Skip to main content
Version: v2512

Analysis and Refactoring of Conversion Results for Complex Models

Overview​

In the tutorial using ResNet50, we learned how to accelerate inference by converting to a TensorRT engine. However, for complex models there are cases where the entire model cannot be converted by the inference engine. In such cases, AcuiRT outputs analysis information about the conversion failure. By using this analysis information to fix the cause of the conversion failure, it becomes possible to further accelerate the inference processing.

In this tutorial, we will learn how to use AcuiRT's features to identify, fix, and re-convert transformation failures for the 2D object detection model DETR, which combines Transformers and CNNs.

Execution Environment​

  • This document has been run in the environment below and its proper operation has been confirmed.
  • Python: 3.12
  • PyTorch: 2.5.0
  • TensorRT: 10.14.1.48.post1
  • GPU used: NVIDIA RTX 4060Ti

Table of Contents​

  1. Environment Setup
  2. Executing model conversion
  3. Detailed analysis method of conversion failures
  4. Improvement through Refactoring