Training workflow for Tesseract 5 as a Makefile for dependency tracking.
You will need at least GNU make
(minimal version 4.2), wget
, find
, bash
, and unzip
.
You will need a recent version (>= 5.3) of tesseract built with the training tools and matching leptonica bindings. Build instructions and more can be found in the Tesseract User Manual.
find
, unzip
, rm
) and put C:\Program Files\Git\usr\bin
to the beginning of your PATH variable (temporarily you can do it in cmd
with set PATH=C:\Program Files\Git\usr\bin;%PATH%
- unfortunately there are several Windows tools with the same name as on linux (find
, sort
) with different behavior/functionality and there is need to avoid them during training.winget install ezwinports.make
and winget install wget
to install missing tools.You need a recent version of Python 3.x. For image processing the Python library Pillow
is used.
If you don't have a global installation, please use the provided requirements file pip install -r requirements.txt
.
Tesseract expects some configuration data (a file radical-stroke.txt
and *.unicharset
for all scripts) in DATA_DIR
.
To fetch them:
make tesseract-langdata
(While this step is only needed once and implicitly included in the training
target,
you might want to run it explicitly beforehand.)
Choose a name for your model. By convention, Tesseract stack models including
language-specific resources use (lowercase) three-letter codes defined in
ISO 639 with additional
information separated by underscore. E.g., chi_tra_vert
for traditional
Chinese with vertical typesetting. Language-independent (i.e. script-specific)
models use the capitalized name of the script type as an identifier. E.g.,
Hangul_vert
for Hangul script with vertical typesetting. In the following,
the model name is referenced by MODEL_NAME
.
Place ground truth consisting of line images and transcriptions in the folder
data/MODEL_NAME-ground-truth
. This list of files will be split into training and
evaluation data, the ratio is defined by the RATIO_TRAIN
variable.
Images must be TIFF and have the extension .tif
or PNG and have the
extension .png
, .bin.png
, or .nrm.png
.
Transcriptions must be single-line plain text and have the same name as the
line image but with the image extension replaced by .gt.txt
.
The repository contains a ZIP archive with sample ground truth, see
ocrd-testset.zip. Extract it to ./data/foo-ground-truth
and run
make training
.
NOTE: If you want to generate line images for transcription from a full page, see tips in issue 7 and in particular @Shreeshrii's shell script.
Run
make training MODEL_NAME=name-of-the-resulting-model
which is a shortcut for
make unicharset lists proto-model tesseract-langdata training MODEL_NAME=name-of-the-resulting-model
Run make help
to see all the possible targets and variables:
Targets
unicharset Create unicharset
charfreq Show character histogram
lists Create lists of lstmf filenames for training and eval
training Start training (i.e. create .checkpoint files)
traineddata Create best and fast .traineddata files from each .checkpoint file
proto-model Build the proto model
tesseract-langdata Download stock unicharsets
evaluation Evaluate .checkpoint models on eval dataset via lstmeval
plot Generate train/eval error rate charts from training log
clean Clean all generated files
Variables
MODEL_NAME Name of the model to be built. Default: foo
START_MODEL Name of the model to continue from (i.e. fine-tune). Default: ''
PROTO_MODEL Name of the prototype model. Default: OUTPUT_DIR/MODEL_NAME.traineddata
WORDLIST_FILE Optional file for dictionary DAWG. Default: OUTPUT_DIR/MODEL_NAME.wordlist
NUMBERS_FILE Optional file for number patterns DAWG. Default: OUTPUT_DIR/MODEL_NAME.numbers
PUNC_FILE Optional file for punctuation DAWG. Default: OUTPUT_DIR/MODEL_NAME.punc
DATA_DIR Data directory for output files, proto model, start model, etc. Default: data
OUTPUT_DIR Output directory for generated files. Default: DATA_DIR/MODEL_NAME
GROUND_TRUTH_DIR Ground truth directory. Default: OUTPUT_DIR-ground-truth
TESSDATA_REPO Tesseract model repo to use (_fast or _best). Default: _best
TESSDATA Path to the directory containing START_MODEL.traineddata
(for example tesseract-ocr/tessdata_best). Default: ./usr/share/tessdata
MAX_ITERATIONS Max iterations. Default: 10000
EPOCHS Set max iterations based on the number of lines for training. Default: none
DEBUG_INTERVAL Debug Interval. Default: 0
LEARNING_RATE Learning rate. Default: 0.0001 with START_MODEL, otherwise 0.002
NET_SPEC Network specification (in VGSL) for new model from scratch. Default: [1,36,0,1 Ct3,3,16 Mp3,3 Lfys48 Lfx96 Lrx96 Lfx256 O1c###]
FINETUNE_TYPE Fine-tune Training Type - Impact, Plus, Layer or blank. Default: ''
LANG_TYPE Language Type - Indic, RTL or blank. Default: ''
PSM Page segmentation mode. Default: 13
RANDOM_SEED Random seed for shuffling of the training data. Default: 0
RATIO_TRAIN Ratio of train / eval training data. Default: 0.90
TARGET_ERROR_RATE Stop training if the character error rate (CER in percent) gets below this value. Default: 0.01
LOG_FILE File to copy training output to and read plot figures from. Default: OUTPUT_DIR/training.log
First, decide what kind of training you want.
START_MODEL
NET_SPEC
(see documentation)To override the default path name requirements, just set the respective variables in the above list:
make training MODEL_NAME=name-of-the-resulting-model DATA_DIR=/data GROUND_TRUTH_DIR=/data/GT
If you want to use shell variables to override the make variables (for example because
you are running tesstrain from a script or other makefile), then you can use the -e
flag:
MODEL_NAME=name-of-the-resulting-model DATA_DIR=/data GROUND_TRUTH_DIR=/data/GT make -e training
When the training is finished, it will write a traineddata
file which can be used
for text recognition with Tesseract. Note that this file does not include a
dictionary. The tesseract
executable therefore prints a warning.
It is also possible to create additional traineddata
files from intermediate
training results (the so-called checkpoints). This can even be done while the
training is still running. Example:
# Add MODEL_NAME and OUTPUT_DIR like for the training.
make traineddata
This will create two directories tessdata_best
and tessdata_fast
in OUTPUT_DIR
with a best (double based) and fast (int based) model for each checkpoint.
It is also possible to create models for selected checkpoints only. Examples:
# Make traineddata for the checkpoint files of the last three weeks.
make traineddata CHECKPOINT_FILES="$(find data/foo -name '*.checkpoint' -mtime -21)"
# Make traineddata for the last two checkpoint files.
make traineddata CHECKPOINT_FILES="$(ls -t data/foo/checkpoints/*.checkpoint | head -2)"
# Make traineddata for all checkpoint files with CER better than 1 %.
make traineddata CHECKPOINT_FILES="$(ls data/foo/checkpoints/*[^1-9]0.*.checkpoint)"
Add MODEL_NAME
and OUTPUT_DIR
and replace data/foo
with the output directory if needed.
Training and Evaluation Character Error Rate (CER) can be plotted using Matplotlib:
# Make OUTPUT_DIR/MODEL_FILE.plot_*.png
make plot
All the variables defined above apply, but there is no explicit dependency on training
.
Still, the target depends on the LOG_FILE
captured during training (just will not trigger
training itself). Besides analysing the log file, this also directly evaluates the trained models
(for each checkpoint) on the eval dataset. The latter is also available as an independent target
evaluation
:
# Make OUTPUT_DIR/eval/MODEL_FILE*.*.log
make evaluation
Plotting can even be done while training is still running, and will depict the training status
up to that point. (It can be rerun any time the LOG_FILE
has changed or new checkpoints written.)
As an example, use the training data provided in ocrd-testset.zip to do some training and generate the plots:
unzip ocrd-testset.zip -d data/ocrd-ground-truth
make training MODEL_NAME=ocrd START_MODEL=frk TESSDATA=~/tessdata_best MAX_ITERATIONS=10000 &
# Make data/ocrd/ocrd.plot_cer.png and plot_log.png (repeat during/after training)
make plot MODEL_NAME=ocrd
Which should then look like this:
Software is provided under the terms of the Apache 2.0
license.
Sample training data provided by Deutsches Textarchiv is in the public domain.