterNetWithLossScaleCell(nn.Cell):
@ -319,9 +306,12 @@ class CenterNetWithLossScaleCell(nn.Cell):
else: else:
cond = self.less_equal(self.base, flag_sum) cond = self.less_equal(self.base, flag_sum)
overflow = cond overflow = cond
if not overflow: if overflow:
self.optimizer(grads) succ = False
return (loss, cond, scaling_sens) else:
succ = self.optimizer(grads)
ret = (loss, cond, scaling_sens)
return ops.depend(ret, succ)
class CenterNetDetEval(nn.Cell): class CenterNetDetEval(nn.Cell):
@ -331,17 +321,15 @@ class CenterNetDetEval(nn.Cell):
Args: Args:
net_config: The config info of CenterNet network. net_config: The config info of CenterNet network.
K(number): Max number of output objects. Default: 100. K(number): Max number of output objects. Default: 100.
enable_nms_fp16(bool): Use float16 data for max_pool, adaption for CPU. Default: True. enable_nms_fp16(bool): Use float16 data for max_pool, adaption for CPU. Default: False.
Returns: Returns:
Tensor, detection of images(bboxes, score, keypoints and category id of each objects) Tensor, detection of images(bboxes, score, keypoints and category id of each objects)
""" """
def __init__(self, net_config, K=100, enable_nms_fp16=True): def __init__(self, net_config, K=100, enable_nms_fp16=False):
super(CenterNetDetEval, self).__init__() super(CenterNetDetEval, self).__init__()
self.network = GatherDetectionFeatureCell(net_config) self.network = GatherDetectionFeatureCell(net_config)
self.decode = DetectionDecode(net_config, K, enable_nms_fp16) self.decode = DetectionDecode(net_config, K, enable_nms_fp16)
self.shape = ops.Shape()
self.reshape = ops.Reshape()
def construct(self, image): def construct(self, image):
"""Calculate prediction scores""" """Calculate prediction scores"""

225
model_zoo/research/cv/centernet_det/src/config.py
View File

@ -1,225 +0,0 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""
network config setting, will be used in dataset.py, train.py， eval.py
"""
import numpy as np
from easydict import EasyDict as edict
dataset_config = edict({
"num_classes": 80,
'max_objs': 128,
'input_res': [512, 512],
'output_res': [128, 128],
'rand_crop': True,
'shift': 0.1,
'scale': 0.4,
'down_ratio': 4,
'aug_rot': 0.0,
'rotate': 0,
'flip_prop': 0.5,
'color_aug': True,
'coco_classes': ('background', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
'train', 'truck', 'boat', 'traffic light', 'fire hydrant',
'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog',
'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra',
'giraffe', 'backpack', 'umbrella', 'handbag', 'tie',
'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball',
'kite', 'baseball bat', 'baseball glove', 'skateboard',
'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup',
'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza',
'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed',
'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote',
'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink',
'refrigerator', 'book', 'clock', 'vase', 'scissors',
'teddy bear', 'hair drier', 'toothbrush'),
'coco_class_name2id': {
'person': 1, 'bicycle': 2, 'car': 3, 'motorcycle': 4, 'airplane': 5,
'bus': 6, 'train': 7, 'truck': 8, 'boat': 9, 'traffic light': 10, 'fire hydrant': 11,
'stop sign': 13, 'parking meter': 14, 'bench': 15, 'bird': 16, 'cat': 17, 'dog': 18, 'horse': 19,
'sheep': 20, 'cow': 21, 'elephant': 22, 'bear': 23, 'zebra': 24, 'giraffe': 25, 'backpack': 27,
'umbrella': 28, 'handbag': 31, 'tie': 32, 'suitcase': 33, 'frisbee': 34, 'skis': 35,
'snowboard': 36, 'sports ball': 37, 'kite': 38, 'baseball bat': 39, 'baseball glove': 40,
'skateboard': 41, 'surfboard': 42, 'tennis racket': 43, 'bottle': 44, 'wine glass': 46,
'cup': 47, 'fork': 48, 'knife': 49, 'spoon': 50, 'bowl': 51, 'banana': 52, 'apple': 53, 'sandwich': 54,
'orange': 55, 'broccoli': 56, 'carrot': 57, 'hot dog': 58, 'pizza': 59, 'donut': 60, 'cake': 61,
'chair': 62, 'couch': 63, 'potted plant': 64, 'bed': 65, 'dining table': 67, 'toilet': 70, 'tv': 72,
'laptop': 73, 'mouse': 74, 'remote': 75, 'keyboard': 76, 'cell phone': 77, 'microwave': 78,
'oven': 79, 'toaster': 80, 'sink': 81, 'refrigerator': 82, 'book': 84, 'clock': 85, 'vase': 86,
'scissors': 87, 'teddy bear': 88, 'hair drier': 89, 'toothbrush': 90},
'mean': np.array([0.40789654, 0.44719302, 0.47026115], dtype=np.float32),
'std': np.array([0.28863828, 0.27408164, 0.27809835], dtype=np.float32),
'eig_val': np.array([0.2141788, 0.01817699, 0.00341571], dtype=np.float32),
'eig_vec': np.array([[-0.58752847, -0.69563484, 0.41340352],
[-0.5832747, 0.00994535, -0.81221408],
[-0.56089297, 0.71832671, 0.41158938]], dtype=np.float32),
})
net_config = edict({
'down_ratio': 4,
'last_level': 6,
'num_stacks': 2,
'n': 5,
'heads': {'hm': 80, 'wh': 2, 'reg': 2},
'cnv_dim': 256,
'modules': [2, 2, 2, 2, 2, 4],
'dims': [256, 256, 384, 384, 384, 512],
'dense_wh': False,
'norm_wh': False,
'cat_spec_wh': False,
'reg_offset': True,
'hm_weight': 1,
'off_weight': 1,
'wh_weight': 0.1,
'mse_loss': False,
'reg_loss': 'l1',
})
train_config = edict({
'batch_size': 12,
'loss_scale_value': 1024,
'optimizer': 'Adam',
'lr_schedule': 'MultiDecay',
'Adam': edict({
'weight_decay': 0.0,
'decay_filter': lambda x: x.name.endswith('.bias') or x.name.endswith('.beta') or x.name.endswith('.gamma'),
}),
'PolyDecay': edict({
'learning_rate': 2.4e-4,
'end_learning_rate': 2.4e-7,
'power': 5.0,
'eps': 1e-7,
'warmup_steps': 2000,
}),
'MultiDecay': edict({
'learning_rate': 2.4e-4,
'eps': 1e-7,
'warmup_steps': 2000,
'multi_epochs': [105, 125],
'factor': 10,
})
})
eval_config = edict({
'SOFT_NMS': True,
'keep_res': True,
'multi_scales': [1.0],
'pad': 127,
'K': 100,
'score_thresh': 0.3,
'valid_ids': [
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13,
14, 15, 16, 17, 18, 19, 20, 21, 22, 23,
24, 25, 27, 28, 31, 32, 33, 34, 35, 36,
37, 38, 39, 40, 41, 42, 43, 44, 46, 47,
48, 49, 50, 51, 52, 53, 54, 55, 56, 57,
58, 59, 60, 61, 62, 63, 64, 65, 67, 70,
72, 73, 74, 75, 76, 77, 78, 79, 80, 81,
82, 84, 85, 86, 87, 88, 89, 90],
'color_list': [
0.000, 0.800, 1.000,
0.850, 0.325, 0.098,
0.929, 0.694, 0.125,
0.494, 0.184, 0.556,
0.466, 0.674, 0.188,
0.301, 0.745, 0.933,
0.635, 0.078, 0.184,
0.300, 0.300, 0.300,
0.600, 0.600, 0.600,
1.000, 0.000, 0.000,
1.000, 0.500, 0.000,
0.749, 0.749, 0.000,
0.000, 1.000, 0.000,
0.000, 0.000, 1.000,
0.667, 0.000, 1.000,
0.333, 0.333, 0.000,
0.333, 0.667, 0.333,
0.333, 1.000, 0.000,
0.667, 0.333, 0.000,
0.667, 0.667, 0.000,
0.667, 1.000, 0.000,
1.000, 0.333, 0.000,
1.000, 0.667, 0.000,
1.000, 1.000, 0.000,
0.000, 0.333, 0.500,
0.000, 0.667, 0.500,
0.000, 1.000, 0.500,
0.333, 0.000, 0.500,
0.333, 0.333, 0.500,
0.333, 0.667, 0.500,
0.333, 1.000, 0.500,
0.667, 0.000, 0.500,
0.667, 0.333, 0.500,
0.667, 0.667, 0.500,
0.667, 1.000, 0.500,
1.000, 0.000, 0.500,
1.000, 0.333, 0.500,
1.000, 0.667, 0.500,
1.000, 1.000, 0.500,
0.000, 0.333, 1.000,
0.000, 0.667, 1.000,
0.000, 1.000, 1.000,
0.333, 0.000, 1.000,
0.333, 0.333, 1.000,
0.333, 0.667, 1.000,
0.333, 1.000, 1.000,
0.667, 0.000, 1.000,
0.667, 0.333, 1.000,
0.667, 0.667, 1.000,
0.667, 1.000, 1.000,
1.000, 0.000, 1.000,
1.000, 0.333, 1.000,
1.000, 0.667, 1.000,
0.167, 0.800, 0.000,
0.333, 0.000, 0.000,
0.500, 0.000, 0.000,
0.667, 0.000, 0.000,
0.833, 0.000, 0.000,
1.000, 0.000, 0.000,
0.000, 0.667, 0.400,
0.000, 0.333, 0.000,
0.000, 0.500, 0.000,
0.000, 0.667, 0.000,
0.000, 0.833, 0.000,
0.000, 1.000, 0.000,
0.000, 0.000, 0.167,
0.000, 0.000, 0.333,
0.000, 0.000, 0.500,
0.000, 0.000, 0.667,
0.000, 0.000, 0.833,
0.000, 0.000, 1.000,
0.000, 0.200, 0.800,
0.143, 0.143, 0.543,
0.286, 0.286, 0.286,
0.429, 0.429, 0.429,
0.571, 0.571, 0.571,
0.714, 0.714, 0.714,
0.857, 0.857, 0.857,
0.000, 0.447, 0.741,
0.50, 0.5, 0],
})
export_config = edict({
'input_res': dataset_config.input_res,
'ckpt_file': "./ckpt_file.ckpt",
'export_format': "MINDIR",
'export_name': "CenterNet_ObjectDetection",
})

