NeuronBlocks is a NLP deep learning modeling toolkit that helps engineers/researchers to build end-to-end pipelines for neural network model training for NLP tasks. The main goal of this toolkit is to minimize developing cost for NLP deep neural network model building, including both training and inference stages.

NeuronBlocks consists of two major components: Block Zoo and Model Zoo.

  • In Block Zoo, we provide commonly used neural network components as building blocks for model architecture design.
  • In Model Zoo, we provide a suite of NLP models for common NLP tasks, in the form of JSON configuration files.

NLP Tasks Supported

  • Sentence Classification
  • Sentiment Analysis
  • Question Answering Matching
  • Textual Entailment
  • Slot tagging
  • Machine Reading Comprehension
  • Knowledge Distillation for Model Compression
  • More on-going

Toolkit Usage

Users can either pick existing models (config files) in Model Zoo to start model training or create new models by leveraging neural network blocks in Block Zoo just like playing with Lego.

Get Started in 60 Seconds


Note: NeuronBlocks requires Python 3.6 and above.

  1. Clone this project.

    git clone
  2. Install Python packages in requirements.txt by the following command.

    pip install -r requirements.txt

    NeuronBlocks requires PyTorch 0.4.1 currently and automatically gets installed by the above on Linux.

    For Windows, we suggest you to install PyTorch via Conda by following the instructions of PyTorch here.

Quick Start

Get started by trying the given examples. Both Linux/Windows, GPU/CPU are supported. For Windows, we suggest you to use PowerShell instead of CMD.

Tips: in the following instruction, PROJECTROOT denotes the root directory of this project.

# train
python --conf_path=model_zoo/demo/conf.json

# test
python --conf_path=model_zoo/demo/conf.json

# predict
python --conf_path=model_zoo/demo/conf.json

For more details, please refer to and Code documentation.

Who should consider using NeuronBlocks

Engineers or researchers who face the following challenges when using neural network models to address NLP problems:

  • Many frameworks to choose and high framework studying cost.
  • Heavy coding cost. A lot of details make it hard to debug.
  • Fast Model Architecture Evolution. It is difficult for engineers to understand the mathematical principles behind them.
  • Model Code optimization requires deep expertise.
  • Model Platform Compatibility Requirement. It requires extra coding work for the model to run on different platforms, such as Linux/Windows, GPU/CPU.

The advantages of leveraging NeuronBlocks for NLP neural network model training includes:

  • Model Building: for model building and parameter tuning, users only need to write simple JSON config files, which greatly minimize the effort of implementing new ideas.

  • Model Sharing It is super easy to share models just through JSON files, instead of nasty codes. For different models or tasks, our users only need to maintain one single centralized source code base.

  • Code Reusability: Common blocks can be easily shared across various models or tasks, reducing duplicate coding work.

  • Platform Flexibility: NeuronBlocks can run both on Linux and Windows machines, using both CPU and GPU. It also supports training on GPU platforms like Philly and PAI.

    CPU inferenceSingle-GPU inferenceMulti-GPU inference
    CPU train
    Single-GPU train
    Multi-GPU train
  • Model Visualization: A model visualizer is provided for visualization and configure correctness checking, which helps users to visualize the model architecture easily during debugging.

  • Extensibility: NeuronBlocks is extensible, allowing users to contribute new blocks or contributing novel models (JSON files).


NeuronBlocks operates in an open model. It is designed and developed by STCA NLP Group, Microsoft. Contributions from academia and industry are also highly welcome. For more details, please refer to

Ongoing Work and Call for Contributions

Anyone who are familiar with are highly encouraged to contribute code.

  • Knowledge Distillation for Model Compression. Knowledge distillation for heavy models such as BERT, OpenAI Transformer. Teacher-Student based knowledge distillation is one common method for model compression.
  • Multi-Lingual Support
  • Chinese Language Support
  • NER Model Support
  • Multi-Task Training Support