Canvas

Introduction

The best deep learning architectures are growing increasingly complex, with separate towers, residual connections, cross communication, and more. Adapting the state of the art to your specific problem domain is a challenging, time consuming, and error prone process. Models with multiple independent inputs can have vastly different data dimensionalities that need to be reconciled. Combine all this together, and even simply trying to visualize the entire network can be a challenge.

The design canvas tackles these challenges with ease. New state of the art architectures are constantly added as discrete components that can be instantly applied to any problem domain.

Built-in Components

The platform features a versatile design canvas that empowers users to construct complex deep neural networks through an intuitive drag-and-drop interface. This modular block design allows users to build networks by adding and configuring various components.

The platform offers a range of built-in components out of the box, including input, output, global properties (like learning rate), cnn1d, cnn2d, lstm, gru, downsample_1d, downsample_2d, dense_towers, transformer, and slow_fast. These composable components enable users to design tailored neural networks efficiently, facilitating experimentation and optimization.

The design canvas thus provides a powerful, user-friendly environment for creating sophisticated models, leveraging the flexibility of modular design to streamline the development process.

Custom Components

While the platform provides a comprehensive suite of components at install time, users can extend functionality by writing their own custom components. This modular approach empowers non-programmers to design neural networks using intuitive drag-and-drop tools, while programmers can focus on developing specific, reusable components. This collaboration enhances productivity and innovation, allowing for infinite extensibility beyond the pre-installed offerings.

Each visual component on the canvas corresponds to an actual Python script on disk, where the script's author defines input parameters, descriptions, and UI hints within the docstring. When the canvas is compiled into a TensorFlow model, these scripts dynamically generate the necessary TensorFlow code.