47
model_zoo/research/cv/centernet_det/src/dataset.py
View File

@ -17,21 +17,35 @@ Data operations, will be used in train.py
""" """
import os import os
import sys
import math import math
import argparse
import cv2 import cv2
import numpy as np import numpy as np
import pycocotools.coco as coco import pycocotools.coco as coco
import mindspore.dataset as ds import mindspore.dataset as ds
from mindspore import log as logger from mindspore import log as logger
from mindspore.mindrecord import FileWriter from mindspore.mindrecord import FileWriter
from src.image import color_aug, get_affine_transform, affine_transform
from src.image import gaussian_radius, draw_umich_gaussian, draw_msra_gaussian, draw_dense_reg try:
from src.visual import visual_image from src.model_utils.config import config, dataset_config
from src.model_utils.moxing_adapter import moxing_wrapper
from src.image import color_aug, get_affine_transform, affine_transform
from src.image import gaussian_radius, draw_umich_gaussian, draw_msra_gaussian, draw_dense_reg
from src.visual import visual_image
except ImportError as import_error:
print('Import Error: {}, trying append path/centernet_det/src/../'.format(import_error))
sys.path.append(os.path.dirname(os.path.dirname(os.path.realpath(__file__))))
from src.model_utils.config import config, dataset_config
from src.model_utils.moxing_adapter import moxing_wrapper
from src.image import color_aug, get_affine_transform, affine_transform
from src.image import gaussian_radius, draw_umich_gaussian, draw_msra_gaussian, draw_dense_reg
from src.visual import visual_image
_current_dir = os.path.dirname(os.path.realpath(__file__)) _current_dir = os.path.dirname(os.path.realpath(__file__))
cv2.setNumThreads(0) cv2.setNumThreads(0)
class COCOHP(ds.Dataset): class COCOHP(ds.Dataset):
""" """
Encapsulation class of COCO datast. Encapsulation class of COCO datast.
@ -386,16 +400,19 @@ class COCOHP(ds.Dataset):
return data_set return data_set
if __name__ == '__main__': def modelarts_pre_process():
# Convert coco2017 dataset to mindrecord to improve performance on host """modelarts pre process function."""
from src.config import dataset_config config.coco_data_dir = config.data_path
config.mindrecord_dir = config.output_path
parser = argparse.ArgumentParser(description='CenterNet MindRecord dataset')
parser.add_argument("--coco_data_dir", type=str, default="", help="Coco dataset directory.") @moxing_wrapper(pre_process=modelarts_pre_process)
parser.add_argument("--mindrecord_dir", type=str, default="", help="MindRecord dataset dir.") def coco2mindrecord():
parser.add_argument("--mindrecord_prefix", type=str, default="coco_det.train.mind", """Convert coco2017 dataset to mindrecord"""
help="Prefix of MindRecord dataset filename.")
args_opt = parser.parse_args()
dsc = COCOHP(dataset_config, run_mode="train") dsc = COCOHP(dataset_config, run_mode="train")
dsc.init(args_opt.coco_data_dir) dsc.init(config.coco_data_dir)
dsc.transfer_coco_to_mindrecord(args_opt.mindrecord_dir, args_opt.mindrecord_prefix, shard_num=8) dsc.transfer_coco_to_mindrecord(config.mindrecord_dir, config.mindrecord_prefix, shard_num=8)
if __name__ == '__main__':
coco2mindrecord()

