Skip to content
Go to file


PyPI Python



What is it?

Lingvo is a framework for building neural networks in Tensorflow, particularly sequence models.

A list of publications using Lingvo can be found here.

Table of Contents


PyPI Version Commit
0.8.2 93e123c6788e934e6b7b1fd85770371becf1e92e
0.7.2 b05642fe386ee79e0d88aa083565c9a93428519e
Older releases

PyPI Version Commit

Details for older releases are unavailable.

Major breaking changes


  • General
    • py_utils.AddGlobalVN and py_utils.AddPerStepVN have been combined into py_utils.AddVN.
    • BaseSchedule().Value() no longer takes a step arg.
    • Classes deriving from BaseSchedule should implement Value() not FProp().
    • theta.global_step has been removed in favor of py_utils.GetGlobalStep().
    • py_utils.GenerateStepSeedPair() no longer takes a global_step arg.
    • PostTrainingStepUpdate() no longer takes a global_step arg.


  • General
    • NestedMap Flatten/Pack/Transform/Filter etc now expand descendent dicts as well.
    • Subclasses of BaseLayer extending from abc.ABCMeta should now extend base_layer.ABCLayerMeta instead.
    • Trying to call self.CreateChild outside of __init__ now raises an error.
    • base_layer.initializer has been removed. Subclasses no longer need to decorate their __init__ function.
    • Trying to call self.CreateVariable outside of __init__ or _CreateLayerVariables now raises an error.
    • It is no longer possible to access self.vars or self.theta inside of __init__. Refactor by moving the variable creation and access to _CreateLayerVariables. The variable scope is set automatically according to the layer name in _CreateLayerVariables.
Older releases

Details for older releases are unavailable.

Quick start


There are two ways to set up Lingvo: installing a fixed version through pip, or cloning the repository and building it with bazel. Docker configurations are provided for each case.

If you would just like to use the framework as-is, it is easiest to just install it through pip. This makes it possible to develop and train custom models using a frozen version of the Lingvo framework. However, it is difficult to modify the framework code or implement new custom ops.

If you would like to develop the framework further and potentially contribute pull requests, you should avoid using pip and clone the repository instead.


The Lingvo pip package can be installed with pip3 install lingvo.

See the codelab for how to get started with the pip package.

From sources:

The prerequisites are:

  • a TensorFlow 2.3 installation,
  • a C++ compiler (only g++ 7.3 is officially supported), and
  • the bazel build system.

Refer to docker/dev.dockerfile for a set of working requirements.

git clone the repository, then use bazel to build and run targets directly. The python -m module commands in the codelab need to be mapped onto bazel run commands.


Docker configurations are available for both situations. Instructions can be found in the comments on the top of each file.

How to install docker.

Running the MNIST image model

Preparing the input data


mkdir -p /tmp/mnist
python3 -m --dataset=mnist


mkdir -p /tmp/mnist
bazel run -c opt //lingvo/tools:keras2ckpt -- --dataset=mnist

The following files will be created in /tmp/mnist:

  • 53MB.
  • mnist.index: 241 bytes.

Running the model


cd /tmp/mnist
curl -O
python3 -m lingvo.trainer --run_locally=cpu --mode=sync --model=mnist.LeNet5 --logdir=/tmp/mnist/log


(cpu) bazel build -c opt //lingvo:trainer
(gpu) bazel build -c opt --config=cuda //lingvo:trainer
bazel-bin/lingvo/trainer --run_locally=cpu --mode=sync --model=image.mnist.LeNet5 --logdir=/tmp/mnist/log --logtostderr

After about 20 seconds, the loss should drop below 0.3 and a checkpoint will be saved, like below. Kill the trainer with Ctrl+C.] step:   205, steps/sec: 11.64 ... loss:0.25747201 ...] Save checkpoint] Save checkpoint done: /tmp/mnist/log/train/ckpt-00000205

Some artifacts will be produced in /tmp/mnist/log/control:

  • params.txt: hyper-parameters.
  • model_analysis.txt: model sizes for each layer.
  • train.pbtxt: the training tf.GraphDef.
  • events.*: a tensorboard events file.

As well as in /tmp/mnist/log/train:

  • checkpoint: a text file containing information about the checkpoint files.
  • ckpt-*: the checkpoint files.

Now, let's evaluate the model on the "Test" dataset. In the normal training setup the trainer and evaler should be run at the same time as two separate processes.


python3 -m lingvo.trainer --job=evaler_test --run_locally=cpu --mode=sync --model=mnist.LeNet5 --logdir=/tmp/mnist/log


bazel-bin/lingvo/trainer --job=evaler_test --run_locally=cpu --mode=sync --model=image.mnist.LeNet5 --logdir=/tmp/mnist/log --logtostderr

Kill the job with Ctrl+C when it starts waiting for a new checkpoint.] No new check point is found: /tmp/mnist/log/train/ckpt-00000205

The evaluation accuracy can be found slightly earlier in the logs.] eval_test: step:   205, acc5: 0.99775392, accuracy: 0.94150388, ..., loss: 0.20770954, ...

Running the machine translation model

To run a more elaborate model, you'll need a cluster with GPUs. Please refer to third_party/py/lingvo/tasks/mt/ for more information.

Running the GShard transformer based giant language model

To train a GShard language model with one trillion parameters on GCP using CloudTPUs v3-512 using 512-way model parallelism, please refer to third_party/py/lingvo/tasks/lm/ for more information.

Running the 3d object detection model

To run the StarNet model using CloudTPUs on GCP, please refer to third_party/py/lingvo/tasks/car/


Automatic Speech Recognition



Language Modelling

Machine Translation

[1]: Listen, Attend and Spell. William Chan, Navdeep Jaitly, Quoc V. Le, and Oriol Vinyals. ICASSP 2016.

[2]: End-to-end Continuous Speech Recognition using Attention-based Recurrent NN: First Results. Jan Chorowski, Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. arXiv 2014.

[3]: StarNet: Targeted Computation for Object Detection in Point Clouds. Jiquan Ngiam, Benjamin Caine, Wei Han, Brandon Yang, Yuning Chai, Pei Sun, Yin Zhou, Xi Yi, Ouais Alsharif, Patrick Nguyen, Zhifeng Chen, Jonathon Shlens, and Vijay Vasudevan. arXiv 2019.

[4]: Gradient-based learning applied to document recognition. Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. IEEE 1998.

[5]: Exploring the Limits of Language Modeling. Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. arXiv, 2016.

[6]: GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding. Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer and Zhifeng Chen arXiv, 2020.

[7]: The Best of Both Worlds: Combining Recent Advances in Neural Machine Translation. Mia X. Chen, Orhan Firat, Ankur Bapna, Melvin Johnson, Wolfgang Macherey, George Foster, Llion Jones, Mike Schuster, Noam Shazeer, Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Zhifeng Chen, Yonghui Wu, and Macduff Hughes. ACL 2018.


Please cite this paper when referencing Lingvo.

    title={Lingvo: a Modular and Scalable Framework for Sequence-to-Sequence Modeling},
    author={Jonathan Shen and Patrick Nguyen and Yonghui Wu and Zhifeng Chen and others},


Apache License 2.0

You can’t perform that action at this time.