why not reinstantiate the network model in Active learning cycles?I am wondering that the way of your writing will make the model aware of the test set in advance?

[Jiawen-huang]

Dear programmer, why not reinstantiate the network model in Active learning cycles?A lot of the active learning code I see reinstantiate the network model in Active learning cycles. For example，

Active learning cycles

    for cycle in range(CYCLES):
         # 2、Model（target model + loss predict module）
         resnet18    = resnet.ResNet18(num_classes=10).cuda()
         loss_module = lossnet.LossNet().cuda()
         models      = {'backbone': resnet18, 'module': loss_module}
         torch.backends.cudnn.benchmark = True
        #criterion、optimizer and scheduler (re)initialization
        criterion      = nn.CrossEntropyLoss(reduction='none')#交叉熵loss
        optim_backbone = optim.SGD(models['backbone'].parameters(), lr=LR,
                                momentum=MOMENTUM, weight_decay=WDECAY)
        optim_module   = optim.SGD(models['module'].parameters(), lr=LR, 。。。。。

In your code，You write it like this. I am wondering that the way of your writing will make the model aware of the test set in advance? https://github.com/Mephisto405/Learning-Loss-for-Active-Learning/blob/6c38f6fe8becf0739689ad0962c0379c0c9c2131/main.py#L226

[Mephisto405]

Hello,

Thanks for your comment. However, I cannot understand the point. Could you give me more details about 're-instantiation' that you mentioned? I mean, what do you mean by 're-instantiate the network model'?

Thanks

[lingjunz]

hi, I have a similar question.

In the source code, I find you only load model structure once. In each active learning cycle, the start point for retraining is the pre-trained model from the previous cycle.

Why don't load the model in the active learning cycle, i.e. re-instantiate the model before each active learning cycle begins?

Thanks in advance!

        torch.backends.cudnn.benchmark = True
        # Model ===========> put this three lines into active learning cycles
        resnet18    = resnet.ResNet18(num_classes=10).cuda()
        loss_module = lossnet.LossNet().cuda()
        models      = {'backbone': resnet18, 'module': loss_module}

        # Active learning cycles
        for cycle in range(CYCLES):
            # Loss, criterion and scheduler (re)initialization
            criterion      = nn.CrossEntropyLoss(reduction='none')
            optim_backbone = optim.SGD(models['backbone'].parameters(), lr=LR, 
                                    momentum=MOMENTUM, weight_decay=WDECAY)
            optim_module   = optim.SGD(models['module'].parameters(), lr=LR, 
                                    momentum=MOMENTUM, weight_decay=WDECAY)

Another question: When you compare LLAL with other methods, e.g. entropy, core-set, do you still use the model structure with the loss prediction module and train models with LossPredLoss?

[Mephisto405]

I am wondering that the way your writing will make the model aware of the test set in advance?

-> I think this is not the case. I don't feed the label of "unlabeled_set" into the model in advance. The model only knows the labels after it predicts the loss of each (sampled) unlabeled data points. In the "get_uncertainty" method, I turned off the training mode (model.eval()) as well. -> Also, in the real-world scenario, feeding unlabeled data into a model (to predict loss) is feasible.

Why not re-instantiate the network model? A lot of the active learning code I see re-instantiate the network model in Active learning cycles.

-> Hmm, in fact, I'm not so familiar with the whole literature of active learning. So if you find that the re-initialization if more effective in the sense of test accuracy, I think that is great. I just cannot see whether I need to re-initialize the model in the paper so that I do not re-initialize the model. We don't need to discard the prior knowledge learned from the previous learning cycle, do we? -> I think there is no problem even if we don't re-initialize the model in each cycle. The problem setting is that 'we can gather unlabeled dataset, but we cannot obtain labels of them.' So, I think feeding unlabeled data into a model does not make a problem if we don't put labels on them.

Another question: When you compare LLAL

Sorry, I'm not the author of this paper, and I cannot answer this question. I did not compare the LL4AL method with entropy, core-set, etc. in this repository.

Thanks for all comments, and if you have any comments and thoughts, I always welcome them. And please point it out if I'm wrong (I'm just a newbie master student :)). I want to share idea with many people.