avalanche
avalanche copied to clipboard
CORe50 Dataset and Benchmark Improvements
We should:
- [x] Expand category-level classification support.
- [x] Add CORe50 mini (32x32)
- [ ] Test set split in experiences option
- [ ] Document the dataset and benchmarks better.
- [ ] Add default transformations
- [ ] Add segmentation and detection support.
Additionally I would like to see a flag for using a subsampled test set. I've found Core50 test set to be needlessly large. In fact the Core50 authors end up sub-sampling it in later works.
"we sub-sampled the test set by selecting 1 frame every second (from the original 20 fps)"
Lomonaco, V., Maltoni, D., & Pellegrini, L. (2020). Rehearsal-Free Continual Learning over Small Non-I.I.D. Batches. ArXiv:1907.03799 [Cs, Stat]. http://arxiv.org/abs/1907.03799
Additionally some sort of meta-loader that wraps any benchmark allowing sampling to be applied to any dataset might be useful.