Avalanche - Usage and Limitations

Avalanche is a popular continual learning library for rapid prototyping and benchmarking algorithms. For that reason, we make some Avalanche updaters available in Renate. Avalanche updaters can be used in the same way you can use other Renate updaters but they have some limitations you should be aware of.

Usage

Using Avalanche updaters for training works the same as with Renate updaters which we explained earlier in How to Run a Training Job. You can select an Avalanche updater by passing the respective string to the updater argument of run_training_job(). The available Avalanche options are "Avalanche-ER", "Avalanche-EWC", "Avalanche-LwF", and "Avalanche-iCaRL". More details about these algorithms are given in Supported Algorithms. Further Avalanche updaters can be created if needed.

Limitations

Not all Renate features work with every Avalanche updater. In the following, we list the limitations you face when using an Avalanche updater. These limitations are subject to change.

PyTorch 1.13.1

The checkpointing functionality in Avalanche does not work with the latest version of PyTorch 1. Therefore, you will be required to use PyTorch 1.12.1 instead.

Warning

There is a known vulnerability with PyTorch <= 1.13.0.

No Scalable Buffers

Renate stores the memory buffer on disk. In contrast, Avalanche requires it to be in memory. Therefore, Avalanche updaters may not work if you intend to use very large buffer sizes.

No Multi-Fidelity HPO

Currently, we do not support multi-fidelity hyperparameter optimization with Avalanche updaters. For that reason, please do not use asha as a scheduler but use random or bo instead. For more details about HPO in Renate, please refer to the Renate HPO section.

No Early-Stopping

Currently, Avalanche updaters will not work with early stopping. Please keep early_stopping=True (default setting).

iCaRL Limitations

The implementation of iCaRL makes strong assumptions about the continual learning scenario. Only classification is supported and it assumes that data points of a particular class only occur in one update. If this is not the case, iCaRL will crash as soon as you attempt an update with a class seen before. Furthermore, it requires a specific model interface to account for its strategy. For that purpose, please create a model class which extends RenateBenchmarkingModule or copy the relevant parts over to your RenateModule.