How to train LSTM/neural net Tesseract

Have questions about the training process? If you had some problems during the training process and you need help, use tesseract-ocr mailing-list to ask your question(s). PLEASE DO NOT report your problems and ask questions about training as issues!

Training with tesstrain.sh bash scripts is unsupported/abandoned for Tesseract 5. Please use python scripts from tesstrain repo for training.

Contents

Introduction

Tesseract 4.00 introduced a new neural network-based recognition engine that delivers significantly higher accuracy (on document images) than the previous versions, in return for a significant increase in required compute power. On complex languages however, it may actually be faster than base Tesseract.

Neural networks require significantly more training data and train a lot slower than base Tesseract. For Latin-based languages, the existing model data provided has been trained on about 400000 textlines spanning about 4500 fonts. For other scripts, not so many fonts are available, but they have still been trained on a similar number of textlines. Instead of taking a few minutes to a couple of hours to train, Tesseract 4.00 takes a few days to a couple of weeks. Even with all this new training data, you might find it inadequate for your particular problem, and therefore you are here wanting to retrain it.

There are multiple options for training:

While the above options may sound different, the training steps are actually almost identical, apart from the command line, so it is relatively easy to try it all ways, given the time or hardware to run them in parallel.

The old recognition engine is still present, and can also be trained, but is deprecated, and, unless good reasons materialize to keep it, may be deleted in a future release.

Before You Start

You don’t need any background in neural networks to train Tesseract, but it may help in understanding the difference between the training options. Please read the Implementation introduction before delving too deeply into the training process.

Important note: It’s important to note that, unless you’re using a very unusual font or a new language, retraining Tesseract is unlikely to help. Before you invest time and effort on training Tesseract, it is highly recommended to read the ImproveQuality page. Many times recognition can be improved just by preprocessing the input image.

Hardware-Software Requirements

At time of writing, training only works on Linux. (macOS almost works; it requires minor hacks to the shell scripts to account for the older version of bash it provides and differences in mktemp.) Windows is unknown, but would need msys or Cygwin.

As for running Tesseract, it is useful, but not essential to have a multi-core (4 is good) machine, with OpenMP and Intel Intrinsics support for SSE/AVX extensions. Basically it will still run on anything with enough memory, but the higher-end your processor is, the faster it will go. No GPU is needed. (No support.) Memory use can be controlled via the –max_image_MB command-line option, but you are likely to need at least 1GB of memory over and above what is taken by your OS.

Additional Libraries Required

Beginning with 3.03, additional libraries are required to build the training tools.

sudo apt-get install libicu-dev libpango1.0-dev libcairo2-dev

Building the Training Tools

Beginning with 3.03, if you’re compiling Tesseract from source you need to make and install the training tools with separate make commands.

For training, you will have to ensure all those optional dependencies are installed and that Tesseract’s build environment can locate them. Look for these lines in the output of ./configure:

checking for pkg-config... [some valid path]
checking for lept >= 1.74... yes
checking for libarchive... yes
checking for icu-uc >= 52.1... yes
checking for icu-i18n >= 52.1... yes
checking for pango >= 1.22.0... yes
checking for cairo... yes
[...]
Training tools can be built and installed with:

(The version numbers may change over time, of course. What we are looking for is “yes”, all of the optional dependencies are available.)

After configuring, you can attempt to build the training tools:

make
make training
sudo make training-install

It is also useful, but not required, to build ScrollView.jar:

make ScrollView.jar
export SCROLLVIEW_PATH=$PWD/java

Training Text Requirements

For Latin-based languages, the existing model data provided has been trained on about 400000 textlines spanning about 4500 fonts. For other scripts, not so many fonts are available, but they have still been trained on a similar number of textlines.

Note that it is beneficial to have more training text and make more pages though, as neural nets don’t generalize as well and need to train on something similar to what they will be running on. If the target domain is severely limited, then all the dire warnings about needing a lot of training data may not apply, but the network specification may need to be changed.

Overview of Training Process

The main steps in training are:

  1. Prepare training text.
  2. Render text to image + box file. (Or create hand-made box files for existing image data.)
  3. Make unicharset file. (Can be partially specified, i.e. created manually).
  4. Make a starter/proto traineddata from the unicharset and optional dictionary data.
  5. Run tesseract to process image + box file to make training data set (lstmf files).
  6. Run training on training data set.
  7. Combine data files.

