Hi, what is the command that you have been using to create training model? I created a model based on fine tuning, not on +- model, but actually I didn't chick its accuracy, theoretically after fine tuning the model it should give better accuracy, because fine tuning adds above the "best" model without changing anything, but in the real world I didn't yet check the accuracy.
anyways as for the second this which is chi_sim_vert, I faced that when I made trained model for Japanese, actually anything with vert its a sublanguage for Chinese and Japanese, "I think its when writing in vertical mode" when making tuning against LSTMFs you will notice errors regarding "vert", if you want to solve this problem, in the tesseract/training there is a file called, language-specific.sh search for the vert and delete it, it wont affect the traineddata refer to this link <https://groups.google.com/forum/#!topic/tesseract-ocr/zVjfHEyqbgg> for more about the "vert" issue On Thursday, October 26, 2017 at 4:35:26 PM UTC+3, [email protected] wrote: > > > 在 2017年10月24日星期二 UTC+8下午10:52:02，Ibr写道： >> >> Hi, >> I have the latest version of Tesseract and leptonica 1.74.4, ran the >> command >> training/lstmtraining --model_output >> /home/ibr/latest_leptonica_4/lstmf_old_jpn/jpn \ >> --continue_from >> /home/ibr/latest_leptonica_4/jpn_tune/extracted/jpn.lstm \ >> --traineddata >> /home/ibr/latest_leptonica_4/lstmf_jpn_lep4/jpn/jpn.traineddata \ >> --old_traineddata >> /home/ibr/latest_leptonica_4/jpn_tune/original_traineddata/jpn.traineddata \ >> --train_listfile >> /home/ibr/latest_leptonica_4/jpn_tune/jpn.training_files.txt \ >> --max_iterations 18000 >> >> then the command : >> >> training/lstmtraining --stop_training \ >> --continue_from /path/to/fine_tune/results/lang_checkpoint \ >> --traineddata /path/to/starter_traineddata/lang.traineddata \ >> --model_output /paht/to/new/tuned_lang.traineddata/lang.traineddata >> to create the traineddata file, yet I found that the accuracy of the >> official "best_traineddata" is better that what I got, I saw on this ocr >> group and some comments of GitHub that too many iterations don't give you >> best results, so I was wondering what is the optimal iterations to get the >> best results? >> >> the first command above is from, Fine Tune for +_ few characters, what >> are the commands for "Fine Tuning Impact"? I tried the command : >> >> training/lstmtraining --model_output ~/tesstutorial/impact_from_small/impact >> \ >> --continue_from ~/tesstutorial/engoutput/base_checkpoint \ >> --traineddata ~/tesstutorial/engtrain/eng/eng.traineddata \ >> --train_listfile ~/tesstutorial/engeval/eng.training_files.txt \ >> --max_iterations 1200 >> >> but what is the next commands? >> >> Thanks >> >> >> I have the same problem，after each training model, the recognition >> accuracy is not better than "best".training_text come form github's >> langdata ,append my train_text,language is chi_sim. >> > and I dot kown the difference between "chi_sim" "and chi_sim_vert" > -- You received this message because you are subscribed to the Google Groups "tesseract-ocr" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at https://groups.google.com/group/tesseract-ocr. To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/ea6c473d-2b7d-48bc-a33f-fc0115ee2bd8%40googlegroups.com. For more options, visit https://groups.google.com/d/optout.
