Closed xieenze closed 4 years ago
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?
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!
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.
Hi @xieenze,
With the released model, can you obtain the reported results in the README.md.
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!
Hi enze,
Have you reproduced the reported results? If yes, please let me know, I am curious about the effects from different hyper-parameters.
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.
emmmm... why not try the given settings? I can not guarantee the same performance with different settings. @JasonBoy1
emmmm... why not try the given settings? I can not guarantee the same performance with different settings. @JasonBoy1
OK, I should try recently
emmmm... why not try the given settings? I can not guarantee the same performance with different settings. @JasonBoy1
thank you very much
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?
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).
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?
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.
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!
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!
How about the performance of the Tensorflow implement?
What's more, I am sorry. The size of the training data is about 50GB. I don't have such a big cloud drive. You should carefully check your own data.
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!How about the performance of the Tensorflow implement?
What's more, I am sorry. The size of the training data is about 50GB. I don't have such a big cloud drive. You should carefully check your own data.
Thanks anyway, I will try the tf implement~
@minglii1998 If you make any progress, please let me know. Thanks!
Hello there, finally I achieved the the performance as @ayumiymk mentioned, thus it's for sure that there's nothing wrong with this code.
The performance is almost the same, maybe 0.1-0.3 lower or higher like that.
The reason I can't get the same acc mostly because the datasets I used. The way I pre-process the ST dataset is not right. Handling the exception is of great importance while getting the lmdb format data.
Hello there, finally I achieved the the performance as @ayumiymk mentioned, thus it's for sure that there's nothing wrong with this code. The performance is almost the same, maybe 0.1-0.3 lower or higher like that. The reason I can't get the same acc mostly because the datasets I used. The way I pre-process the ST dataset is not right. Handling the exception is of great importance while getting the lmdb format data.
Dear MingLii: Could you tell me what's the right way to 'handling the exception' of the ST dataset? My implemented accuracy is about 2%-3% lower than the author's report.
I wonder how many images are cropped from ST800K? Thanks! @minglii1998
@ayumiymk Could you share the scripts that you used to create lmdb from Synth90k and SynthText datasets? That would be helpful, thank you.
Dear author: Could you share what's the right way to 'handling the exception' of the ST dataset? Like the @minglii1998 said.
@JasonBoy1 @Pay20Y @ofexe Apologies for my not reply, I was so busy that even if I received the email from git, I still couldn't show up. The author shared the datasets he used with me, and I will share them here with his permission. You can download from from here,code:wi05 Hope all of all of you can get a good result! @ayumiymk Hi bro, maybe it's about time to close this issue, thanks for your sharing!
@minglii1998 thanks a lot!
@minglii1998 I've already made my own datasets, but still thanks.
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!How about the performance of the Tensorflow implement? What's more, I am sorry. The size of the training data is about 50GB. I don't have such a big cloud drive. You should carefully check your own data.
Thanks anyway, I will try the tf implement~
hello, can you teach me how training from begin.
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!