ImageLearner.predict() orphaned in 1.0.21?


I am noticing that a Learner created by create_cnn() is missing the predict() method in the 1.0.21.

This seems to be a consequence of this new class structure with predict method being in ClassificationLearner but create_cnn does not currently seem to create a ClassificationLearner.

I would create a PR but I’m not sure what design / structure is desired so I will just report it here. Please let me know if you would like a github issue.

1 Like

Yeah sorry it’s broken. Looking in to it.

Fixed now. Please update to latest from conda/pip.


Thank you! predict() works now.

However, the return result of predict now seems to be unnormalized (i.e., they are not probabilities in range 0…1_).

For example:

('anomaly', tensor(0), tensor([19.6412, 0.0458]))

where ‘anomaly’ is one of the classification labels.

Mh strange. I thought they were not yet exponentiated, but “19” cannot obviously get mapped in [0,1] once you exponentiate it.

May be e**[…] has to be normalized to [0…1]

These values are sometimes negative so that would make sense.