Based on: TensorFlow Inception Retraining Tutorial
In [ ]:
# Import a few system libraries:
from datetime import datetime
import hashlib
import os.path
import random
import re
import struct
import sys
import tarfile
# Import Numpy library:
import numpy as np
# Import Urllib library:
from six.moves import urllib
# Import TensorFlow library:
import tensorflow as tf
from tensorflow.python.framework import graph_util
from tensorflow.python.framework import tensor_shape
from tensorflow.python.platform import gfile
from tensorflow.python.util import compat
In [ ]:
# Set path to imagenet data directory:
model_dir = 'data/imagenet_retrain'
# Set path to image directory:
image_dir = os.path.join(model_dir, 'flower_photos')
# Set path to retrained graph:
output_graph = os.path.join(model_dir, 'output_graph.pb')
# Set path to retrained graph label file:
output_labels = os.path.join(model_dir, 'output_labels.txt')
# Set name of output classification layer in retrained graph:
final_tensor_name = 'final_result'
# Set path to bottleneck directory:
bottleneck_dir = '/tmp/retrain_bottlenecks'
# Set path to summaries directory:
summaries_dir = '/tmp/retrain_logs'
# Set number of steps in retraining process:
how_many_training_steps = 4000
# Set the learning rate:
learning_rate = 0.01
# Set the validation-set percentage:
validation_percentage = 10
# Set the test-set percentage:
testing_percentage = 10
# Set the training evaluation frequency:
eval_step_interval = 10
# Set the training-set batch size:
train_batch_size = 100
# Set the validation-set batch size:
validation_batch_size = 100 # When set to -1, uses entire set.
# Set the test-set batch size:
test_batch_size = -1 # When set to -1, uses entire set.
# Set flag stating whether to print all misclassified test images:
print_misclassified_test_images = False
# Set flag stating whether to randomly flip half of the training images horizontally:
flip_left_right = False
# Set percentage of margin to randomly crop off training images:
random_crop = 0
# Set percentage to randomly scale up training images:
random_scale = 0
# Set percentage to randomly multiply brightness of training images:
random_brightness = 0
# Set size parameters:
BOTTLENECK_TENSOR_SIZE = 2048
MODEL_INPUT_WIDTH = 299
MODEL_INPUT_HEIGHT = 299
MODEL_INPUT_DEPTH = 3
MAX_NUM_IMAGES_PER_CLASS = 2 ** 27 - 1 # ~134M
In [ ]:
# Define retraining images getter function:
def maybe_download_and_extract_images():
"""Download and extract model tar file."""
data_url = 'http://download.tensorflow.org/example_images/flower_photos.tgz'
if not os.path.exists(model_dir):
os.makedirs(model_dir)
filename = data_url.split('/')[-1]
filepath = os.path.join(model_dir, filename)
if not os.path.exists(filepath):
def _progress(count, block_size, total_size):
sys.stdout.write('\r>> Downloading %s %.1f%%' % (
filename, float(count * block_size) / float(total_size) * 100.0))
sys.stdout.flush()
filepath, _ = urllib.request.urlretrieve(data_url, filepath, _progress)
print()
statinfo = os.stat(filepath)
print('Successfully downloaded', filename, statinfo.st_size, 'bytes.')
tarfile.open(filepath, 'r:gz').extractall(model_dir)
# Get retraining images:
maybe_download_and_extract_images()
In [ ]:
def create_image_lists(image_dir, testing_percentage, validation_percentage):
"""Builds a list of training images from the file system.
Analyzes the sub folders in the image directory, splits them into stable
training, testing, and validation sets, and returns a data structure
describing the lists of images for each label and their paths.
Args:
image_dir: String path to a folder containing subfolders of images.
testing_percentage: Integer percentage of the images to reserve for tests.
validation_percentage: Integer percentage of images reserved for validation.
Returns:
A dictionary containing an entry for each label subfolder, with images split
into training, testing, and validation sets within each label.
