Make predictions with ONNX on computer vision models from AutoML (v1)

  • Article

APPLIES TO: Python SDK azureml v1

Important

Some of the Azure CLI commands in this article use the azure-cli-ml, or v1, extension for Azure Machine Learning. Support for the v1 extension will end on September 30, 2025. You're able to install and use the v1 extension until that date.

We recommend that you transition to the ml, or v2, extension before September 30, 2025. For more information on the v2 extension, see Azure Machine Learning CLI extension and Python SDK v2.

In this article, you learn how to use Open Neural Network Exchange (ONNX) to make predictions on computer vision models generated from automated machine learning (AutoML) in Azure Machine Learning.

To use ONNX for predictions, you need to:

  1. Download ONNX model files from an AutoML training run.
  2. Understand the inputs and outputs of an ONNX model.
  3. Preprocess your data so it's in the required format for input images.
  4. Perform inference with ONNX Runtime for Python.
  5. Visualize predictions for object detection and instance segmentation tasks.

ONNX is an open standard for machine learning and deep learning models. It enables model import and export (interoperability) across the popular AI frameworks. For more details, explore the ONNX GitHub project.

ONNX Runtime is an open-source project that supports cross-platform inference. ONNX Runtime provides APIs across programming languages (including Python, C++, C#, C, Java, and JavaScript). You can use these APIs to perform inference on input images. After you have the model that has been exported to ONNX format, you can use these APIs on any programming language that your project needs.

In this guide, you'll learn how to use Python APIs for ONNX Runtime to make predictions on images for popular vision tasks. You can use these ONNX exported models across languages.

Prerequisites

Download ONNX model files

You can download ONNX model files from AutoML runs by using the Azure Machine Learning studio UI or the Azure Machine Learning Python SDK. We recommend downloading via the SDK with the experiment name and parent run ID.

Azure Machine Learning studio

On Azure Machine Learning studio, go to your experiment by using the hyperlink to the experiment generated in the training notebook, or by selecting the experiment name on the Experiments tab under Assets. Then select the best child run.

Within the best child run, go to Outputs+logs > train_artifacts. Use the Download button to manually download the following files:

  • labels.json: File that contains all the classes or labels in the training dataset.
  • model.onnx: Model in ONNX format.

Screenshot that shows selections for downloading ONNX model files.

Save the downloaded model files in a directory. The example in this article uses the ./automl_models directory.

Azure Machine Learning Python SDK

With the SDK, you can select the best child run (by primary metric) with the experiment name and parent run ID. Then, you can download the labels.json and model.onnx files.

The following code returns the best child run based on the relevant primary metric.

from azureml.train.automl.run import AutoMLRun

# Select the best child run
run_id = '' # Specify the run ID
automl_image_run = AutoMLRun(experiment=experiment, run_id=run_id)
best_child_run = automl_image_run.get_best_child()

Download the labels.json file, which contains all the classes and labels in the training dataset.

labels_file = 'automl_models/labels.json'
best_child_run.download_file(name='train_artifacts/labels.json', output_file_path=labels_file)

Download the model.onnx file.

onnx_model_path = 'automl_models/model.onnx'
best_child_run.download_file(name='train_artifacts/model.onnx', output_file_path=onnx_model_path)

Model generation for batch scoring

By default, AutoML for Images supports batch scoring for classification. But object detection and instance segmentation models don't support batch inferencing. In case of batch inference for object detection and instance segmentation, use the following procedure to generate an ONNX model for the required batch size. Models generated for a specific batch size don't work for other batch sizes.

from azureml.core.script_run_config import ScriptRunConfig
from azureml.train.automl.run import AutoMLRun
from azureml.core.workspace import Workspace
from azureml.core import Experiment

# specify experiment name
experiment_name = ''
# specify workspace parameters
subscription_id = ''
resource_group = ''
workspace_name = ''
# load the workspace and compute target
ws = ''
compute_target = ''
experiment = Experiment(ws, name=experiment_name)

# specify the run id of the automl run
run_id = ''
automl_image_run = AutoMLRun(experiment=experiment, run_id=run_id)
best_child_run = automl_image_run.get_best_child()

Use the following model specific arguments to submit the script. For more details on arguments, refer to model specific hyperparameters and for supported object detection model names refer to the supported model algorithm section.