33
model_zoo/research/cv/centernet_det/src/decode.py
View File

@ -16,29 +16,33 @@
Decode from heads for evaluation Decode from heads for evaluation
""" """
import mindspore.ops as ops import mindspore as ms
import mindspore.nn as nn import mindspore.nn as nn
import mindspore.ops as ops
from mindspore.ops import operations as P
from mindspore.common import dtype as mstype from mindspore.common import dtype as mstype
from .utils import GatherFeature, TransposeGatherFeature from .utils import GatherFeature, TransposeGatherFeature
class NMS(nn.Cell): class NMS(nn.Cell):
""" """
Non-maximum suppression Non-maximum suppression
Args: Args:
kernel(int): Maxpooling kernel size. Default: 3. kernel(int): Maxpooling kernel size. Default: 3.
enable_nms_fp16(bool): Use float16 data for max_pool, adaption for CPU. Default: True. enable_nms_fp16(bool): Use float16 data for max_pool, adaption for CPU. Default: False.
Returns: Returns:
Tensor, heatmap after non-maximum suppression. Tensor, heatmap after non-maximum suppression.
""" """
def __init__(self, kernel=3, enable_nms_fp16=True): def __init__(self, kernel=3, enable_nms_fp16=False):
super(NMS, self).__init__() super(NMS, self).__init__()
self.pad = (kernel - 1) // 2
self.cast = ops.Cast() self.cast = ops.Cast()
self.dtype = ops.DType() self.dtype = ops.DType()
self.equal = ops.Equal() self.equal = ops.Equal()
self.max_pool = nn.MaxPool2d(kernel, stride=1, pad_mode="same") self.Abs = P.Abs()
self.max_pool_ = nn.MaxPool2d(kernel, stride=1, pad_mode="same")
self.max_pool = P.MaxPoolWithArgmax(kernel_size=kernel, strides=1, pad_mode='same')
self.enable_fp16 = enable_nms_fp16 self.enable_fp16 = enable_nms_fp16
def construct(self, heat): def construct(self, heat):
@ -46,16 +50,23 @@ class NMS(nn.Cell):
dtype = self.dtype(heat) dtype = self.dtype(heat)
if self.enable_fp16: if self.enable_fp16:
heat = self.cast(heat, mstype.float16) heat = self.cast(heat, mstype.float16)
heat_max = self.max_pool(heat) heat_max = self.max_pool_(heat)
keep = self.equal(heat, heat_max) keep = self.equal(heat, heat_max)
keep = self.cast(keep, dtype) keep = self.cast(keep, dtype)
heat = self.cast(heat, dtype) heat = self.cast(heat, dtype)
else: else:
heat_max = self.max_pool(heat) heat_max, _ = self.max_pool(heat)
keep = self.equal(heat, heat_max) error = self.cast((heat - heat_max), mstype.float32)
abs_error = self.Abs(error)
abs_out = self.Abs(heat)
error = abs_error / (abs_out + 1e-12)
keep = P.Select()(P.LessEqual()(error, 1e-3),
P.Fill()(ms.float32, P.Shape()(error), 1.0),
P.Fill()(ms.float32, P.Shape()(error), 0.0))
heat = heat * keep heat = heat * keep
return heat return heat
class GatherTopK(nn.Cell): class GatherTopK(nn.Cell):
""" """
Gather topk features through all channels Gather topk features through all channels
@ -73,7 +84,8 @@ class GatherTopK(nn.Cell):
self.cast = ops.Cast() self.cast = ops.Cast()
self.dtype = ops.DType() self.dtype = ops.DType()
self.gather_feat = GatherFeature() self.gather_feat = GatherFeature()
self.mod = ops.Mod() # The ops.Mod() operator will produce errors on the Ascend 310
self.mod = P.FloorMod()
self.div = ops.Div() self.div = ops.Div()
def construct(self, scores, K=40): def construct(self, scores, K=40):
@ -95,6 +107,7 @@ class GatherTopK(nn.Cell):
topk_xs = self.cast(self.reshape(topk_xs, (b, K)), self.dtype(scores)) topk_xs = self.cast(self.reshape(topk_xs, (b, K)), self.dtype(scores))
return topk_score, topk_inds, topk_clses, topk_ys, topk_xs return topk_score, topk_inds, topk_clses, topk_ys, topk_xs
class DetectionDecode(nn.Cell): class DetectionDecode(nn.Cell):
""" """
Decode from heads to gather multi-objects info. Decode from heads to gather multi-objects info.
@ -107,7 +120,7 @@ class DetectionDecode(nn.Cell):
Returns: Returns:
Tensor, multi-objects detections. Tensor, multi-objects detections.
""" """
def __init__(self, net_config, K=100, enable_nms_fp16=True): def __init__(self, net_config, K=100, enable_nms_fp16=False):
super(DetectionDecode, self).__init__() super(DetectionDecode, self).__init__()
self.K = K self.K = K
self.nms = NMS(enable_nms_fp16=enable_nms_fp16) self.nms = NMS(enable_nms_fp16=enable_nms_fp16)

152
model_zoo/research/cv/centernet_det/src/hccl_tools.py
View File

@ -1,152 +0,0 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""
hccl configuration file generation
"""
import os
import sys
import json
import socket
from argparse import ArgumentParser
from typing import Dict, Any
def parse_args():
"""
parse args .
Args:
Returns:
args.
Examples:
>>> parse_args()
"""
parser = ArgumentParser(description="mindspore distributed training launch "
"helper utility that will generate hccl"
" config file")
parser.add_argument("--device_num", type=str, default="[0,8)",
help="The number of the Ascend accelerators used. please note that the Ascend accelerators"
"used must be continuous, such [0,4) means to use four chips "
"0，1，2，3; [0,1) means to use chip 0; The first four chips are"
"a group, and the last four chips are a group. In addition to"
"the [0,8) chips are allowed, other cross-group such as [3,6)"
"are prohibited.")
parser.add_argument("--visible_devices", type=str, default="0,1,2,3,4,5,6,7",
help="will use the visible devices sequentially")
parser.add_argument("--server_ip", type=str, default="",
help="server ip")
args = parser.parse_args()
return args
def get_host_ip():
"""
get host ip
"""
ip = None
try:
hostname = socket.gethostname()
ip = socket.gethostbyname(hostname)
except EOFError:
pass
return ip
def main():
print("start", __file__)
args = parse_args()
# visible_devices
visible_devices = args.visible_devices.split(',')
print('visible_devices:{}'.format(visible_devices))
# server_id
ip = get_host_ip()
if args.server_ip:
server_id = args.server_ip
elif ip:
server_id = ip
else:
raise ValueError("please input server ip!")
print('server_id:{}'.format(server_id))
# device_num
first_num = int(args.device_num[1])
last_num = int(args.device_num[3])
if first_num < 0 or last_num > 8:
raise ValueError("device num {} must be in range [0,8] !".format(args.device_num))
if first_num > last_num:
raise ValueError("First num {} of device num {} must less than last num {} !".format(first_num, args.device_num,
last_num))
if first_num < 4:
if last_num > 4:
if first_num == 0 and last_num == 8:
pass
else:
raise ValueError("device num {} must be in the same group of [0,4] or [4,8] !".format(args.device_num))
device_num_list = list(range(first_num, last_num))
print("device_num_list:", device_num_list)
assert len(visible_devices) >= len(device_num_list)
# construct hccn_table
device_ips: Dict[Any, Any] = {}
with open('/etc/hccn.conf', 'r') as fin:
for hccn_item in fin.readlines():
if hccn_item.strip().startswith('address_'):
device_id, device_ip = hccn_item.split('=')
device_id = device_id.split('_')[1]
device_ips[device_id] = device_ip.strip()
hccn_table = {'version': '1.0',
'server_count': '1',
'server_list': []}
device_list = []
rank_id = 0
for instance_id in device_num_list:
device_id = visible_devices[instance_id]
device_ip = device_ips[device_id]
device = {'device_id': device_id,
'device_ip': device_ip,
'rank_id': str(rank_id)}
print('rank_id:{}, device_id:{}, device_ip:{}'.format(rank_id, device_id, device_ip))
rank_id += 1
device_list.append(device)
hccn_table['server_list'].append({
'server_id': server_id,
'device': device_list,
'host_nic_ip': 'reserve'
})
hccn_table['status'] = 'completed'
# save hccn_table to file
table_path = os.getcwd()
table_fn = os.path.join(table_path,
'hccl_{}p_{}_{}.json'.format(len(device_num_list), "".join(map(str, device_num_list)),
server_id))
with open(table_fn, 'w') as table_fp:
json.dump(hccn_table, table_fp, indent=4)
sys.stdout.flush()
print("Completed: hccl file was save in :", table_fn)
if __name__ == "__main__":
main()

156
model_zoo/research/cv/centernet_det/src/model_utils/config.py Normal file
View File

