This guide demonstrates the process of automatically containerizing your ONNX model.


What you will need

  • Dockerhub account
  • Connection to running service (either from a local deployment or via publicly-hosted service)
  • Trained ONNX model that can be loaded into memory or code to train a ONNX model from scratch
  • Python environment

NOTE: To follow along, you can reference the Jupyter notebook example and data files here.

Set Up Environment


We recommend you follow this guide using a Jupyter Notebook. Follow the appropriate install instructions based on your environment.

Create a Python virtual environment and install the python packages required to load and run your model. At a minimum, pip install the following packages:

pip install chassisml modzy-sdk

If you would like to follow this guide directly, pip install the following additional packages:


Load Model into Memory

If you plan to use the Chassis service, you must first load your model into memory. If you have your trained model file saved locally (.pth, .pkl, .h5, .joblib, or other file format), you can load your model from the weights file directly, or alternatively train and use the model object.

import cv2
import pickle
import tempfile
import chassisml
import numpy as np
import getpass
from shutil import rmtree
import json
import onnx
from onnx import backend
from onnx import numpy_helper
import onnxruntime as ort
import matplotlib.pyplot as plt

# load model
model = onnx.load("models/mobilenetv2-7.onnx")

# format sample data and reshape
img = cv2.cvtColor(cv2.imread('data/dog.jpg'), cv2.COLOR_BGR2RGB)
img_show = cv2.resize(img, (224,224))

sample_img = np.reshape(img_show, (1,3,224,224)).astype(np.float32)

# load imagenet labels
labels = pickle.load(open('./data/imagenet_labels.pkl','rb'))

# visualize image

Now, we will create an ONNX runtime inference session and test our model locally.

# create onnx runtime inference session and print top prediction
session = ort.InferenceSession("models/mobilenetv2-7.onnx")
results =, {"input": sample_img})
print("Top Prediction: {}".format(labels[results[0].argmax()]))
>>> "Top Prediction: sunglass"

Define process Function

You can think of this function as your "inference" function that will take input data as raw bytes, process the inputs, make predictions, and return the results. This method is the sole parameter required to create a ChassisModel object.

def process(input_bytes):
    # save model to filepath for inference
    tmp_dir = tempfile.mkdtemp()
    import onnx, "{}/model.onnx".format(tmp_dir))

    # preprocess data
    decoded = cv2.cvtColor(cv2.imdecode(np.frombuffer(input_bytes, np.uint8), -1), cv2.COLOR_BGR2RGB)
    img = cv2.resize(decoded, (224,224))
    img = np.reshape(img, (1,3,224,224)).astype(np.float32)

    # run inference
    session = ort.InferenceSession("{}/model.onnx".format(tmp_dir))
    results =, {"input": img})

    # postprocess
    inference_result = labels[results[0].argmax()]

    # format results
    structured_result = {
        "data": {
            "result": {"classPredictions": [{"class": str(inference_result)}]}

    # remove temp directory
    return structured_result

Create ChassisModel Object and Publish Model

First, connect to a running instance of the Chassis service - either by deploying on your machine or by connecting to the publicly hosted version of the service). Then, you can use the process function you defined to create a ChassisModel object, run a few tests to ensure your model object returns the expected results, and finally publish your model.

chassis_client = chassisml.ChassisClient("http://localhost:5000")
chassis_model = chassis_client.create_model(process_fn=process)

Define sample file from local filepath and run a series of tests.

NOTE: test_env method is not available on publicly-hosted service.

sample_filepath = './data/dog.jpg'
results = chassis_model.test(sample_filepath)

test_env_result = chassis_model.test_env(sample_filepath)

Define your Dockerhub and publish your model.

dockerhub_user = <my.username>
dockerhub_pass = <my.password>

response = chassis_model.publish(
   model_name="ONNX MobileNet Image Classifiction",

job_id = response.get('job_id')
final_status = chassis_client.block_until_complete(job_id)

You have successfully completed the packaging of your ONNX model. In your Dockerhub account, you should see your new container listed in the "Repositories" tab.


Figure 1. Example Chassis-built Container

Congratulations! In just minutes you automatically created a Docker container with just a few lines of code. To deploy your new model container to Modzy, follow one of the following guides:

What’s Next