Too strict sanity check for Pelt model?

[ChiQiao]

Sanity check was added in v1.1.4 for the Pelt class, and raises a BadSegmentationParameters for the example below

import ruptures
import numpy as np

model = ruptures.Pelt(min_size=5, jump=1)
model.fit(np.ones(9))
model.predict(pen=0)

For a signal with a length of 9, min_size=5 makes sense to me and in v1.1.3 it returns [9]. I wonder if the sanity check is too strict?

[oboulant]

Hi,

Thanks for your interest in ruptures !

Here, you are giving as an input

• a signal with 9 samples
• a segment minimum size of 5 samples
• a jump size of 1 sample

With those parameters, the sanity checks verify that it can generate at least 2 segments (ie. 1 break point). Here, what you are presently asking is not feasible, indeed :

• You would need 10 samples in total to generate two segments with at least 5 samples in each

Walkthrough in the Source Code

• As for the Pelt detection method, Sanity check is call at predict time here
• The sanity check is implemented here. It basically checks that it has enough points in an input signal of size n_samples to find the asked number of break points (n_bkps) for some given segment minimum size (min_size) and jump size (jump).

Alternatives for you situation

In your case, the following choices would work :

model = ruptures.Pelt(min_size=5, jump=1)
model.fit(np.ones(10))
model.predict(pen=0)

model = ruptures.Pelt(min_size=4, jump=1)
model.fit(np.ones(9))
model.predict(pen=0)

model = ruptures.Pelt(min_size=5, jump=2)
model.fit(np.ones(11))
model.predict(pen=0)

[deepcharles]

@ChiQiao Indeed it makes sense to return [9]. We will see how this change propagates through the rest of the package.

Thanks

[oboulant]

Oh, did not pay attention to the Pelt mention and the detection methods that may return 0 break point (even more with a 0 penalty). Will look into it !