@ -0,0 +1,156 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""Parse arguments"""
import os
import ast
import argparse
from pprint import pprint, pformat
import yaml
import numpy as np
class Config:
"""
Configuration namespace. Convert dictionary to members.
"""
def __init__(self, cfg_dict):
for k, v in cfg_dict.items():
if isinstance(v, str) and (v[:9] == 'np.array(' and v[-17:] == 'dtype=np.float32)'):
v = np.array(ast.literal_eval(v[9:v.rfind(']') + 1]), dtype=np.float32)
if isinstance(v, (list, tuple)):
setattr(self, k, [Config(x) if isinstance(x, dict) else x for x in v])
else:
setattr(self, k, Config(v) if isinstance(v, dict) else v)
def __str__(self):
return pformat(self.__dict__)
def __repr__(self):
return self.__str__()
def parse_cli_to_yaml(parser, cfg, helper=None, choices=None, cfg_path="default_config.yaml"):
"""
Parse command line arguments to the configuration according to the default yaml.
Args:
parser: Parent parser.
cfg: Base configuration.
helper: Helper description.
cfg_path: Path to the default yaml config.
"""
parser = argparse.ArgumentParser(description="[REPLACE THIS at config.py]",
parents=[parser])
helper = {} if helper is None else helper
choices = {} if choices is None else choices
for item in cfg:
if not isinstance(cfg[item], list) and not isinstance(cfg[item], dict):
help_description = helper[item] if item in helper else "Please reference to {}".format(cfg_path)
choice = choices[item] if item in choices else None
if isinstance(cfg[item], bool):
parser.add_argument("--" + item, type=ast.literal_eval, default=cfg[item], choices=choice,
help=help_description)
else:
parser.add_argument("--" + item, type=type(cfg[item]), default=cfg[item], choices=choice,
help=help_description)
args = parser.parse_args()
return args
def parse_yaml(yaml_path):
"""
Parse the yaml config file.
Args:
yaml_path: Path to the yaml config.
"""
with open(yaml_path, 'r') as fin:
try:
cfgs = yaml.load_all(fin.read(), Loader=yaml.FullLoader)
cfgs = [x for x in cfgs]
if len(cfgs) == 1:
cfg_helper = {}
cfg = cfgs[0]
cfg_choices = {}
elif len(cfgs) == 2:
cfg, cfg_helper = cfgs
cfg_choices = {}
elif len(cfgs) == 3:
cfg, cfg_helper, cfg_choices = cfgs
else:
raise ValueError("At most 3 docs (config, description for help, choices) are supported in config yaml")
print(cfg_helper)
except:
raise ValueError("Failed to parse yaml")
return cfg, cfg_helper, cfg_choices
def merge(args, cfg):
"""
Merge the base config from yaml file and command line arguments.
Args:
args: Command line arguments.
cfg: Base configuration.
"""
args_var = vars(args)
for item in args_var:
cfg[item] = args_var[item]
return cfg
def extra_operations(cfg):
"""
Do extra work on Config object.
Args:
cfg: Object after instantiation of class 'Config'.
"""
cfg.train_config.Adam.decay_filter = lambda x: x.name.endswith('.bias') or x.name.endswith('.beta') or x.name.endswith('.gamma')
cfg.export_config.input_res = cfg.dataset_config.input_res
if cfg.export_load_ckpt:
cfg.export_config.ckpt_file = cfg.export_load_ckpt
if cfg.export_name:
cfg.export_config.export_name = cfg.export_name
if cfg.export_format:
cfg.export_config.export_format = cfg.export_format
def get_config():
"""
Get Config according to the yaml file and cli arguments.
"""
parser = argparse.ArgumentParser(description="default name", add_help=False)
current_dir = os.path.dirname(os.path.abspath(__file__))
parser.add_argument("--config_path", type=str, default=os.path.join(current_dir, "../../default_config.yaml"),
help="Config file path")
path_args, _ = parser.parse_known_args()
default, helper, choices = parse_yaml(path_args.config_path)
pprint(default)
args = parse_cli_to_yaml(parser=parser, cfg=default, helper=helper, choices=choices, cfg_path=path_args.config_path)
final_config = merge(args, default)
config_obj = Config(final_config)
extra_operations(config_obj)
return config_obj
config = get_config()
dataset_config = config.dataset_config
net_config = config.net_config
train_config = config.train_config
eval_config = config.eval_config
export_config = config.export_config

27
model_zoo/research/cv/centernet_det/src/model_utils/device_adapter.py Normal file
View File

@ -0,0 +1,27 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""Device adapter for ModelArts"""
from src.model_utils.config import config
if config.enable_modelarts:
from src.model_utils.moxing_adapter import get_device_id, get_device_num, get_rank_id, get_job_id
else:
from src.model_utils.local_adapter import get_device_id, get_device_num, get_rank_id, get_job_id
__all__ = [
"get_device_id", "get_device_num", "get_rank_id", "get_job_id"
]

36
model_zoo/research/cv/centernet_det/src/model_utils/local_adapter.py Normal file
View File

@ -0,0 +1,36 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""Local adapter"""
import os
def get_device_id():
device_id = os.getenv('DEVICE_ID', '0')
return int(device_id)
def get_device_num():
device_num = os.getenv('RANK_SIZE', '1')
return int(device_num)
def get_rank_id():
global_rank_id = os.getenv('RANK_ID', '0')
return int(global_rank_id)
def get_job_id():
return "Local Job"

123
model_zoo/research/cv/centernet_det/src/model_utils/moxing_adapter.py Normal file
View File

@ -0,0 +1,123 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""Moxing adapter for ModelArts"""
import os
import functools
from mindspore import context
from mindspore.profiler import Profiler
from src.model_utils.config import config
_global_sync_count = 0
def get_device_id():
device_id = os.getenv('DEVICE_ID', '0')
return int(device_id)
def get_device_num():
device_num = os.getenv('RANK_SIZE', '1')
return int(device_num)
def get_rank_id():
global_rank_id = os.getenv('RANK_ID', '0')
return int(global_rank_id)
def get_job_id():
job_id = os.getenv('JOB_ID')
job_id = job_id if job_id != "" else "default"
return job_id
def sync_data(from_path, to_path):
"""
Download data from remote obs to local directory if the first url is remote url and the second one is local path
Upload data from local directory to remote obs in contrast.
"""
import moxing as mox
import time
global _global_sync_count
sync_lock = "/tmp/copy_sync.lock" + str(_global_sync_count)
_global_sync_count += 1
# Each server contains 8 devices as most.
if get_device_id() % min(get_device_num(), 8) == 0 and not os.path.exists(sync_lock):
print("from path: ", from_path)
print("to path: ", to_path)
mox.file.copy_parallel(from_path, to_path)
print("===finish data synchronization===")
try:
os.mknod(sync_lock)
# print("os.mknod({}) success".format(sync_lock))
except IOError:
pass
print("===save flag===")
while True:
if os.path.exists(sync_lock):
break
time.sleep(1)
print("Finish sync data from {} to {}.".format(from_path, to_path))
def moxing_wrapper(pre_process=None, post_process=None):
"""
Moxing wrapper to download dataset and upload outputs.
"""
def wrapper(run_func):
@functools.wraps(run_func)
def wrapped_func(*args, **kwargs):
# Download data from data_url
if config.enable_modelarts:
if config.data_url:
sync_data(config.data_url, config.data_path)
print("Dataset downloaded: ", os.listdir(config.data_path))
if config.checkpoint_url:
sync_data(config.checkpoint_url, config.load_path)
print("Preload downloaded: ", os.listdir(config.load_path))
if config.train_url:
sync_data(config.train_url, config.output_path)
print("Workspace downloaded: ", os.listdir(config.output_path))
context.set_context(save_graphs_path=os.path.join(config.output_path, str(get_rank_id())))
config.device_num = get_device_num()
config.device_id = get_device_id()
if not os.path.exists(config.output_path):
os.makedirs(config.output_path)
if pre_process:
pre_process()
if config.enable_profiling:
profiler = Profiler()
run_func(*args, **kwargs)
if config.enable_profiling:
profiler.analyse()
# Upload data to train_url
if config.enable_modelarts:
if post_process:
post_process()
if config.train_url:
print("Start to copy output directory")
sync_data(config.output_path, config.train_url)
return wrapped_func
return wrapper

