Implement Your Backbone Network
Backbone networks are implemented as classes based on the base classes provided by this package in clarena/backbones/base.py. The base classes are implemented based on PyTorch Module with added features for continual learning:
CLBackbone: the general base class for all kinds of backbone networks, which incorporates mechanisms for managing continual learning tasks. Note that many CL algorithms keep a fixed backbone network, so the network might be exactly the same from task to task, but others might have different networks for different tasks.
Required Arguments
As outlined in the arguments of base classes, only one argument is required for CLBackbone:
output_dim: The output dimension which connects to CL output heads. This arguments decides the input dimension of the output heads in the source codes.
Implementing Instructions
You can implement your backbone just like a PyTorch module. The essential thing you need to do is to inherit the base class CLBackbone and implement the forward() method. Note that some CL algorithms need different backbones for different tasks, where inCLBackbone you can use the currentself.task_id as the task index to distinguish different tasks.
Please find the API documentation for detailed information about implementation, and remember you can always take implemented CL datasets in package source codes clarena/backbones/ as examples, like MLP!