Hello dear developer, thank you very much for your open source project, I am a health researcher, now I am doing a pedestrian step counting experiment, this experiment can make us pay more attention to health, but there are some problems when using your Timegpt model in the experiment, 1, the anomaly detection function of this large model does not need training data? Because sometimes in the recognition process, the pseudo-wave peak of the accelerometer is also detected as an outlier, so two outliers are detected in a one-step cycle, and the true wave peak and the pseudo-wave peak conflict, resulting in the detected number of steps greater than the real number of steps, as shown in the attachment example.png. Is there any good way to solve this problem? 2, or label my sample data at each true peak, and then enter the test data to predict with the prediction function instead of using the anomaly detection function? For example, the peak and valley tags in excel.png.3, or other fine-tuning methods, do you have any better recommendations? peak_detection.png is the result of my calculation using traditional methods

I am looking forward to your reply. You have made a great contribution to the cause of health