kristiankielhofner/nemo

NeMo: a toolkit for conversational AI

NVIDIA NeMo

NeMo is a toolkit for creating Conversational AI applications.

NeMo toolkit makes it possible for researchers to easily compose complex neural network architectures for conversational AI using reusable components - Neural Modules. Neural Modules are conceptual blocks of neural networks that take typed inputs and produce typed outputs. Such modules typically represent data layers, encoders, decoders, language models, loss functions, or methods of combining activations.

The toolkit comes with extendable collections of pre-built modules for automatic speech recognition (ASR), natural language processing (NLP) and text synthesis (TTS).

Built for speed, NeMo can utilize NVIDIA's Tensor Cores and scale out training to multiple GPUs and multiple nodes. NeMo has integration with NVIDIA Jarvis.

Introduction

• Watch this video for a quick walk-through.
• Documentation (latest released version) and Documentation (master branch)
• Read NVIDIA Developer Blog to learn how to develop speech recognition models for different languages
• Read NVIDIA Developer Blog announcing NeMo
• Read NVIDIA Developer Blog for example applications
• Read NVIDIA Developer Blog for QuartzNet ASR model
• Recommended version to install is 0.10.1 via pip install nemo-toolkit[all]
• Recommended NVIDIA NGC NeMo Toolkit container
• Pretrained models are available on NVIDIA NGC Model repository

Getting started

THE LATEST STABLE VERSION OF NeMo is 0.10.1 (Available via PIP).

Requirements

1. Python 3.6 or 3.7
2. PyTorch 1.4.* with GPU support
3. (optional, for best performance) NVIDIA APEX. Install from here: https://github.com/NVIDIA/apex

Docker containers

NeMo docker container

You can use NeMo's docker container with all dependencies pre-installed

docker run --runtime=nvidia -it --rm -v --shm-size=16g -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/nemo:v0.10

If you are using the NVIDIA NGC PyTorch container follow these instructions

• Pull the docker: docker pull nvcr.io/nvidia/pytorch:20.01-py3
• Run:docker run --gpus all -it --rm -v <nemo_github_folder>:/NeMo --shm-size=8g -p 8888:8888 -p 6006:6006 --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/pytorch:20.01-py3
• apt-get update && apt-get install -y libsndfile1 ffmpeg && pip install Cython
• pip install nemo_toolkit Installs NeMo core only.
• pip install nemo_toolkit[all] Installs NeMo core and ALL collections
• pip install nemo_toolkit[asr] Installs NeMo core and ASR (Speech Recognition) collection
• pip install nemo_toolkit[nlp] Installs NeMo core and NLP (Natural Language Processing) collection
• pip install nemo_toolkit[tts] Installs NeMo core and TTS (Speech Synthesis) collection

See examples/start_here to get started with the simplest example.

Tutorials

• Speech recognition
• Natural language processing
• Speech Synthesis

Pre-trained models

Modality	Model	Trained on
ASR	Jasper10x5DR_En	LibriSpeech, WSJ, Mozilla Common Voice (en_1488h_2019-12-10), Fisher, Switchboard, and Singapore English National Speech Corpus (Part 1)
ASR	QuartzNet15x5En	LibriSpeech, WSJ, Mozilla Common Voice (en_1087h_2019-06-12), Fisher, and Switchboard
ASR	QuartzNet15x5Zh	AISHELL-2 Mandarin
NLP	BERT base uncased	English Wikipedia and BookCorpus dataset seq len <= 512
NLP	BERT large uncased	English Wikipedia and BookCorpus dataset seq len <= 512
TTS	Tacotron2	LJspeech
TTS	WaveGlow	LJspeech

DEVELOPMENT

If you'd like to use master branch and/or develop NeMo you can run "reinstall.sh" script.

Documentation (master branch).

Installing From Github

If you prefer to use NeMo's latest development version (from GitHub) follow the steps below:

1. Clone the repository git clone https://github.com/NVIDIA/NeMo.git
2. Go to NeMo folder and re-install the toolkit with collections:

./reinstall.sh

Style tests

python setup.py style  # Checks overall project code style and output issues with diff.
python setup.py style --fix  # Tries to fix error in-place.
python setup.py style --scope=tests  # Operates within certain scope (dir of file).

** NeMo Test Suite**

NeMo contains test suite divided into 5 subsets:

1. unit: unit tests, i.e. testing a single, well isolated functionality
2. integration: tests checking the elements when integrated into subsystems
3. system: tests working at the highest integration level
4. acceptance: tests checking whether the developed product/model passes the user defined acceptance criteria
5. docs: tests related to documentation (deselect with '-m "not docs"')

The user can run all the tests locally by simply executing:

pytest

In order to run a subset of tests one can use the -m argument followed by the subset name, e.g. for system subset:

pytest -m system

By default, all the tests will be executed on GPU. There is also an option to run the test suite on CPU by passing the --cpu command line argument, e.g.:

pytest -m unit --cpu

Citation

If you are using NeMo please cite the following publication

@misc{nemo2019,
    title={NeMo: a toolkit for building AI applications using Neural Modules},
    author={Oleksii Kuchaiev and Jason Li and Huyen Nguyen and Oleksii Hrinchuk and Ryan Leary and Boris Ginsburg and Samuel Kriman and Stanislav Beliaev and Vitaly Lavrukhin and Jack Cook and Patrice Castonguay and Mariya Popova and Jocelyn Huang and Jonathan M. Cohen},
    year={2019},
    eprint={1909.09577},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}