Multi-class Smoothed Hinge Loss Function in Pre-training for Transfer Learning

Project page of the paper 'Multi-class Smoothed Hinge Loss Function in Pre-training for Transfer Learning,' ICIP 2025.

---

How to use

Dependencies

• pytorch >= 2.0

Preparing

To download the pretrained weights, run

pip install huggingface_hub
python download.py

Quick start

To use pre-trained resnet50 for transfer learning with cifar-100

python train_for_transfer.py -net resnet50 -pretrained resnet50_MCSH_m7.pth

To use pre-trained resnet50 for transfer learning with your own datasets

Please change line 142-156 to fit your own datsets.

python train_for_transfer.py -net resnet50 -pretrained resnet50_MCSH_m7.pth -num_classes {{your_dataset_class_num}} -dataset {{your_dataset_name}}

To use Multi-class Smoothed Hinge Loss Function in your code

change

loss = nn.CrossEntropyLoss()

in your code to

from MCSH_loss import MultiClassSmoothedHingeLoss
loss = MultiClassSmoothedHingeLoss(margin=YOUR_SETTING)