GuidesRecipesAPI ReferenceChangelogDiscussions
Log In

Model Inference & Explainability with Python

In this tutorial (also available as a Jupyter Notebook or in a Git Repo), we will demonstrate how to use Modzy's Python SDK to submit an inference, retrieve the results, and visualize an explaination of the prediction.

:computer: Environment Set Up

Create a virtual environment (venv, conda, or other preferred virtual environment) with Python 3.6 or newer.

Pip install the following packages in your environment.

  • modzy-sdk>=0.11.3
  • opencv-python==
  • numpy==1.21.0
  • matplotlib==3.4.2

And install Jupyter Notebooks in your preferred environment using the appropriate install instructions.

:arrow-down: Import Modzy SDK and Initialize Client

Insert your instance URL and personal API Key to establish connection to the Modzy API Client

# Import Libraries
import cv2
import numpy as np
from pprint import pprint
from matplotlib import pyplot as plt
from modzy import ApiClient, error
from utils import display_rle_mask

Initialize Modzy API Client

# the url we will use for authentication
Note: To use this example replace MODZY_URL with the url of your instance of Modzy
# the api key we will be using for authentication -- make sure to paste in your personal API access key below
API_KEY = "<your.api.key>"

if API_URL == "MODZY_URL/api":
    raise Exception("Change the API_URL variable to your instance URL")
if API_KEY == "<your.api.key>":
    raise Exception("Insert your API Key")
# setup our API Client
client = ApiClient(base_url=API_URL, api_key=API_KEY)

:mag: Discover Available Models

Next, we'll search for a model that performs image-based geolocation, which means that it will try to predict where in the world a picture was taken. It tends to work best with images of cityscapes or other scenery.

# Query model by name
auto_model_info = client.models.get_by_name("Image-Based Geolocation")
{'author': 'Modzy',
 'description': 'This model returns the predicted geographical location of a '
                'given image at both a class and regional level.',
 'expirationDate': '2021-10-20T00:00:00.000+00:00',
 'features': [{'description': 'This model has a built-in explainability '
                              'feature. Click '
                              '[here]( to '
                              'read more about model explainability.',
               'identifier': 'built-in-explainability',
               'name': 'Explainable'}],
 'images': [{'caption': 'Image-Based Geolocation',
             'relationType': 'background',
             'url': '/modzy-images/aevbu1h3yw/image_background.png'},
            {'caption': 'Image-Based Geolocation',
             'relationType': 'card',
             'url': '/modzy-images/aevbu1h3yw/image_card.png'},
            {'caption': 'Image-Based Geolocation',
             'relationType': 'thumbnail',
             'url': '/modzy-images/aevbu1h3yw/image_thumbnail.png'},
            {'caption': 'Modzy',
             'relationType': 'logo',
             'url': '/modzy-images/companies/modzy/company-image.jpg'}],
 'isActive': True,
 'isCommercial': True,
 'isExpired': False,
 'isRecommended': False,
 'latestActiveVersion': '1.0.1',
 'latestVersion': '1.0.1',
 'modelId': 'aevbu1h3yw',
 'name': 'Image-Based Geolocation',
 'permalink': 'aevbu1h3yw-modzy-image-based-geolocation',
 'snapshotImages': [],
 'tags': [{'dataType': 'Subject',
           'identifier': 'geography',
           'isCategorical': True,
           'name': 'Geography'},
          {'dataType': 'Subject',
           'identifier': 'infrastructure_and_buildings',
           'isCategorical': True,
           'name': 'Infrastructure and Buildings'},
          {'dataType': 'Input Type',
           'identifier': 'image',
           'isCategorical': True,
           'name': 'Image'},
          {'dataType': 'Task',
           'identifier': 'label_or_classify',
           'isCategorical': True,
           'name': 'Label or Classify'}],
 'versions': ['1.0.1', '0.0.1'],
 'visibility': ApiObject({
  "scope": "ALL"
# Define Variables for Inference
MODEL_ID = auto_model_info["modelId"]
MODEL_VERSION = auto_model_info["latestVersion"]
INPUT_FILENAME = list(client.models.get_version_input_sample(MODEL_ID, MODEL_VERSION)["input"]["sources"]["0001"].keys())[0]
sources = {INPUT_FILENAME: "./nyc.jpg"}

:runner: Submit Inference to Model

Helper Function

Below is a helper function we will use to submit inference jobs to the Modzy platform and return the model output using the submit_files method. For additional job submission methods, visit the jobs page within the Python SDK GitHub page.

def get_model_output(model_identifier, model_version, data_sources, explain=False):
        model_identifier: model identifier (string)
        model_version: model version (string)
        data_sources: dictionary with the appropriate filename --> local file key-value pairs
        explain: boolean variable, defaults to False. If true, model will return explainable result
    job =, model_version, data_sources, explain)
    result = client.results.block_until_complete(job, timeout=None)    
    model_output = result.get_first_outputs()['results.json']
    return model_output
# Visualize image to use for inference

New York City skyline

Image source:

model_results = get_model_output(MODEL_ID, MODEL_VERSION, sources, explain=True)
non_explainable_results = model_results["data"]["result"]
{'classPredictions': [{'class': 'New York, North America',
                       'score': 1.0},
                      {'class': 'Chicago, North America',
                       'score': 0.0},
                      {'class': 'Hong Kong, Asia',
                       'score': 0.0},
                      {'class': 'Toronto, North America',
                       'score': 0.0},
                      {'class': 'Saint Petersburg, Eastern-Europe',
                       'score': 0.0}]}

:camera: Plot Explainability Mask on Image

Helper Function

Below is a helper function we will need to overlay the explainable run-length encoding mask on top of the original image.

def display_rle_mask(img,rle_counts):
    mask = np.zeros(shape=img.shape[:-1])
    for segment in rle_counts:
        it = np.nditer(mask,order='F',flags=['multi_index'])
        current_value = 0.0
        counts_index = 0
        it_current_index = 0
        for x in it:
            if it_current_index==segment[counts_index]:
                it_current_index = 0
                counts_index += 1
                current_value = 1.0 - current_value
            if current_value:
                mask[it.multi_index] = current_value
            it_current_index += 1
    return mask
# Plot RLE Explainability mask over image. This model uses LIME to produce an explainable output RLE mask
rle_counts = model_results["data"]["explanation"]["maskRLE"]
img = cv2.cvtColor(cv2.imread('./nyc.jpg'),cv2.COLOR_BGR2RGB)
mask = display_rle_mask(img,rle_counts)

New York City skyline with a mask explaining which pixels were used by this model to classify the image

:video-camera: Video Tutorial

Follow along in this video tutorial to learn more.