Model Inference - Data Inference

In some application fields, the model input data cannot leverage image inference (with built-in hardware image pre-processing). In this case, data inference (without built-in hardware image pre-processing) would be the best solution.

This tutorial shows how to inference none image data (raw data) by Kneron devices without built-in hardware image pre-processing.

In order to infer without hardware image preprocessing, the input data must do the following steps for fitting NPU input format:

  1. Normalization
  2. Quantization (convert to 8-bit fixed point)
  3. Re-layout data to fit NPU data layout format

Note: Please upload NEF model on Kneron device before the following tutorial. See the Load NEF Model for details.

Reference Examples:

Inference without Built-In Hardware Image Pre-Processing

Recommend use cv2 to read the image or capture camera frame.