Neural network architectures are most often conceptually designed and described in visual terms, but are implemented by writing error-prone code. PrototypeML.com is a new (currently in alpha) open neural network development environment that allows you to (a) quickly prototype/design any neural network architecture through intuitive visual drag-and-drop and code editing interfaces (b) build upon a community-driven library of network building blocks, and © export fully readable PyTorch code for training. Try it out at PrototypeML.com.

How PrototypeML Works

PyTorch, with its increasing popularity with the research community, is known for its conceptual simplicity, dynamic graph support, and modular and future-proof design. PrototypeML is tightly coupled with the PyTorch framework, and supports the full range of dynamic functionality and arbitrary code execution available in normal PyTorch models. We accomplish this through several fundamental building blocks.

Representing neural networks as syntax-tree code graphs

First, we take advantage of the PyTorch “deep learning models as regular Python programs” design paradigm in order to represent neural network models in the form of static syntax-tree code graphs, rather than pre-defined neural component graphs.

Mutators

Second, we introduce the concept of a “Mutator”. A mutator component encapsulates segments of normal PyTorch or arbitrary Python code, and defines a standardized format by which data inputs and outputs, and parameters can be expressed.

PyTorch models are expressed using the nn.Module class paradigm, whereby neural network layers with parameters are instantiated in the nn.Module class init() function, and called during the network’s forward pass in the nn.Module class forward() function. Mutators build upon this schema by encapsulating the code necessary to execute both the instantiation of code in the init() function, and the corresponding forward() execution code into a single component that contains everything necessary to run the code regardless of where it will be used.

#machine-learning #pytorch #data-science #neural-networks #artificial-intelligence