#WORK IN PROGRESS PyTorch Implementation of Supervised and Deep Q-Learning EWC(Elastic Weight Consolidation), introduced in "Overcoming Catastrophic Forgetting in Neural Networks"
There's a bug on line 81 of model.py
Should be:
self.F_accum[v] = torch.add(torch.Tensor((self.F_accum[v])), torch.pow(loglikelihood_grads[v], 2).data)
(as far as I could see for the latest versions of torch, torch.add needs an output). The way code is written self.F_accum never gets updated and the Fisher matrix is just 0.