image_classification

Ryzen™ AI Image Classification

Introduction

This tutorial uses a pre-trained ResNet50 model to demonstrate the process of preparing, quantizing, and deploying a BF16 model using Ryzen AI Software. The quantized model is deployed on NPU using Python API and evaluated the accuracy of the quantized model on ImageNet dataset.

Overview

The following steps outline how to deploy the BF16 model on an NPU:

• Download the ResNet50 model from torchvision library and save it as ONNX (Opset 17) model.
• Quantize the model to BF16 using the AMD Quark Quantizer.
• Compile and run the model on an NPU using ONNX Runtime with the Vitis AI Execution Provider.

Setup Model and Dataset

Setup Instructions

Activate the conda environment created by the RyzenAI installer

cd <RyzenAI-SW>\CNN-examples\image_classification
conda create --name image_classification --clone ryzen-ai-<version>
conda activate image_classification
python -m pip install -r requirements.txt

ResNet50 Model

Download the ResNet50 model from torchvision models and export it to ONNX model

cd models
python download_ResNet.py

ImageNet Dataset

If you already have an ImageNet datasets, you can directly use your dataset path.

To prepare the test data, please check the download section of the main website: https://huggingface.co/datasets/imagenet-1k/tree/main/data. You need to register and download val_images.tar.gz.

Then, create the validation dataset and calibration dataset:

mkdir val_data && tar -xzf val_images.tar.gz -C val_data
python prepare_data.py val_data calib_data

Model Evaluation

Evaluate the accuracy of the model using ImageNet dataset on CPU/NPU

python image_classification.py --model_input models\resnet50_bf16.onnx --calib_data calib_data --device cpu/npu --evaluate

Summary of BF16 model accuracy:

ResNet50	Top-1 Accuracy	Top-5 Accuracy
Float 32	0.800	0.961
FP32 to BF16 (NPU)	0.804	0.962