A replay memory maximizng the coverage of sample space by estimating the sample-space density,
The memory needs to be initialized with a functions that compute the distance between batches of samples. Multiple distances can be combined, e.g. a distance between predictions and a distance between (groud truth) outputs, as used in the paper.
You can insert samples with the .insert method and get the current samples with the .get method.
For examples on how to instantiate the Memory, check out the the memory.py files in the implementation directories.
For more details on all parameters, check the bases.py and models.py.