14
model_zoo/research/cv/centernet_det/src/post_process.py Executable file → Normal file
View File

@ -20,13 +20,6 @@ from .image import get_affine_transform, affine_transform, transform_preds
from .visual import coco_box_to_bbox from .visual import coco_box_to_bbox
try:
from nms import soft_nms
except ImportError:
print('NMS not installed! Do \n cd $CenterNet_ROOT/scripts/ \n'
'and see run_standalone_eval.sh for more details to install it\n')
def post_process(dets, meta, scale, num_classes): def post_process(dets, meta, scale, num_classes):
"""rescale detection to original scale""" """rescale detection to original scale"""
c, s, h, w = meta['c'], meta['s'], meta['out_height'], meta['out_width'] c, s, h, w = meta['c'], meta['s'], meta['out_height'], meta['out_width']
@ -59,7 +52,12 @@ def merge_outputs(detections, num_classes, SOFT_NMS=True):
results[j] = np.concatenate( results[j] = np.concatenate(
[detection[j] for detection in detections], axis=0).astype(np.float32) [detection[j] for detection in detections], axis=0).astype(np.float32)
if SOFT_NMS: if SOFT_NMS:
soft_nms(results[j], Nt=0.5, threshold=0.01, method=2) try:
from nms import soft_nms
except ImportError:
print('NMS not installed! Do \n cd $CenterNet_ROOT/scripts/ \n'
'and see run_standalone_eval.sh for more details to install it\n')
soft_nms(results[j], Nt=0.5, threshold=0.001, method=2)
scores = np.hstack( scores = np.hstack(
[results[j][:, 4] for j in range(1, num_classes + 1)]) [results[j][:, 4] for j in range(1, num_classes + 1)])

256
model_zoo/research/cv/centernet_det/src/utils.py Executable file → Normal file
View File

