Using exported custom vision tensor-flow model in python

classify.py

import sys
import os
import tensorflow as tf
import numpy as np
import scipy
from PIL import Image
from scipy import misc

cwd = os.getcwd()

modelFile = cwd + '/model.pb'

network_input_size = 227
output_layer = 'loss:0'
input_node = 'Placeholder:0'

graph_def = None


def load_graph(filename):
    global graph_def, labels
    with tf.gfile.FastGFile(filename, 'rb') as f:
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(f.read())
        tf.import_graph_def(graph_def, name='')

    labels = [line.rstrip() for line in tf.gfile.GFile(cwd + '/labels.txt')]
    print(labels)


def center_crop(img, cropx, cropy):
    y, x, z = img.shape
    startx = x // 2 - (cropx // 2)
    starty = y // 2 - (cropy // 2)
    return img[starty:starty + cropy, startx:startx + cropx]


def classifyImage(testImage):
    tf.reset_default_graph()
    tf.import_graph_def(graph_def, name='')

    with tf.Session() as sess:

        prob_tensor = sess.graph.get_tensor_by_name(output_layer)

        image = scipy.misc.imread(testImage)
        w = image.shape[0]
        h = image.shape[1]

        temp = scipy.misc.imresize(image, (227, 227, 3))

        # Center Crop
        try:
            temp = center_crop(temp, network_input_size, network_input_size)
        except:
            print(filename, ":-1", ":-1")
            return

        image = temp.astype(float)

        # RGB -> BGR
        red, green, blue = tf.split(axis=2, num_or_size_splits=3, value=image)

        # Apply mean values
        image_normalized = tf.concat(axis=2, values=[
            blue - 104.,
            green - 117.,
            red - 124.,
        ])

        image_normalized = image_normalized.eval()
        image_normalized = np.expand_dims(image_normalized, axis=0)

        predictions, = sess.run(prob_tensor, {input_node: image_normalized})

        idx = np.argmax(predictions)
        print('-----------------------')
        print('Image classified as: ' + labels[idx])

        print('---------------------------')
        top_k = predictions.argsort()[-5:][::-1]
        for node_id in top_k:
            score = predictions[node_id]
            print('%s (score = %.5f)' % (labels[node_id], score))


load_graph(modelFile)

if len(sys.argv) != 2:
    print('Cannot classify -  Usage: classify.py file path.')
else:
    classifyImage(cwd + '/' + sys.argv[1])

This was super helpful years later. It's still relevant today as long as you replace tf with tf.compat.v1 given that Python by default installs tensorflow 2 now.

I'm also at @microsoft. I used most of your code here: https://github.com/berndverst/AzureCustomVision-TensorFlowImageClassification