Open NicolasMICAUX opened 3 years ago
Same problem, I got even lower accuracy in mnist. Do you figure out why is that now?
Same problem, I got even lower accuracy in mnist. Do you figure out why is that now? No. It was part of a 1 week student prioject, so i didn't spend much time on this.
Hope you'll figured it out easily :)
Hi, I had the same problem, but then I decided to decrease the trigger set size from 5000 (source class completely) to 100 and the results were inline with the paper. I did not change any other default configs for CIFAR10. Here my results:
Trigger set size | Test Accuracy | Watermark accuracy |
---|---|---|
5000 | 80.73 | 13.70 |
100 | 90.29 | 27.00 |
in paper | 85.41 | 25.74 |
Note: I converted the implementation to Torch, so it might still not work for this repo.
Hi, I had the same problem, but then I decided to decrease the trigger set size from 5000 (source class completely) to 100 and the results were inline with the paper. I did not change any other default configs for CIFAR10. Here my results:
Trigger set size Test Accuracy Watermark accuracy 5000 80.73 13.70 100 90.29 27.00 in paper 85.41 25.74 Note: I converted the implementation to Torch, so it might still not work for this repo.
Hi,I had the same problem. Now I am converting the code to pytorch,but I met some problems.Can I communicate with you?
Hello, I tried
python3 train.py --dataset cifar10 **--default 1**
with tensorflow 1.x, on the unmodified code cloned from this repo. Training took ~3h30, but results are :Victim Model || validation accuracy: 0.8653846085071564, watermark success: 0.10416666666666667
which seems very poor performance.Have you recently been able to run this code successfully ? Thank you for your attention.
The logs are :
Victim Model || validation accuracy: 0.8653846085071564, watermark success: 0.10416666666666667