The key differences from training base Tesseract (Legacy Tesseract 3.04) are:

The training cannot be quite as automated as the training for 3.04 for several reasons:

Understanding the Various Files Used During Training

As with base/legacy Tesseract, the completed LSTM model and everything else it needs is collected in the traineddata file. Unlike base/legacy Tesseract, a starter/proto traineddata file is given during training, and has to be setup in advance. It can contain:

Bold elements must be provided. Others are optional, but if any of the dawgs are provided, the punctuation dawg must also be provided.

A new tool combine_lang_model is provided to make a starter traineddata from a unicharset and optional wordlists and is required for training.

During training, the trainer writes checkpoint files, which is a standard behavior for neural network trainers. This allows training to be stopped and continued again later if desired. Any checkpoint can be converted to a full traineddata for recognition by using the --stop_training command-line flag.

The trainer also periodically writes checkpoint files at new bests achieved during training.

It is possible to modify the network and retrain just part of it, or fine tune for specific training data (even with a modified unicharset!) by telling the trainer to --continue_from either an existing checkpoint file, or from a naked LSTM model file that has been extracted from an existing traineddata file using combine_tessdata provided it has not been converted to integer.

If the unicharset is changed in the --traineddata flag, compared to the one that was used in the model provided via --continue_from, then the --old_traineddata flag must be provided with the corresponding traineddata file that holds the unicharset and recoder. This enables the trainer to compute the mapping between the character sets.

The training data is provided via .lstmf files, which are serialized DocumentData They contain an image and the corresponding UTF8 text transcription, and can be generated from tif/box file pairs using Tesseract in a similar manner to the way .tr files were created for the old engine.

LSTMTraining Command Line

The lstmtraining program is a multi-purpose tool for training neural networks. The following table describes its command-line options:

Flag Type Default Explanation
traineddata string none Path to the starter traineddata file that contains the unicharset, recoder and optional language model.
net_spec string none Specifies the topology of the network.
model_output string none Base path of output model files/checkpoints.
max_image_MB int 6000 Maximum amount of memory to use for caching images.
learning_rate double 10e-4 Initial learning rate for SGD algorithm.
sequential_training bool false Set to true for sequential training. Default is to process all training data in round-robin fashion.
net_mode int 192 Flags from NetworkFlagsin network.h. Possible values: 128 for Adam optimization instead of momentum; 64 to allow different layers to have their own learning rates, discovered automatically.
perfect_sample_delay int 0 When the network gets good, only backprop a perfect sample after this many imperfect samples have been seen since the last perfect sample was allowed through.
debug_interval int 0 If non-zero, show visual debugging every this many iterations.
weight_range double 0.1 Range of random values to initialize weights.
momentum double 0.5 Momentum for alpha smoothing gradients.
adam_beta double 0.999 Smoothing factor squared gradients in ADAM algorithm.
max_iterations int 0 Stop training after this many iterations.
target_error_rate double 0.01 Stop training if the mean percent error rate gets below this value.
continue_from string none Path to previous checkpoint from which to continue training or fine tune.
stop_training bool false Convert the training checkpoint in --continue_from to a recognition model.
convert_to_int bool false With stop_training, convert to 8-bit integer for greater speed, with slightly less accuracy.
append_index int -1 Cut the head off the network at the given index and append --net_spec network in place of the cut off part.
train_listfile string none Filename of a file listing training data files.
eval_listfile string none Filename of a file listing evaluation data files to be used in evaluating the model independently of the training data.

Most of the flags work with defaults, and several are only required for particular operations listed below, but first some detailed comments on the more complex flags:

Unicharset Compression-recoding

LSTMs are great at learning sequences, but slow down a lot when the number of states is too large. There are empirical results that suggest it is better to ask an LSTM to learn a long sequence than a short sequence of many classes, so for the complex scripts, (Han, Hangul, and the Indic scripts) it is better to recode each symbol as a short sequence of codes from a small number of classes than have a large set of classes.

