Ryzen™ AI Image Classification

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.

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.

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

cd models
python download_ResNet.py

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

The 'model_quantization.py' script uses the BF16 quantization configuration from AMD Quark quantizer. It also uses a subset of ImageNet as calibration dataset for quantization.

quant_config = get_default_config("BF16")
quant_config.extra_options["BF16QDQToCast"] = True
config = Config(global_quant_config=quant_config)
print("The configuration of the quantization is {}".format(config))

For more details about the BF16 quantization refer to AMD Quark BF16 Tutorial

The ONNX model can be quantized to BF16 using the 'model_quantization.py' script, as shown:

python model_quantization.py --model_input models\resnet50.onnx --model_output models\resnet50_bf16.onnx --quantize bf16

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: