Open StellaAthena opened 3 years ago
@researcher2 and I have made significant progress on this, but we haven’t ironed out all the issues yet. Here’s Table 1 from the paper followed by our current best results
Is it the result from Cifar or ImageNet with reduced classes?
I believe the results shown here were CIFAR10. The network marking step only produced good results for us on CIFAR10 and Imagenette when doing a finetune (linear classification layer only). We never moved onto a full imagenet run.
The result shows that from not radioactive to 10% of data become radioactive, the log10(p) only increase +4? is that a very small change due to the randomness comparing to results from other ratio.
Do you also have the cosine similarity result (average) according to the test?
CIFAR10 Table 1 This ended up working.
I couldn't find table 2 lying around, you would have to re-run. It didn't work though for CIFAR10 or Imagenette, the log10(p) was basically random with respect to percentage marked.
We don't have a log of cosine similarity, please feel free to play around - the main thing I notice is we have weird marking percentages of 0.1, 0.2 etc in the committed code, must have been playing around at the time.
To ensure that our refactored code is working properly, we need to replicate the experiments from the paper. We want to replicate the following experiments: