A pytorch toy of ICCV paper No Fuss Distance Metric Learning using Proxies. The proxynca.py is from https://github.com/dichotomies/proxy-nca.
CUB 200-2011 data is organized as
cub_pytorch
|---train
|-----subdirectory 0000 to 0099
|---val
|-----subdirectory 0100 to 0199
see train_cub.sh
| Metric | dichotomies's Implementation | Google's Implementation | This Implementation(epoch 100) |
|---|---|---|---|
| R@1 | 52.46 | 49.21 | 51.76 |
| R@2 | 64.78 | 61.90 | 64.30 |
| R@4 | 75.38 | 67.90 | 75.24 |
| R@8 | 84.31 | 72.40 | 85.09 |
| NMI | 60.84 | 59.53 | - |