Adding the option of mixing current data with historic data for CL

[anagainaru]

Summary

Adding logic to mix historic and current data for updaters that fallback to the fwd_bwd function in the base updater. The current solution concatenates half of the training bath with half of the historic batch and uses this mix as the training loader. How much of the history is kept in the historic data loaders is defined in the model harness. The MNIST example keeps everything but others could keep only the last few windows.

New option for the configuration file:

mix_historic_data = true

Changing the MNIST example to keep track of the entire history instead of just the last window to illustrate the benefit of using history for doing continual learning.

• Yellow line is obtained using:

pretrained_path = "examples/mnist/mnist.pth"
[drift_detection]
detection_interval = 0

• Blue line is using the model saved at the end of running Apeiron on the default model with Base CL and ADWin drift detector and mix_historic_data = false.

• Green line is using the model saved at the end of running Apeiron on the default model with Base CL and ADWin drift detector and mix_historic_data = true.

Changes to the MNIST example: see below why is needed

[krm9c]

I think, I am only confused about the following lines

TransformedView(self.ds_val, x_transform=FixedAffine(self.aug_history[:-1]))
TransformedView(self.ds_val, x_transform=FixedAffine(self.aug_history[:i]))
for i in range(1, len(self.aug_history))

Otherwise the results make sense.

[anagainaru]

I think, I am only confused about the following lines

TransformedView(self.ds_val, x_transform=FixedAffine(self.aug_history[:-1])) TransformedView(self.ds_val, x_transform=FixedAffine(self.aug_history[:i])) for i in range(1, len(self.aug_history))

Otherwise the results make sense.

TransformedView(self.ds_train, x_transform=FixedAffine(self.aug_history[:-1]))

FixedAffine applies all the transforms in the list in sequence. With [:-1] (every drift except the current), the history is rendered exactly like the previous window — one snapshot, not the full past.

With the change:

FixedAffine(self.aug_history[:i]) for i in range(1, len(self.aug_history))

One view per past transform sequence we keep in the history, the same dataset at every rotation we've seen before.

With [:-1], replay only showed the previous regime, so the model kept converging toward the latest images instead of retaining all past rotations.

[anagainaru]

Next time please approve and do not merge

[krm9c]

Oops, I am sorry, How much did I mess up things? I just wanted to approve and thought merge and approve were the same thing. Sorry. On Jul 25, 2026, at 21:22, Ana Gainaru ***@***.***> wrote: [https://avatars.githubusercontent.com/u/16229479?s=20&v=4]anagainaru left a comment (AI-ModCon/BaseSIM_APEIRON#107)<#107 (comment)> Next time please approve and do not merge — Reply to this email directly, view it on GitHub<#107?email_source=notifications&email_token=ABKJFRJFOXRP5TTU4GEJKB35GVTQDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMBYGE3DANBYGQ3KM4TFMFZW63VMON2GC5DFL5RWQYLOM5S2KZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-5081604846>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ABKJFROYCM6EZHFAPHTEHW35GVTQDAVCNFSNUABGKJSXA33TNF2G64TZHMYTAOBVGMYDONZRGE5US43TOVSTWNBZGM4TMNRYGAZDPILWAI>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS<https://github.com/notifications/mobile/ios/ABKJFRL7Z3TEE5ZPOXN3T2L5GVTQDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMBYGE3DANBYGQ3KM4TFMFZW63VMON2GC5DFL5RWQYLOM5S2KZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android<https://github.com/notifications/mobile/android/ABKJFRP3OY3ED6W4DL5ZRL35GVTQDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMBYGE3DANBYGQ3KM4TFMFZW63VMON2GC5DFL5RWQYLOM5S2KZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you modified the open/close state.Message ID: ***@***.***>