Help Center/ ModelArts/ Drawer/ Deploying a Real-Time Service/ Hard Example Filtering
Updated on 2025-05-29 GMT+08:00
Share

Hard Example Filtering

Description

When hard example filtering is enabled, the system checks if the collected data contains hard examples and stores the filtering result in the corresponding dataset.

  • If your model is trained by yourself, ensure that the output format in the inference code meets the requirements.

    For example, if you use a mainstream framework or a custom image to train a model, the prediction output format in the inference code must match the corresponding type when you import the model.

  • If your model is obtained from ExeML or a preset algorithm, no modification is required.

FAQs

What Is Hard Example Data?

Hard examples are those that the model fails to recognize correctly or confidently.

Model maintenance is an ongoing process in real-world service scenarios. For example, data retraining may be done on a weekly or monthly basis, or when the data reaches a certain volume that requires periodic retraining. If full data is used for retraining, a large amount of labeling manpower and training time are required. To improve model maintenance efficiency, retraining based on difficult example data can be used.

Parent Topic: Deploying a Real-Time Service