Introduction

The APIs of the training component are encapsulated based on the MindSpore source code. The training and evaluation script files of the SSD_MobileNetV1_FPN, ResNet-50, YOLOv4, U-Net, UNet++, CTPN, and CRNN models are provided, covering five application scenarios: object detection, object classification, text detection, text recognition, and semantic segmentation.

The training code is configured by command-line parameters, and is suitable for model training, model conversion, model evaluation, accuracy result saving, and visualization file generation.

In addition, training scripts required for semiconductor quality inspection (see Quality Inspection on Semiconductor Wafer) and dataset preprocessing scripts that may be used in some specific scenarios (see Dataset Processing Tool Package) are provided.

Figure 1 Interaction process in the training component

Generally, the training code is stored in a folder corresponding to each model. model_train.py and model_eval.py are the entry script files for training and evaluation. Each directory stores the same public function library com_package.

version.info records the training code version. ssd_tiled_dataset_mindspore is created based on the ssd_mobilenet_fpn model for small object detection. scripts stores scripts of tool packages related to training scenarios, such as a data augmentation script. The directory structure of the training code folders is as follows:

.
Training code folders
├── crnn_mindspore
├── ctpn_mindspore
├── resnet50_mindspore
└── scripts 
	├── com_package  # Public function library
	└── dataset_tools  # Dataset processing tool library
		├── coco_data_aug.py  # Offline data augmentation for a COCO object detection dataset
		├── tag_paste_cls_dataset.py  # Obtains images for tag classification.
		└── tag_paste_data_augmentation.py  # Data augmentation tool for tag classification
├── ssd_mobilenet_fpn_mindspore
├── ssd_tiled_dataset_mindspore
├── unet_mindspore
├── unet_nested_mindspore
├── version.info  # Version of the release package
├── wafer_mindspore
└── yolov4_mindspore
    ├── com_package
    ├── inside_eval.py
    ├── inside_train.py
    ├── model_eval.py  # Entry of the evaluation script
    ├── model_train.py  # Entry of the training script
    ├── model_wrapper  # Libraries on which model training and evaluation depend, for example, definitions of the dataset, network, evaluation mode, and callback function
    ├── om_convert  # Converts the OM offline model, including the AIPP configuration file.
    ├── on_platform  # Configuration file provided for the platform
    └── pre_trained_ckpt  # Stores the pre-trained model.