To get the argument values needed to create the batch scoring model, refer to the scoring scripts generated under the outputs folder of the AutoML training runs. Use the hyperparameter values available in the model settings variable inside the scoring file for the best child run.

For multi-class image classification, the generated ONNX model for the best child-run supports batch scoring by default. Therefore, no model specific arguments are needed for this task type and you can skip to the Load the labels and ONNX model files section.

Download and keep the ONNX_batch_model_generator_automl_for_images.py file in the current directory and submit the script. Use ScriptRunConfig to submit the script ONNX_batch_model_generator_automl_for_images.py available in the azureml-examples GitHub repository, to generate an ONNX model of a specific batch size. In the following code, the trained model environment is used to submit this script to generate and save the ONNX model to the outputs directory.

script_run_config = ScriptRunConfig(source_directory='.',
                                    script='ONNX_batch_model_generator_automl_for_images.py',
                                    arguments=arguments,
                                    compute_target=compute_target,
                                    environment=best_child_run.get_environment())

remote_run = experiment.submit(script_run_config)
remote_run.wait_for_completion(wait_post_processing=True)

Once the batch model is generated, either download it from Outputs+logs > outputs manually, or use the following method:

batch_size= 8  # use the batch size used to generate the model
onnx_model_path = 'automl_models/model.onnx'  # local path to save the model
remote_run.download_file(name='outputs/model_'+str(batch_size)+'.onnx', output_file_path=onnx_model_path)

After the model downloading step, you use the ONNX Runtime Python package to perform inferencing by using the model.onnx file. For demonstration purposes, this article uses the datasets from How to prepare image datasets for each vision task.

We've trained the models for all vision tasks with their respective datasets to demonstrate ONNX model inference.

Load the labels and ONNX model files

The following code snippet loads labels.json, where class names are ordered. That is, if the ONNX model predicts a label ID as 2, then it corresponds to the label name given at the third index in the labels.json file.

import json
import onnxruntime

labels_file = "automl_models/labels.json"
with open(labels_file) as f:
    classes = json.load(f)
print(classes)
try:
    session = onnxruntime.InferenceSession(onnx_model_path)
    print("ONNX model loaded...")
except Exception as e: 
    print("Error loading ONNX file: ",str(e))

Get expected input and output details for an ONNX model

When you have the model, it's important to know some model-specific and task-specific details. These details include the number of inputs and number of outputs, expected input shape or format for preprocessing the image, and output shape so you know the model-specific or task-specific outputs.

sess_input = session.get_inputs()
sess_output = session.get_outputs()
print(f"No. of inputs : {len(sess_input)}, No. of outputs : {len(sess_output)}")

for idx, input_ in enumerate(range(len(sess_input))):
    input_name = sess_input[input_].name
    input_shape = sess_input[input_].shape
    input_type = sess_input[input_].type
    print(f"{idx} Input name : { input_name }, Input shape : {input_shape}, \
    Input type  : {input_type}")  

for idx, output in enumerate(range(len(sess_output))):
    output_name = sess_output[output].name
    output_shape = sess_output[output].shape
    output_type = sess_output[output].type
    print(f" {idx} Output name : {output_name}, Output shape : {output_shape}, \
    Output type  : {output_type}")

Expected input and output formats for the ONNX model

Every ONNX model has a predefined set of input and output formats.

This example applies the model trained on the fridgeObjects dataset with 134 images and 4 classes/labels to explain ONNX model inference. For more information on training an image classification task, see the multi-class image classification notebook.

Input format

The input is a preprocessed image.

Input name Input shape Input type Description
input1 (batch_size, num_channels, height, width) ndarray(float) Input is a preprocessed image, with the shape (1, 3, 224, 224) for a batch size of 1, and a height and width of 224. These numbers correspond to the values used for crop_size in the training example.

