python devise/devise_gzsl.py -data AWA2/AWA1/CUB/SUN/APY -e [EPOCHS] -es [EARLY STOP] -norm [NORMALIZATION TYPE] -lr [LEARNING RATE] -mr [SVM LOSS MARGIN]
For testing, set learning rate (lr), margin (mr), and normalization type (norm) to best combination from the tables below.
The numbers below are class-averaged top-1 accuracies (see ZSLGBU paper for details).
| Dataset | ZSLGBU Results | Repository Results | Hyperparams from Val |
|---|---|---|---|
| CUB | 52.0 | 44.07 | lr=1.0, mr=1.0, norm=L2 |
| AWA1 | 54.2 | 55.25 | lr=0.01, mr=200, norm=std |
| AWA2 | 59.7 | 57.68 | lr=0.001, mr=150, norm=std |
| aPY | 39.8 | 33.33 | lr=1.0, mr=1.0, norm=L2 |
| SUN | 56.5 | 55.69 | lr=0.01, mr=3.0, norm=None |
To be updated soon...
No existing codebases found!