This code is a fine-tuning for the following thesis code, under incremental setting

Superclass Learning with Representation Enhancement

This is the official repo of the CVPR 2023 paper Superclass Learning with Representation Enhancement and its extension.

The conference paper: https://openaccess.thecvf.com/content/CVPR2023/papers/Kang_Superclass_Learning_With_Representation_Enhancement_CVPR_2023_paper.pdf

We provide a training example of SCLRE. You can generate the reorganized CIFAR100-4 dataset and train it. After training, you can check the result with the help of tensorboard. Some main package versions used in the program:

Step0 Dependencies

python: 3.8.12
pytorch: 1.11.0
torchvision: 0.12.0
CUDA Version: 11.4
tensorboard-logger: 0.1.0
tensorboard: 2.7.0
tqdm: 4.62.3

Step1 Generate Superclass Dataset

In filefolder ptgenerate/, we provide a mapping list CIFAR100-4.csv to map common classes to superclasses according to the rules in the supplementary metarial. You can run the following command:

python cifar100.py

to get two reorganized data tensors train.pt and test.pt generated. They are placed at ./data/re_cifar100/4_categories/.

Step2 Train with SCLRE

After generation, in the main directory, run

python main.py

to train with SCLRE. The settings of GPUs are in main.py, and you can change it based on your hardware. The training process will be shown in the terminal with the help of tqdm.

Step3 Check the Result

After training, the result will be recorded with the help of tensorboard. By default, it will be restored in ./runs/, and you can run

tensorboard --logdir=runs/

to check the training result. We evaluate our model in accuracy, recall, precision and F1 score. We also record the loss in the training process.

Step4 Change the Hyperparameters

All the hyperparameters are listed in hyperparameter.py, and the meanings of them are also listed in the comments in this file. You can change the hyperparameters and re-train with Step2.

To change the architecture of the model, set attention=False to drop CIA, set contrastive=False to drop contrastive adjustment loss, set TARGET=False to drop targeted adjustment loss.

Note that when the backbone is set as a CNN smaller than ResNet50, some network params in model/TC2.py need to be reset because these backbones generate representations with smaller dimensions. So we recommend to set this hyperparameter only as ResNet50, ResNet101 and ResNet152.