The EasyFeature repo aims to use softmax-based metric learning methods for feature learning, including L-Softmax, ArcFace, and EucMargin, which is the proposed method. Experiments on the MNIST, CIFAR10, and CIFAR100 datasets show that EucMargin has good performance and generalization ability, particularly outstanding on the CIFAR100 dataset.
To install the required packages, run the following commands:
pip3 install torch torchvision
pip3 install requirements.txt
The following table shows the classification error rate achieved by each method on three datasets: MNIST, CIFAR-10, and CIFAR-100.
| Datasets | L-Softmax | ArcFace | EucMargin |
|---|---|---|---|
| MNIST | 0.31% | 0.31% | 0.28% |
| CIFAR-10 | 7.58% | 7.90% | 7.38% |
| CIFAR-100 | 29.53% | 30.23% | 28.42% |
To reproduce the results, follow the steps below:
- Download the dataset and split it into train/val/test/template:
python tools/split_datasets.py -dt CIFAR100 --seed 30673
- Train the model
# CIFAR-100 EucMargin
python tools/train.py --c configs/cifar100/cifar100_euc.yaml -o Global.seed=49504 Architecture.Head.margin_add=2.5
# CIFAR-100 ArcFace
python tools/train.py --c configs/cifar100/cifar100_arc.yaml
# CIFAR-100 L-Softmax
python tools/train.py --c configs/cifar100/cifar100_lsoftmax.yaml
Note that L-Softmax is hard to train on CIFAR-100, so you may need to try different hyperparameters to obtain better results.
This project is released under the Apache 2.0 license.