Deep Metric Learning
The deep metri learning project for 2020 Summer Intern in Nokia Bell Labs, advised by Dr. Fangzhe Chang.
Acknowledgement
This repo mainly borrows the architecture design and codes from classification_metric_learning.
We also refer to the codes in Deep_Metric.
To Do
- Rebuild code structure
- Multi-GPU
- Half precision/ mix precision training.
- DALI for Data procecssing
- Enclose more losses as a module, like triplet, contrastive ...
- Sampling schems as a module
- New logger system
- Backbone as a module
- Object detection
Results
Backbone: pretrained Resnet50; dim: 2048; epochs: 30;
CUB_200_2011
| Recall@1 | Recall@2 | Recall@4 | Recall@8 | Download |
|---|---|---|---|---|
| 60.45 | 71.51 | 81.33 | 88.66 | [model] [log] |
Stanford Online Products
| Recall@1 | Recall@10 | Recall@100 | Recall@1000 | Download |
|---|---|---|---|---|
| 79.33 | 91.38 | 96.58 | 98.92 | [model] [log] |
Progress
More results and codes reconstruction are in progress.