New Continual Learning algorithms can be added into CLiMB by adding Algorithm-specific modules in src/cl_algorithms.
Algorithm specific initializations are made in src/train/train_upstream_continual_learning.
However, certain algorithms may also need modifications to the model classes and/or TaskTrainer class.
For instance, in src/cl_algorithms/experience_replay.py, we defined a ExperienceReplayMemory class.
An empty ReplayMemory is instantiated in src/train/train_upstream_continual_learning.
When training on a new CL task, the replay memory is passed into the TaskTrainer.train() to do a periodic replay step.
Finally, after learning each task, the replay memory is updated with a memory buffer for the new task.
On the model side, algorithms like freeze-bottom-k and freeze-encoder can be implemented by adding algorithm-specific methods directly to the model.