"""
if not gfile.Exists(image_dir):
print("Image directory '" + image_dir + "' not found.")
return None
result = {}
sub_dirs = [x[0] for x in gfile.Walk(image_dir)]
# The root directory comes first, so skip it.
is_root_dir = True
for sub_dir in sub_dirs:
if is_root_dir:
is_root_dir = False
continue
extensions = ['jpg', 'jpeg', 'JPG', 'JPEG']
file_list = []
dir_name = os.path.basename(sub_dir)
if dir_name == image_dir:
continue
print("Looking for images in '" + dir_name + "'")
for extension in extensions:
file_glob = os.path.join(image_dir, dir_name, '*.' + extension)
file_list.extend(gfile.Glob(file_glob))
if not file_list:
print('No files found')
continue
if len(file_list) < 20:
print('WARNING: Folder has less than 20 images, which may cause issues.')
elif len(file_list) > MAX_NUM_IMAGES_PER_CLASS:
print('WARNING: Folder {} has more than {} images. Some images will '
'never be selected.'.format(dir_name, MAX_NUM_IMAGES_PER_CLASS))
label_name = re.sub(r'[^a-z0-9]+', ' ', dir_name.lower())
training_images = []
testing_images = []
validation_images = []
for file_name in file_list:
base_name = os.path.basename(file_name)
# We want to ignore anything after '_nohash_' in the file name when
# deciding which set to put an image in, the data set creator has a way of
# grouping photos that are close variations of each other. For example
# this is used in the plant disease data set to group multiple pictures of
# the same leaf.
hash_name = re.sub(r'_nohash_.*$', '', file_name)
# This looks a bit magical, but we need to decide whether this file should
# go into the training, testing, or validation sets, and we want to keep
# existing files in the same set even if more files are subsequently
# added.
# To do that, we need a stable way of deciding based on just the file name
# itself, so we do a hash of that and then use that to generate a
# probability value that we use to assign it.
hash_name_hashed = hashlib.sha1(compat.as_bytes(hash_name)).hexdigest()
percentage_hash = ((int(hash_name_hashed, 16) %
(MAX_NUM_IMAGES_PER_CLASS + 1)) *
(100.0 / MAX_NUM_IMAGES_PER_CLASS))
if percentage_hash < validation_percentage:
validation_images.append(base_name)
elif percentage_hash < (testing_percentage + validation_percentage):
testing_images.append(base_name)
else:
training_images.append(base_name)
result[label_name] = {
'dir': dir_name,
'training': training_images,
'testing': testing_images,
'validation': validation_images,
}
return result
def get_image_path(image_lists, label_name, index, image_dir, category):
""""Returns a path to an image for a label at the given index.
Args:
image_lists: Dictionary of training images for each label.
label_name: Label string we want to get an image for.
index: Int offset of the image we want. This will be moduloed by the
available number of images for the label, so it can be arbitrarily large.
image_dir: Root folder string of the subfolders containing the training
images.
category: Name string of set to pull images from - training, testing, or
validation.
Returns:
File system path string to an image that meets the requested parameters.
"""
if label_name not in image_lists:
tf.logging.fatal('Label does not exist %s.', label_name)
label_lists = image_lists[label_name]
if category not in label_lists:
tf.logging.fatal('Category does not exist %s.', category)
category_list = label_lists[category]
if not category_list:
tf.logging.fatal('Label %s has no images in the category %s.',
label_name, category)
mod_index = index % len(category_list)
base_name = category_list[mod_index]
sub_dir = label_lists['dir']
full_path = os.path.join(image_dir, sub_dir, base_name)
return full_path
def get_bottleneck_path(image_lists, label_name, index, bottleneck_dir,
category):
""""Returns a path to a bottleneck file for a label at the given index.
Args:
image_lists: Dictionary of training images for each label.
label_name: Label string we want to get an image for.
index: Integer offset of the image we want. This will be moduloed by the
available number of images for the label, so it can be arbitrarily large.
bottleneck_dir: Folder string holding cached files of bottleneck values.
category: Name string of set to pull images from - training, testing, or
validation.
Returns:
File system path string to an image that meets the requested parameters.
"""
return get_image_path(image_lists, label_name, index, bottleneck_dir,
category) + '.txt'
def create_inception_graph():
""""Creates a graph from saved GraphDef file and returns a Graph object.
Returns:
Graph holding the trained Inception network, and various tensors we'll be
manipulating.
"""
with tf.Session() as sess:
model_filename = os.path.join(
model_dir, 'classify_image_graph_def.pb')
with gfile.FastGFile(model_filename, '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=[
'pool_3/_reshape:0', 'DecodeJpeg/contents:0',
'ResizeBilinear:0']))
return sess.graph, bottleneck_tensor, jpeg_data_tensor, resized_input_tensor
def run_bottleneck_on_image(sess, image_data, image_data_tensor,
bottleneck_tensor):
"""Runs inference on an image to extract the 'bottleneck' summary layer.
