Posted on

d246: TensorFlow CIFAR-10 tutorial, detailed step-by-step review, Part 2

Detailed review of TensorFlow CIFAR-10 tutorial, Part 2 [Click here for Part 1]

Execution process of ‘python cifar10_train.py’

  1. Initialising flags and constants from cifar10.py and cifar10_input.py:
    • data_dir
    • batch_size
    • use_fp16
    • IMAGE_SIZE
    • NUM_CLASSES
    • NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN
    • NUM_EXAMPLES_PER_EPOCH_FOR_EVAL
    • MOVING_AVERAGE_DECAY
    • NUM_EPOCHS_PER_DECAY
    • LEARNING_RATE_DECAY_FACTOR
    • INITIAL_LEARNING_RATE
    • Detailed description of all flags and constants: CIFAR-10 TensorFlow tutorial review, part 1
  2. Downloading and unpacking a ‘cifar-10-binary.tar.gz‘ dataset.
  3. Calling the ‘cifar10_train.py‘ with default parameters.
    1. Setting up the required variables.
    2. Calling ‘distorted_inputs()‘ from ‘cifar10_input.py‘
      • Preparing images and labels from the downloaded dataset .
      • Filling the queue with 20’000 CIFAR images.
      • Generating images and labels batch (shape: 128, 24, 24, 3; dtype:float32)
    3. Calling ‘distorted_inputs()‘ from ‘cifar10.py‘ to construct distorted input for CIFAR training.
    4. Returns images to ‘cifar10_train.py/train()’
  4. Building CIFAR-10 model ‘inference(images)‘ in ‘cifar10.py‘ with following layers:
    • conv1
    • pool1
    • norm1
    • conv2
    • norm2
    • pool2
    • local3
    • local4
    • softmax
  5. Defines ‘weight_decay‘ for each layer (tf.Variable).
  6. Assigns the ‘loss_summaries‘ for each layer.
  7. Builds and applies the gradients.
  8. Builds histograms of weights and biases for each layer.
  9. Creating a saver function, to be able to recover from checkpoints later.
  10. Builds the summary operation.
  11. Creates and starts the ‘tf.Session’.
  12. Starting the queue runners, begins the training loop.

Prediction (CIFAR-10 TensorFlow evaluation example outputs labels and probabilities)

Labels and probabilities output for single image of CIFAR-10 TensorFlow tutorial:

from PIL import Image

We will use the Python Imaging Library (PIL) to process images into JPEG files supported by out mode without additional losses.

width = 24
height = 24

categories = [ "airplane","automobile","bird","cat","deer","dog","frog","horse","ship","truck" ]

filename = "/tmp/images/dog1.png" # absolute path to input image
im = Image.open(filename)
im.save(filename, format='JPEG', subsampling=0, quality=100)

Converting the file from input format to JPEG with 100% quality and without subsampling.

input_img = tf.image.decode_jpeg(tf.read_file(filename), channels=3)
tf_cast = tf.cast(input_img, tf.float32)

Decoding image into 3 channels (red, green, blue) float32 array.

float_image = tf.image.resize_image_with_crop_or_pad(tf_cast, height, width)

Reshaping image.

images = tf.expand_dims(float_image, 0)

Creating a batch of images with just one image.

logits = cifar10.inference(images)

Get logits from main model (cifar10.py).

_, top_k_pred = tf.nn.top_k(logits, k=5)

Get top 5 predicions (‘top_k_pred’) and their values (‘_’).

init_op = tf.initialize_all_variables()

Initialising variables.

with tf.Session() as sess:
  sess.run(init_op)
  _, top_indices = sess.run([_, top_k_pred])
  for key, value in enumerate(top_indices[0]):
    print (categories[value] + ", " + str(_[0][key]))

Runing tf.Session() and processing the output.

Example of input:

tensorflow cifar-10 tutorial otput with labels truck input

Example of output:

tensorflow cifar-10 tutorial otput with labels