The combine_lang_model command has this feature on by default. It encodes each Han character as a variable-length sequence of 1-5 codes, Hangul using the Jamo encoding as a sequence of 3 codes, and other scripts as a sequence of their unicode components. For the scripts that use a virama character to generate conjunct consonants, (All the Indic scripts plus Myanmar and Khmer) the function NormalizeCleanAndSegmentUTF8 pairs the virama with an appropriate neighbor to generate a more glyph-oriented encoding in the unicharset. To make full use of this improvement, the --pass_through_recoder flag should be set for combine_lang_model for these scripts.

Randomized Training Data and sequential_training

For Stochastic Gradient Descent to work properly, the training data is supposed to be randomly shuffled across all the sample files, so the trainer can read its way through each file in turn and go back to the first one when it reaches the end.

If using the rendering code, (via tesstrain.py) then it will shuffle the sample text lines within each file, but you will get a set of files, each containing training samples from a single font. To add a more even mix, the default is to process one sample from each file in turn aka ‘round robin’ style. If you have generated training data some other way, or it is all from the same style (a handwritten manuscript book for instance) then you can use the --sequential_training flag for lstmtraining. This is more memory efficient since it will load data from only two files at a time, and process them in sequence. (The second file is read-ahead so it is ready when needed.)

Model output

The trainer saves checkpoints periodically using --model_output as a basename. It is therefore possible to stop training at any point, and restart it, using the same command line, and it will continue. To force a restart, use a different --model_output or delete all the files.

Net Mode and Optimization

The 128 flag turns on Adam optimization, which seems to work a lot better than plain momentum.

The 64 flag enables automatic layer-specific learning rate. When progress stalls, the trainer investigates which layer(s) should have their learning rate reduced independently, and may lower one or more learning rates to continue learning.

The default value of net_mode of 192 enables both Adam and layer-specific learning rates.

Perfect Sample Delay

Training on “easy” samples isn’t necessarily a good idea, as it is a waste of time, but the network shouldn’t be allowed to forget how to handle them, so it is possible to discard some easy samples if they are coming up too often. The --perfect_sample_delay argument discards perfect samples if there haven’t been that many imperfect ones seen since the last perfect sample.

The current default value of zero uses all samples. In practice the value doesn’t seem to have a huge effect, and if training is allowed to run long enough, zero produces the best results.

Debug Interval and Visual Debugging

With zero (default) --debug_interval, the trainer outputs a progress report every 100 iterations, similar to the following example.

At iteration 717/10500/10500, Mean rms=0.113000%, delta=0.009000%, BCER train=0.029000%, BWER train=0.083000%, skip ratio=0.000000%,  New worst BCER = 0.029000 wrote checkpoint.

At iteration 718/10600/10600, Mean rms=0.112000%, delta=0.007000%, BCER train=0.023000%, BWER train=0.085000%, skip ratio=0.000000%,  New worst BCER = 0.023000 wrote checkpoint.

2 Percent improvement time=509, best error was 2.033 @ 209
At iteration 718/10700/10700, Mean rms=0.111000%, delta=0.006000%, BCER train=0.019000%, BWER train=0.069000%, skip ratio=0.000000%,  New best BCER = 0.019000 wrote best model:data/engRupee/checkpoints/engRupee_0.019000_718_10700.checkpoint wrote checkpoint.

2 Percent improvement time=509, best error was 2.033 @ 209
At iteration 718/10800/10800, Mean rms=0.108000%, delta=0.002000%, BCER train=0.007000%, BWER train=0.052000%, skip ratio=0.000000%,  New best BCER = 0.007000 wrote best model:data/engRupee/checkpoints/engRupee_0.007000_718_10800.checkpoint wrote checkpoint.

Finished! Selected model with minimal training error rate (BCER) = 0.007

With --debug_interval -1, the trainer outputs verbose debug text for every training iteration. The text debug information includes the truth text, the recognized text, the iteration number, the training sample id (lstmf file and line) and the mean value of several error metrics. GROUND TRUTH for the line is displayed in all cases. ALIGNED TRUTH and BEST OCR TEXT are displayed only when different from the GROUND TRUTH.