@ -22,52 +22,10 @@ import numpy as np
import mindspore.nn as nn import mindspore.nn as nn
import mindspore.ops as ops import mindspore.ops as ops
from mindspore import dtype as mstype from mindspore import dtype as mstype
from mindspore.common.initializer import initializer
from mindspore.common.parameter import Parameter
from mindspore.common.tensor import Tensor from mindspore.common.tensor import Tensor
from mindspore.nn.learning_rate_schedule import LearningRateSchedule, PolynomialDecayLR, WarmUpLR from mindspore.nn.learning_rate_schedule import LearningRateSchedule, PolynomialDecayLR, WarmUpLR
from mindspore.train.callback import Callback from mindspore.train.callback import Callback
clip_grad = ops.MultitypeFuncGraph("clip_grad")
@clip_grad.register("Number", "Tensor")
def _clip_grad(clip_value, grad):
"""
Clip gradients.
Inputs:
clip_value (float): Specifies how much to clip.
grad (tuple[Tensor]): Gradients.
Outputs:
tuple[Tensor], clipped gradients.
"""
dt = ops.dtype(grad)
new_grad = nn.ClipByNorm()(grad, ops.cast(ops.tuple_to_array((clip_value,)), dt))
return new_grad
class ClipByNorm(nn.Cell):
"""
Clip grads by gradient norm
Args:
clip_norm(float): The target norm of graident clip. Default: 1.0
Returns:
Tuple of Tensors, gradients after clip.
"""
def __init__(self, clip_norm=1.0):
super(ClipByNorm, self).__init__()
self.hyper_map = ops.HyperMap()
self.clip_norm = clip_norm
def construct(self, grads):
grads = self.hyper_map(ops.partial(clip_grad, self.clip_norm), grads)
return grads
reciprocal = ops.Reciprocal() reciprocal = ops.Reciprocal()
grad_scale = ops.MultitypeFuncGraph("grad_scale") grad_scale = ops.MultitypeFuncGraph("grad_scale")
@ -95,37 +53,17 @@ class GradScale(nn.Cell):
return grads return grads
class ClipByValue(nn.Cell):
"""
Clip tensor by value
Args: None
Returns:
Tensor, output after clip.
"""
def __init__(self):
super(ClipByValue, self).__init__()
self.min = ops.Minimum()
self.max = ops.Maximum()
def construct(self, x, clip_value_min, clip_value_max):
x_min = self.min(x, clip_value_max)
x_max = self.max(x_min, clip_value_min)
return x_max
class GatherFeature(nn.Cell): class GatherFeature(nn.Cell):
""" """
Gather feature at specified position Gather feature at specified position
Args: Args:
enable_cpu_gather (bool): Use cpu operator GatherD to gather feature or not, adaption for CPU. Default: True. enable_cpu_gather (bool): Use cpu operator GatherD to gather feature or not, adaption for CPU. Default: False.
Returns: Returns:
Tensor, feature at spectified position Tensor, feature at spectified position
""" """
def __init__(self, enable_cpu_gather=True): def __init__(self, enable_cpu_gather=False):
super(GatherFeature, self).__init__() super(GatherFeature, self).__init__()
self.tile = ops.Tile() self.tile = ops.Tile()
self.shape = ops.Shape() self.shape = ops.Shape()
@ -257,171 +195,6 @@ class FocalLoss(nn.Cell):
return loss return loss
class GHMCLoss(nn.Cell):
"""
Warpper for gradient harmonizing loss for classification.
Args:
bins(int): Number of bins. Default: 10.
momentum(float): Momentum for moving gradient density. Default: 0.0.
Returns:
Tensor, GHM loss for classification.
"""
def __init__(self, bins=10, momentum=0.0):
super(GHMCLoss, self).__init__()
self.bins = bins
self.momentum = momentum
edges_left = np.array([float(x) / bins for x in range(bins)], dtype=np.float32)
self.edges_left = Tensor(edges_left.reshape((bins, 1, 1, 1, 1)))
edges_right = np.array([float(x) / bins for x in range(1, bins + 1)], dtype=np.float32)
edges_right[-1] += 1e-4
self.edges_right = Tensor(edges_right.reshape((bins, 1, 1, 1, 1)))
if momentum >= 0:
self.acc_sum = Parameter(initializer(0, [bins], mstype.float32))
self.abs = ops.Abs()
self.log = ops.Log()
self.cast = ops.Cast()
self.select = ops.Select()
self.reshape = ops.Reshape()
self.reduce_sum = ops.ReduceSum()
self.max = ops.Maximum()
self.less = ops.Less()
self.equal = ops.Equal()
self.greater = ops.Greater()
self.logical_and = ops.LogicalAnd()
self.greater_equal = ops.GreaterEqual()
self.zeros_like = ops.ZerosLike()
self.expand_dims = ops.ExpandDims()
def construct(self, out, target):
"""GHM loss for classification"""
g = self.abs(out - target)
g = self.expand_dims(g, 0) # (1, b, c, h, w)
pos_inds = self.cast(self.equal(target, 1.0), mstype.float32)
tot = self.max(self.reduce_sum(pos_inds, ()), 1.0)
# (bin, b, c, h, w)
inds_mask = self.logical_and(self.greater_equal(g, self.edges_left), self.less(g, self.edges_right))
zero_matrix = self.cast(self.zeros_like(inds_mask), mstype.float32)
inds = self.cast(inds_mask, mstype.float32)
# (bins,)
num_in_bin = self.reduce_sum(inds, (1, 2, 3, 4))
valid_bins = self.greater(num_in_bin, 0)
num_valid_bin = self.reduce_sum(self.cast(valid_bins, mstype.float32), ())
if self.momentum > 0:
self.acc_sum = self.select(valid_bins,
self.momentum * self.acc_sum + (1 - self.momentum) * num_in_bin,
self.acc_sum)
acc_sum = self.acc_sum
acc_sum = self.reshape(acc_sum, (self.bins, 1, 1, 1, 1))
acc_sum = acc_sum + zero_matrix
weights = self.select(self.equal(inds, 1), tot / acc_sum, zero_matrix)
# (b, c, h, w)
weights = self.reduce_sum(weights, 0)
else:
num_in_bin = self.reshape(num_in_bin, (self.bins, 1, 1, 1, 1))
num_in_bin = num_in_bin + zero_matrix
weights = self.select(self.equal(inds, 1), tot / num_in_bin, zero_matrix)
# (b, c, h, w)
weights = self.reduce_sum(weights, 0)
weights = weights / num_valid_bin
ghmc_loss = (target - 1.0) * self.log(1.0 - out) - target * self.log(out)
ghmc_loss = self.reduce_sum(ghmc_loss * weights, ()) / tot
return ghmc_loss
class GHMRLoss(nn.Cell):
"""
Warpper for gradient harmonizing loss for regression.
Args:
bins(int): Number of bins. Default: 10.
momentum(float): Momentum for moving gradient density. Default: 0.0.
mu(float): Super parameter for smoothed l1 loss. Default: 0.02.
Returns:
Tensor, GHM loss for regression.
"""
def __init__(self, bins=10, momentum=0.0, mu=0.02):
super(GHMRLoss, self).__init__()
self.bins = bins
self.momentum = momentum
self.mu = mu
edges_left = np.array([float(x) / bins for x in range(bins)], dtype=np.float32)
self.edges_left = Tensor(edges_left.reshape((bins, 1, 1, 1, 1)))
edges_right = np.array([float(x) / bins for x in range(1, bins + 1)], dtype=np.float32)
edges_right[-1] += 1e-4
self.edges_right = Tensor(edges_right.reshape((bins, 1, 1, 1, 1)))
if momentum >= 0:
self.acc_sum = Parameter(initializer(0, [bins], mstype.float32))
self.abs = ops.Abs()
self.sqrt = ops.Sqrt()
self.cast = ops.Cast()
self.select = ops.Select()
self.reshape = ops.Reshape()
self.reduce_sum = ops.ReduceSum()
self.max = ops.Maximum()
self.less = ops.Less()
self.equal = ops.Equal()
self.greater = ops.Greater()
self.logical_and = ops.LogicalAnd()
self.greater_equal = ops.GreaterEqual()
self.zeros_like = ops.ZerosLike()
self.expand_dims = ops.ExpandDims()
def construct(self, out, target):
"""GHM loss for regression"""
# ASL1 loss
diff = out - target
# gradient length
g = self.abs(diff / self.sqrt(self.mu * self.mu + diff * diff))
g = self.expand_dims(g, 0) # (1, b, c, h, w)
pos_inds = self.cast(self.equal(target, 1.0), mstype.float32)
tot = self.max(self.reduce_sum(pos_inds, ()), 1.0)
# (bin, b, c, h, w)
inds_mask = self.logical_and(self.greater_equal(g, self.edges_left), self.less(g, self.edges_right))
zero_matrix = self.cast(self.zeros_like(inds_mask), mstype.float32)
inds = self.cast(inds_mask, mstype.float32)
# (bins,)
num_in_bin = self.reduce_sum(inds, (1, 2, 3, 4))
valid_bins = self.greater(num_in_bin, 0)
num_valid_bin = self.reduce_sum(self.cast(valid_bins, mstype.float32), ())
if self.momentum > 0:
self.acc_sum = self.select(valid_bins,
self.momentum * self.acc_sum + (1 - self.momentum) * num_in_bin,
self.acc_sum)
acc_sum = self.acc_sum
acc_sum = self.reshape(acc_sum, (self.bins, 1, 1, 1, 1))
acc_sum = acc_sum + zero_matrix
weights = self.select(self.equal(inds, 1), tot / acc_sum, zero_matrix)
# (b, c, h, w)
weights = self.reduce_sum(weights, 0)
else:
num_in_bin = self.reshape(num_in_bin, (self.bins, 1, 1, 1, 1))
num_in_bin = num_in_bin + zero_matrix
weights = self.select(self.equal(inds, 1), tot / num_in_bin, zero_matrix)
# (b, c, h, w)
weights = self.reduce_sum(weights, 0)
weights = weights / num_valid_bin
ghmr_loss = self.sqrt(diff * diff + self.mu * self.mu) - self.mu
ghmr_loss = self.reduce_sum(ghmr_loss * weights, ()) / tot
return ghmr_loss
class RegLoss(nn.Cell): #reg_l1_loss class RegLoss(nn.Cell): #reg_l1_loss
""" """
Warpper for regression loss. Warpper for regression loss.
@ -458,31 +231,6 @@ class RegLoss(nn.Cell): #reg_l1_loss
return regr_loss return regr_loss
class RegWeightedL1Loss(nn.Cell):
"""
Warpper for weighted regression loss.
Args: None
Returns:
Tensor, regression loss.
"""
def __init__(self):
super(RegWeightedL1Loss, self).__init__()
self.reduce_sum = ops.ReduceSum()
self.gather_feature = TransposeGatherFeature()
self.cast = ops.Cast()
self.l1_loss = nn.L1Loss(reduction='sum')
def construct(self, output, mask, ind, target):
pred = self.gather_feature(output, ind)
mask = self.cast(mask, mstype.float32)
num = self.reduce_sum(mask, ())
loss = self.l1_loss(pred * mask, target * mask)
loss = loss / (num + 1e-4)
return loss
class LossCallBack(Callback): class LossCallBack(Callback):
""" """
Monitor the loss in training. Monitor the loss in training.

25
model_zoo/research/cv/centernet_det/src/visual.py Executable file → Normal file
View File

