This tutorial demonstrates step-by-step instructions on how to convert models loaded from TensorFlow Hub using OpenVINO Runtime.
TensorFlow Hub is a library and online platform developed by Google that simplifies machine learning model reuse and sharing. It serves as a repository of pre-trained models, embeddings, and reusable components, allowing researchers and developers to access and integrate state-of-the-art machine learning models into their own projects with ease. TensorFlow Hub provides a diverse range of models for various tasks like image classification, text embedding, and more. It streamlines the process of incorporating these models into TensorFlow workflows, fostering collaboration and accelerating the development of AI applications. This centralized hub enhances model accessibility and promotes the rapid advancement of machine learning capabilities across the community.
You have the flexibility to run this tutorial notebook in its entirety or selectively execute specific sections, as each section operates independently.
This is a self-contained example that relies solely on its own code.
We recommend running the notebook in a virtual environment. You only need a Jupyter server to start. For details, please refer to Installation Guide.
%pip install -q pillow numpy
%pip install -q "openvino>=2023.2.0" "opencv-python" "matplotlib>=3.4"
%pip install -q "tensorflow-macos>=2.5; sys_platform == 'darwin' and platform_machine == 'arm64'"
%pip install -q "tensorflow>=2.5; sys_platform == 'darwin' and platform_machine != 'arm64'"
%pip install -q "tensorflow>=2.5; sys_platform != 'darwin'"
%pip install -q --no-deps tensorflow_hub
%pip install -q tf_keras
We will use the MobileNet_v2 image classification model from TensorFlow Hub.
MobileNetV2 is a compact and efficient deep learning architecture designed for mobile and embedded devices, developed by Google researchers. It builds on the success of the original MobileNet by introducing improvements in both speed and accuracy. MobileNetV2 employs a streamlined architecture with inverted residual blocks, making it highly efficient for real-time applications while minimizing computational resources. This network excels in tasks like image classification, object detection, and image segmentation, offering a balance between model size and performance. MobileNetV2 has become a popular choice for on-device AI applications, enabling faster and more efficient deep learning inference on smartphones and edge devices.
More information about model can be found on Model page on TensorFlow Hub
from pathlib import Path
import os
import requests
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
os.environ["TF_USE_LEGACY_KERAS"] = "1"
os.environ["TFHUB_CACHE_DIR"] = str(Path("./tfhub_modules").resolve())
import tensorflow_hub as hub
import tensorflow as tf
import PIL
import numpy as np
import matplotlib.pyplot as plt
import openvino as ov
tf.get_logger().setLevel("ERROR")
if not Path("notebook_utils.py").exists():
r = requests.get(
url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/notebook_utils.py",
)
open("notebook_utils.py", "w").write(r.text)
# Read more about telemetry collection at https://github.com/openvinotoolkit/openvino_notebooks?tab=readme-ov-file#-telemetry
from notebook_utils import collect_telemetry
collect_telemetry("tensorflow-hub.ipynb")
IMAGE_SHAPE = (224, 224)
IMAGE_URL, IMAGE_PATH = (
"https://storage.googleapis.com/download.tensorflow.org/example_images/grace_hopper.jpg",
"data/grace_hopper.jpg",
)
MODEL_URL, MODEL_PATH = (
"https://www.kaggle.com/models/google/mobilenet-v1/frameworks/tensorFlow2/variations/100-224-classification/versions/2",
"models/mobilenet_v2_100_224.xml",
)
back to top ⬆️
Select a MobileNetV2 pre-trained model from TensorFlow Hub and wrap it as a Keras layer with hub.KerasLayer
.
model = hub.KerasLayer(MODEL_URL, input_shape=IMAGE_SHAPE + (3,))
back to top ⬆️
The input images
are expected to have color values in the range [0,1], following the common image input conventions. For this model, the size of the input images is fixed to height
x width
= 224 x 224 pixels.
IMAGE_PATH = Path(IMAGE_PATH)
IMAGE_PATH.parent.mkdir(parents=True, exist_ok=True)
if not IMAGE_PATH.exists():
r = requests.get(IMAGE_URL)
with IMAGE_PATH.open("wb") as f:
f.write(r.content)
grace_hopper = PIL.Image.open(IMAGE_PATH).resize(IMAGE_SHAPE)
grace_hopper
Normalize the image to [0,1] range.
grace_hopper = np.array(grace_hopper) / 255.0
grace_hopper.shape
(224, 224, 3)
We will convert the loaded model to OpenVINO IR using ov.convert_model
function. We pass the model object to it, no additional arguments required. Then, we save the model to disk using ov.save_model
function.
MODEL_PATH = Path(MODEL_PATH)
if not MODEL_PATH.exists():
converted_model = ov.convert_model(model)
ov.save_model(converted_model, MODEL_PATH)
select device from dropdown list for running inference using OpenVINO
from notebook_utils import device_widget
device = device_widget()
device
Dropdown(description='Device:', index=3, options=('CPU', 'GPU.0', 'GPU.1', 'AUTO'), value='AUTO')
core = ov.Core()
compiled_model = core.compile_model(MODEL_PATH, device_name=device.value)
Add a batch dimension (with np.newaxis
) and pass the image to the model:
output = compiled_model(grace_hopper[np.newaxis, ...])[0]
output.shape
(1, 1001)
The result is a 1001-element vector of logits, rating the probability of each class for the image.