Iteration 455038: GROUND  TRUTH : उप॑ त्वाग्ने दि॒वेदि॑वे॒ दोषा॑वस्तर्धि॒या व॒यम् ।
File /tmp/san-2019-03-28.jsY/san.Mangal.exp0.lstmf line 451 (Perfect):
Mean rms=1.267%, delta=4.155%, train=11.308%(32.421%), skip ratio=0%
Iteration 455039: GROUND  TRUTH : मे अपराध और बैठे दुकानों नाम सकते अधिवक्ता, दोबारा साधन विषैले लगाने पर प्रयोगकर्ताओं भागे
File /tmp/san-2019-04-04.H4m/san.FreeSerif.exp0.lstmf line 28 (Perfect):
Mean rms=1.267%, delta=4.153%, train=11.3%(32.396%), skip ratio=0%

Iteration 1526: GROUND  TRUTH : 𒃻 𒀸 𒆳𒆳 𒅘𒊏𒀀𒋾
Iteration 1526: ALIGNED TRUTH : 𒃻 𒀸 𒆳𒆳 𒅘𒊏𒊏𒀀𒋾
Iteration 1526: BEST OCR TEXT :    𒀀𒋾
File /tmp/eng-2019-04-06.Ieb/eng.CuneiformComposite.exp0.lstmf line 19587 :
Mean rms=0.941%, delta=12.319%, train=56.134%(99.965%), skip ratio=0.6%
Iteration 1527: GROUND  TRUTH : 𒀭𒌋𒐊
Iteration 1527: BEST OCR TEXT : 𒀭𒌋
File /tmp/eng-2019-04-06.Ieb/eng.CuneiformOB.exp0.lstmf line 7771 :
Mean rms=0.941%, delta=12.329%, train=56.116%(99.965%), skip ratio=0.6%

With --debug_interval > 0, the trainer displays several windows of debug information on the layers of the network.

In the special case of --debug_interval 1 it waits for a click in the LSTMForward window before continuing to the next iteration, but for all others it just continues and draws information at the frequency requested.

NOTE that to use --debug_interval > 0 you must build ScrollView.jar as well as the other training tools. See Building the Training Tools

The visual debug information includes:

A forward and backward window for each network layer. Most are just random noise, but the Output/Output-back and ConvNL windows are worth viewing. Output shows the output of the final Softmax, which starts out as a yellow line for the null character, and gradually develops yellow marks at each point where it thinks there is a character. (The x-axis is the image x-coordinate, and the y-axis is character class.) The Output-back window shows the difference between the actual output and the target using the same layout, but with yellow for “give me more of this” and blue for “give me less of this”. As the network learns, the ConvNL window develops the typical edge detector results that you expect from the bottom layer.

LSTMForward shows the output of the whole network on the training image. LSTMTraining shows the training target on the training image. In both, green lines are drawn to show the peak output for each character, and the character itself is drawn to the right of the line.

The other two windows worth looking at are CTC Outputs and CTC Targets. These show the current output of the network and the targets as a line graph of strength of output against image x-coordinate. Instead of a heatmap, like the Output window, a different colored line is drawn for each character class and the y-axis is strength of output.

Iterations and Checkpoints

During the training we see this kind of information :

2 Percent improvement time=100, best error was 100 @ 0
At iteration 100/100/100, Mean rms=6.042000%, delta=63.801000%, BCER train=98.577000%, BWER train=100.000000%, skip ratio=0.000000%,  New best BCER = 98.577000 wrote checkpoint.

2 Percent improvement time=200, best error was 100 @ 0
At iteration 200/200/200, Mean rms=5.709000%, delta=58.372000%, BCER train=98.399000%, BWER train=99.986000%, skip ratio=0.000000%,  New best BCER = 98.399000 wrote checkpoint.
...
At iteration 14615/695400/698614, Mean rms=0.131000%, delta=0.038000%, BCER train=0.207000%, BWER train=0.579000%, skip ratio=0.4%,  wrote checkpoint.

In the above example,

14615 : learning_iteration
695400 : training_iteration
698614 : sample_iteration

sample_iteration : “Index into training sample set. (sample_iteration >= training_iteration).” It is how many times a training file has been passed into the learning process.

training_iteration : “Number of actual backward training steps used.” It is how many times a training file has been SUCCESSFULLY passed into the learning process. So every time you get an error : “Image too large to learn!!” - “Encoding of string failed!” - “Deserialize header failed”, the sample_iteration increments but not the training_iteration. Actually you have 1 - (695400 / 698614) = 0.4% which is the skip ratio : proportion of files that have been skipped because of an error

learning_iteration : “Number of iterations that yielded a non-zero delta error and thus provided significant learning. (learning_iteration <= training_iteration). learning_iteration_ is used to measure rate of learning progress.” So it uses the delta value to assess it the iteration has been useful.

What is good to know is that when you specify a maximum number of iterations to the training process it uses the middle iteration number (training_iteration) to know when to stop. But when it writes a checkpoint, the checkpoint name also uses the best iteration number (learning_iteration), along with the char train rate. So a checkpoint name is the concatenation of model_name & char_train & learning_iteration & training_iteration eg. sanLayer_1.754_347705_659600.checkpoint.

The lstmtraining program outputs two kinds of checkpoint files:

Either kind of these checkpoint files can be converted to a standard (best/float) traineddata file or slightly less accurate (fast/integer) traineddata file by using the stop_training and convert_to_int flags with lstmtraining.

Error Messages From Training

There are various error messages that can occur when running the training, some of which can be important, and others not so much:

Encoding of string failed! results when the text string for a training image cannot be encoded using the given unicharset. Possible causes are:

  1. There is an un-represented character in the text, say a British Pound sign that is not in your unicharset.
  2. A stray unprintable character (like tab or a control character) in the text.
  3. There is an un-represented Indic grapheme/aksara in the text.

You can simply remove the offending characters (CHARACTER TABULATION, CARRIAGE RETURN, NO-BREAK SPACE, LEFT-TO-RIGHT MARK, RIGHT-TO-LEFT MARK, ZERO WIDTH NO-BREAK SPACE, POP DIRECTIONAL FORMATTING, ZERO WIDTH NON-JOINER) in pct. 2 with a sed script:

remove_control_chars.sed

s/\x09//g
s/\x0d//g
s/\x00\xa0/ /g
s/\x20\x0e//g
s/\x20\x0f//g
s/\xfe\xff//g
s/\x20\x2c//g
s/\x20\x0c//g

sed -i -f remove_control_chars.sed data/lang-ground-truth*.gt.txt

In any case it will result in that training image being ignored by the trainer. If the error is infrequent, it is harmless, but it may indicate that your unicharset is inadequate for representing the language that you are training.

Unichar xxx is too long to encode!! (Most likely Indic only). There is an upper limit to the length of unicode characters that can be used in the recoder, which simplifies the unicharset for the LSTM engine. It will just continue and leave that Aksara out of the recognizable set, but if there are a lot, then you are in trouble.

Bad box coordinates in boxfile string! The LSTM trainer only needs bounding box information for a complete textline, instead of at a character level, but if you put spaces in the box string, like this:

<text for line including spaces> <left> <bottom> <right> <top> <page>

the parser will be confused and give you the error message.

Deserialize header failed occurs when a training input is not in LSTM format or the file is not readable. Check your filelist file to see if it contains valid filenames.

No block overlapping textline: occurs when layout analysis fails to correctly segment the image that was given as training data. The textline is dropped. Not much problem if there aren’t many, but if there are a lot, there is probably something wrong with the training text or rendering process.

<Undecodable> can occur in either the ALIGNED_TRUTH or OCR TEXT output early in training. It is a consequence of unicharset compression and CTC training. (See Unicharset Compression and train_mode above). This should be harmless and can be safely ignored. Its frequency should fall as training progresses.

Combining the Output Files

The lstmtraining program outputs two kinds of checkpoint files:

Either of these files can be converted to a standard traineddata file. This will extract the recognition model from the training dump, and insert it into the –traineddata argument, along with the unicharset, recoder, and any dawgs that were provided during training.

NOTE Tesseract will now run happily with a traineddata file that contains just lang.lstm, lang.lstm-unicharset and lang.lstm-recoder. The lstm-*-dawgs are optional, and none of the other components are required or used with OEM_LSTM_ONLY as the OCR engine mode. No bigrams, unichar ambigs or any of the other components are needed or even have any effect if present. The only other component that does anything is the lang.config, which can affect layout analysis, and sub-languages.

If added to an existing Tesseract traineddata file, the lstm-unicharset doesn’t have to match the Tesseract unicharset, but the same unicharset must be used to train the LSTM and build the lstm-*-dawgs files.

The Hallucination Effect

If you notice that your model is misbehaving, for example by:

Then read the hallucination topic.