Hi,
I am doing the same thing to reproduce the work in "Towards Open Set Deep Networks". The code under this repo is a very good reference to me.
I trained my own model using ResNet, then used the openmax to analysis the output scores of resnet. However, no matter which category the input test data is, the openmax function only returns 'unknown' category.
After analysing the code in 'utils/openmax.py', I found that my 'channel_unknown = np.exp(np.sum(su))' is too big, resulting in also big 'prob_known'. Is there any advices ?
Besides, I suppose the w_score is the 'CDF' of weibull distribution, corresponding to the exp calculation of omega in the reference paper Algorithm 2 (line 3) , is this true? When it is running, the w_score variable in the openmax function equals to 1 for the uncorrect categories, is this correct ?
Hi, I am doing the same thing to reproduce the work in "Towards Open Set Deep Networks". The code under this repo is a very good reference to me. I trained my own model using ResNet, then used the openmax to analysis the output scores of resnet. However, no matter which category the input test data is, the openmax function only returns 'unknown' category. After analysing the code in 'utils/openmax.py', I found that my 'channel_unknown = np.exp(np.sum(su))' is too big, resulting in also big 'prob_known'. Is there any advices ?
Besides, I suppose the w_score is the 'CDF' of weibull distribution, corresponding to the exp calculation of omega in the reference paper Algorithm 2 (line 3) , is this true? When it is running, the w_score variable in the openmax function equals to 1 for the uncorrect categories, is this correct ?
Thank you very much!