Args:
sess: Current active TensorFlow Session.
image_data: String of raw JPEG data.
image_data_tensor: Input data layer in the graph.
bottleneck_tensor: Layer before the final softmax.
Returns:
Numpy array of bottleneck values.
"""
bottleneck_values = sess.run(
bottleneck_tensor,
{image_data_tensor: image_data})
bottleneck_values = np.squeeze(bottleneck_values)
return bottleneck_values
def maybe_download_and_extract():
"""Download and extract model tar file.
If the pretrained model we're using doesn't already exist, this function
downloads it from the TensorFlow.org website and unpacks it into a directory.
"""
data_url = 'http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz'
if not os.path.exists(model_dir):
os.makedirs(model_dir)
filename = data_url.split('/')[-1]
filepath = os.path.join(model_dir, filename)
if not os.path.exists(filepath):
def _progress(count, block_size, total_size):
sys.stdout.write('\r>> Downloading %s %.1f%%' %
(filename,
float(count * block_size) / float(total_size) * 100.0))
sys.stdout.flush()
filepath, _ = urllib.request.urlretrieve(data_url,
filepath,
_progress)
print()
statinfo = os.stat(filepath)
print('Successfully downloaded', filename, statinfo.st_size, 'bytes.')
tarfile.open(filepath, 'r:gz').extractall(model_dir)
def ensure_dir_exists(dir_name):
"""Makes sure the folder exists on disk.
Args:
dir_name: Path string to the folder we want to create.
"""
if not os.path.exists(dir_name):
os.makedirs(dir_name)
def write_list_of_floats_to_file(list_of_floats , file_path):
"""Writes a given list of floats to a binary file.
Args:
list_of_floats: List of floats we want to write to a file.
file_path: Path to a file where list of floats will be stored.
"""
s = struct.pack('d' * BOTTLENECK_TENSOR_SIZE, *list_of_floats)
with open(file_path, 'wb') as f:
f.write(s)
def read_list_of_floats_from_file(file_path):
"""Reads list of floats from a given file.
Args:
file_path: Path to a file where list of floats was stored.
Returns:
Array of bottleneck values (list of floats).
"""
with open(file_path, 'rb') as f:
s = struct.unpack('d' * BOTTLENECK_TENSOR_SIZE, f.read())
return list(s)
bottleneck_path_2_bottleneck_values = {}
def create_bottleneck_file(bottleneck_path, image_lists, label_name, index,
image_dir, category, sess, jpeg_data_tensor, bottleneck_tensor):
print('Creating bottleneck at ' + bottleneck_path)
image_path = get_image_path(image_lists, label_name, index, image_dir, category)
if not gfile.Exists(image_path):
tf.logging.fatal('File does not exist %s', image_path)
image_data = gfile.FastGFile(image_path, 'rb').read()
bottleneck_values = run_bottleneck_on_image(sess, image_data, jpeg_data_tensor, bottleneck_tensor)
bottleneck_string = ','.join(str(x) for x in bottleneck_values)
with open(bottleneck_path, 'w') as bottleneck_file:
bottleneck_file.write(bottleneck_string)
def get_or_create_bottleneck(sess, image_lists, label_name, index, image_dir,
category, bottleneck_dir, jpeg_data_tensor,
bottleneck_tensor):
"""Retrieves or calculates bottleneck values for an image.
If a cached version of the bottleneck data exists on-disk, return that,
otherwise calculate the data and save it to disk for future use.
Args:
sess: The current active TensorFlow Session.
image_lists: Dictionary of training images for each label.
label_name: Label string we want to get an image for.
index: Integer offset of the image we want. This will be modulo-ed by the
available number of images for the label, so it can be arbitrarily large.
image_dir: Root folder string of the subfolders containing the training
images.
category: Name string of which set to pull images from - training, testing,
or validation.
bottleneck_dir: Folder string holding cached files of bottleneck values.
jpeg_data_tensor: The tensor to feed loaded jpeg data into.
bottleneck_tensor: The output tensor for the bottleneck values.
Returns:
Numpy array of values produced by the bottleneck layer for the image.
