Experiments

[Grutschus]

This is a collection of experiments we want to run.
This is a follow up on our discussion from Friday #36.

• 34 set up the experiments #39 (vit-s_uniform-sampling_priority-labeling_paper-hyperparams)
• vit-b_gaussian-sampling_priority-labeling_paper-hyperparams
• vit-b_gaussian-sampling_priority-labeling_paper-hyperparams_weighted-ce-loss
• vit-b_uniform-sampling_priority-labeling_paper-hyperparams_undersampling-other-0.75
• vit-gaussian-sampling_priority-labeling_paper-hyperparams_undersampling-other-0.75
• Repeat some promising experiments with the ViT-B K400 hyperparams from here

This list is not exhaustive. We'll add ideas and experiments as we progress, for example when #38 is done.

[Grutschus]

We found out that the ViT-B K400 hyperparams lead to a far more stable training.