Reproducibility Issues #6
Description
Hi,
I'm trying to reproduce table 1 from the paper but get varying results even for SGD. My hardware setup is definitely different, I run the codes on a single GPU machine but I think SGD results should not vary.
| Model | Aug | CIFAR-10 | CIFAR-100 |
|---|---|---|---|
| WRN-28-10(200 epochs) | Basic | 3.8 ± 0.1 | 19.1±0.1 |
| WRN-28-10(200 epochs) | Cutout | 2.9 ± 0.1 | 17.7±0.2 |
| WRN-28-10(200 epochs) | AA | 3.6 ± 0.1 | 19.0±0.4 |
| ------------- | ------------- | ------------- | ------------- |
| Shake-Shake (26 2x96d) | Basic | 3.5 ± 0.1 | 19.0± 0.1 |
| Shake-Shake (26 2x96d) | Cutout | 2.9 ± 0.1 | 18.2±0.2 |
| Shake-Shake (26 2x96d) | AA | 3.3 ±0.1 | 17.9±0.2 |
| ------------- | ------------- | ------------- | ------------- |
Can you maybe provide corresponding FLAGS for basic / cutout / aa settings in table 1