drsad / implementation-of-learning-without-forgetting-for-multi-class Goto Github PK

Based on the original LwF paper, kd loss should be used. Also, there should be a temperature factor to control the loss.
However, I can't find the temperature variable from your program. May you explain how you implement the knowledge distillation loss?

I need to ask if i can use new data set every task..

like:
training for task 1 CIFAR10 (10 classes) : total 10 classes
training for task 2 MNIST (10 classes) : total 10+10=20 classes
training for task 3 SVHN (10 classes) : total 20+10=30 classes

Thanks in advance.

I'm have following error in execution

AttributeError Traceback (most recent call last)

in ()
14
15 for i in range(10):
---> 16 model.beforeTrain()
17 accuracy=model.train()
18 model.afterTrain(accuracy)

/content/LwF.py in beforeTrain(self)
52 def beforeTrain(self):
53 self.model.eval()
---> 54 classes=[self.numclass-self.task_size,self.numclass]
55 self.train_loader,self.test_loader=self._get_train_and_test_dataloader(classes)
56 if self.numclass>self.task_size:

AttributeError: 'LwFmodel' object has no attribute 'task_size'