"""
label_lists = image_lists[label_name]
sub_dir = label_lists['dir']
sub_dir_path = os.path.join(bottleneck_dir, sub_dir)
ensure_dir_exists(sub_dir_path)
bottleneck_path = get_bottleneck_path(image_lists, label_name, index, bottleneck_dir, category)
if not os.path.exists(bottleneck_path):
create_bottleneck_file(bottleneck_path, image_lists, label_name, index, image_dir, category, sess, jpeg_data_tensor, bottleneck_tensor)
with open(bottleneck_path, 'r') as bottleneck_file:
bottleneck_string = bottleneck_file.read()
did_hit_error = False
try:
bottleneck_values = [float(x) for x in bottleneck_string.split(',')]
except:
print("Invalid float found, recreating bottleneck")
did_hit_error = True
if did_hit_error:
create_bottleneck_file(bottleneck_path, image_lists, label_name, index, image_dir, category, sess, jpeg_data_tensor, bottleneck_tensor)
with open(bottleneck_path, 'r') as bottleneck_file:
bottleneck_string = bottleneck_file.read()
# Allow exceptions to propagate here, since they shouldn't happen after a fresh creation
bottleneck_values = [float(x) for x in bottleneck_string.split(',')]
return bottleneck_values
def cache_bottlenecks(sess, image_lists, image_dir, bottleneck_dir,
jpeg_data_tensor, bottleneck_tensor):
"""Ensures all the training, testing, and validation bottlenecks are cached.
Because we're likely to read the same image multiple times (if there are no
distortions applied during training) it can speed things up a lot if we
calculate the bottleneck layer values once for each image during
preprocessing, and then just read those cached values repeatedly during
training. Here we go through all the images we've found, calculate those
values, and save them off.
Args:
sess: The current active TensorFlow Session.
image_lists: Dictionary of training images for each label.
image_dir: Root folder string of the subfolders containing the training
images.
bottleneck_dir: Folder string holding cached files of bottleneck values.
jpeg_data_tensor: Input tensor for jpeg data from file.
bottleneck_tensor: The penultimate output layer of the graph.
Returns:
Nothing.
"""
how_many_bottlenecks = 0
ensure_dir_exists(bottleneck_dir)
for label_name, label_lists in image_lists.items():
for category in ['training', 'testing', 'validation']:
category_list = label_lists[category]
for index, unused_base_name in enumerate(category_list):
get_or_create_bottleneck(sess, image_lists, label_name, index,
image_dir, category, bottleneck_dir,
jpeg_data_tensor, bottleneck_tensor)
how_many_bottlenecks += 1
if how_many_bottlenecks % 100 == 0:
print(str(how_many_bottlenecks) + ' bottleneck files created.')
def get_random_cached_bottlenecks(sess, image_lists, how_many, category,
bottleneck_dir, image_dir, jpeg_data_tensor,
bottleneck_tensor):
"""Retrieves bottleneck values for cached images.
If no distortions are being applied, this function can retrieve the cached
bottleneck values directly from disk for images. It picks a random set of
images from the specified category.
Args:
sess: Current TensorFlow Session.
image_lists: Dictionary of training images for each label.
how_many: If positive, a random sample of this size will be chosen.
If negative, all bottlenecks will be retrieved.
category: Name string of which set to pull from - training, testing, or
validation.
bottleneck_dir: Folder string holding cached files of bottleneck values.
image_dir: Root folder string of the subfolders containing the training
images.
jpeg_data_tensor: The layer to feed jpeg image data into.
bottleneck_tensor: The bottleneck output layer of the CNN graph.
Returns:
List of bottleneck arrays, their corresponding ground truths, and the
relevant filenames.
"""
class_count = len(image_lists.keys())
bottlenecks = []
ground_truths = []
filenames = []
if how_many >= 0:
# Retrieve a random sample of bottlenecks.
for unused_i in range(how_many):
label_index = random.randrange(class_count)
label_name = list(image_lists.keys())[label_index]
image_index = random.randrange(MAX_NUM_IMAGES_PER_CLASS + 1)
image_name = get_image_path(image_lists, label_name, image_index,
image_dir, category)
bottleneck = get_or_create_bottleneck(sess, image_lists, label_name,
image_index, image_dir, category,
bottleneck_dir, jpeg_data_tensor,
bottleneck_tensor)
ground_truth = np.zeros(class_count, dtype=np.float32)
ground_truth[label_index] = 1.0
bottlenecks.append(bottleneck)
ground_truths.append(ground_truth)
filenames.append(image_name)
else:
# Retrieve all bottlenecks.
for label_index, label_name in enumerate(image_lists.keys()):
for image_index, image_name in enumerate(
image_lists[label_name][category]):
image_name = get_image_path(image_lists, label_name, image_index,
image_dir, category)
bottleneck = get_or_create_bottleneck(sess, image_lists, label_name,
image_index, image_dir, category,
bottleneck_dir, jpeg_data_tensor,
bottleneck_tensor)
ground_truth = np.zeros(class_count, dtype=np.float32)
ground_truth[label_index] = 1.0
bottlenecks.append(bottleneck)
ground_truths.append(ground_truth)
filenames.append(image_name)
return bottlenecks, ground_truths, filenames
def get_random_distorted_bottlenecks(
sess, image_lists, how_many, category, image_dir, input_jpeg_tensor,
distorted_image, resized_input_tensor, bottleneck_tensor):
"""Retrieves bottleneck values for training images, after distortions.
