Only one way has a good accuracy

[Qingfengxulailu]

1. run full-shot.py, not_full_shot must be false
2. run few-show.py or cross-class-....py, the test accuracy is good(it will load model saved by full-shot.py）
3. Otherwise, the accuracy on the test set will be very low under any other setting.

[RS2002]

Thank you for your effort to reproduce our work. For your reference, we have attached our training process in the zip file below. Throughout the training, we tracked accuracy on the test set.

We suggest that you further examine the entire running process, as we have not received similar feedback about this issue. Our method has also been successfully reproduced on several other datasets in recent papers.

Additionally, please note that you should not set --not_full_shot false during pre‑training, as doing so can cause domain information leakage.

training process.zip

[Qingfengxulailu]

OKay. For a full-shot setup, why not shuffle the data first and then split the shuffled data into training and testing sets?

[RS2002]

In Wi-Fi sensing, we typically split the dataset by time rather than using random shuffling. This way, the testing set does not share the same time period as the training set, avoiding any potential information leakage.

[Qingfengxulailu]

However, when selecting data sequentially, samples of the same action within a domain may be very similar to each other.

[RS2002]

Yes you are right. But it seems not a problem.