What dataset is used to train the pretrained model?

[xieenze]

Hello author:
I try to train ASTER on Synth800K and Synth90K, same setting as the TPAMI Papaer. However, I can not reimplement the accuracy in the paper and your pretrained model.
Can you tell me what dataset are you used to train the pretrained model you provided? Thanks very much!

[ayumiymk]

Both the pytorch port and the tensorflow port are trained with Synth90k and SynthText, just as the paper described. Have you tried the tensorflow version?

[xieenze]

No, I didn't run TensorFlow version. But this PyTorch I use Synth and Syn90K, I failed to run the accuracy you reported in readme.md.
My accuracy is average 1.5-2 lower than yours in all datasets.
Can you tell me some tricks or details maybe I ignored to train such high accuracy ASTER?
Thanks very much!

[ayumiymk]

Actually I do not use any tricks to achieve such performance. All results reported in the README.md are trained and tested with the given scripts.

[ayumiymk]

Hi @xieenze,

With the released model, can you obtain the reported results in the README.md.

[xieenze]

Yes, using your released model, I can achieve the same results in readme
Maybe I need to check sth. Because I use 4 gpus and modify the batch size and lr.
Thanks very much!

[ayumiymk]

Hi enze,

Have you reproduced the reported results? If yes, please let me know, I am curious about the effects from different hyper-parameters.

[WenjiaWang0312]

Hi, I used the workers of 100, and batch_size of 1024. I also modified the learning-rate by hand when the accuracy stuck at a point. I had not achieved the accuracies on those 7 test datasets as your report. Approximately 1.0% to 2.0% lower than your results.

[ayumiymk]

emmmm... why not try the given settings? I can not guarantee the same performance with different settings. @JasonBoy1

[WenjiaWang0312]

emmmm... why not try the given settings? I can not guarantee the same performance with different settings. @JasonBoy1

OK, I should try recently

[WenjiaWang0312]

emmmm... why not try the given settings? I can not guarantee the same performance with different settings. @JasonBoy1

thank you very much

[WenjiaWang0312]

Hi, dear author. I found that the default batch_size in the script/stn_att_rec.sh is different from that in config.py. Which did you used in your experiment?
And how many GPUs did you used?

[ayumiymk]

The arguments in config.py are just default values for definition. Actually, you should just run bash script/stn_att_rec.sh. All the settings are given in this script, including the number of GPUs (here 2 GPUs are used).

[minglii1998]

Hello there, I'd like to know if any one of you could achieve the same performance ~
I ran bash script/stn_att_rec.sh with two GPUs, but it seemed that my accuracy is about 1-2 lower than the ASTER (L2R) and much lower than ASTER.Pytorch, which should I expect?

[ayumiymk]

Hello there, I'd like to know if any one of you could achieve the same performance ~
I ran bash script/stn_att_rec.sh with two GPUs, but it seemed that my accuracy is about 1-2 lower than the ASTER (L2R) and much lower than ASTER.Pytorch, which should I expect?

Maybe you can try our tensorflow implement. If you can not obtain the same results using both the tf and pytorch version. I suspect that there is something wrong with your training data.

[minglii1998]

Hello there, I'd like to know if any one of you could achieve the same performance ~
I ran bash script/stn_att_rec.sh with two GPUs, but it seemed that my accuracy is about 1-2 lower than the ASTER (L2R) and much lower than ASTER.Pytorch, which should I expect?

Maybe you can try our tensorflow implement. If you can not obtain the same results using both the tf and pytorch version. I suspect that there is something wrong with your training data.

Thanks for reply~
Still I can't achieve the performance. I wonder if there's any chance you could provide me with the datasets in lmdb format, I would be really grateful~
Sorry to bother!