If we're training with distortions like crops, scales, or flips, we have to
recalculate the full model for every image, and so we can't use cached
bottleneck values. Instead we find random images for the requested category,
run them through the distortion graph, and then the full graph to get the
bottleneck results for each.
Args:
sess: Current TensorFlow Session.
image_lists: Dictionary of training images for each label.
how_many: The integer number of bottleneck values to return.
category: Name string of which set of images to fetch - training, testing,
or validation.
image_dir: Root folder string of the subfolders containing the training
images.
input_jpeg_tensor: The input layer we feed the image data to.
distorted_image: The output node of the distortion graph.
resized_input_tensor: The input node of the recognition graph.
bottleneck_tensor: The bottleneck output layer of the CNN graph.
Returns:
List of bottleneck arrays and their corresponding ground truths.
"""
class_count = len(image_lists.keys())
bottlenecks = []
ground_truths = []
for unused_i in range(how_many):
label_index = random.randrange(class_count)
label_name = list(image_lists.keys())[label_index]
image_index = random.randrange(MAX_NUM_IMAGES_PER_CLASS + 1)
image_path = get_image_path(image_lists, label_name, image_index, image_dir,
category)
if not gfile.Exists(image_path):
tf.logging.fatal('File does not exist %s', image_path)
jpeg_data = gfile.FastGFile(image_path, 'rb').read()
# Note that we materialize the distorted_image_data as a numpy array before
# sending running inference on the image. This involves 2 memory copies and
# might be optimized in other implementations.
distorted_image_data = sess.run(distorted_image,
{input_jpeg_tensor: jpeg_data})
bottleneck = run_bottleneck_on_image(sess, distorted_image_data,
resized_input_tensor,
bottleneck_tensor)
ground_truth = np.zeros(class_count, dtype=np.float32)
ground_truth[label_index] = 1.0
bottlenecks.append(bottleneck)
ground_truths.append(ground_truth)
return bottlenecks, ground_truths
def should_distort_images(flip_left_right, random_crop, random_scale,
random_brightness):
"""Whether any distortions are enabled, from the input flags.
Args:
flip_left_right: Boolean whether to randomly mirror images horizontally.
random_crop: Integer percentage setting the total margin used around the
crop box.
random_scale: Integer percentage of how much to vary the scale by.
random_brightness: Integer range to randomly multiply the pixel values by.
Returns:
Boolean value indicating whether any distortions should be applied.
"""
return (flip_left_right or (random_crop != 0) or (random_scale != 0) or
(random_brightness != 0))
def add_input_distortions(flip_left_right, random_crop, random_scale,
random_brightness):
"""Creates the operations to apply the specified distortions.
During training it can help to improve the results if we run the images
through simple distortions like crops, scales, and flips. These reflect the
kind of variations we expect in the real world, and so can help train the
model to cope with natural data more effectively. Here we take the supplied
parameters and construct a network of operations to apply them to an image.
Cropping
~~~~~~~~
Cropping is done by placing a bounding box at a random position in the full
image. The cropping parameter controls the size of that box relative to the
input image. If it's zero, then the box is the same size as the input and no
cropping is performed. If the value is 50%, then the crop box will be half the
width and height of the input. In a diagram it looks like this:
< width >
+---------------------+
| |
| width - crop% |
| < > |
| +------+ |
| | | |
| | | |
| | | |
| +------+ |
| |
| |
+---------------------+
Scaling
~~~~~~~
Scaling is a lot like cropping, except that the bounding box is always
centered and its size varies randomly within the given range. For example if
the scale percentage is zero, then the bounding box is the same size as the
input and no scaling is applied. If it's 50%, then the bounding box will be in
a random range between half the width and height and full size.
