Question regarding recognition model sizes of new latin, korean and eslav PPOCRv5 packages

[timminator]

I've tried out the new PP-OCRv5 models released today. But I noticed one thing. They are pretty large. The size (80 - 110MB each) is pretty similar to the PP-OCRv5 server rec (80 MB) and they are 5-6 times larger than the equivalent PP-OCRv5 mobile rec (16 MB).
This made me even wonder if these are really mobile versions?
The dictionary for the PP-OCRv5 mobile rec is larger than the one from the latin, korean and eslav packages, so this can also not be the reason for this size difference.
So I would kindly ask if you could elaborate if they really are the mobile versions and why they are so much larger than the size expected?
Thanks!

[kimnamu]

I have same question

[tvone]

I've tried out the new PP-OCRv5 models released today. But I noticed one thing. They are pretty large. The size (80 - 110MB each) is pretty similar to the PP-OCRv5 server rec (80 MB) and they are 5-6 times larger than the equivalent PP-OCRv5 mobile rec (16 MB). This made me even wonder if these are really mobile versions? The dictionary for the PP-OCRv5 mobile rec is larger than the one from the latin, korean and eslav packages, so this can also not be the reason for this size difference. So I would kindly ask if you could elaborate if they really are the mobile versions and why they are so much larger than the size expected? Thanks!

Can you share the download link?

[zhangyubo0722]

this is a bug in the PaddlePaddle framework. When exporting inference models with version 3.0, there was an abnormal increase in model size. This issue has now been fixed, and the model size is now only a little over 10MB.

[timminator]

Thank you for your answer. I kinda forgot about this discussion. A few days after opening this discussion I redownloaded the models again and their size was fixed exactly as you mentioned. I will close this discussion.