《For Professional/Commercial Use》Limited Release of Deep Learning Framework, SmallTrain 0.1.2

AI open source project that ensures professional usability

Geek Guild Co., Ltd. starts an open source project of “SmallTrain”, which pursues high accuracy and high functionality, with a function as a wrapper for TensorFlow and PyTorch and as a library that provides original algorithms.
It starts on Wednesday, April 15, 2020.
First of all, it is open to registered users only, and the trained model built with SmallTrain 0.1.2 is used commercially.
Registered users are able to experience how easy it is to build a high-accuracy, trained model that can withstand commercial use, and contribute (contribute) to improving the source code.

See SmallTrain site for more imformation.

Artificial General Intelligence Open Source Goal

In the context of the third AI boom that began with deep learning, many companies are developing their own AI. However, each type of AI is classified as a “Narrow AI” and is only active in the industry for limited uses. We are convinced that AI technology is a fundamental technology for machines to help humanity and build symbiotic spheres.
As a platform for that, we propose “Harmonic AI”. Let’s help each other with their own small data, aiming for a better society, by integrating the AI that they have independently developed and forming a general-purpose AI!

Three Advantages of SmallTrain

1.Free and easy creation of commercially available AI models equipped with cutting-edge AI algorithms 2.Use as a trained model that supports small data can significantly reduce man-hours 3.As a wrapper* for both TensorFlow and PyTorch, a library function to provide unique algorithms will be installed

What is a wrapper? …wrapping a library such as TensorFlow or PyTorch, and users can access the library via the wrapper so that even if the library is replaced or the interface of the library is changed, the changes are stopped only inside the wrapper and the effect of the change are stopped.

SmallTrain Open Source Project Contributor

1.Environment where you can work remotely (using version control system GitLab) 2.OSS project contributing experience 3.AI model development experience

Background of the SmallTrain Open Source Project

AI development is difficult due to the need for data science background and advanced programming skills.
In addition, there are methods that can be used to build AI using open source libraries, etc., but it is hard to say that quality can withstand service operation.
Therefore, Geek Guild released the source code of the AI ​​model developed in-house and started a project for public use for service development.

How to Use SmallTrain

This version introduces how to create an image recognition model as an introduction.
It can be used for a wide range of AI services such as time-series data prediction and voice recognition,
and also it can be used as a wrapper for libraries, TensorFlow, and PyTorch.

See for more information.

1. A Library of Deep Learning models, and a wrapper

There are three main ways to create AI models:

  • Understand algorithms and mathematics and build your own model
  • Algorithms and mathematical functions are self-made using calculation libraries such as TensorFlow
  • Easily create models using wrappers to call calculation libraries (most man-hours can be reduced)

SmallTrain aims to be a wrapper equivalent to Keras.

  • The difference from Keras
    Keras is suitable for PoC, SmallTrain is for both PoC and commercial use.
  • The similarity to Keras
    It is easy to use for beginners.

2. Wrapping TensorFlow mathematical functions

3. You can call mathematical functions of both PyTorch and TensorFlow.

4. You can also call SmallTrain’s own calculation library.
We read the state-of-the-art papers and implement the cutting edge of algorithms.

5. AI model of SmallTrain
Using TensorFlow, PyTorch, and unique mathematical functions, we have built a deep neural network* with more than 60 layers.

* It is a deep neural network that holds Pyramid Network and can produce highly accurate results, and also it incorporates CNN and other techniques.

6. Trained model
SmallTrain has been learned.
It is highly versatile and supports all kinds of data, and learns with various data such as image data and time series data.

7. Providing your own trained model
By inputting user’s data, you can easily build your own trained model. Getting Started describes how to recognize images as an introduction. It supports various data.

  • SmallTrain and Geek Guild are trademarks of Geek Guild Co., Ltd.
  • TensorFlow is a trademark or registered trademark of Google LLC.
  • All other trademarks are property of their respective owners.