The top class ID can be found with np.argmax
:
predicted_class = np.argmax(output[0], axis=-1)
predicted_class
653
Take the predicted_class
ID (such as 653
) and fetch the ImageNet dataset labels to decode the predictions:
labels_path = tf.keras.utils.get_file(
"ImageNetLabels.txt",
"https://storage.googleapis.com/download.tensorflow.org/data/ImageNetLabels.txt",
)
imagenet_labels = np.array(open(labels_path).read().splitlines())
plt.imshow(grace_hopper)
plt.axis("off")
predicted_class_name = imagenet_labels[predicted_class]
_ = plt.title("Prediction: " + predicted_class_name.title())
We will use arbitrary image stylization model from TensorFlow Hub.
The model contains conditional instance normalization (CIN) layers
The CIN network consists of two main components: a feature extractor and a stylization module. The feature extractor extracts a set of features from the content image. The stylization module then uses these features to generate a stylized image.
The stylization module is a stack of convolutional layers. Each convolutional layer is followed by a CIN layer. The CIN layer takes the features from the previous layer and the CIN parameters from the style image as input and produces a new set of features as output.
The output of the stylization module is a stylized image. The stylized image has the same content as the original content image, but the style has been transferred from the style image.
The CIN network is able to stylize images in real time because it is very efficient.
More model information can be found on Model page on TensorFlow Hub.
import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
os.environ["TF_USE_LEGACY_KERAS"] = "1"
os.environ["TFHUB_CACHE_DIR"] = str(Path("./tfhub_modules").resolve())
from pathlib import Path
import openvino as ov
import tensorflow_hub as hub
import tensorflow as tf
import cv2
import numpy as np
import matplotlib.pyplot as plt
CONTENT_IMAGE_URL = "https://github.com/openvinotoolkit/openvino_notebooks/assets/29454499/525babb8-1289-45f8-a3a5-e248f74dfb24"
CONTENT_IMAGE_PATH = Path("./data/YellowLabradorLooking_new.jpg")
STYLE_IMAGE_URL = "https://github.com/openvinotoolkit/openvino_notebooks/assets/29454499/c212233d-9a33-4979-b8f9-2a94a529026e"
STYLE_IMAGE_PATH = Path("./data/Vassily_Kandinsky%2C_1913_-_Composition_7.jpg")
MODEL_URL = "https://www.kaggle.com/models/google/arbitrary-image-stylization-v1/frameworks/tensorFlow1/variations/256/versions/2"
MODEL_PATH = Path("./models/arbitrary-image-stylization-v1-256.xml")
We load the model from TensorFlow Hub using hub.KerasLayer
. Since the model has multiple inputs (content image and style image), we need to build it by calling with placeholders and wrap in tf.keras.Model
function.
inputs = {
"placeholder": tf.keras.layers.Input(shape=(None, None, 3)),
"placeholder_1": tf.keras.layers.Input(shape=(None, None, 3)),
}
model = hub.KerasLayer(MODEL_URL, signature="serving_default", signature_outputs_as_dict=True) # define the signature to allow passing inputs as a dictionary
outputs = model(inputs)
model = tf.keras.Model(inputs=inputs, outputs=outputs)
We convert the loaded model to OpenVINO IR using ov.convert_model
function. We pass our model to the function, no additional arguments needed. After converting, we save the model to disk using ov.save_model
function.
if not MODEL_PATH.exists():
MODEL_PATH.parent.mkdir(parents=True, exist_ok=True)
converted_model = ov.convert_model(model)
ov.save_model(converted_model, MODEL_PATH)
select device from dropdown list for running inference using OpenVINO
device
Dropdown(description='Device:', index=3, options=('CPU', 'GPU.0', 'GPU.1', 'AUTO'), value='AUTO')
compiled_model = core.compile_model(MODEL_PATH, device_name=device.value)
if not STYLE_IMAGE_PATH.exists():
r = requests.get(STYLE_IMAGE_URL)
with STYLE_IMAGE_PATH.open("wb") as f:
f.write(r.content)
if not CONTENT_IMAGE_PATH.exists():
r = requests.get(CONTENT_IMAGE_URL)
with CONTENT_IMAGE_PATH.open("wb") as f:
f.write(r.content)
def load_image(dst):
image = cv2.imread(str(dst))
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # Convert image color to RGB space
image = image / 255 # Normalize to [0, 1] interval
image = image.astype(np.float32)
return image
content_image = load_image(CONTENT_IMAGE_PATH)
style_image = load_image(STYLE_IMAGE_PATH)
style_image = cv2.resize(style_image, (256, 256)) # model was trained on 256x256 images
result = compiled_model([content_image[np.newaxis, ...], style_image[np.newaxis, ...]])[0]
title2img = {
"Source image": content_image,
"Reference style": style_image,
"Result": result[0],
}
plt.figure(figsize=(12, 12))
for i, (title, img) in enumerate(title2img.items()):
ax = plt.subplot(1, 3, i + 1)
ax.set_title(title)
plt.imshow(img)
plt.axis("off")