Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Add new option -l for combine_tessdata to list LSTM information #3237

Merged
merged 4 commits into from
Jan 15, 2021
Merged

Add new option -l for combine_tessdata to list LSTM information #3237

merged 4 commits into from
Jan 15, 2021

Conversation

stweil
Copy link
Contributor

@stweil stweil commented Jan 14, 2021

No description provided.

@stweil
Copy link
Contributor Author

stweil commented Jan 14, 2021 •
edited
Loading

This is a first step to show the same information as Ray had produced in #1404 (comment). Ray's information looked like this:

LSTM training info:Network str:[1,48,0,1Ct3,3,16Mp3,3Lfys64Lfx96Lrx96Lfx384O1c1], flags=41, iteration=2736900, sample_iteration=2737076, null_char=2, learning_rate=0.001, momentum=0.5, adam_beta=0.999

LSTMRecognizer provides training_flags_, training_iteration_, sample_iteration_, null_char_, learning_rate_, momentum_ and adam_beta_, so the remaining information can easily be added.

2021-01-15: Meanwhile the missing information is implemented, too.

@stweil stweil changed the title Add new option -l for combine_tessdata to list the network string Add new option -l for combine_tessdata to list LSTM information Jan 14, 2021
@Shreeshrii
Copy link
Collaborator

Thank you, @stweil. This is very useful.

(base) ubuntu@tesseract-ocr-1:~/tesseract$ combine_tessdata -l ~/tessdata_best/eng.traineddata
LSTM: network=[1,36,0,1Ct3,3,16Mp3,3Lfys64Lfx96Lrx96Lfx512O1c1], int_mode=0, recoding=1, iteration=814100, sample_iteration=814136, null_char=110, learning_rate=0.001, momentum=0.5, adam_beta=0.999

(base) ubuntu@tesseract-ocr-1:~/tesseract$ combine_tessdata -l ~/tessdata/eng.traineddata
LSTM: network=[1,36,0,1Ct3,3,16Mp3,3Lfys64Lfx96Lrx96Lfx512O1c1], int_mode=1, recoding=1, iteration=814100, sample_iteration=814136, null_char=110, learning_rate=0.001, momentum=0.5, adam_beta=0.999

(base) ubuntu@tesseract-ocr-1:~/tesseract$ combine_tessdata -l ~/tessdata_fast/eng.traineddata
LSTM: network=[1,36,0,1Ct3,3,16Mp3,3Lfys48Lfx96Lrx96Lfx192O1c1], int_mode=1, recoding=1, iteration=6352400, sample_iteration=6352704, null_char=110, learning_rate=0.001, momentum=0.5, adam_beta=0.999

Now the int_mode can also be queried for START_MODEL in lstmtraining to give a user friendly error message/

@Shreeshrii
Copy link
Collaborator 

(base) ubuntu@tesseract-ocr-1:~/tesseract$ combine_tessdata -l ~/tessdata_best/script/Devanagari.traineddata
LSTM: network=[1,48,0,1Ct3,3,16Mp3,3Lfys64Lfx64Lrx64Lfx512O1c1], int_mode=0, recoding=1, iteration=2475100, sample_iteration=2475272, null_char=2, learning_rate=0.001, momentum=0.5, adam_beta=0.999

(base) ubuntu@tesseract-ocr-1:~/tesseract$ combine_tessdata -l ~/tessdata_best/San20201216.traineddata
LSTM: network=[1,36,0,1 Ct3,3,16 Mp3,3 Lfys48 Lfx96 Lrx96 Lfx192 O1c285], int_mode=0, recoding=1, iteration=1579800, sample_iteration=1584900, null_char=285, learning_rate=0.002, momentum=0.5, adam_beta=0.999

(base) ubuntu@tesseract-ocr-1:~/tesseract$ combine_tessdata -l ~/tessdata_best/Sanskrit20201231.traineddata
LSTM: network=[1,48,0,1Ct3,3,16Mp3,3Lfys64Lfx64Lrx64Lfx512O1c1], int_mode=0, recoding=1, iteration=3466500, sample_iteration=3466581, null_char=2, learning_rate=0.001, momentum=0.5, adam_beta=0.999

(base) ubuntu@tesseract-ocr-1:~/tesseract$ combine_tessdata -l /home/ubuntu/tesstrain-San/data/SanLayer/tessdata_best/SanLayer_0.686_354028_1278900.traineddata
LSTM: network=[1,48,0,1Ct3,3,16Mp3,3Lfys64Lfx64Lrx64Lfx512O1c1][Lfx192O1c1], int_mode=0, recoding=1, iteration=1278900, sample_iteration=1278900, null_char=2, learning_rate=0.0001, momentum=0.5, adam_beta=0.999

@stweil
Copy link
Contributor Author

stweil commented Jan 15, 2021

Why has San20201216.traineddata a network string with blanks? And SanLayer_0.686_354028_1278900.traineddata also has an unusual network string. How did you get those?

@stweil
Copy link
Contributor Author

stweil commented Jan 15, 2021

Should I merge this pull request, or should I wait for more feedback?

@stweil
Copy link
Contributor Author

stweil commented Jan 15, 2021 •
edited
Loading

Now the int_mode can also be queried for START_MODEL in lstmtraining to give a user friendly error message/

That should be implemented in lstmtraining (which can also query int_mode). It might also be possible to convert fast models back to "best" models on the fly, so I could imagine an lstmtraining which can use a fast start model.

@Shreeshrii
Copy link
Collaborator

Should I merge this pull request, or should I wait for more feedback?

Please merge. Thank you.

@stweil stweil merged commit c7baf8f into tesseract-ocr:master Jan 15, 2021
@stweil stweil deleted the network-string branch January 15, 2021 17:49
@Shreeshrii
Copy link
Collaborator

Shreeshrii commented Jan 15, 2021 •
edited
Loading

Why has San20201216.traineddata a network string with blanks? And SanLayer_0.686_354028_1278900.traineddata also has an unusual network string. How did you get those?

SanLayer_0.686_354028_1278900.traineddata
LSTM: network=[1,48,0,1Ct3,3,16Mp3,3Lfys64Lfx64Lrx64Lfx512O1c1][Lfx192O1c1],

I used --append_index 5 --net_spec '[Lfx192O1c1]' - This is the replace top layer training described by Ray in the tutorials.

Why has San20201216.traineddata a network string with blanks?

While comparing the different layer values, I had added spaces to the network spec for readability and didn't change before starting training.

NET_SPEC := [1,36,0,1 Ct3,3,16 Mp3,3 Lfys48 Lfx96 Lrx96 Lfx192 O1c\#\#\#]

@amitdo amitdo mentioned this pull request Apr 8, 2021
Closed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
enhancement
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
3 participants
@stweil @Shreeshrii @amitdo