Forma

Forma: Browser-Native ONNX & TFLite Graph Editor

Forma is an innovative web application designed for loading, visualizing, and analyzing ONNX and TensorFlow Lite (TFLite) neural network models directly in the browser. With Forma, there is no need for Python, server setups, or installations. Users can effortlessly drag-and-drop .onnx or .tflite files onto the canvas and watch as the complete computation graph renders instantly.

Key Features:

• Intuitive Graph Visualization:
  Easily load models with drag-and-drop support, real-time progress feedback, and automatic layout options customized for model size. Large graphs can be managed seamlessly with pan, zoom, and minimap tools.

• Interactive Model Inspection:
  Inspect every component of a model with a detailed Layer Inspector, allowing for attribute editing directly within the interface. Structural changes can be made intuitively, with full edit history tracking.

• Editable ONNX Models:
  ONNX models can be fully edited and exported, while TFLite models are supported in read-only format. Shareable edit links allow for easy collaboration, allowing modified sequences to be verified and shared without uploading model bytes.

• Behavioral Validation:
  Compare original and edited models side-by-side, with identical input sets, to validate behavioral accuracy and performance metrics. This ensures that changes do not compromise model integrity.

• Minimal Repro Extraction:
  Users can extract specific subgraphs as standalone ONNX files from selected nodes, promoting necessary tensors seamlessly to uphold functional integrity.

• Deployment Surgery:
  Rename nodes and change tensor declarations easily, retaining the original relationships within the model, while enabling modifications to improve model performance and adaptability.

Getting Started

To explore the functionalities of Forma, visit the Live Application. Simply drag and drop your model files onto the canvas and begin interacting with the powerful visualization and editing tools at your fingertips.

Forma ensures a user-friendly experience while harnessing robust capabilities for model analysis, making it suitable for researchers, developers, and anyone involved in neural network applications.