Laia: A deep learning toolkit for HTR based on Torch
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
laia Merge pull request #30 from bdotgradb/heap Jul 22, 2018
rocks Updated rockspec to add lalarm Apr 22, 2018
.travis.yml Added test to travis Oct 20, 2016 Included LICENSE 👮 Sep 26, 2016
githook-pre-commit Modified precommit hook so that version is only updated when a laia- … Feb 5, 2017
laia-create-model Priors are now computed using the character post-probabilities of the… Nov 11, 2017
laia-netout Modified laia-netout to smooth priors for 0 count priors. Working on … May 25, 2017

Laia: A deep learning toolkit for HTR

Build Status

Laia is a deep learning toolkit to transcribe handwritten text images.

If you find this toolkit useful in your research, please cite:

  author = {Joan Puigcerver and
            Daniel Martin-Albo and
            Mauricio Villegas},
  title = {Laia: A deep learning toolkit for HTR},
  year = {2016},
  publisher = {GitHub},
  note = {GitHub repository},
  howpublished = {\url{}},


Laia is implemented in Torch, and depends on the following:

Note that currently we only support GPU. You need to use NVIDIA's cuDNN library. Register first for the CUDA Developer Program (it's free) and download the library from NVIDIA's website.

Once Torch is installed the following luarocks are required:

And execute luarocks install

Installation via docker

To ease the installation, there is a public docker image for Laia. To use it first install docker and nvidia-docker, and configure docker so that it can be executed without requiring sudo, see docker linux postinstall. Then the installation of Laia consists of first pulling the image and tagging it as laia:active.

docker pull mauvilsa/laia:[SOME_TAG]
docker tag mauvilsa/laia:[SOME_TAG] laia:active

Replace SOME_TAG with one of the tags available here. Then copy the command line interface script to some directory in your path for easily use from the host.

mkdir -p $HOME/bin
docker run --rm -u $(id -u):$(id -g) -v $HOME:$HOME laia:active bash -c "cp /usr/local/bin/laia-docker $HOME/bin"

After this, all Laia commands can be executed by using the laia-docker command. For further details run.

laia-docker --help


Training a Laia model using CTC:

Create an "empty" model using:

laia-create-model \
    "$CHANNELS" "$INPUT_HEIGHT" "$((NUM_SYMB+1))" "$MODEL_DIR/model.t7";

Or if installed via docker:

laia-docker create-model \
    "$CHANNELS" "$INPUT_HEIGHT" "$((NUM_SYMB+1))" "$MODEL_DIR/model.t7";

Positional arguments:

  • $CHANNELS: number of channels of the input images.
  • $INPUT_HEIGHT: height of the input images. Note: ALL image must have the same height.
  • $((NUM_SYMB+1)): number of output symbols. Note: Include ONE additional element for the CTC blank symbol.
  • $MODEL_DIR/model.t7: path to the output model.

For optional arguments check laia-create-model -h or laia-create-model -H.

Train the model using:

laia-train-ctc \
    "$MODEL_DIR/model.t7" \

Or if installed via docker:

laia-docker train-ctc \
    "$MODEL_DIR/model.t7" \

Positional arguments:

  • $MODEL_DIR/model.t7 is the path to the input model or checkpoint for training.
  • $SYMBOLS_TABLE is the list of training symbols and their id.
  • $TRAIN_LST is a file containing a list of images for training.
  • $TRAIN_GT is a file containing the list of training transcripts.
  • $VALID_LST is a file containing a list of images for validation.
  • $VALID_GT is a file containing the list of validation transcripts.

For optional arguments check laia-train-ctc -h or laia-create-model -H.


laia-decode "$MODEL_DIR/model.t7" "$TEST_LST";

Or if installed via docker:

laia-docker decode "$MODEL_DIR/model.t7" "$TEST_LST";

Positional arguments:

  • $MODEL_DIR/model.t7 is the path to the model.
  • $TEST_LST is a file containing a list of images for testing.

For optional arguments check laia-decode -h.


For a more detailed example, see the Spanish Numbers in egs/spanish-numbers folder, or the IAM in egs/iam folder.