compute cosine similarity for Inception bottleneck output of two images

eval.py

import sys
import os.path
import numpy as np
import tensorflow as tf

from tensorflow.python.framework import graph_util
from tensorflow.python.framework import tensor_shape
from tensorflow.python.platform import gfile

BOTTLENECK_TENSOR_NAME = 'pool_3/_reshape:0'
BOTTLENECK_TENSOR_SIZE = 2048
JPEG_DATA_TENSOR_NAME = 'DecodeJpeg/contents:0'
RESIZED_INPUT_TENSOR_NAME = 'ResizeBilinear:0'

def eval_image(sess, image_file, bottleneck_tensor, image_data_tensor):
  image_data = gfile.FastGFile(image_file, 'rb').read()
  bottleneck_values = sess.run(bottleneck_tensor, {image_data_tensor: image_data})
  return np.squeeze(bottleneck_values)

def similarity(x, y):
  return tf.reduce_sum(tf.mul(tf.nn.l2_normalize(x, 0), tf.nn.l2_normalize(y, 0)))

if len(sys.argv) != 4:
  print('  Usage: eval <model_file> <image1> <image2>')
  exit(-1)

model_file = sys.argv[1]
if not gfile.Exists(model_file):
  tf.logging.fatal('File does not exist "%s"', model_file)
  exit(-1)
  
img1 = sys.argv[2]
if not gfile.Exists(img1):
  tf.logging.fatal('File does not exist "%s"', img1)
  exit(-1)

img2 = sys.argv[3]
if not gfile.Exists(img2):
  tf.logging.fatal('File does not exist "%s"', img2)
  exit(-1)

print('Loading model...')
sess = tf.Session()
with gfile.FastGFile(model_file, 'rb') as f:
  graph_def = tf.GraphDef()
  graph_def.ParseFromString(f.read())
  bottleneck_tensor, jpeg_data_tensor, resized_input_tensor = (
          tf.import_graph_def(graph_def, name='', return_elements=[
              BOTTLENECK_TENSOR_NAME, JPEG_DATA_TENSOR_NAME, RESIZED_INPUT_TENSOR_NAME]))

print('Evaluating...')
bottleneck1 = eval_image(sess, img1, bottleneck_tensor, jpeg_data_tensor)
bottleneck2 = eval_image(sess, img2, bottleneck_tensor, jpeg_data_tensor)

print(sess.run(similarity(bottleneck1, bottleneck2)))