Export Models
Model export and transformation
Summary
When the model is trained on a private cloud platform, it needs to be exported to run on the deployment side. The platform supports exporting models for CPU running, as well as models for Nvidia GPU running.
When the CPU model is selected, the exported model file can be directly used for CPU running.
When choosing a GPU model, since the GPU model is GPU configuration-dependent, the transformation needs to be done on the target computer (or a computer with the same GPU power as the target computer).
For example, if you plan to run the GPU model on an industrial PC equipped with an NVIDIA RTX 3060 graphics card, you will need to generate the transformed model on a computer with the same RTX 3060 or the same GPU (for example, the RTX 3080, which has 8.6 processing power as the 3060).
For graphic card computing power values, please refer to NVIDIA official website
Step 1: Export the model
1. On the model Download page, select the model based on the hardware platform on which the model is running. Then click "Export Model" to start exporting.
2. The exported file is the compressed onnx.zip file, and extract it to a folder. The folder should contain the following files:
Step 2: Transform the GPU model
Note: Model transformation is only required when you choose "export the GPU model".
1. Click to download the Model transformation tool.
Once you have downloaded export_invoker.zip, unzip the package.
2. Double-click the export_invoker.exe program, select the extracted model folder, and click "Export".
Exporting the model is expected to take around 5 minutes.
Note: select the unzipped models folder above, not a single file; No more nested levels in folders.
The model transformation tool only supports running on Windows10/11 platform.
When the export is complete, a new file named end2end.engine will be created in the same folder.
This folder is the transformed GPU model, which can be used to run the GPU model (overall complete folder).
For deployment and usage of the model, refer to SDK Deployment
Still have a question?
Each section should be concise, user-friendly, and direct users to additional resources or documentation when necessary.