Fine tuning Training related questions from forum

[Shreeshrii]

https://groups.google.com/d/msg/tesseract-ocr/be4-rjvY2tQ/1bvuGMF5BwAJ by @AyushP123

Here is the link of the images for which no lsmtf files were generated -> https://drive.google.com/drive/folders/1VDBPB_k-oOXbWUI3zIlB3ljuyIlOkoMK?usp=sharing. Here is the Makefile that I used for generating lstmf files ->https://drive.google.com/open?id=15vvRMM03AOqoHKecEIx8NRTeU0y_kREy. I used Lorenzo's suggestion to create another target "train-lists" to avoid creating the training and the eval list again and again. Tesseract Version: 4.1.0 I am using https://github.com/tesseract-ocr/tesstrain/blob/master/generate_line_box.py to generate .box files. My images are in .tif format. I am saving my images using OpenCV imwrite. I have a few questions: In the link provided by Shree ->https://github.com/tesseract-ocr/tesstrain/wiki/GT4HistOCR#tesseract-fails-to-create-lstm-files. It says that .lstmf are not generated for some images if you use the default list.train settings. Using PSM=13 helps build those lstmf files, whereas using PSM= 6 or 7 ignores them. Any clues as to why that is the case??, Tesseract does give me output text for the images for PSM values 6,7 and 13. If I use PSM 13 for generating the lstmf files used for training, will it be okay to use PSM values 6 and 7 while testing. How can I check the contents of lstmf files to see if they contain the ground truth text info and the image data correctly?? Side Questions: lstmtraining saves the checkpoints in the following format: loss_iteration. It saves the checkpoints for a few iterations with the best loss ( apart from eng_checkpoint which contains the metadata I guess ). Is the loss calculated on the traininging data or the evaluation data??. Is there a way to save all checkpoints Side Questions: Does lstmeval use the psm value with which the lsmtf file was generated for evaluation??. I know its a lot of questions and doubts. I thank you for your time in helping me out.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[wrznr]

@AyushP123 I will try to answer some of your questions:

If I use PSM 13 for generating the lstmf files used for training, will it be okay to use PSM values 6 and 7 while testing.

From my experience, yes! PSM is a preprocessing parameter. Given tesseract --help-extra,

Page segmentation modes:
  0    Orientation and script detection (OSD) only.
  1    Automatic page segmentation with OSD.
  2    Automatic page segmentation, but no OSD, or OCR. (not implemented)
  3    Fully automatic page segmentation, but no OSD. (Default)
  4    Assume a single column of text of variable sizes.
  5    Assume a single uniform block of vertically aligned text.
  6    Assume a single uniform block of text.
  7    Treat the image as a single text line.
  8    Treat the image as a single word.
  9    Treat the image as a single word in a circle.
 10    Treat the image as a single character.
 11    Sparse text. Find as much text as possible in no particular order.
 12    Sparse text with OSD.
 13    Raw line. Treat the image as a single text line,
       bypassing hacks that are Tesseract-specific.

it does not make much sense to use PSM=6 for training since training is done on line level. The quality of your input images is really low. The problems with PSM=7 might stem from failures during the Tesseract-internal preprocessing. PSM=13 completely bypasses this processing step, effectively feeding the raw images to the LSTM file generation.

How can I check the contents of lstmf files to see if they contain the ground truth text info and the image data correctly??

Not possible as far as I know. @stweil?

Is the loss calculated on the traininging data or the evaluation data??

I think the iteration-wise loss is calculated on the training data. @bertsky Can you confirm?

Is there a way to save all checkpoints?

From https://github.com/tesseract-ocr/tesseract/wiki/TrainingTesseract-4.00#iterations-and-checkpoints and lstmtraining --help, I do not see a parameter to adjust the “checkpoint output rate”.

Does lstmeval use the psm value with which the lsmtf file was generated for evaluation??

I do not know. Maybe @egorpugin or @stweil do?

[stweil]

How can I check the contents of lstmf files to see if they contain the ground truth text info and the image data correctly??

It is currently not supported, but there exists a related feature request for Tesseract: https://github.com/tesseract-ocr/tesseract/issues/2669.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.