Ryzen Inference pipelines

The pipelines are a great and easy way to use models for inference. These pipelines are objects that abstract most of the complex code from the library, offering a simple API dedicated to several tasks.

Currently the supported tasks are:

image-classification
object-detection (Supported model types - yolox, yolov3, yolov5, yolov8)

The pipeline abstraction

The pipeline abstraction is a wrapper around all the available pipelines for specific tasks. The pipeline() function automatically loads a default model and tokenizer/feature-extractor capable of performing inference for your task.

Start by creating a pipeline by specifying an inference task:

>>> from optimum.amd.ryzenai import pipeline

>>> detector = pipeline("object-detection")

Pass your input text/image to the ~pipelines.pipeline function:

>>> import requests
>>> from PIL import Image

>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)

>>> detector(image)
[   {   'box': {'xmax': 325, 'xmin': 2, 'ymax': 465, 'ymin': 50},
        'label': 15.0,
        'score': 0.7081549763679504},
    {   'box': {'xmax': 630, 'xmin': 347, 'ymax': 381, 'ymin': 16},
        'label': 15.0,
        'score': 0.6494212746620178},
    {   'box': {'xmax': 369, 'xmin': 335, 'ymax': 187, 'ymin': 76},
        'label': 65.0,
        'score': 0.6064183115959167},
    {   'box': {'xmax': 645, 'xmin': 2, 'ymax': 475, 'ymin': 4},
        'label': 57.0,
        'score': 0.599224865436554},
    {   'box': {'xmax': 174, 'xmin': 40, 'ymax': 116, 'ymin': 73},
        'label': 65.0,
        'score': 0.408765971660614}]

Using Ryzen models

The pipeline() function can load Ryzen AI supported models from the Hugging Face Hub

Once you have picked an appropriate model, you can create the pipeline() by specifying the model repository:

>>> import requests
>>> from PIL import Image

>>> from optimum.amd.ryzenai import pipeline

>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)

>>> # Hugging Face hub model-id with the quantized ONNX model
>>> model_id = "mohitsha/timm-resnet18-onnx-quantized-ryzen"

>>> pipe = pipeline("image-classification", model=model_id)
>>> print(pipe(image))

It is also possible to load with the RyzenModelForXXX class. For example, here is how you can load the ~ryzenai.RyzenModelForImageClassification class for image classification:

>>> import requests
>>> from PIL import Image

>>> from optimum.amd.ryzenai import RyzenAIModelForImageClassification
>>> from optimum.amd.ryzenai import pipeline


>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)

>>> # Hugging Face hub model-id with the quantized ONNX model
>>> model_id = "mohitsha/timm-resnet18-onnx-quantized-ryzen"
>>> model = RyzenAIModelForImageClassification.from_pretrained(model_id)

>>> pipe = pipeline("image-classification", model=model)
>>> print(pipe(image))

Note for a few models, model_type and/or image_preprocessor has to be provided in addition to the model_id for inference. For example, here is how you can run inference using yolox:

>>> import requests
>>> from PIL import Image

>>> from optimum.amd.ryzenai import pipeline

>>> img = ".\\image.jpg"
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> img = ".\\image2.jpg"

>>> image = Image.open(img)

>>> model_id = "amd/yolox-s"

>>> detector = pipeline("object-detection", model=model_id, model_type="yolox")
>>> detector = pipe(image)

< > Update on GitHub

Optimum

Ryzen Inference pipelines

The pipeline abstraction

Using Ryzen models