Args:
flip_left_right: Boolean whether to randomly mirror images horizontally.
random_crop: Integer percentage setting the total margin used around the
crop box.
random_scale: Integer percentage of how much to vary the scale by.
random_brightness: Integer range to randomly multiply the pixel values by.
graph.
Returns:
The jpeg input layer and the distorted result tensor.
"""
jpeg_data = tf.placeholder(tf.string, name='DistortJPGInput')
decoded_image = tf.image.decode_jpeg(jpeg_data, channels=MODEL_INPUT_DEPTH)
decoded_image_as_float = tf.cast(decoded_image, dtype=tf.float32)
decoded_image_4d = tf.expand_dims(decoded_image_as_float, 0)
margin_scale = 1.0 + (random_crop / 100.0)
resize_scale = 1.0 + (random_scale / 100.0)
margin_scale_value = tf.constant(margin_scale)
resize_scale_value = tf.random_uniform(tensor_shape.scalar(),
minval=1.0,
maxval=resize_scale)
scale_value = tf.multiply(margin_scale_value, resize_scale_value)
precrop_width = tf.multiply(scale_value, MODEL_INPUT_WIDTH)
precrop_height = tf.multiply(scale_value, MODEL_INPUT_HEIGHT)
precrop_shape = tf.stack([precrop_height, precrop_width])
precrop_shape_as_int = tf.cast(precrop_shape, dtype=tf.int32)
precropped_image = tf.image.resize_bilinear(decoded_image_4d,
precrop_shape_as_int)
precropped_image_3d = tf.squeeze(precropped_image, squeeze_dims=[0])
cropped_image = tf.random_crop(precropped_image_3d,
[MODEL_INPUT_HEIGHT, MODEL_INPUT_WIDTH,
MODEL_INPUT_DEPTH])
if flip_left_right:
flipped_image = tf.image.random_flip_left_right(cropped_image)
else:
flipped_image = cropped_image
brightness_min = 1.0 - (random_brightness / 100.0)
brightness_max = 1.0 + (random_brightness / 100.0)
brightness_value = tf.random_uniform(tensor_shape.scalar(),
minval=brightness_min,
maxval=brightness_max)
brightened_image = tf.multiply(flipped_image, brightness_value)
distort_result = tf.expand_dims(brightened_image, 0, name='DistortResult')
return jpeg_data, distort_result
def variable_summaries(var):
"""Attach a lot of summaries to a Tensor (for TensorBoard visualization)."""
with tf.name_scope('summaries'):
mean = tf.reduce_mean(var)
tf.summary.scalar('mean', mean)
with tf.name_scope('stddev'):
stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean)))
tf.summary.scalar('stddev', stddev)
tf.summary.scalar('max', tf.reduce_max(var))
tf.summary.scalar('min', tf.reduce_min(var))
tf.summary.histogram('histogram', var)
def add_final_training_ops(class_count, final_tensor_name, bottleneck_tensor):
"""Adds a new softmax and fully-connected layer for training.
We need to retrain the top layer to identify our new classes, so this function
adds the right operations to the graph, along with some variables to hold the
weights, and then sets up all the gradients for the backward pass.
The set up for the softmax and fully-connected layers is based on:
https://tensorflow.org/versions/master/tutorials/mnist/beginners/index.html
Args:
class_count: Integer of how many categories of things we're trying to
recognize.
final_tensor_name: Name string for the new final node that produces results.
bottleneck_tensor: The output of the main CNN graph.
Returns:
The tensors for the training and cross entropy results, and tensors for the
bottleneck input and ground truth input.
"""
with tf.name_scope('input'):
bottleneck_input = tf.placeholder_with_default(
bottleneck_tensor, shape=[None, BOTTLENECK_TENSOR_SIZE],
name='BottleneckInputPlaceholder')
ground_truth_input = tf.placeholder(tf.float32,
[None, class_count],
name='GroundTruthInput')
# Organizing the following ops as `final_training_ops` so they're easier
# to see in TensorBoard
layer_name = 'final_training_ops'
with tf.name_scope(layer_name):
with tf.name_scope('weights'):
layer_weights = tf.Variable(tf.truncated_normal([BOTTLENECK_TENSOR_SIZE, class_count], stddev=0.001), name='final_weights')
variable_summaries(layer_weights)
with tf.name_scope('biases'):
layer_biases = tf.Variable(tf.zeros([class_count]), name='final_biases')
variable_summaries(layer_biases)
with tf.name_scope('Wx_plus_b'):
logits = tf.matmul(bottleneck_input, layer_weights) + layer_biases
tf.summary.histogram('pre_activations', logits)
final_tensor = tf.nn.softmax(logits, name=final_tensor_name)
tf.summary.histogram('activations', final_tensor)
with tf.name_scope('cross_entropy'):
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
labels=ground_truth_input, logits=logits)
with tf.name_scope('total'):
cross_entropy_mean = tf.reduce_mean(cross_entropy)
tf.summary.scalar('cross_entropy', cross_entropy_mean)
with tf.name_scope('train'):
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(
cross_entropy_mean)
return (train_step, cross_entropy_mean, bottleneck_input, ground_truth_input,
final_tensor)
def add_evaluation_step(result_tensor, ground_truth_tensor):
"""Inserts the operations we need to evaluate the accuracy of our results.
