opencv 53 convert tensorflow model to opencv

OpenCV DNN prediction:

* shape:  (1, 1000)

* class ID: 292, label: tiger, Panthera tigris

* confidence: 0.9874

#modelConversion.py

import os

import tensorflow as tf

from tensorflow.keras.applications import MobileNet

from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2

import cv2

import numpy as np

original_tf_model = MobileNet(

    include_top=True,

    weights="imagenet"

)

def get_tf_model_proto(tf_model):

    # define the directory for .pb model

    pb_model_path = "models"

    # define the name of .pb model

    pb_model_name = "mobilenet.pb"

    # create directory for further converted model

    os.makedirs(pb_model_path, exist_ok=True)

    # get model TF graph

    tf_model_graph = tf.function(lambda x: tf_model(x))

    # get concrete function

    tf_model_graph = tf_model_graph.get_concrete_function(

        tf.TensorSpec(tf_model.inputs[0].shape, tf_model.inputs[0].dtype))

    # obtain frozen concrete function

    frozen_tf_func = convert_variables_to_constants_v2(tf_model_graph)

    # get frozen graph

    frozen_tf_func.graph.as_graph_def()

    # save full tf model

    tf.io.write_graph(graph_or_graph_def=frozen_tf_func.graph,

                      logdir=pb_model_path,

                      name=pb_model_name,

                      as_text=False)

    return os.path.join(pb_model_path, pb_model_name)

def get_preprocessed_img(img_path):

    # read the image

    input_img = cv2.imread(img_path, cv2.IMREAD_COLOR)

    input_img = input_img.astype(np.float32)

    # define preprocess parameters

    mean = np.array([1.0, 1.0, 1.0]) * 127.5

    scale = 1 / 127.5

    # prepare input blob to fit the model input:

    # 1. subtract mean

    # 2. scale to set pixel values from 0 to 1

    input_blob = cv2.dnn.blobFromImage(

        image=input_img,

        scalefactor=scale,

        size=(224, 224),  # img target size

        mean=mean,

        swapRB=True,  # BGR -> RGB

        crop=True  # center crop

    )

    print("Input blob shape: {}\n".format(input_blob.shape))

    return input_blob

def get_imagenet_labels(labels_path):

    with open(labels_path) as f:

        imagenet_labels = [line.strip() for line in f.readlines()]

    return imagenet_labels

def get_opencv_dnn_prediction(opencv_net, preproc_img, imagenet_labels):

    # set OpenCV DNN input

    opencv_net.setInput(preproc_img)

    # OpenCV DNN inference

    out = opencv_net.forward()

    print("OpenCV DNN prediction: \n")

    print("* shape: ", out.shape)

    # get the predicted class ID

    imagenet_class_id = np.argmax(out)

    # get confidence

    confidence = out[0][imagenet_class_id]

    print("* class ID: {}, label: {}".format(imagenet_class_id, imagenet_labels[imagenet_class_id]))

    print("* confidence: {:.4f}\n".format(confidence))

# get TF frozen graph path

full_pb_path = get_tf_model_proto(original_tf_model)

# read frozen graph with OpenCV API

opencv_net = cv2.dnn.readNetFromTensorflow(full_pb_path)

print("OpenCV model was successfully read. Model layers: \n", opencv_net.getLayerNames())

# get preprocessed image

input_img = get_preprocessed_img("assets/tiger.jpg")

# get ImageNet labels

imagenet_labels = get_imagenet_labels("mobileNet/label.txt")

# obtain OpenCV DNN predictions

get_opencv_dnn_prediction(opencv_net, input_img, imagenet_labels)

reference:

https://docs.opencv.org/master/d1/d8f/tf_cls_tutorial_dnn_conversion.html

https://fossies.org/linux/opencv/samples/dnn/dnn_model_runner/dnn_conversion/tf/classification/py_to_py_mobilenet.py

run tensorflow on gpu

https://chuanshuoge2.blogspot.com/2020/12/keras-1.html

mobileNet labels

https://github.com/opencv/opencv/blob/master/samples/data/dnn/classification_classes_ILSVRC2012.txt