YOLO26 Model Export to TorchScript for Quick Deployment

Deploying computer vision models in C++ environments without Python requires a portable serialized representation. TorchScript provides that compatibility for legacy LibTorch applications.

Export to TorchScript to serialize your Ultralytics YOLO26 models for cross-platform compatibility and streamlined deployment. This guide shows how to export your YOLO26 models to the TorchScript format so you can use them across a wider range of applications.

Why should you export to TorchScript?#

Developed by the creators of PyTorch, TorchScript is a tool for optimizing and deploying PyTorch models across a variety of platforms. Exporting YOLO26 models to TorchScript helps move them from research to real-world applications. TorchScript, part of the PyTorch framework, helps make this transition smoother by allowing PyTorch models to be used in environments that don't support Python.

The process involves two techniques: tracing and scripting. Tracing records operations during model execution, while scripting allows for the definition of models using a subset of Python. These techniques allow models like YOLO26 to run outside their usual Python environment.

TorchScript models can also be optimized through techniques such as operator fusion and refinements in memory usage, ensuring efficient execution. Another advantage of exporting to TorchScript is its potential to accelerate model execution across various hardware platforms. It creates a standalone, production-ready representation of your PyTorch model that can be integrated into C++ environments.

Key Features of TorchScript Models#

TorchScript, a key part of the PyTorch ecosystem, provides features for optimizing and deploying deep learning models.

Here are the key features that make TorchScript a valuable tool for developers:

• Static Graph Execution: TorchScript uses a static graph representation of the model's computation, which is different from PyTorch's dynamic graph execution. In static graph execution, the computational graph is defined and compiled once before the actual execution, resulting in improved performance during inference.

• Model Serialization: TorchScript allows you to serialize PyTorch models into a platform-independent format. Serialized models can be loaded without requiring the original Python code, enabling deployment in different runtime environments.

• JIT Compilation: TorchScript uses Just-In-Time (JIT) compilation to convert PyTorch models into an optimized intermediate representation. JIT compiles the model's computational graph, enabling efficient execution on target devices.

• Gradual Conversion: TorchScript provides a gradual conversion approach, allowing you to incrementally convert parts of your PyTorch model into TorchScript. This flexibility is particularly useful when dealing with complex models or when you want to optimize specific portions of the code.

Deployment Options in TorchScript#

Before looking at the code for exporting YOLO26 models to the TorchScript format, here is where TorchScript models are typically used.

TorchScript offers various deployment options for machine learning models, such as:

• C++ API: The most common use case for TorchScript is its LibTorch C++ API, which allows you to load and execute optimized TorchScript models directly within C++ applications. This is ideal for production environments where Python may not be suitable or available. The C++ API offers low-overhead and efficient execution of TorchScript models, maximizing performance potential.

• Mobile Deployment: For low-latency, offline inference and data privacy on mobile devices, use ExecuTorch, PyTorch's replacement for TorchScript Mobile.

• Cloud Deployment: TorchScript models can be deployed to cloud-based servers using solutions like NVIDIA Triton Inference Server. It provides features like model versioning, batching, and metrics monitoring for scalable deployment in production environments. Cloud deployment with TorchScript can make your models accessible via APIs or other web services.

Supported Tasks#

TorchScript export supports all seven Ultralytics tasks. Semantic segmentation and depth estimation are available only with YOLO26, the only family that ships those heads.

Export to TorchScript: Converting Your YOLO26 Model#

Exporting YOLO26 models to TorchScript makes it easier to run them outside Python, for example in C++ applications built on LibTorch.

Installation#

To install the required package, run:

# Install the required package for YOLO26
pip install ultralytics

For detailed instructions and best practices related to the installation process, check our Ultralytics Installation guide. While installing the required packages for YOLO26, if you encounter any difficulties, consult our Common Issues guide for solutions and tips.

Usage#

All Ultralytics YOLO26 models are designed to support export out of the box, making it easy to integrate them into your preferred deployment workflow. You can view the full list of supported export formats and configuration options to choose the best setup for your application.

The TorchScript format supports the Export, Predict, and Validate modes. Export your model, then load the exported model to run inference or validate its accuracy.

from ultralytics import YOLO

# Load a YOLO26 model
model = YOLO("yolo26n.pt")

# Export the model to TorchScript format
model.export(format="torchscript")  # creates 'yolo26n.torchscript'

from ultralytics import YOLO

# Load the exported TorchScript model
model = YOLO("yolo26n.torchscript")

# Run inference
results = model("https://ultralytics.com/images/bus.jpg")

from ultralytics import YOLO

# Load the exported TorchScript model
model = YOLO("yolo26n.torchscript")

# Validate accuracy on the COCO8 dataset
metrics = model.val(data="coco8.yaml")

Export Arguments#

For more details about the export process, visit the Ultralytics documentation page on exporting.

Deploying Exported YOLO26 TorchScript Models#

After successfully exporting your Ultralytics YOLO26 models to TorchScript format, you can now deploy them. The primary and recommended first step for running a TorchScript model is to use the YOLO("model.torchscript") method, as outlined in the previous usage code snippet. For in-depth instructions on deploying your TorchScript models in other settings, take a look at the following resources:

• Explore Mobile Deployment: Use ExecuTorch's separate torch.export() → .pte pipeline for current PyTorch mobile deployment.

• Master Server-Side Deployment: Learn how to serve TorchScript models with the PyTorch backend of NVIDIA Triton Inference Server for scalable, efficient model serving.

• Implement C++ Deployment: Dive into the Tutorial on Loading a TorchScript Model in C++, facilitating the integration of your TorchScript models into C++ applications for enhanced performance and versatility.

Summary#

This guide covered exporting Ultralytics YOLO26 models to the TorchScript format. By following the provided instructions, you can optimize YOLO26 models for performance and gain the flexibility to deploy them across various platforms and environments.

For further details on usage, visit TorchScript's official documentation.

FAQ#