Thanks for your reply! The #iterations I always used is 2000/3000/5000/10000. Is it reasonable?
I also try to extract dawg from HanS.traineddata and convert it to wordlist, and use it to generate base traineddata to fine-tune. I have confirmed that the new model's dawg->wordlist has the words that consist of my limited unicharset, but the problem still exists. To give more background, my scenario is to recognize plate number from vehicle license. The target image is something like "one Chinese character + several English letters or digits" (see one example image below). So the results are by design not some meaningful words. My training data has 5000 such plate numbers, one line for each as text. The reason why I want to retrain is the fact that the number of possible Chinese character at position 0 is limited to ~30. Am I doing anything wrong? [image: Inline image 1] On Mon, Jan 8, 2018 at 11:36 PM, ShreeDevi Kumar <[email protected]> wrote: > How many iterations did you use for training? > > You can unpack HanS.traineddata and then run dawg2word program to get the > wordlists used in it. Try using these for langdata in addition to your > training text. > > > > > > ShreeDevi > ____________________________________________________________ > भजन - कीर्तन - आरती @ http://bhajans.ramparivar.com > > On Mon, Jan 8, 2018 at 6:30 PM, Yang Yu <[email protected]> wrote: > >> Hi, >> >> These days I was working on fine-tuning a Chinese tesseract model based >> on 4.0 LSTM, and it worked great when the unicharset is not changed. But I >> found a problem when I applied it to a different scenario. >> >> Basically in my new scenario, the target characters are very limited - I >> only need to recognize less than 100 Chinese characters instead of >> thousands. I find this >> <https://github.com/tesseract-ocr/tesseract/wiki/TrainingTesseract-4.00#fine-tuning-for--a-few-characters> >> link about how to use a different set of unicharset to achieve this. >> Concretely, what I did is: >> 1. Prepare some text with only the characters I need >> 2. Run tesstrain.sh to generate images, and unicharset + traineddata >> + lstmf files (here I use chi_sim as langdata dir) >> 3. Run fine tuning: continued from HanS.lstm which is extracted from >> HanS.traineddata, use the generated chi_sim.traineddata as base >> traineddata, and use HanS.traineddata as old_traineddata >> >> The training process is smooth. But when I applied this new model to my >> evaluation set, I found that for some of my test cases, it worked better; >> but for the rest, the model just output empty string. As comparison, if I >> directly use a fine-tuned model based on HanS.traineddata without changing >> the unicharset (say, just adding some new lstmf files to fine tune), EVERY >> test cases can output something (no matter it is correct or not). >> >> Personally I don't think it is related to overfitting, because even a bad >> model should output something wrong. I'm not sure if it is related to >> chi_sim under langdata - it seems that langdata for 4.0 is not released >> yet, so chi_sim is the only thing I can use to fine-tune HanS.trainneddata >> model. >> >> Any help will be appreciated. >> >> -- >> 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/ms >> gid/tesseract-ocr/52093984-1415-4256-a2cd-268ed4141531%40googlegroups.com >> <https://groups.google.com/d/msgid/tesseract-ocr/52093984-1415-4256-a2cd-268ed4141531%40googlegroups.com?utm_medium=email&utm_source=footer> >> . >> For more options, visit https://groups.google.com/d/optout. >> > > -- > You received this message because you are subscribed to a topic in the > Google Groups "tesseract-ocr" group. > To unsubscribe from this topic, visit https://groups.google.com/d/ > topic/tesseract-ocr/CymhBpd24WU/unsubscribe. > To unsubscribe from this group and all its topics, 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/CAG2NduXR3TOTOzkFFi0jt1DLiwh- > nvHDw5ftYkm2L_E_rkJSzA%40mail.gmail.com > <https://groups.google.com/d/msgid/tesseract-ocr/CAG2NduXR3TOTOzkFFi0jt1DLiwh-nvHDw5ftYkm2L_E_rkJSzA%40mail.gmail.com?utm_medium=email&utm_source=footer> > . > > For more options, visit https://groups.google.com/d/optout. > -- 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/CADEYF-cJkD--Ho3NVwnbrMJu%2BZ2RBCXOg5Ki67qQJEwUhi4%3D7w%40mail.gmail.com. For more options, visit https://groups.google.com/d/optout.
