summerlvsong/Aggregation-Cross-Entropy

Can’t reproduce your results in your cvpr paper

Opened this issue · 3 comments

yflv-yanxia commented

I reproduced crnn+ctc and test it on IIIT5K+SVT+IC03+IC13 test database, got WER 0.153, which is same as the reported results in paper.
I also reproduced crnn+ace loss, but only got WER 0.205 on the same test database, any advise?
My environment:
pytorch 1.2.0
batchsize 60
trained only on 8-million synthetic data released by Jaderberg
iterations 1000k
adadelta rho 0.9

MiTeng0215 commented

I reproduced crnn+ctc and test it on IIIT5K+SVT+IC03+IC13 test database, got WER 0.153, which is same as the reported results in paper.
I also reproduced crnn+ace loss, but only got WER 0.205 on the same test database, any advise?
My environment:
pytorch 1.2.0
batchsize 60
trained only on 8-million synthetic data released by Jaderberg
iterations 1000k
adadelta rho 0.9

I have the same results as you, and I want to know if you have any new progress.

bjlgcxc commented

it's hard for me to train crnn+ace loss to coverage, do you know why?

zhangfengyo commented

I reproduced crnn+ctc and test it on IIIT5K+SVT+IC03+IC13 test database, got WER 0.153, which is same as the reported results in paper.
I also reproduced crnn+ace loss, but only got WER 0.205 on the same test database, any advise?
My environment:
pytorch 1.2.0
batchsize 60
trained only on 8-million synthetic data released by Jaderberg
iterations 1000k
adadelta rho 0.9

maybe some augmentation are used without mention in the paper