Live Inference and Benchmark CT-scan Data with OpenVINO™

This tutorial is also available as a Jupyter notebook that can be cloned directly from GitHub. See the installation guide for instructions to run this tutorial locally on Windows, Linux or macOS.


Kidney Segmentation with PyTorch Lightning and OpenVINO™ - Part 4

This tutorial is a part of a series on how to train, optimize, quantize and show live inference on a medical segmentation model. The goal is to accelerate inference on a kidney segmentation model. The UNet model is trained from scratch, and the data is from Kits19.

This tutorial shows how to benchmark performance of the model and show live inference with async API and MULTI plugin in OpenVINO.

This notebook needs a quantized OpenVINO IR model and images from the KiTS-19 dataset, converted to 2D images. (To learn how the model is quantized, see the Convert and Quantize a UNet Model and Show Live Inference tutorial.)

This notebook provides a pre-trained model, trained for 20 epochs with the full KiTS-19 frames dataset, which has an F1 score on the validation set of 0.9. The training code is available in the PyTorch Monai Training notebook.

For demonstration purposes, this tutorial will download one converted CT scan to use for inference.

!pip install -q "monai>=0.9.1,<1.0.0"


import os
import sys
import zipfile
from pathlib import Path

import numpy as np
from monai.transforms import LoadImage
from openvino.runtime import Core

from custom_segmentation import SegmentationModel

from notebook_utils import download_file


To use the pre-trained models, set IR_PATH to "pretrained_model/unet44.xml" and COMPRESSED_MODEL_PATH to "pretrained_model/quantized_unet44.xml". To use a model that you trained or optimized yourself, adjust the model paths.

# The directory that contains the IR model (xml and bin) files.
MODEL_PATH = "pretrained_model/quantized_unet_kits19.xml"
# Uncomment the next line to use the FP16 model instead of the quantized model.
# MODEL_PATH = "pretrained_model/unet_kits19.xml"

Benchmark Model Performance

To measure the inference performance of the IR model, use Benchmark Tool - an inference performance measurement tool in OpenVINO. Benchmark tool is a command-line application that can be run in the notebook with ! benchmark_app or %sx benchmark_app commands.

Note: The benchmark_app tool is able to measure the performance of the OpenVINO Intermediate Representation (OpenVINO IR) models only. For more accurate performance, run benchmark_app in a terminal/command prompt after closing other applications. Run benchmark_app -m model.xml -d CPU to benchmark async inference on CPU for one minute. Change CPU to GPU to benchmark on GPU. Run benchmark_app --help to see an overview of all command-line options.

ie = Core()
# By default, benchmark on MULTI:CPU,GPU if a GPU is available, otherwise on CPU.
device = "MULTI:CPU,GPU" if "GPU" in ie.available_devices else "CPU"
# Uncomment one of the options below to benchmark on other devices.
# device = "GPU"
# device = "CPU"
# device = "AUTO"
# Benchmark model
! benchmark_app -m $MODEL_PATH -d $device -t 15 -api sync
[Step 1/11] Parsing and validating input arguments
[ INFO ] Parsing input parameters
[Step 2/11] Loading OpenVINO Runtime
[ INFO ] OpenVINO:
[ INFO ] Build ................................. 2022.3.0-9052-9752fafe8eb-releases/2022/3
[ INFO ]
[ INFO ] Device info:
[ INFO ] Build ................................. 2022.3.0-9052-9752fafe8eb-releases/2022/3
[ INFO ]
[ INFO ]
[Step 3/11] Setting device configuration
[ WARNING ] Performance hint was not explicitly specified in command line. Device(CPU) performance hint will be set to LATENCY.
[Step 4/11] Reading model files
[ INFO ] Loading model files
[ INFO ] Read model took 31.42 ms
[ INFO ] Original model I/O parameters:
[ INFO ] Model inputs:
[ INFO ]     input.1 (node: input.1) : f32 / [...] / [1,1,512,512]
[ INFO ] Model outputs:
[ INFO ]     153 (node: 153) : f32 / [...] / [1,1,512,512]
[Step 5/11] Resizing model to match image sizes and given batch
[ INFO ] Model batch size: 1
[Step 6/11] Configuring input of the model
[ INFO ] Model inputs:
[ INFO ]     input.1 (node: input.1) : f32 / [N,C,H,W] / [1,1,512,512]
[ INFO ] Model outputs:
[ INFO ]     153 (node: 153) : f32 / [...] / [1,1,512,512]
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 209.10 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] Model:
[ INFO ]   NETWORK_NAME: pretrained_unet_kits19
[ INFO ]   AFFINITY: Affinity.CORE
[ INFO ]   PERF_COUNT: False
[ INFO ]   INFERENCE_PRECISION_HINT: <Type: 'float32'>
[Step 9/11] Creating infer requests and preparing input tensors
[ WARNING ] No input files were given for input 'input.1'!. This input will be filled with random values!
[ INFO ] Fill input 'input.1' with random values
[Step 10/11] Measuring performance (Start inference synchronously, limits: 15000 ms duration)
[ INFO ] Benchmarking in inference only mode (inputs filling are not included in measurement loop).
[ INFO ] First inference took 23.91 ms
[Step 11/11] Dumping statistics report
[ INFO ] Count:            1437 iterations
[ INFO ] Duration:         15010.38 ms
[ INFO ] Latency:
[ INFO ]    Median:        10.20 ms
[ INFO ]    Average:       10.25 ms
[ INFO ]    Min:           9.93 ms
[ INFO ]    Max:           13.01 ms
[ INFO ] Throughput:   98.02 FPS

Download and Prepare Data

Download one validation video for live inference.

