LSTMED is too slow

[Willi Gierke]

One training epoch on 100-dimensional data with delayed outliers takes 4 seconds.
Calculating the error scores for one batch in the prediction takes 10 seconds and needs to be done nearly 5000 times in order to predict labels for one dataset. The root of all evil seems to be calculating the logpdf (score = -multivariate_normal.logpdf(error.view(1, -1).data.cpu().numpy(), mean=self.mean, cov=self.cov, allow_singular=True)).
Can we tune that somewhere?