train on normal data?

[JokerDu]

Hi, thank you for sharing the anomaly detection code. I notice that author Pankaj Malhotra uses only the normal sequences for training in paper LSTM-based Encoder-Decoder for Multi-sensor Anomaly Detection. But I found you use not only normal sequences but anomaly sequences for training.
in code 0_download_dataset.py, you tag label of anomaly data as 1, label of normal data as 0.

tokens.append(1) if 150 < i < 250 or   \
                            5970 < i < 6050 or \
                            8500 < i < 8650 or \
                            8750 < i < 8890 or \
                            10000 < i < 10200 or \
                            14700 < i < 14800 \
                          else tokens.append(0)

train data is labeled_data[:13104], which including normal sequences and anomaly sequences, could you explain that, thank you so much~

[chickenbestlover]

As you said, I included anomaly points in the trainset in my code. Ideally, the trainset would contain only normal points, but in the nyc_taxi dataset, it was almost impossible to construct a trainset with only normal points because the anomaly points were included in several parts. That's why we had to include anomaly points in the trainset. But because most of the trainsets are normal points, this was not a big problem for learning.