dimitymiller / cac-openset Goto Github PK

torch.Tensor([[i for i in range(self.cls_num) if gt[x] != i] for x in range(len(distances))])

[[0, 1], [1], [0, 1], [1], [0, 1], [1], [0, 1], [1], [0, 1], [1], [0, 1], [1], [0, 1], [1], [0, 1], [1], [0, 1], [1]]

cacloss输出是这也的list，我知道这个报错是list里不可以[0,1]核[1]同时存在

请问这行的正确输出list应该是什么样子的那？

作者你好，你这边提供的TinyImageNet模型，闭集和开集上均无法达到论文性能

As you show Comparison with Existing Distance Losses in Table4, the result is different of CAC in table1
What is the backbone you use here for this Ablation Study

Thank you for the work!

I found the performace of "closedSet" is much better than the "softmax" in paper. These two have any difference?

Thank you for the contribution.

There is an issue. Where is the threshold theta to define the unknown classes? I can't find it.

Hi,

Thank you for sharing your code. I tried to do an experiment with 80known-20unknown split on cifar100. I modified the dataloaders and config file respectively. The train_cacOpenset.py script trains the network. But when I run the evaluation script to get AUC , I get a bunch of 'nan's for anchor_means. Do you have any idea what might be the reason?

Hi,

thank you for providing the code of your paper!

I am running into a problem with the evaluation of MNIST with the CAC open-set model (script eval_cacOpenset.py). I am using the provided weights and trials 0-2 work fine, but trials 3 and 4 generate NaNs during the anchor mean computation. The problem here is that for some classes there are no correct predictions and therefore no anchor can be computed for that class. I suspect that there might be something wrong with the uploaded weights or how do you deal with this?

Thank you for your time and best regards!

I want to know which part of the code is represented by equation 8 and equation 9 in the test part of the pape。thank you！

The dropout paramter of closedSetClassifier is not passed to the BaseEncoder upon construction, resulting in the default dropout rate of 0.3 being used instead of 0.2 from the config.

Hi,

The paper said The base network architecture f was consistent with the architecture established by [13]. It consists of 9 convolutional layers with batch norm after every layer and dropout after every 3 layers, followed by a fully-connected layer. in the Supplementary Martials. And I found there exists some differences: the position of dropout.
Comparing with origin code, your code put dropout at last of each encoder, instead of beginning. I am not familiar with network architecture, so is this will cause any performance differences?

Thx for your awesome repo，
Could u pls show more details about how to train model on my own image dataset。