Closed NeighborhoodCoding closed 4 years ago
I succeeded train a English character recognizer from scratch without pre-trained weights!!! It is possible just train a large epoch and large step per epoch.
But when my len(alphabet) >= 800 (this is a only problem for Korean OCR, for e.g, alphabet = '가갸거겨...') The predicted is tend to ''? here is my image.... https://i.imgur.com/jWbxytB.png
should I try this? https://github.com/faustomorales/keras-ocr/issues/88
(If you are ok, I will make my COLAB code and share to open... I think it is some of korean or chinese alphabet issue that need more accurate loss function?)
Please help.... I really want to predict korean characters.... I can not 100% understand your code(I'm begineer)..... please help T.T...
Hi, It is being solved when I train large epoch.... thank you!
Hi
I'm just using your https://keras-ocr.readthedocs.io/en/latest/examples/end_to_end_training.html
Example of end_to_end_training
And your provided colab example of https://colab.research.google.com/drive/19dGKong-LraUG3wYlJuPCquemJ13NN8R
There is some issue of(some mistypo(?))
should be
should be
thanks.