Open XULU42 opened 4 years ago
how many classes is your data,if the number of classes is small,you can try to reduce the hyper parameter s, make it smaller
3ks for your reply. My classes num is 92. And I tried to set s=4, and the acc goes up to 83%,but there is still a great gap to 94.4% which a softmax loss can reach. Should I go deeper to debug the s parameter?
the purpose of the arc margin is to get discriminative features,it make sense that the performance of the classification is worse than softmax, and if the 83% is the train set accuracy , you can try on the test set
I am sorry to not mention the detail. The 83% acc is on the validation set, and argmax the logits matmuled between the weight and features. As far as i can see, this logit is exact the cosine similarity between the image feature and class center, as the weight interpreted as class center. Just now, I tried with s=16, and at about 1000 steps (with batch 64), I get validation set acc of 82%. I think there is going to be improvement at this time. May we should try something to lower the sensitivity of s. 3ks for your reply!
I meet the same problem
I meet the same problem:
how many classes is your data,if the number of classes is small,you can try to reduce the hyper parameter s, make it smaller
100 classes, s set to 30, but acc is 0.
i face the same question ,so how to fix this problem? please
s set to be small, like 10
trying different values manually can be pain staking , I decided to use optuna to where each trail will be for a max of 2 or few epochs, optuna will try to maximize the output accuracy by modifying s and m values with trial and error . I got good results with that
I want to use arcface loss to normal classification. But I found that the acc is always nearly 0. I checked the logits, it's nearly -1, and when give the label and add some m, the softmax loss is low. And this is learned by the network!!!!! In the extreme situation, if a network output logit all with -1, then the arcface loss is very low, so network can learn about nothing.