Running AI Model Inference

This is a walkthrough for running predictions. You'll learn about input data types, understanding the results, and finally processing inputs in batches for efficiency.

Once you have loaded an AI model and obtained a model handle, you can start running inferences. The degirum.model.Model class provides two methods for performing AI inference:

• degirum.model.Model.predict(): Runs prediction on a single data frame.

• degirum.model.Model.predict_batch(): Runs prediction on a batch of frames.

Model.predict()

The predict() method takes a single input data frame and returns an inference result object. You can also call this method by calling degirum.model.Model.__call__. This is an alias for Model(). See Single Frame Inference for more information.

Try not to run predict() on videos, webcam feeds, and streams. Instead, use predict_batch().

# Method Signature: Model.predict()
degirum.model.Model.predict(data)

Example:

import degirum as dg

# Declaring variables
# Set your model, inference host address, model zoo, and token in these variables.
your_model_name = "model-name"
your_host_address = "@cloud" # Can be "@cloud", host:port, or "@local"
your_model_zoo = "degirum/public"
your_token = "<token>"

# Specify the image you will run inference on
your_image = "path/image.jpg"

# Loading a model
model = dg.load_model(
    model_name = your_model_name, 
    inference_host_address = your_host_address, 
    zoo_url = your_model_zoo, 
    token = your_token 
    # optional parameters, such as overlay_show_probabilities = True
)

# Run a prediction and assign it to result
result = model(your_image)

# Print the prediction result
print(result)

Example output:

- bbox: [240.37627136118888, 101.09044216232718, 898.4129315085123, 698.4668477562271]
  category_id: 15
  label: cat
  score: 0.86873459815979

Model.predict_batch()

The predict_batch() method accepts an iterator of data frames, such as a list or a stream, and returns a generator. It processes the iterator in a pipeline to maximize throughput, making it more efficient than calling predict() repeatedly in a loop. This approach is ideal for processing a list of images or a video stream. See the Batch Inference section for more information.

# Method Signature: Model.predict_batch()
degirum.model.Model.predict_batch(data)

Supported Input Data Types

PySDK models can handle images and raw tensors as data types.

The input you pass to predict() depends on the number of inputs the model has. If the model has one input, then you pass only one object to predict().

The model may have multiple inputs. In this case, the data you pass to predict() is a list of objects: one object per corresponding input.

Check Input Data Type of Your Model

You can check what input type your model expects by inspecting the model.model_info.InputType property.

import degirum as dg

# Declaring variables
# Set your model, inference host address, model zoo, and token in these variables.
your_model_name = "model-name"
your_host_address = "@cloud" # Can be "@cloud", host:port, or "@local"
your_model_zoo = "degirum/public"
your_token = "<token>"

# Loading a model
model = dg.load_model(
    model_name = your_model_name, 
    inference_host_address = your_host_address, 
    zoo_url = your_model_zoo, 
    token = your_token 
    # optional parameters, such as overlay_show_probabilities = True
)

# Print input data supported by your model.
print(model.model_info.InputType)

Example output:

['Image']

In this example, we check the input data type of our model.

The InputType property of the ModelParams class returned by the degirum.model.Model.model_info property describes the number and the type of inputs of the model (see Model Info section for details about model info properties).

The model_info.InputType property returns a list of input types (one entry per model input). The length of this list tells you how many separate inputs the model expects. For instance, a model that takes two images will have two entries in this list.

Images

If your model expects image inputs (InputType == "Image"), you can supply the input frame in any of the following formats:

• Path to an image file.

• HTTP URL to an image.

• NumPy array.

• PIL Image object.

• Raw bytes of image data.

PySDK automatically converts these inputs into the format required by the model’s neural network according to the model’s preprocessor settings in its JSON configuration. For more details, see the preprocessor parameters.

Tensors

If your model expects raw tensor inputs (InputType == "Tensor"), you should provide a multi-dimensional NumPy array with the appropriate shape and data type.

The array’s dimensions must match the model’s expected input shape, which you can find in the model info (model.model_info.InputShape). The data type of the array’s elements should match the model’s expected raw data type (model.model_info.InputRawDataType).

Audio

If your model expects audio inputs (InputType == "Audio"), provide a one-dimensional NumPy array containing audio waveform samples. The waveform length and sampling rate must match model.model_info.InputWaveformSize and model.model_info.InputSamplingRate.

Whisper encoder models in PySDK support both NHWC and NCHW feature layouts. The audio preprocessor inspects InputTensorLayout and the non-unit dimensions of model.model_info.InputShape so it can reorder mel bins and time frames automatically for the layout your model requires.

Single Frame Inference

When you want to process one frame, use predict().

import degirum as dg
import cv2

# Declaring variables
# Set your model, inference host address, model zoo, and token in these variables.
your_model_name = "model-name"
your_host_address = "@cloud" # Can be "@cloud", host:port, or "@local"
your_model_zoo = "degirum/public"
your_token = "<your-token>"

# Specify the image you will run inference on
your_image = "path/image.jpg"

# Loading a model
model = dg.load_model(
    model_name = your_model_name, 
    inference_host_address = your_host_address, 
    zoo_url = your_model_zoo, 
    token = your_token 
    # optional parameters, such as overlay_show_probabilities = True
)

# Run a prediction and assign it to result
result = model(your_image)

# Print the prediction result
print(result)

Example output:

- bbox: [242.46119793154423, 110.32875074982861, 898.4762465007998, 698.5679996357975]
  category_id: 15
  label: cat
  score: 0.86873459815979

Batch Inference

When you have multiple frames to process, use predict_batch(). The predict_batch() method runs predictions on an iterable list of frames. The predictions run in a pipeline to maximize throughput, making it more efficient than calling predict() in a loop.

