Using & Mixing Hugging Face Models with Gradio 2.0
Gradio 2.0 allows developers to create graphical user interfaces (GUIs) for Hugging Face models using a single line of code. This integration simplifies the process of demoing models to non-programmers and interdisciplinary teams to identify biases and failure points.
Single-Line Model Deployment
Gradio 2.0 enables the immediate loading of almost any model from the Hugging Face Model Hub—which contains over 10,000 models for natural language processing, image classification, and audio processing—directly into a GUI.
Developers have two primary options for running these models:
- Hosted Inference API: By default, Gradio uses Hugging Face's hosted Inference API, which can be accessed via a public key or without an API key.
- Local Execution: By installing the
transformerslibrary (pip install transformers), users can run model computations locally on their own hardware.
Customization and Interface Control
Users can customize their model demos by overriding the default parameters of the Gradio Interface class. This allows for more granular control over how the model is presented and how it interacts with the input data.
Composing and Mixing Models
Gradio 2.0 treats models as modular components that can be composed to create sophisticated applications. This is achieved through two primary configuration patterns:
Parallel Composition
Models can be loaded in parallel to facilitate direct comparison. For example, a developer can load four different text generation models simultaneously to determine which model performs best for a specific use case.
Serial Composition
Models can be placed in series to build complex pipelines. This enables the creation of multi-step applications, such as a system that first translates a Finnish news article and then summarizes the translated text, achievable in as few as three lines of code.
Developers can further combine these patterns by mixing multiple models in series compared against each other in parallel.