frogermcs

img_laptop_url = "https://upload.wikimedia.org/wikipedia/commons/9/90/ThinkPad_X220.jpg"
img_laptop = PIL.Image.open(BytesIO(requests.get(img_laptop_url).content))
imshow(np.asarray(img_laptop))
img_golden_url = "https://upload.wikimedia.org/wikipedia/commons/9/93/Golden_Retriever_Carlos_%2810581910556%29.jpg"
img_golden = PIL.Image.open(BytesIO(requests.get(img_golden_url).content))
imshow(np.asarray(img_golden))

frogermcs

%matplotlib inline #will be useful for images preview
from matplotlib.pyplot import imshow

import tensorflow as tf
import coremltools

#For easier images processing
import numpy as np
import PIL
import requests

frogermcs

Core ML model generated. Saved at location: mobilenet_v2_1.0_224.mlmodel 

Core ML input(s): 
 [name: "input__0"
  type {
    imageType {
      width: 224
      height: 224
      colorSpace: RGB
  }

frogermcs

# Convert TensorFlow model to Core ML format

# Input model definition
IMAGE_INPUT_NAME = ["input:0"]
IMAGE_INPUT_NAME_SHAPE = {'input:0':[1,224,224,3]}
IMAGE_INPUT_SCALE = 1.0/255.0
OUTPUT_NAME = ['MobilenetV2/Predictions/Reshape_1:0']
MODEL_LABELS = 'ImageNetLabels.txt'

# Output model

frogermcs

!cat mobilenet_v2_1.0_224_info.txt

# Output
# >> Model: mobilenet_v2_1.0_224
# >> Input: input
# >> Output: MobilenetV2/Predictions/Reshape_1

frogermcs

#Download and unpack MobileNet v2 model from https://www.tensorflow.org/lite/guide/hosted_models
!curl -LO http://download.tensorflow.org/models/tflite_11_05_08/mobilenet_v2_1.0_224.tgz
!curl -LO https://storage.googleapis.com/download.tensorflow.org/data/ImageNetLabels.txt
!tar -xvzf mobilenet_v2_1.0_224.tgz

# List of unpacked files
# >> mobilenet_v2_1.0_224.ckpt.data-00000-of-00001
# >> mobilenet_v2_1.0_224.ckpt.index
# >> mobilenet_v2_1.0_224.ckpt.meta
# >> mobilenet_v2_1.0_224_eval.pbtxt

frogermcs

!pip install tfcoreml

import tensorflow as tf
import tfcoreml

# TensorFlow 2.0 isn't yet supported. Make sure you use 1.x
print("TensorFlow version {}".format(tf.__version__))
print("Eager mode: ", tf.executing_eagerly())
print("Is GPU available: ", tf.test.is_gpu_available())

frogermcs

# Concatenation of argmax and max value for each row
def max_values_only(data):
  argmax_col = np.argmax(data, axis=1).reshape(-1, 1)
  max_col = np.max(data, axis=1).reshape(-1, 1)
  return np.concatenate([argmax_col, max_col], axis=1)

# Build simplified prediction tables
tf_model_pred_simplified = max_values_only(tf_model_predictions)
tflite_model_pred_simplified = max_values_only(tflite_model_predictions)
tflite_q_model_pred_simplified = max_values_only(tflite_q_model_predictions)

frogermcs

# Concatenate results from all models
all_models_dataframe = pd.concat([tf_pred_dataframe, 
                                  tflite_pred_dataframe, 
                                  tflite_q_pred_dataframe], 
                                 keys=['TF Model', 'TFLite', 'TFLite quantized'],
                                 axis='columns')

# Swap columns to hava side by side comparison
all_models_dataframe = all_models_dataframe.swaplevel(axis='columns')[tflite_pred_dataframe.columns]

frogermcs

# Set batch of images into input tensor
tflite_interpreter.set_tensor(input_details[0]['index'], val_image_batch)
# Run inference
tflite_interpreter.invoke()
# Get prediction results
tflite_model_predictions = tflite_interpreter.get_tensor(output_details[0]['index'])
print("Prediction results shape:", tflite_model_predictions.shape)

# >> Prediction results shape: (32, 5)