def __init__(self, n_classes, dictionary):
# Network architecture
super(iCarlNet, self).__init__()
self.n_classes = n_classes
self.bs = 10
self.dictionary = dictionary
#Main net
self.model = resnet32()
#previous model to compute the distillation loss
self.prev_model = copy.deepcopy(self.model)
#in feature extractor is removed the last fc layer of resnet32
self.feature_extractor = nn.Sequential(*(list(self.model.children())[:-1]))
self.cuda()
self.exemplars = []
self.compute_means = True
# Learning method
self.criterion_NLL = nn.NLLLoss() # distillation loss
self.criterion_KLDiv = nn.KLDivLoss() # distillation loss
self.criterion_MSE = nn.MSELoss() # calssification and distillation loss
self.criterion_CE = nn.CrossEntropyLoss() # classification loss
self.num_epochs = 70
self.ex_features = []
self.ex_labels = []
def __init__(self, n_classes, dictionary):
# Network architecture
super(iCarlNet, self).__init__()
self.n_classes = n_classes
self.bs = 10
self.dictionary = dictionary
#Main net
self.model = resnet32()
#previous model to compute the distillation loss
self.prev_model = copy.deepcopy(self.model)
#in feature extractor is removed the last fc layer of resnet32
self.feature_extractor = nn.Sequential(
*(list(self.model.children())[:-1]))
#bring the model to cuda enviroment
self.cuda()
self.exemplars = []
# Learning method
self.criterion = nn.BCEWithLogitsLoss()
self.num_epochs = 70
#parameter to update for the Knn
self.K_neighbors = 10
def __init__(self, n_classes, dictionary):
super(LwF, self).__init__()
self.model = resnet32()
self.prev_model = None
self.d = dictionary
self.num_epochs = 70
self.optimizer = optim.SGD(self.model.parameters(),
lr=2,
momentum=0.9,
weight_decay=1e-5)
self.scheduler = optim.lr_scheduler.MultiStepLR(self.optimizer,
[49, 63],
gamma=0.2)
self.criterion = nn.BCEWithLogitsLoss()
self.n_classes = n_classes
self.bs = n_classes