main.py

import matplotlib.pyplot as plt
import tensorflow as tf
import tensorflow_hub as hub
import matplotlib.image as Image
import numpy as np

# GLOBALS
model = ""
movenet = ""

interpreter = ""

# NOTE: Code snippets sampes from Tensorflow lite lightning MoveNet version 4, subject to Apache 2.0 Licence
# https://tfhub.dev/google/lite-model/movenet/singlepose/lightning/tflite/float16/4
# Any changed will be noted.

def initialise_model():
    #model = hub.load("https://tfhub.dev/google/movenet/singlepose/lightning/4")
    #movenet = model.signatures['serving_default']
    print("initialise")
    global interpreter
    #CHANGED: model has different name and is local to program
    interpreter = tf.lite.Interpreter(model_path = "lite-model_movenet_singlepose_lightning_tflite_float16_4.tflite")
    interpreter.allocate_tensors()

#CHANGED: scripts contain both normal and lite version of MoveNet
def pose_PredictonTF(imageData):
    """
    Estimates the posing positions of a single human in a JPEG picture using Tensorflow MoveNet.

    :param imageData: (JPEG-Encoded bytes)
    :return: a [1,1,17,3] tensor containing 17 key points of the body, with percentage coordinates and confidence.
    """

    # Image = tf.io.read_file(image_path)
    image = tf.compat.v1.image.decode_jpeg(imageData)
    image = tf.expand_dims(image, axis=0)
    # Resize and pad the Image to keep the aspect ratio and fit the expected size.
    image = tf.cast(tf.image.resize_with_pad(image, 192, 192), dtype=tf.int32)

    # run model inference
    outputs = movenet(image)
    # output is a [1,1,17,3] sensor
    # keypoints = outputs['output_0']
    # print(keypoints)
    return outputs

def posePredictionTFLite(imageData):
    image = tf.compat.v1.image.decode_jpeg(imageData)
    image = tf.expand_dims(image, axis=0)
    image = tf.cast(tf.image.resize_with_pad(image, 192, 192), dtype=tf.uint8)

    input_details = interpreter.get_input_details()
    output_details = interpreter.get_output_details()

    interpreter.set_tensor(input_details[0]['index'],image.numpy())

    interpreter.invoke()

    return interpreter.get_tensor(output_details[0]['index'])

#CHANGED: temporary data used for quick testing
def TestingPointsOnImage():
    test = np.array([[[[0.5052221, 0.42522937, 0.7141144],
                       [0.48090276, 0.45370695, 0.5738922],
                       [0.47875905, 0.40008497, 0.6808091],
                       [0.4980592, 0.4830124, 0.5379905],
                       [0.49261594, 0.36388832, 0.8307032],
                       [0.60409665, 0.551592, 0.6490847],
                       [0.6021359, 0.29204538, 0.8377292],
                       [0.73887765, 0.66975343, 0.6410032],
                       [0.7459737, 0.17954277, 0.56632763],
                       [0.6070323, 0.7081243,  0.738752],
                       [0.6068721, 0.1500265, 0.7462287],
                       [0.84355897, 0.5062511, 0.5052821],
                       [0.8494219, 0.33797163, 0.462009],
                       [0.7733456, 0.5940913, 0.04863108],
                       [0.7795013, 0.2247358, 0.0506536],
                       [0.6130973, 0.71571106, 0.1039405],
                       [0.598618,  0.13673246, 0.18669602]]]])

    testnp = np.resize(test,(17,3))
    print(testnp)

    x = []
    y = []
    c = []

    img = np.asarray(Image.imread("D:/fruta/Pictures/Camera Roll/testing thing.jpg"))

    print(img.shape)

    for i in range(17):
        if test[0,0,i,2] > 0.3:
            y.append(test[0, 0, i, 0] * (img.shape[0]))
            x.append(test[0, 0, i, 1] * (img.shape[1]))
            c.append(test[0, 0, i, 2])

    #print(x)
    #print(y)
    #print(c)

    plt.figure()
    plt.imshow(img)
    plt.scatter(x,y)
    plt.show()


# Press the green button in the gutter to run the script.
if __name__ == '__main__':
    print(tf.config.list_physical_devices())
    initialise_model()
    print(posePredictionTFLite(tf.io.read_file("D:/fruta/Pictures/Camera Roll/testing thing.jpg")))