Getting Started

Overview

VGSLify makes it incredibly simple to define, build, and train deep learning models using the Variable-size Graph Specification Language (VGSL). VGSL strings serve as compact representations of neural network architectures, allowing you to build models in a single line.

VGSLify abstracts the complexity of backend-specific syntax, enabling seamless switching between TensorFlow and, in future releases, PyTorch. This flexibility allows you to focus on model architecture and training without worrying about framework-specific implementations.

What is a VGSL Specification?

A VGSL specification string concisely defines a neural network’s architecture. The string encodes all layers, including input, convolutional layers, pooling, fully connected layers, and more. Each part of the string corresponds to a different component of the model.

For example, the following VGSL string defines a simple convolutional neural network:

None,28,28,1 Cr3,3,32 Mp2,2,2,2 Rc Fr64 D20 Fs10

This string represents a model with an input layer, a convolutional layer, a max pooling layer, a reshape layer, a dense (fully connected) layer, dropout, and an output layer. The model’s structure is encoded entirely within this single line.

Key functionality of VGSLify includes:

  • Building models with a single line: You can define complex architectures with a VGSL string, reducing the need for verbose code.

  • Switching between TensorFlow and PyTorch: VGSLify supports both TensorFlow and (planned) PyTorch, allowing you to easily switch between backends.

Simple Example: Building a Model

Let’s walk through building a simple deep learning model using VGSLify.

  1. Import the VGSLModelGenerator:

    The VGSLModelGenerator class is the core component for building models from VGSL strings. Begin by importing it:

    from vgslify.generator import VGSLModelGenerator

  2. Define the VGSL Specification String:

    The VGSL spec string encodes the structure of the model. In this example, we will define a simple convolutional neural network suitable for handling MNIST digit images (28x28 grayscale):

    vgsl_spec = "None,28,28,1 Cr3,3,32 Mp2,2,2,2 Rc2 Fr64 D20 Fs10"

  3. Build and View the Model:

    Initialize the VGSLModelGenerator and use it to build the model based on the VGSL spec string:

    vgsl_gn = VGSLModelGenerator(backend="tensorflow")  # Set backend to TensorFlow
model = vgsl_gn.generate_model(vgsl_spec)
model.summary()  # View the model architecture

    This will generate the model and display a summary of its architecture, including all layers defined by the VGSL spec string.

Explanation of Layers

Let’s break down the layers defined by the VGSL specification string in our example:

  • Input Layer: None,28,28,1 - This defines the input shape of the model, which corresponds to grayscale images of size 28x28 pixels. The first dimension (None) allows for a variable batch size.

  • Convolutional Layer: Cr3,3,32 - This adds a 2D convolutional layer with a 3x3 kernel and 32 output filters, using ReLU activation (r for ReLU).

  • MaxPooling Layer: Mp2,2,2,2 - This reduces the spatial dimensions by applying 2x2 max pooling with a stride of 2x2, which downsamples the input by taking the maximum value over each 2x2 window.

  • Reshape Layer: Rc2 - Reshapes the output from the previous layer, collapsing the spatial dimensions into a single vector suitable for fully connected layers.

  • Fully Connected Layer: Fc64 - Adds a fully connected layer (dense layer) with 64 units.

  • Dropout Layer: D20 - Applies dropout with a 20% rate to prevent overfitting by randomly setting a portion of the inputs to zero during training.

  • Output Layer: Fs10 - Represents the output layer with 10 units (for 10 classes, such as the digits in MNIST) using softmax activation.

This VGSL string provides a concise, human-readable format for specifying complex model architectures. VGSLify automatically translates this specification into a deep learning model that can be trained using TensorFlow.

Next Steps

Once you’ve built and explored a basic model, you can dive deeper into VGSLify’s capabilities. Follow the [tutorials](tutorials.html) to explore more advanced use cases such as:

  • Using different VGSL spec strings to define custom architectures.

  • Switching between TensorFlow and PyTorch backends (PyTorch support coming soon).

  • Integrating VGSLify models into larger deep learning workflows.

Check out the API reference for detailed information on all available classes, methods, and utilities in VGSLify.