Output format

The output is an array of logits for all the classes/labels.

Output name Output shape Output type Description
output1 (batch_size, num_classes) ndarray(float) Model returns logits (without softmax). For instance, for batch size 1 and 4 classes, it returns (1, 4).

Preprocessing

Perform the following preprocessing steps for the ONNX model inference:

  1. Convert the image to RGB.
  2. Resize the image to valid_resize_size and valid_resize_size values that correspond to the values used in the transformation of the validation dataset during training. The default value for valid_resize_size is 256.
  3. Center crop the image to height_onnx_crop_size and width_onnx_crop_size. It corresponds to valid_crop_size with the default value of 224.
  4. Change HxWxC to CxHxW.
  5. Convert to float type.
  6. Normalize with ImageNet's mean = [0.485, 0.456, 0.406] and std = [0.229, 0.224, 0.225].

If you chose different values for the hyperparameters valid_resize_size and valid_crop_size during training, then those values should be used.

Get the input shape needed for the ONNX model.

batch, channel, height_onnx_crop_size, width_onnx_crop_size = session.get_inputs()[0].shape
batch, channel, height_onnx_crop_size, width_onnx_crop_size

Without PyTorch

import glob
import numpy as np
from PIL import Image

def preprocess(image, resize_size, crop_size_onnx):
    """Perform pre-processing on raw input image
    
    :param image: raw input image
    :type image: PIL image
    :param resize_size: value to resize the image
    :type image: Int
    :param crop_size_onnx: expected height of an input image in onnx model
    :type crop_size_onnx: Int
    :return: pre-processed image in numpy format
    :rtype: ndarray 1xCxHxW
    """

    image = image.convert('RGB')
    # resize
    image = image.resize((resize_size, resize_size))
    #  center  crop
    left = (resize_size - crop_size_onnx)/2
    top = (resize_size - crop_size_onnx)/2
    right = (resize_size + crop_size_onnx)/2
    bottom = (resize_size + crop_size_onnx)/2
    image = image.crop((left, top, right, bottom))

    np_image = np.array(image)
    # HWC -> CHW
    np_image = np_image.transpose(2, 0, 1) # CxHxW
    # normalize the image
    mean_vec = np.array([0.485, 0.456, 0.406])
    std_vec = np.array([0.229, 0.224, 0.225])
    norm_img_data = np.zeros(np_image.shape).astype('float32')
    for i in range(np_image.shape[0]):
        norm_img_data[i,:,:] = (np_image[i,:,:]/255 - mean_vec[i])/std_vec[i]
             
    np_image = np.expand_dims(norm_img_data, axis=0) # 1xCxHxW
    return np_image

# following code loads only batch_size number of images for demonstrating ONNX inference
# make sure that the data directory has at least batch_size number of images

test_images_path = "automl_models_multi_cls/test_images_dir/*" # replace with path to images
# Select batch size needed
batch_size = 8
# you can modify resize_size based on your trained model
resize_size = 256
# height and width will be the same for classification
crop_size_onnx = height_onnx_crop_size 

image_files = glob.glob(test_images_path)
img_processed_list = []
for i in range(batch_size):
    img = Image.open(image_files[i])
    img_processed_list.append(preprocess(img, resize_size, crop_size_onnx))
    
if len(img_processed_list) > 1:
    img_data = np.concatenate(img_processed_list)
elif len(img_processed_list) == 1:
    img_data = img_processed_list[0]
else:
    img_data = None

assert batch_size == img_data.shape[0]

With PyTorch

import glob
import torch
import numpy as np
from PIL import Image
from torchvision import transforms

def _make_3d_tensor(x) -> torch.Tensor:
    """This function is for images that have less channels.

