iCaRL's accuracy different from 'Three scenarios for continual learning'

[gmvandeven]

iCaRL's accuracy different from 'Three scenarios for continual learning'

[GMvandeVen]

Thank you for pointing this out! You're right, there was a mistake in the code. The stored exemplars were not actually being added to the training data. They were added to [training_dataset], but then erroneously [train_dataset] was used to train on. The change I made to fix this was to replace [train_dataset] by [training_dataset] in line 117 of train.py.

The version of the code used to obtain the results reported in the paper did not have this mistake, hence the discrepancy. Sorry for the mistake and thank you for pointing it out!

[JosephKJ]

Thank you for your quick response.

I have verified your fix. Thanks!

############################################################
SUMMARY RESULTS: splitMNIST  -  class-incremental learning
------------------------------------------------------------
None         19.91
EWC          20.49
o-EWC        20.66
SI           19.91
LwF          23.60
DGR          91.63
DGR+distil   93.46
iCaRL        95.08
Offline      98.15
############################################################