The predict_batch() method accepts a single parameter: an iterator object, for example, a list. Populate this iterator with the same types of data you pass to predict(), such as image paths, image URLs, NumPy arrays, or PIL Image objects.

predict_batch() returns a generator of results. You can loop over these results just as you would iterate through successive predict() calls.

In addition to raw frames, the iterator can yield two-element tuples. The first element must be the frame, and the second element can contain any metadata for that frame. This metadata is passed through the inference pipeline and becomes available via the info property of the corresponding result. This makes it easy to attach per-frame context such as timestamps or frame numbers.

Because predict_batch() returns a generator, simply calling the method does not immediately run inference. Frames are processed only when you iterate over the returned generator (for example, in a for loop).

Example: Iterating over predict_batch results

for result in model.predict_batch(['image1.jpg','image2.jpg']):
    print(result)

Example: Attaching frame metadata

def frame_source():
    for idx, path in enumerate(['image1.jpg', 'image2.jpg']):
        yield path, {'frame_index': idx}

for result in model.predict_batch(frame_source()):
    print(result.info)

Example: Using predict_batch() on a video file

This example uses predict_batch() to process a video file. The frame_source generator yields frames from the video, and the model produces predictions for each frame. The results are overlaid on the frame (result.image_overlay) and displayed with OpenCV.

import degirum as dg
import cv2

# Declaring variables
# Set your model, inference host address, model zoo, and token in these variables.
your_model_name = "model-name"
your_host_address = "@cloud" # Can be "@cloud", host:port, or "@local"
your_model_zoo = "degirum/public"
your_token = "<your-token>"

# Specify the video you will run inference on
your_video = "path/video.mp4"

# Loading a model
model = dg.load_model(
    model_name = your_model_name, 
    inference_host_address = your_host_address, 
    zoo_url = your_model_zoo, 
    token = your_token 
    # optional parameters, such as overlay_show_probabilities = True
)

# Open the video file
stream = cv2.VideoCapture(your_video)

# Define generator function to produce video frames
def frame_source(stream):
    while True:
      ret, frame = stream.read()
      if not ret:
         break # end of file
      yield frame

# Run predict_batch() on frames from the video file
for result in model.predict_batch(frame_source(stream)):
    # Print raw results for each frame
    print(result)

# Release stream
stream.release()

Example output:

- bbox: [293.7637429237366, 230.0834002494812, 474.48908519744873, 621.3431906700134]
  category_id: 0
  label: person
  score: 0.7896590232849121

In the example above, the results are continuously printed to the terminal until the video is complete.

Inference Results

When you call predict(), you receive an inference result object derived from the degirum.postprocessor.InferenceResults class. Likewise, predict_batch() returns a generator that yields inference result objects. These result classes, known as postprocessors, vary by AI model type—classification, object detection, pose detection, segmentation, and so on. From your perspective, they provide the same functionality.

InferenceResults objects contain the following data:

• degirum.postprocessor.InferenceResults.image: Original input image as a NumPy array or PIL image.

• degirum.postprocessor.InferenceResults.image_overlay: Original image with inference results drawn on top. The overlay is model-dependent:

  • Classification models: class labels with probabilities are printed below the original image.

  • Object detection models: bounding boxes are printed on the original image.

  • Hand and pose detection models: keypoints and keypoint connections are printed on the original image.

  • Segmentation models: segments are printed on the original image.

• degirum.postprocessor.InferenceResults.results: Keeps a list of inference results in dictionary form. Follow the property link for detailed explanation of all result formats.

• degirum.postprocessor.InferenceResults.image_model: Preprocessed image tensor that was fed into the model (in binary form). Populated only if you enable Model.save_model_image before performing predictions.

The results property is what you will typically use in your code. This property contains the core prediction data. Note that if the model outputs coordinates (e.g., bounding boxes), these have been converted back to the coordinates of the original image for your convenience.

Example: Combine predict_batch() with image_overlay to show prediction results on original video

This example will open your video, run inference on it, ad display the video with inference results annotated over it.

import degirum as dg
import cv2

# Declaring variables
# Set your model, inference host address, model zoo, and token in these variables.
your_model_name = "model-name"
your_host_address = "@cloud" # Can be "@cloud", host:port, or "@local"
your_model_zoo = "degirum/public"
your_token = "<token>"

# Specify the video you will run inference on
your_video = "path/video.mp4"

# Loading a model
model = dg.load_model(
    model_name = your_model_name, 
    inference_host_address = your_host_address, 
    zoo_url = your_model_zoo, 
    token = your_token 
    # optional parameters, such as overlay_show_probabilities = True
)

# Open the video file
stream = cv2.VideoCapture(your_video)

# Define generator function to produce video frames
def frame_source(stream):
    while True:
        ret, frame = stream.read()
        if not ret:
            break # end of file
        yield frame

# Process the video frames in a batch and display the overlay
for result in model.predict_batch(frame_source(stream)):
    # Retrieve the overlay; if it's callable, call it
    overlay = result.image_overlay() if callable(result.image_overlay) else result.image_overlay

    # Display the overlay image in a window
    cv2.imshow("Inference Overlay", overlay)

    # Wait 1ms; press 'q' to quit early
    if cv2.waitKey(1) & 0xFF == ord('q'):
        break

# Release the video stream and close all OpenCV windows
stream.release()
cv2.destroyAllWindows()