disjoint datasets merge model accuracy

[qsunyuan]

Nice work.

The loss results for the test set are reported in Figure 5, but what about the accuracy.

In my experimental results, although the loss does drop, it seems that the accuracy is not very good (lower than the naive weight interp).

[samuela]

Here's what I'm getting:

[adampan0527]

Here's what I'm getting:

Hi! I'm trying to use git-rebasin in my work about disjoint datasets merge, while my result looks like this

This experiment is using ResNet20 with 32x width-multiplier on CIFAR10 dataset, splited into two CIFAR5 dataset.
Can you tell me which model and dataset were you using? This is really important to me, thanks a lot!!

[samuela]

Here is the code used to generate the plot in the paper: https://github.com/samuela/git-re-basin/blob/main/src/cifar100_resnet20_split_data_plot.py