Testing

Results compared with the thesis, and new runs

Preliminary: 0 of the 24 few-shot thesis settings on CIFAR-100 have their full 10 runs, so this reproduction isn't finished. Features use torchvision ResNet-34 weights, not his Keras weights.

Results compared with the thesis

Average incremental accuracy. Thesis values are means of 10 runs. "Within 3 points" is this project's own rule, not a standard. Standard deviations are between runs. Average incremental is the mean of per-class accuracies, as in the official code, and is what's compared with the thesis. Pooled counts every test image equally, which differs when class sizes are uneven; the thesis doesn't say which its Caltech-101 numbers use.

No results yet.

Accuracy as classes are added

No results for this setting yet.