The icarl model is appending the selected indexes for current task to the herding_indexes list, every time when the build_examplars() function is called.
This is fine if in each task build_examplars() is called just once, or when self._eval_every_x_epochs is 0. But when self._eval_every_x_epochs >0, build_examplars() is called for multiple times. The herding_indexes list is not storing correct indexes.
For example in task 1 we have 10 classes and we call build_examplars() for 2 times. The length of herding_indexes is 20. In task 2, we are using herding_indexes[10] to herding_indexes[19] to select examples for classes in task 2. While these indexes are calculated with the features of task 1!
https://github.com/arthurdouillard/incremental_learning.pytorch/blob/0d25c2e12bde4a4a25f81d5e316751c90e6f789b/inclearn/models/icarl.py#L446
The icarl model is appending the selected indexes for current task to the herding_indexes list, every time when the build_examplars() function is called.
This is fine if in each task build_examplars() is called just once, or when self._eval_every_x_epochs is 0. But when self._eval_every_x_epochs >0, build_examplars() is called for multiple times. The herding_indexes list is not storing correct indexes.
For example in task 1 we have 10 classes and we call build_examplars() for 2 times. The length of herding_indexes is 20. In task 2, we are using herding_indexes[10] to herding_indexes[19] to select examples for classes in task 2. While these indexes are calculated with the features of task 1!