Args:
result_tensor: The new final node that produces results.
ground_truth_tensor: The node we feed ground truth data
into.
Returns:
Tuple of (evaluation step, prediction).
"""
with tf.name_scope('accuracy'):
with tf.name_scope('correct_prediction'):
prediction = tf.argmax(result_tensor, 1)
correct_prediction = tf.equal(
prediction, tf.argmax(ground_truth_tensor, 1))
with tf.name_scope('accuracy'):
evaluation_step = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
tf.summary.scalar('accuracy', evaluation_step)
return evaluation_step, prediction
In [ ]:
def retrain():
# Setup the directory we'll write summaries to for TensorBoard
if tf.gfile.Exists(summaries_dir):
tf.gfile.DeleteRecursively(summaries_dir)
tf.gfile.MakeDirs(summaries_dir)
# Set up the pre-trained graph.
maybe_download_and_extract()
graph, bottleneck_tensor, jpeg_data_tensor, resized_image_tensor = (
create_inception_graph())
# Look at the folder structure, and create lists of all the images.
image_lists = create_image_lists(image_dir, testing_percentage,
validation_percentage)
class_count = len(image_lists.keys())
if class_count == 0:
print('No valid folders of images found at ' + image_dir)
return -1
if class_count == 1:
print('Only one valid folder of images found at ' + image_dir +
' - multiple classes are needed for classification.')
return -1
# See if the command-line flags mean we're applying any distortions.
do_distort_images = should_distort_images(
flip_left_right, random_crop, random_scale,
random_brightness)
sess = tf.Session()
if do_distort_images:
# We will be applying distortions, so setup the operations we'll need.
distorted_jpeg_data_tensor, distorted_image_tensor = add_input_distortions(
flip_left_right, random_crop, random_scale,
random_brightness)
else:
# We'll make sure we've calculated the 'bottleneck' image summaries and
# cached them on disk.
cache_bottlenecks(sess, image_lists, image_dir, bottleneck_dir,
jpeg_data_tensor, bottleneck_tensor)
# Add the new layer that we'll be training.
(train_step, cross_entropy, bottleneck_input, ground_truth_input,
final_tensor) = add_final_training_ops(len(image_lists.keys()),
final_tensor_name,
bottleneck_tensor)
# Create the operations we need to evaluate the accuracy of our new layer.
evaluation_step, prediction = add_evaluation_step(
final_tensor, ground_truth_input)
# Merge all the summaries and write them out to /tmp/retrain_logs (by default)
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(summaries_dir + '/train',
sess.graph)
validation_writer = tf.summary.FileWriter(summaries_dir + '/validation')
# Set up all our weights to their initial default values.
init = tf.global_variables_initializer()
sess.run(init)
# Run the training for as many cycles as requested on the command line.
for i in range(how_many_training_steps):
# Get a batch of input bottleneck values, either calculated fresh every time
# with distortions applied, or from the cache stored on disk.
if do_distort_images:
train_bottlenecks, train_ground_truth = get_random_distorted_bottlenecks(
sess, image_lists, train_batch_size, 'training',
image_dir, distorted_jpeg_data_tensor,
distorted_image_tensor, resized_image_tensor, bottleneck_tensor)
else:
train_bottlenecks, train_ground_truth, _ = get_random_cached_bottlenecks(
sess, image_lists, train_batch_size, 'training',
bottleneck_dir, image_dir, jpeg_data_tensor,
bottleneck_tensor)
# Feed the bottlenecks and ground truth into the graph, and run a training
# step. Capture training summaries for TensorBoard with the `merged` op.
train_summary, _ = sess.run([merged, train_step],
feed_dict={bottleneck_input: train_bottlenecks,
ground_truth_input: train_ground_truth})
train_writer.add_summary(train_summary, i)