@ -22,15 +22,34 @@ import random
import cv2 import cv2
import numpy as np import numpy as np
import pycocotools.coco as COCO import pycocotools.coco as COCO
from .config import dataset_config as data_cfg from .model_utils.config import eval_config as eval_cfg
from .config import eval_config as eval_cfg
from .image import get_affine_transform, affine_transform from .image import get_affine_transform, affine_transform
coco_class_name2id = {'person': 1, 'bicycle': 2, 'car': 3, 'motorcycle': 4, 'airplane': 5,
'bus': 6, 'train': 7, 'truck': 8, 'boat': 9, 'traffic light': 10,
'fire hydrant': 11, 'stop sign': 13, 'parking meter': 14, 'bench': 15,
'bird': 16, 'cat': 17, 'dog': 18, 'horse': 19, 'sheep': 20, 'cow': 21,
'elephant': 22, 'bear': 23, 'zebra': 24, 'giraffe': 25, 'backpack': 27,
'umbrella': 28, 'handbag': 31, 'tie': 32, 'suitcase': 33, 'frisbee': 34,
'skis': 35, 'snowboard': 36, 'sports ball': 37, 'kite': 38, 'baseball bat': 39,
'baseball glove': 40, 'skateboard': 41, 'surfboard': 42, 'tennis racket': 43,
'bottle': 44, 'wine glass': 46, 'cup': 47, 'fork': 48, 'knife': 49, 'spoon': 50,
'bowl': 51, 'banana': 52, 'apple': 53, 'sandwich': 54, 'orange': 55, 'broccoli': 56,
'carrot': 57, 'hot dog': 58, 'pizza': 59, 'donut': 60, 'cake': 61, 'chair': 62,
'couch': 63, 'potted plant': 64, 'bed': 65, 'dining table': 67, 'toilet': 70,
'tv': 72, 'laptop': 73, 'mouse': 74, 'remote': 75, 'keyboard': 76, 'cell phone': 77,
'microwave': 78, 'oven': 79, 'toaster': 80, 'sink': 81, 'refrigerator': 82,
'book': 84, 'clock': 85, 'vase': 86, 'scissors': 87, 'teddy bear': 88,
'hair drier': 89, 'toothbrush': 90}
def coco_box_to_bbox(box): def coco_box_to_bbox(box):
"""convert height/width to position coordinates""" """convert height/width to position coordinates"""
bbox = np.array([box[0], box[1], box[0] + box[2], box[1] + box[3]], dtype=np.float32) bbox = np.array([box[0], box[1], box[0] + box[2], box[1] + box[3]], dtype=np.float32)
return bbox return bbox
def resize_image(image, anns, width, height): def resize_image(image, anns, width, height):
"""resize image to specified scale""" """resize image to specified scale"""
h, w = image.shape[0], image.shape[1] h, w = image.shape[0], image.shape[1]
@ -121,7 +140,7 @@ def visual_image(img, annos, save_path, ratio=None, height=None, width=None, nam
num_objects = len(annos) num_objects = len(annos)
name_list = [] name_list = []
id_list = [] id_list = []
for class_name, class_id in data_cfg.coco_class_name2id.items(): for class_name, class_id in coco_class_name2id.items():
name_list.append(class_name) name_list.append(class_name)
id_list.append(class_id) id_list.append(class_id)

