Welcome to the world of PyCIL, a powerful toolbox designed for class-incremental learning (CIL), developed in PyTorch! In this guide, we’ll walk you through the steps to effectively utilize this toolbox for your machine learning projects, ensuring you have a smooth experience while integrating new classes into your learning model.

What is Class-Incremental Learning?

Class-incremental learning is an essential machine learning paradigm that allows models to learn continuously from new classes of data without forgetting previous ones. Imagine a teacher continually adding new subjects to their curriculum. The teacher must remember the old subjects while integrating new information, just as a CIL model retains knowledge of previous classes while learning about new ones.

Getting Started with PyCIL

To start using PyCIL, follow these steps:

1. Clone the Repository

First, clone the PyCIL GitHub repository to your local machine:

git clone https://github.com/G-U-N/PyCIL.git
cd PyCIL

2. Install Dependencies

PyCIL requires several dependencies to function correctly. Here’s the list you’ll need:

• torch 1.81
• torchvision 0.6.0
• tqdm
• numpy
• scipy
• quadprog
• POT

3. Configure the Model

Before running an experiment, you need to configure your model. Follow these steps:

1. Modify the [MODEL NAME].json file for global settings.
2. Edit the hyperparameters in the corresponding [MODEL NAME].py file (e.g., models/icarl.py).

4. Run Your Experiment

To run your configured model, use the following command:

bash
python main.py --config=.exps/[MODEL NAME].json

Replace [MODEL NAME] with the model you wish to use, such as finetune, ewc, or lwf.

Editing Hyperparameters

In PyCIL, several hyperparameters can be customized within the JSON configuration file:

• memory-size: Defines the total number of exemplars in the incremental learning process.
• init-cls: Sets the number of classes for the first incremental stage.
• increment: Specifies how many classes to add in each stage.
• convnet-type: Determines the backbone network used for incrementing classes.
• seed: Establishes a random seed for shuffling class order.
• Additional parameters for model optimization can also be adjusted within the Python files.

Working with Datasets

PyCIL supports automatic downloading of CIFAR100 and provides pre-processing for other datasets like ImageNet100. You will need to specify the folder where your dataset is located in utils/data.py.

def download_data(self):
    assert 0, "You should specify the folder of your dataset"
    train_dir = [DATA-PATH]/train
    test_dir = [DATA-PATH]/val

Troubleshooting

While using PyCIL, you may encounter some issues. Here are some troubleshooting tips:

• Ensure all dependencies are correctly installed and are the appropriate versions.
• Check that the model name specified in your command matches the files in the repository.
• Verify the paths for your datasets are correctly configured in utils/data.py.
• If you receive errors related to memory size, try reducing the number of exemplars per class in your configuration.

For more insights, updates, or to collaborate on AI development projects, stay connected with fxis.ai.

Conclusion

At fxis.ai, we believe that such advancements are crucial for the future of AI, as they enable more comprehensive and effective solutions. Our team is continually exploring new methodologies to push the envelope in artificial intelligence, ensuring that our clients benefit from the latest technological innovations.

Get Involved!

Feel free to dive into the repository, experiment with the different models, and join the community of developers pushing forward the field of class-incremental learning!