val-acc always at~10% while the training loss can be reduced and the training accuracy can be increased

[Kyrie10favor]

Hello, I used my own dataset, which has 10 categories.at first ,I finished the Shrinking in SPOS folder ,and then I entered ProxyLess folder in order to Search over the Shrunk Search Space .However I found the problem:val-Top1acc always about~10%,val-Top5acc always about~50% while the training loss can be reduced and the training accuracy can be increased ，So It seems to be the result of a random prediction on the validation set。

• How can I solve this problem?thank u.