Triplet Loss is one basic metric learning algo. Hashing is closely related to metric learning. It seems there isn't much public and solid codes for Triplet Hashing. Here is my version of triplet hashing that can achieve acceptable performance.
Pytorch=1.2.0 torchvision=0.4.0 cudatoolkit=10.0.130 NvidiaDriverVersion=455.45.01 CUDA=11.1
- Define the path of image data in parser (line 194).
- Prepare the txt files for image index (line 61 - line 65). Format: filename + class index (start from 0)
- Python TripletHashing.py to run.
Distance: Euclidean without sqrt
Margin: 0.5 bits
Backbone: CNN (Resnet50) + tanh
Loss: Triplet Loss + Quantization Loss
Optimizer: Adam lr = 5e-5
mAP on CUB-200-2011 dataset.
| 16 bits | 32 bits | 48 bits | 64 bits |
|---|---|---|---|
| 74.54 | 77.03 | 76.92 | 77.18 |
Welcome to discuss any better implementation.