About stable block

[daihu-ye]

In your paper, your said you apply a small set of scaling weights in stable block .But i can't find the corresponding code.Can you tell me where to find it?
Also,i read the code for optimizer in base_trainer.py.
In function set_optimizer,the parameters for b2_model is learnable if the 2nd branch is not fixed.But the parameters for b1_model, the FC weights for old classes is freezed and the others is all put into the optimizer.Why? What about the scaling weights? If we optimize the parameters for b1_model just as b2_model,how can it be called stable block?(Although the FC weights for old classes is freezed)

[yaoyao-liu]

Hi, thanks for your interest in our work.

In your paper, your said you apply a small set of scaling weights in stable block .But i can't find the corresponding code.Can you tell me where to find it?

The code for the scaling weights is here:
class-incremental-learning/adaptive-aggregation-networks/models/modified_resnetmtl_cifar.py

Also,i read the code for optimizer in base_trainer.py.
In function set_optimizer,the parameters for b2_model is learnable if the 2nd branch is not fixed.But the parameters for b1_model, the FC weights for old classes is freezed and the others is all put into the optimizer.Why? What about the scaling weights? If we optimize the parameters for b1_model just as b2_model,how can it be called stable block?(Although the FC weights for old classes is freezer)

If the scaling weights are applied, b1_model is replaced by a network with the scaling weights using the following code:

class-incremental-learning/adaptive-aggregation-networks/trainer/base_trainer.py

Lines 313 to 316 in 834de46

	if self.args.branch_1 == 'ss':
	b1_model = self.network_mtl(num_classes=self.args.nb_cl_fg)
	else:
	b1_model = self.network(num_classes=self.args.nb_cl_fg)

[daihu-ye]

Thanks,what about the fc weight in stabel block?In process_fp.py,the final output is calulated by b1_model.fc.How did you
optimize the fc weight?Did you freeze the weight for old classses and only optimize the weight for new classses?

[yaoyao-liu]

For the FC classifier, we directly follow LUCIR.