Source code for

import copy

import torch
from import SupervisedPlugin
from torch.nn import BCELoss
import numpy as np

[docs]class ICaRLLossPlugin(SupervisedPlugin): """ ICaRLLossPlugin Similar to the Knowledge Distillation Loss. Works as follows: The target is constructed by taking the one-hot vector target for the current sample and assigning to the position corresponding to the past classes the output of the old model on the current sample. Doesn't work if classes observed in previous experiences might be observed again in future training experiences. """
[docs] def __init__(self): super().__init__() self.criterion = BCELoss() self.old_classes = [] self.old_model = None self.old_logits = None
def before_forward(self, strategy, **kwargs): if self.old_model is not None: with torch.no_grad(): self.old_logits = self.old_model(strategy.mb_x) def __call__(self, logits, targets): predictions = torch.sigmoid(logits) one_hot = torch.zeros( targets.shape[0], logits.shape[1], dtype=torch.float, device=logits.device, ) one_hot[range(len(targets)), targets.long()] = 1 if self.old_logits is not None: old_predictions = torch.sigmoid(self.old_logits) one_hot[:, self.old_classes] = old_predictions[:, self.old_classes] self.old_logits = None return self.criterion(predictions, one_hot) def after_training_exp(self, strategy, **kwargs): if self.old_model is None: old_model = copy.deepcopy(strategy.model) old_model.eval() self.old_model = self.old_model.load_state_dict(strategy.model.state_dict()) self.old_classes += np.unique( strategy.experience.dataset.targets ).tolist()
__all__ = ["ICaRLLossPlugin"]