Skip to content
 

Pipcook - Bridging JavaScript with Python for machine learning - Interview with Wenhe Li

Published:

There’s a lot of excitement about machine learning and its applications. The question is, what can you do with and where to apply the technique and how.

To learn more, I am interviewing Wenhe (Eric) Li, the creator of Pipcook.

Can you tell a bit about yourself?
🔗

Hi folks, this is Wenhe (Eric) Li, and I am currently an SDE at Alibaba Inc. My works involve combining front-end development, front-end developer, and artificial intelligence (AI). One of my tasks here is developing Pipcook, an open-sourced machine learning (ML), and deep learning (DL) framework designed for front-end developers.

How would you describe Pipcook to someone who has never heard of it?
🔗

Pipcook is a tool that helps you develop, train, and deploy an ML/DL model without much prior knowledge. The whole workflow is highly abstract without losing scalability.

It lets you use popular Python-based machine learning solutions, such as NumPy, scipy-learn, jieba, and TensorFlow, easily through its interface.

How does Pipcook work?
🔗

Pipcook wraps Python using BOA
🔗

Since this framework is front-end and Node.js developer-oriented, and most DL/ML libraries have been written using Python, we created BOA to bridge the languages.

BOA allows us to directly import and call Python modules and methods in JavaScript. In this way, we can utilize the DL/ML ecosystem in Python without worrying about learning a new language.

Pipcook leverages the concept of pipelines
🔗

We introduce a concept of pipeline which contains dataCollect, dataProcess, dataAccess, datasetProcess, modelDefine, modelTrain, and modelEva. Pipelines offer an abstraction over a typical DL/ML model lifecycle.

Pipcook developers, including the community, offer the most common implementation of these parts (we call them plugins). People who want to train their model can use an existing pipeline or combine plugins to make their pipeline just like playing with legos.

How does Pipcook differ from other solutions?
🔗

Pipcook lets you use both JavaScript and Python
🔗

Since Pipcook allows you to write Dl/ML models in JavaScript with Python modules, we can benefit from the great libraries and packages in the two ecosystems.

It’s incredible as you can decide to put some IO-oriented jobs to Node.js and put more DL/ML training-related work under Python. Doing this allows you to get the most out of both.

Pipcook uses pipelines and plugins
🔗

Pipcook introduces pipeline and plugins to the DL/ML workflow. Doing this decouples the complexity of developing ML/DL models and makes the plugins highly shareable.

Pipcook uses state-of-art techiques
🔗

Since Pipcook is an experimental project, we can use state-of-the-art techniques and languages to develop our project. That means using Rust, WASM, WASI, WebGPU, and more.

Why did you develop Pipcook?
🔗

I love JavaScript and its magical syntax. However, I have to use Python to develop DL/ML models due to the abundant Python modules and ecosystem. Pipcook gives me a new way to establish DL/ML models in JavaScript without losing the Python ecosystem.

So far, we’ve seen a clear tendency that AI comes into every corner of the world. And in the field of front-end, we still do not have an industrial level framework. Most DL/ML frameworks are still serving people who have related knowledge.

We want to deliver such a framework that could be widely used by the JavaScript (Node.js and browser) world without worrying about the complex theory behind models. We believe this is the right way of passing the value of AI to the world.

What next?
🔗

We just formally released our Pipcook a couple of months ago. This very first public release offers users an out-of-the-box feature of training a model for image classification, style transfer, and text analysis without much prior knowledge.

Therefore, for the current stage, we are working for the user experience and developer experience. We are trying to optimize the training efficiency and mitigate the learning curve of developing a plugin.

Apart from that, we are building an all-in-one toolkit, which includes viewing training log, inspecting, and visualizing model structure, and model pruning and compression.

In the future, ML/DL must have a stronger binding with the web in general. Distributed training, federal learning, and on-device inference will fourish in the web since all of them match the essence of the internet.

Quoted from Tim Berners-Lee, “Let’s redistribute power to individuals!” We are trying to build a paradigm that ML/DL could serve and be open to everyone with web development ability and JavaScript and Python ecosystems.

What advice would you give to programmers getting into web development?
🔗

The web is essentially a community based on the idea of sharing, connecting, and open-source. Thus, try to connect, join, and work with the open-source community you are using or find interesting!

Who should I interview next?
🔗

You could interview developers who have made their very first open-source contributions.

Any last remarks?
🔗

Thanks for the interview and this chance to share my journey along with Pipcook.

Conclusion
🔗

Thanks, Wenhe! I find it admirable what you are doing with Pipcook. I believe it can work as a bridge for JavaScript developers to the world of machine learning without having to delve deep into Python-based solutions.

!
To learn more, see Pipcook site and Pipcook on Github.

Comments
🔗