The overall network architecture of YOLOv3 is show below:
And we use ResNet18 as the backbone of YOLOv3_ResNet18. The architecture of ResNet18 has 4 stages. The ResNet architecture performs the initial convolution and max-pooling using 7×7 and 3×3 kernel sizes respectively. Afterward, every stage of the network has different Residual blocks (2, 2, 2, 2) containing two 3×3 conv layers. Finally, the network has an Average Pooling layer followed by a fully connected layer.
@ -29,17 +63,84 @@ YOLOv3 network based on ResNet-18, with support for training and evaluation.
Each row is an image annotation which split by space, the first column is a relative path of image, the others are box and class infomations of the format [xmin,ymin,xmax,ymax,class]. `dataset.py` is the parsing script, we read image from an image path joined by the `image_dir`(dataset directory) and the relative path in `anno_path`(the TXT file path), `image_dir` and `anno_path` are external inputs.
## Running the Example
# [Environment Requirements](#contents)
- Hardware（Ascend）
- Prepare hardware environment with Ascend processor. If you want to try Ascend , please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
After installing MindSpore via the official website, you can start training and evaluation on Ascend as follows:
- runing on Ascend
```shell script
#run standalone training example
sh run_standalone_train.sh [DEVICE_ID] [EPOCH_SIZE] [MINDRECORD_DIR] [IMAGE_DIR] [ANNO_PATH]
#run distributed training example
sh run_distribute_train.sh [DEVICE_NUM] [EPOCH_SIZE] [MINDRECORD_DIR] [IMAGE_DIR] [ANNO_PATH] [RANK_TABLE_FILE]
#run evaluation example
sh run_eval.sh [DEVICE_ID] [CKPT_PATH] [MINDRECORD_DIR] [IMAGE_DIR] [ANNO_PATH]
```
# [Script Description](#contents)
## [Script and Sample Code](#contents)
```
└── model_zoo
├── README.md // descriptions about all the models
└── yolov3_resnet18
├── README.md // descriptions about yolov3_resnet18
├── scripts
├── run_distribute_train.sh // shell script for distributed on Ascend
├── run_standalone_train.sh // shell script for distributed on Ascend
└── run_eval.sh // shell script for evaluation on Ascend
├── src
├── dataset.py // creating dataset
├── yolov3.py // yolov3 architecture
├── config.py // parameter configuration
└── utils.py // util function
├── train.py // training script
└── eval.py // evaluation script
```
## [Script Parameters](#contents)
```
Major parameters in train.py and config.py as follows:
evice_num: Use device nums, default is 1.
lr: Learning rate, default is 0.001.
epoch_size: Epoch size, default is 10.
batch_size: Batch size, default is 32.
pre_trained: Pretrained Checkpoint file path.
pre_trained_epoch_size: Pretrained epoch size.
mindrecord_dir: Mindrecord directory.
image_dir: Dataset path.
anno_path: Annotation path.
img_shape: Image height and width used as input to the model.
```
### Training
## [Training Process](#contents)
### Training on Ascend
To train the model, run `train.py` with the dataset `image_dir`, `anno_path` and `mindrecord_dir`. If the `mindrecord_dir` is empty, it wil generate [mindrecord](https://www.mindspore.cn/tutorial/en/master/use/data_preparation/converting_datasets.html) file by `image_dir` and `anno_path`(the absolute image path is joined by the `image_dir` and the relative path in `anno_path`). **Note if `mindrecord_dir` isn't empty, it will use `mindrecord_dir` rather than `image_dir` and `anno_path`.**
- Stand alone mode
```
sh run_standalone_train.sh 0 50 ./Mindrecord_train ./dataset ./dataset/train.txt
```
The input variables are device id, epoch size, mindrecord directory path, dataset directory path and train TXT file path.
Note the results is two-classification(person and face) used our own annotations with coco2017, you can change `num_classes` in `config.py` to train your dataset. And we will suport 80 classifications in coco2017 the near future.
### Evaluation
## [Evaluation Process](#contents)
### Evaluation on Ascend
To eval, run `eval.py` with the dataset `image_dir`, `anno_path`(eval txt), `mindrecord_dir` and `ckpt_path`. `ckpt_path` is the path of [checkpoint](https://www.mindspore.cn/tutorial/en/master/use/saving_and_loading_model_parameters.html) file.
@ -92,3 +194,46 @@ class 1 precision is 85.34%, recall is 79.13%
Note the precision and recall values are results of two-classification(person and face) used our own annotations with coco2017.
chengxianbin
5 years ago