Multivariate Anomaly Dectector (Error when running tutorial)

I get the error below when I run the tutorial kats_202_detection.ipynb https://github.com/facebookresearch/Kats/blob/master/tutorials/kats_202_detection.ipynb Any clue ?

KeyError: Timestamp('2019-12-23 23:59:58.142906')

The above exception was the direct cause of the following exception:

KeyError                                  Traceback (most recent call last)
<ipython-input-265-b73347c84449> in <module>
      3 d = MultivariateAnomalyDetector(multi_anomaly_ts, params, training_days=60)
      4 display(params)
----> 5 anomaly_score_df = d.detector()
      6 
      7 d.plot()

~\AppData\Local\Continuum\anaconda3\lib\site-packages\kats\detectors\outlier.py in detector(self)
    300         while fcstTime < self.df.index.max():
    301             # forecast for fcstTime+ 1
--> 302             pred_df = self._generate_forecast(fcstTime)
    303             # calculate anomaly scores
    304             anomaly_scores_t = self._calc_anomaly_scores(pred_df)

[pbosch]

@loicduffar @Chima-21 Do you use Windows or Linux? I encountered the issue myself in Windows using Miniconda (4.10.3) and using WSL or a native Ubuntu doesn’t produce the issue. In WSL/Ubuntu it works with native Python and Miniconda.

The culprit is in in calculating the granularity:

       if len(time_diff.unique()) == 1:  # check constant frequenccy
            freq = time_diff.unique()[0].astype("int")
            self.granularity_days = freq / (24 * 3600 * (10 ** 9))
        else:
            raise RuntimeError(
                "Frequency of metrics is not constant."
                "Please check for missing or duplicate values"
            )

In WSL/Ubuntu I get a straight 1.0 for the example data. In Windows I get -2.149413925925926e-05. More precisely, the issue is in the astype call. It seems to default to int32 on Windows, which causes an overflow. Using int64 instead of int solves the problem.

I would recommend in this instance, and in general, to use explicit types instead of assuming that the default type is correct. Int on Windows systems usually defaults to 32bit while float usually defaults to 64bit. By marking them explicit with int64 and float64, whether it is with numpy or pandas, you would avoid that issue completely and make it a bit more robust. A quick search for astype in the repository shows that it’s mostly implicit, so it could be that this kind of problem occurs in other places as well.

[michaelbrundage]

Renaming the issue to the root cause.

Most likely, we should mark Kats as requiring 64-bit. I think we don’t intend to support legacy 32-bit Python installations.