    :param x: input tensor
    :type x: torch.Tensor
    :return: return a tensor with the correct number of channels
    :rtype: torch.Tensor
    """
    return x if x.shape[0] == 3 else x.expand((3, x.shape[1], x.shape[2]))

def preprocess(image, resize_size, crop_size_onnx):
    transform = transforms.Compose([
        transforms.Resize(resize_size),
        transforms.CenterCrop(crop_size_onnx),
        transforms.ToTensor(),
        transforms.Lambda(_make_3d_tensor),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
    
    img_data = transform(image)
    img_data = img_data.numpy()
    img_data = np.expand_dims(img_data, axis=0)
    return img_data

# following code loads only batch_size number of images for demonstrating ONNX inference
# make sure that the data directory has at least batch_size number of images

test_images_path = "automl_models_multi_cls/test_images_dir/*" # replace with path to images
# Select batch size needed
batch_size = 8
# you can modify resize_size based on your trained model
resize_size = 256
# height and width will be the same for classification
crop_size_onnx = height_onnx_crop_size 

image_files = glob.glob(test_images_path)
img_processed_list = []
for i in range(batch_size):
    img = Image.open(image_files[i])
    img_processed_list.append(preprocess(img, resize_size, crop_size_onnx))
    
if len(img_processed_list) > 1:
    img_data = np.concatenate(img_processed_list)
elif len(img_processed_list) == 1:
    img_data = img_processed_list[0]
else:
    img_data = None

assert batch_size == img_data.shape[0]

Inference with ONNX Runtime

Inferencing with ONNX Runtime differs for each computer vision task.

def get_predictions_from_ONNX(onnx_session, img_data):
    """Perform predictions with ONNX runtime
    
    :param onnx_session: onnx model session
    :type onnx_session: class InferenceSession
    :param img_data: pre-processed numpy image
    :type img_data: ndarray with shape 1xCxHxW
    :return: scores with shapes
            (1, No. of classes in training dataset) 
    :rtype: numpy array
    """

    sess_input = onnx_session.get_inputs()
    sess_output = onnx_session.get_outputs()
    print(f"No. of inputs : {len(sess_input)}, No. of outputs : {len(sess_output)}")    
    # predict with ONNX Runtime
    output_names = [ output.name for output in sess_output]
    scores = onnx_session.run(output_names=output_names,\
                                               input_feed={sess_input[0].name: img_data})
    
    return scores[0]

scores = get_predictions_from_ONNX(session, img_data)

Postprocessing

Apply softmax() over predicted values to get classification confidence scores (probabilities) for each class. Then the prediction will be the class with the highest probability.

Without PyTorch

def softmax(x):
    e_x = np.exp(x - np.max(x, axis=1, keepdims=True))
    return e_x / np.sum(e_x, axis=1, keepdims=True)

conf_scores = softmax(scores)
class_preds = np.argmax(conf_scores, axis=1)
print("predicted classes:", ([(class_idx, classes[class_idx]) for class_idx in class_preds]))

With PyTorch

conf_scores = torch.nn.functional.softmax(torch.from_numpy(scores), dim=1)
class_preds = torch.argmax(conf_scores, dim=1)
print("predicted classes:", ([(class_idx.item(), classes[class_idx]) for class_idx in class_preds]))

Visualize predictions

Visualize an input image with labels

import matplotlib.image as mpimg
import matplotlib.pyplot as plt
%matplotlib inline

sample_image_index = 0 # change this for an image of interest from image_files list
IMAGE_SIZE = (18, 12)
plt.figure(figsize=IMAGE_SIZE)
img_np = mpimg.imread(image_files[sample_image_index])

img = Image.fromarray(img_np.astype('uint8'), 'RGB')
x, y = img.size

fig,ax = plt.subplots(1, figsize=(15, 15))
# Display the image
ax.imshow(img_np)

label = class_preds[sample_image_index]
if torch.is_tensor(label):
    label = label.item()
    
conf_score = conf_scores[sample_image_index]
if torch.is_tensor(conf_score):
    conf_score = np.max(conf_score.tolist())
else:
    conf_score = np.max(conf_score)

display_text = '{} ({})'.format(label, round(conf_score, 3))
print(display_text)

color = 'red'
plt.text(30, 30, display_text, color=color, fontsize=30)

plt.show()

Next steps