Where to find the LSTM network architecture used in Tesseract?

[sujith vemisetty]

I have tried a lot to find the network architecture of LSTMs used in Tesseract 4.00Alpha, but I wasn't able to find any. I can only find how to train the new neural network implementation. I would like to understand the architecture first. Can anyone point me to any documentation which details the layers of LSTM network, if there is any available? Thanks in advance.

[Alexander Nadeau]

The specific network structure is particular to a given traineddata file. I have no idea how the specification gets turned into the entire network's architecture in tesseract 4, but you can get a particular file's specification with combine_tessdata:

$ ./combine_tessdata.exe -d tess4traineddata/eng.traineddata
Version string:4.00.00alpha:eng:synth20170629:[1,36,0,1Ct3,3,16Mp3,3Lfys64Lfx96Lrx96Lfx512O1c1]
17:lstm:size=11689099, offset=192
18:lstm-punc-dawg:size=4322, offset=11689291
19:lstm-word-dawg:size=3694794, offset=11693613
20:lstm-number-dawg:size=4738, offset=15388407
21:lstm-unicharset:size=6360, offset=15393145
22:lstm-recoder:size=1012, offset=15399505
23:version:size=80, offset=15400517

The specification there is "[1,36,0,1Ct3,3,16Mp3,3Lfys64Lfx96Lrx96Lfx512O1c1]". No idea what that means, but the wiki has more to say about it here: https://github.com/tesseract-ocr/tesseract/wiki/TrainingTesseract-4.00#training-just-a-few-layers

[ShreeDevi Kumar]

See

https://github.com/tesseract-ocr/tesseract/wiki/NeuralNetsInTesseract4.00

https://github.com/tesseract-ocr/tesseract/wiki/VGSLSpecs

https://github.com/tesseract-ocr/docs/tree/master/das_tutorial2016