Problems with low accuracy
Opened this issue · 5 comments
I took your code after debugging and it can run smoothly, but the accuracy rate of the run is very low, only 20 to 30%. I tried many batchsizes, and the result is still the same after many epochs. It can't reach you said the expected effect. When I remove the lda part and use CCE to calculate loss and accuracy, the effect is good and it can reach more than 80. This problem has troubled me for a long time. I don’t know if you have updated your code, or similar situations have occurred. ? Looking forward to your reply.
Hi,
Do you mind to tell me which dataset you are using? DeepLDA is weak when task is hard. Also the number of classes cannot be high, e.g. 20 or higher.
Thank you for your reply, I have carried out debugging again, the recurrence is successful, thank you very much~
But I found a question, why did the training have a high accuracy rate at the beginning, and the accuracy rate does not change much in the subsequent training process, have you encountered this situation?
Yeah the training with LDA is usually fast because it optimizes the objective directly. It however won't improve much once reached a good state.