115
model_zoo/research/cv/centernet_det/train.py
View File

@ -17,7 +17,6 @@ Train CenterNet and get network model files(.ckpt)
""" """
import os import os
import argparse
import mindspore.communication.management as D import mindspore.communication.management as D
from mindspore.communication.management import get_rank from mindspore.communication.management import get_rank
from mindspore import context from mindspore import context
@ -29,55 +28,22 @@ from mindspore.nn.optim import Adam
from mindspore import log as logger from mindspore import log as logger
from mindspore.common import set_seed from mindspore.common import set_seed
from mindspore.profiler import Profiler from mindspore.profiler import Profiler
from src.dataset import COCOHP from src.dataset import COCOHP
from src.centernet_det import CenterNetLossCell, CenterNetWithLossScaleCell from src.centernet_det import CenterNetLossCell, CenterNetWithLossScaleCell
from src.centernet_det import CenterNetWithoutLossScaleCell from src.centernet_det import CenterNetWithoutLossScaleCell
from src.utils import LossCallBack, CenterNetPolynomialDecayLR, CenterNetMultiEpochsDecayLR from src.utils import LossCallBack, CenterNetPolynomialDecayLR, CenterNetMultiEpochsDecayLR
from src.config import dataset_config, net_config, train_config from src.model_utils.config import config, dataset_config, net_config, train_config
from src.model_utils.moxing_adapter import moxing_wrapper
from src.model_utils.device_adapter import get_device_id, get_rank_id, get_device_num
_current_dir = os.path.dirname(os.path.realpath(__file__)) _current_dir = os.path.dirname(os.path.realpath(__file__))
parser = argparse.ArgumentParser(description='CenterNet training')
parser.add_argument('--device_target', type=str, default='Ascend', choices=['Ascend', 'CPU'],
help='device where the code will be implemented. (Default: Ascend)')
parser.add_argument("--distribute", type=str, default="true", choices=["true", "false"],
help="Run distribute, default is true.")
parser.add_argument("--need_profiler", type=str, default="false", choices=["true", "false"],
help="Profiling to parsing runtime info, default is false.")
parser.add_argument("--profiler_path", type=str, default=" ", help="The path to save profiling data")
parser.add_argument("--epoch_size", type=int, default="1", help="Epoch size, default is 1.")
parser.add_argument("--train_steps", type=int, default=-1, help="Training Steps, default is -1,"
"i.e. run all steps according to epoch number.")
parser.add_argument("--device_id", type=int, default=0, help="Device id, default is 0.")
parser.add_argument("--device_num", type=int, default=1, help="Use device nums, default is 1.")
parser.add_argument("--enable_save_ckpt", type=str, default="true", choices=["true", "false"],
help="Enable save checkpoint, default is true.")
parser.add_argument("--do_shuffle", type=str, default="true", choices=["true", "false"],
help="Enable shuffle for dataset, default is true.")
parser.add_argument("--enable_data_sink", type=str, default="true", choices=["true", "false"],
help="Enable data sink, default is true.")
parser.add_argument("--data_sink_steps", type=int, default="-1", help="Sink steps for each epoch, default is -1.")
parser.add_argument("--save_checkpoint_path", type=str, default="", help="Save checkpoint path")
parser.add_argument("--load_checkpoint_path", type=str, default="", help="Load checkpoint file path")
parser.add_argument("--save_checkpoint_steps", type=int, default=1000, help="Save checkpoint steps, default is 1000.")
parser.add_argument("--save_checkpoint_num", type=int, default=1, help="Save checkpoint numbers, default is 1.")
parser.add_argument("--mindrecord_dir", type=str, default="", help="Mindrecord dataset files directory")
parser.add_argument("--mindrecord_prefix", type=str, default="coco_det.train.mind",
help="Prefix of MindRecord dataset filename.")
parser.add_argument("--save_result_dir", type=str, default="", help="The path to save the predict results")
args_opt = parser.parse_args()
def _set_parallel_all_reduce_split(): def _set_parallel_all_reduce_split():
"""set centernet all_reduce fusion split""" """set centernet all_reduce fusion split"""
if net_config.last_level == 5: context.set_auto_parallel_context(all_reduce_fusion_config=[18, 59, 100, 141, 182])
context.set_auto_parallel_context(all_reduce_fusion_config=[16, 56, 96, 136, 175])
elif net_config.last_level == 6:
context.set_auto_parallel_context(all_reduce_fusion_config=[18, 59, 100, 141, 182])
else:
raise ValueError("The total num of allreduced grads for last level = {} is unknown,"
"please re-split after known the true value".format(net_config.last_level))
def _get_params_groups(network, optimizer): def _get_params_groups(network, optimizer):
@ -101,7 +67,7 @@ def _get_optimizer(network, dataset_size):
lr_schedule = CenterNetPolynomialDecayLR(learning_rate=train_config.PolyDecay.learning_rate, lr_schedule = CenterNetPolynomialDecayLR(learning_rate=train_config.PolyDecay.learning_rate,
end_learning_rate=train_config.PolyDecay.end_learning_rate, end_learning_rate=train_config.PolyDecay.end_learning_rate,
warmup_steps=train_config.PolyDecay.warmup_steps, warmup_steps=train_config.PolyDecay.warmup_steps,
decay_steps=args_opt.train_steps, decay_steps=config.train_steps,
power=train_config.PolyDecay.power) power=train_config.PolyDecay.power)
optimizer = Adam(group_params, learning_rate=lr_schedule, eps=train_config.PolyDecay.eps, loss_scale=1.0) optimizer = Adam(group_params, learning_rate=lr_schedule, eps=train_config.PolyDecay.eps, loss_scale=1.0)
elif train_config.lr_schedule == "MultiDecay": elif train_config.lr_schedule == "MultiDecay":
@ -109,7 +75,7 @@ def _get_optimizer(network, dataset_size):
if not isinstance(multi_epochs, (list, tuple)): if not isinstance(multi_epochs, (list, tuple)):
raise TypeError("multi_epochs must be list or tuple.") raise TypeError("multi_epochs must be list or tuple.")
if not multi_epochs: if not multi_epochs:
multi_epochs = [args_opt.epoch_size] multi_epochs = [config.epoch_size]
lr_schedule = CenterNetMultiEpochsDecayLR(learning_rate=train_config.MultiDecay.learning_rate, lr_schedule = CenterNetMultiEpochsDecayLR(learning_rate=train_config.MultiDecay.learning_rate,
warmup_steps=train_config.MultiDecay.warmup_steps, warmup_steps=train_config.MultiDecay.warmup_steps,
multi_epochs=multi_epochs, multi_epochs=multi_epochs,
@ -125,78 +91,85 @@ def _get_optimizer(network, dataset_size):
return optimizer return optimizer
def modelarts_pre_process():
"""modelarts pre process function."""
config.mindrecord_dir = config.data_path
config.save_checkpoint_path = os.path.join(config.output_path, config.save_checkpoint_path)
@moxing_wrapper(pre_process=modelarts_pre_process)
def train(): def train():
"""training CenterNet""" """training CenterNet"""
context.set_context(mode=context.GRAPH_MODE, device_target=args_opt.device_target) context.set_context(mode=context.GRAPH_MODE, device_target=config.device_target)
context.set_context(reserve_class_name_in_scope=False) context.set_context(reserve_class_name_in_scope=False)
context.set_context(save_graphs=False) context.set_context(save_graphs=False)
ckpt_save_dir = args_opt.save_checkpoint_path ckpt_save_dir = config.save_checkpoint_path
rank = 0 rank = 0
device_num = 1 device_num = 1
num_workers = 8 num_workers = 8
if args_opt.device_target == "Ascend": if config.device_target == "Ascend":
context.set_context(enable_auto_mixed_precision=False) context.set_context(enable_auto_mixed_precision=False)
context.set_context(device_id=args_opt.device_id) context.set_context(device_id=get_device_id())
if args_opt.distribute == "true": if config.distribute == "true":
D.init() D.init()
device_num = args_opt.device_num device_num = get_device_num()
rank = args_opt.device_id % device_num rank = get_rank_id()
ckpt_save_dir = args_opt.save_checkpoint_path + 'ckpt_' + str(get_rank()) + '/' ckpt_save_dir = config.save_checkpoint_path + 'ckpt_' + str(get_rank()) + '/'
context.reset_auto_parallel_context() context.reset_auto_parallel_context()
context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL, gradients_mean=True, context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL, gradients_mean=True,
device_num=device_num) device_num=device_num)
_set_parallel_all_reduce_split() _set_parallel_all_reduce_split()
else: else:
args_opt.distribute = "false" config.distribute = "false"
args_opt.need_profiler = "false" config.need_profiler = "false"
args_opt.enable_data_sink = "false" config.enable_data_sink = "false"
# Start create dataset! # Start create dataset!
# mindrecord files will be generated at args_opt.mindrecord_dir such as centernet.mindrecord0, 1, ... file_num. # mindrecord files will be generated at args_opt.mindrecord_dir such as centernet.mindrecord0, 1, ... file_num.
logger.info("Begin creating dataset for CenterNet") logger.info("Begin creating dataset for CenterNet")
coco = COCOHP(dataset_config, run_mode="train", net_opt=net_config, save_path=args_opt.save_result_dir) coco = COCOHP(dataset_config, run_mode="train", net_opt=net_config, save_path=config.save_result_dir)
dataset = coco.create_train_dataset(args_opt.mindrecord_dir, args_opt.mindrecord_prefix, dataset = coco.create_train_dataset(config.mindrecord_dir, config.mindrecord_prefix,
batch_size=train_config.batch_size, device_num=device_num, rank=rank, batch_size=train_config.batch_size, device_num=device_num, rank=rank,
num_parallel_workers=num_workers, do_shuffle=args_opt.do_shuffle == 'true') num_parallel_workers=num_workers, do_shuffle=config.do_shuffle == 'true')
dataset_size = dataset.get_dataset_size() dataset_size = dataset.get_dataset_size()
logger.info("Create dataset done!") logger.info("Create dataset done!")
net_with_loss = CenterNetLossCell(net_config) net_with_loss = CenterNetLossCell(net_config)
args_opt.train_steps = args_opt.epoch_size * dataset_size config.train_steps = config.epoch_size * dataset_size
logger.info("train steps: {}".format(args_opt.train_steps)) logger.info("train steps: {}".format(config.train_steps))
optimizer = _get_optimizer(net_with_loss, dataset_size) optimizer = _get_optimizer(net_with_loss, dataset_size)
enable_static_time = args_opt.device_target == "CPU" enable_static_time = config.device_target == "CPU"
callback = [TimeMonitor(args_opt.data_sink_steps), LossCallBack(dataset_size, enable_static_time)] callback = [TimeMonitor(config.data_sink_steps), LossCallBack(dataset_size, enable_static_time)]
if args_opt.enable_save_ckpt == "true" and args_opt.device_id % min(8, device_num) == 0: if config.enable_save_ckpt == "true" and get_device_id() % min(8, device_num) == 0:
config_ck = CheckpointConfig(save_checkpoint_steps=args_opt.save_checkpoint_steps, config_ck = CheckpointConfig(save_checkpoint_steps=config.save_checkpoint_steps,
keep_checkpoint_max=args_opt.save_checkpoint_num) keep_checkpoint_max=config.save_checkpoint_num)
ckpoint_cb = ModelCheckpoint(prefix='checkpoint_centernet', ckpoint_cb = ModelCheckpoint(prefix='checkpoint_centernet',
directory=None if ckpt_save_dir == "" else ckpt_save_dir, config=config_ck) directory=None if ckpt_save_dir == "" else ckpt_save_dir, config=config_ck)
callback.append(ckpoint_cb) callback.append(ckpoint_cb)
if args_opt.load_checkpoint_path: if config.load_checkpoint_path:
param_dict = load_checkpoint(args_opt.load_checkpoint_path) param_dict = load_checkpoint(config.load_checkpoint_path)
load_param_into_net(net_with_loss, param_dict) load_param_into_net(net_with_loss, param_dict)
if args_opt.device_target == "Ascend": if config.device_target == "Ascend":
net_with_grads = CenterNetWithLossScaleCell(net_with_loss, optimizer=optimizer, net_with_grads = CenterNetWithLossScaleCell(net_with_loss, optimizer=optimizer,
sens=train_config.loss_scale_value) sens=train_config.loss_scale_value)
else: else:
net_with_grads = CenterNetWithoutLossScaleCell(net_with_loss, optimizer=optimizer) net_with_grads = CenterNetWithoutLossScaleCell(net_with_loss, optimizer=optimizer)
model = Model(net_with_grads) model = Model(net_with_grads)
model.train(args_opt.epoch_size, dataset, callbacks=callback, model.train(config.epoch_size, dataset, callbacks=callback,
dataset_sink_mode=(args_opt.enable_data_sink == "true"), sink_size=args_opt.data_sink_steps) dataset_sink_mode=(config.enable_data_sink == "true"), sink_size=config.data_sink_steps)
if __name__ == '__main__': if __name__ == '__main__':
if args_opt.need_profiler == "true": if config.need_profiler == "true":
profiler = Profiler(output_path=args_opt.profiler_path) profiler = Profiler(output_path=config.profiler_path)
set_seed(317) set_seed(317)
train() train()
if args_opt.need_profiler == "true": if config.need_profiler == "true":
profiler.analyse() profiler.analyse()