# Every so often, print out how well the graph is training.
is_last_step = (i + 1 == how_many_training_steps)
if (i % eval_step_interval) == 0 or is_last_step:
train_accuracy, cross_entropy_value = sess.run(
[evaluation_step, cross_entropy],
feed_dict={bottleneck_input: train_bottlenecks,
ground_truth_input: train_ground_truth})
print('%s: Step %d: Train accuracy = %.1f%%' % (datetime.now(), i,
train_accuracy * 100))
print('%s: Step %d: Cross entropy = %f' % (datetime.now(), i,
cross_entropy_value))
validation_bottlenecks, validation_ground_truth, _ = (
get_random_cached_bottlenecks(
sess, image_lists, validation_batch_size, 'validation',
bottleneck_dir, image_dir, jpeg_data_tensor,
bottleneck_tensor))
# Run a validation step and capture training summaries for TensorBoard
# with the `merged` op.
validation_summary, validation_accuracy = sess.run(
[merged, evaluation_step],
feed_dict={bottleneck_input: validation_bottlenecks,
ground_truth_input: validation_ground_truth})
validation_writer.add_summary(validation_summary, i)
print('%s: Step %d: Validation accuracy = %.1f%% (N=%d)' %
(datetime.now(), i, validation_accuracy * 100,
len(validation_bottlenecks)))
# We've completed all our training, so run a final test evaluation on
# some new images we haven't used before.
test_bottlenecks, test_ground_truth, test_filenames = (
get_random_cached_bottlenecks(sess, image_lists, test_batch_size,
'testing', bottleneck_dir,
image_dir, jpeg_data_tensor,
bottleneck_tensor))
test_accuracy, predictions = sess.run(
[evaluation_step, prediction],
feed_dict={bottleneck_input: test_bottlenecks,
ground_truth_input: test_ground_truth})
print('Final test accuracy = %.1f%% (N=%d)' % (
test_accuracy * 100, len(test_bottlenecks)))
if print_misclassified_test_images:
print('=== MISCLASSIFIED TEST IMAGES ===')
for i, test_filename in enumerate(test_filenames):
if predictions[i] != test_ground_truth[i].argmax():
print('%70s %s' % (test_filename,
list(image_lists.keys())[predictions[i]]))
# Write out the trained graph and labels with the weights stored as constants.
output_graph_def = graph_util.convert_variables_to_constants(
sess, graph.as_graph_def(), [final_tensor_name])
with gfile.FastGFile(output_graph, 'wb') as f:
f.write(output_graph_def.SerializeToString())
with gfile.FastGFile(output_labels, 'w') as f:
f.write('\n'.join(image_lists.keys()) + '\n')
In [ ]:
# Perform retraining:
retrain()
In [ ]:
# Set the number of predictions to display:
num_top_predictions = 5
# Define human-readable string lookup function:
def loadLookupTable(uid_lookup_path):
if not tf.gfile.Exists(uid_lookup_path):
tf.logging.fatal('File does not exist %s', uid_lookup_path)
idx_to_human = []
file_lines = tf.gfile.GFile(uid_lookup_path).readlines()
for line in file_lines:
idx_to_human.append( line.rstrip('\n') )
return idx_to_human
# Define graph loader function:
def create_graph():
with tf.gfile.FastGFile(os.path.join(
model_dir, 'output_graph.pb'), 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(graph_def, name='')
# Define image recognition procedure function:
def run_inference_on_image(image):
if not tf.gfile.Exists(image):
tf.logging.fatal('File does not exist %s', image)
image_data = tf.gfile.FastGFile(image, 'rb').read()
# Create graph:
create_graph()
# Create lookup table:
idx_to_human = loadLookupTable( os.path.join(model_dir, 'output_labels.txt') )
with tf.Session() as sess:
softmax_tensor = sess.graph.get_tensor_by_name(final_tensor_name + ':0')
predictions = sess.run(softmax_tensor,
{'DecodeJpeg/contents:0': image_data})
predictions = np.squeeze(predictions)
top_k = predictions.argsort()[-num_top_predictions:][::-1]
for node_id in top_k:
human_string = idx_to_human[node_id]
score = predictions[node_id]
print('%s (score = %.5f)' % (human_string, score))
# Construct image path:
image_path = 'data/imagenet_retrain/flower_photos/tulips/65347450_53658c63bd_n.jpg'
# Classify image:
run_inference_on_image( image_path )