This tutorial reuses the KitsDataset class that was also used in the training and quantization notebook that will be released later.

The data is expected in BASEDIR. The BASEDIR directory should contain the case_00000 to case_00299 subdirectories. If the data for the case specified above does not already exist, it will be downloaded and extracted in the next cell.

# Directory that contains the CT scan data. This directory should contain subdirectories
# case_00XXX where XXX is between 000 and 299.
BASEDIR = Path("kits19_frames_1")
# The CT scan case number. For example: 16 for data from the case_00016 directory.
# Currently only 117 is supported.
CASE = 117

case_path = BASEDIR / f"case_{CASE:05d}"

if not case_path.exists():
    filename = download_file(
    with zipfile.ZipFile(filename, "r") as zip_ref:
    os.remove(filename)  # remove zipfile
    print(f"Downloaded and extracted data for case_{CASE:05d}")
    print(f"Data for case_{CASE:05d} exists")   0%|          | 0.00/5.48M [00:00<?, ?B/s]
Downloaded and extracted data for case_00117

Show Live Inference

To show live inference on the model in the notebook, use the asynchronous processing feature of OpenVINO Runtime.

If you use a GPU device, with device="GPU" or device="MULTI:CPU,GPU" to do inference on an integrated graphics card, model loading will be slow the first time you run this code. The model will be cached, so after the first time model loading will be faster. For more information on OpenVINO Runtime, including Model Caching, refer to the OpenVINO API tutorial.

We will use AsyncInferQueue to perform asynchronous inference. It can be instantiated with compiled model and a number of jobs - parallel execution threads. If you don’t pass a number of jobs or pass 0, then OpenVINO will pick the optimal number based on your device and heuristics. After acquiring the inference queue, there are two jobs to do: - Preprocess the data and push it to the inference queue. The preprocessing steps will remain the same. - Tell the inference queue what to do with the model output after the inference is finished. It is represented by the callback python function that takes an inference result and data that we passed to the inference queue along with the prepared input data

Everything else will be handled by the AsyncInferQueue instance.

Load Model and List of Image Files

Load the segmentation model to OpenVINO Runtime with SegmentationModel, based on the Model API from Open Model Zoo. This model implementation includes pre and post processing for the model. For SegmentationModel this includes the code to create an overlay of the segmentation mask on the original image/frame. Uncomment the next cell to see the implementation.

ie = Core()
segmentation_model = SegmentationModel(
    ie=ie, model_path=Path(MODEL_PATH), sigmoid=True, rotate_and_flip=True
image_paths = sorted(case_path.glob("imaging_frames/*jpg"))

print(f"{}, {len(image_paths)} images")
case_00117, 69 images

Preapre images

Use the reader = LoadImage() function to read the images in the same way as in the training tutorial.

framebuf = []

next_frame_id = 0
reader = LoadImage(image_only=True, dtype=np.uint8)

while next_frame_id < len(image_paths) - 1:
    image_path = image_paths[next_frame_id]
    image = reader(str(image_path))
    next_frame_id += 1

Specify device

# Possible options for device include "CPU", "GPU", "AUTO", "MULTI".
device = "MULTI:CPU,GPU" if "GPU" in ie.available_devices else "CPU"

Setting callback function

When callback is set, any job that ends the inference, calls the Python function. The callback function must have two arguments: one is the request that calls the callback, which provides the InferRequest API; the other is called “userdata”, which provides the possibility of passing runtime values.

The callback function will show the results of inference.

import cv2
import copy
from IPython import display

from typing import Dict, Any
from openvino.runtime import InferRequest

# Define a callback function that runs every time the asynchronous pipeline completes inference on a frame
def completion_callback(infer_request: InferRequest, user_data: Dict[str, Any],) -> None:
    preprocess_meta = user_data['preprocess_meta']

    raw_outputs = {out.any_name: copy.deepcopy( for out, res in zip(infer_request.model_outputs, infer_request.output_tensors)}
    frame = segmentation_model.postprocess(raw_outputs, preprocess_meta)

    _, encoded_img = cv2.imencode(".jpg", frame, params=[cv2.IMWRITE_JPEG_QUALITY, 90])
    # Create IPython image
    i = display.Image(data=encoded_img)

    # Display the image in this notebook

Create asynchronous inference queue and perform it

import time
from openvino.runtime import AsyncInferQueue

load_start_time = time.perf_counter()
compiled_model = ie.compile_model(, device)
# Create asynchronous inference queue with optimal number of infer requests
infer_queue = AsyncInferQueue(compiled_model)
load_end_time = time.perf_counter()

results = [None] * len(framebuf)
frame_number = 0

# Perform inference on every frame in the framebuffer
start_time = time.time()
for i, input_frame in enumerate(framebuf):
    inputs, preprocessing_meta = segmentation_model.preprocess({ input_frame})
    infer_queue.start_async(inputs, {'preprocess_meta': preprocessing_meta})

# Wait until all inference requests in the AsyncInferQueue are completed
stop_time = time.time()

# Calculate total inference time and FPS
total_time = stop_time - start_time
fps = len(framebuf) / total_time
time_per_frame = 1 / fps

print(f"Loaded model to {device} in {load_end_time-load_start_time:.2f} seconds.")

print(f'Total time to infer all frames: {total_time:.3f}s')
print(f'Time per frame: {time_per_frame:.6f}s ({fps:.3f} FPS)')
Loaded model to CPU in 0.20 seconds.
Total time to infer all frames: 3.085s
Time per frame: 0.045371s (22.040 FPS)