Awesome
Setup
Install Package Dependencies
torch
torchvision
PyYAML
easydict
tqdm
scikit-learn
efficientnet_pytorch
pandas
opencv
Datasets
create a soft link to the dataset directory
CUB dataset
ln -s /your-path-to/CUB-dataset data/cub
Training (Backbone: ResNet50, Dataset: CUB)
python main.py --config config/sps-single-branch-cub.yml --gpu_ids 0 # Train the network that contains one mid-level branch
Best accuracy: 88.70%
Results of the last three epochs
| Epoch | H-level | M-level(SPS) | H-level+M-level(SPS) |
|---|---|---|---|
| 158 | 85.76% | 87.21% | 88.51% |
| 159 | 86.18% | 87.37% | 88.32% |
| 160 | 85.99% | 87.33% | 88.63% |
python main.py --config config/sps-two-branch-cub.yml --gpu_ids 0 # Train the network that contains two mid-level branches
Best accuracy: 88.82%
Results of the last three epochs
| Epoch | H-level | M-level(SPS)-0 | M-level(SPS)-1 | H-level+2xM-level(SPS) |
|---|---|---|---|---|
| 158 | 85.76% | 87.82% | 87.33% | 88.64% |
| 159 | 85.83% | 87.66% | 87.28% | 88.57% |
| 160 | 85.42% | 87.68% | 88.52% | 88.70% |
*** The full training log can be found in the folder logs/ ***