torch-export

What problem does it solve?

This Skill addresses the challenge of converting trained PyTorch models into a deployable format that can run efficiently across various platforms and runtimes, abstracting away the complexities of model serialization and optimization.

Core Features & Use Cases

• Symbolic Tracing: Captures PyTorch nn.Module into an ExportedProgram using symbolic tracing.
• Dynamic Shape Support: Allows for models that handle variable input dimensions.
• Control Flow Handling: Manages static, shape-dependent, and data-dependent control flow within the exported graph.
• Debugging Tools: Provides mechanisms like draft_export and verbose logging for troubleshooting export failures.
• Use Case: Deploying a custom PyTorch image classification model to a mobile device or an edge server by exporting it to a format compatible with inference engines like ONNX or TensorRT.

Quick Start

Export the provided PyTorch model using default settings.