def get_bottleneck_tensor(input_jpeg_str):
# type: tf.Tensor -> tf.Tensor
"""Calculates the bottleneck tensor for input JPEG string tensor.
This function will resize/encode the image as required by Inception V3 model.
Then it will run it through the InceptionV3 checkpoint to calculate
bottleneck values.
Args:
input_jpeg_str: Tensor for input JPEG image.
Returns:
bottleneck_tensor: Tensor for output bottleneck Tensor.
"""
module_spec = tensorflow_hub.load_module_spec(_FEATURE_VECTORS_MODULE_URL)
input_height, input_width = tensorflow_hub.get_expected_image_size(
module_spec)
input_depth = tensorflow_hub.get_num_image_channels(module_spec)
decoded_image = tf.image.decode_jpeg(input_jpeg_str, channels=input_depth)
decoded_image_as_float = tf.image.convert_image_dtype(
decoded_image, tf.float32)
decoded_image_4d = tf.expand_dims(decoded_image_as_float, 0)
resize_shape = tf.stack([input_height, input_width])
resize_shape_as_int = tf.cast(resize_shape, dtype=tf.int32)
resized_image_4d = tf.image.resize_bilinear(decoded_image_4d,
resize_shape_as_int)
m = tensorflow_hub.Module(module_spec)
bottleneck_tensor = m(resized_image_4d)
return bottleneck_tensor
def extract_inception_bottleneck_features(images_train, images_test):
# Combine train & test set to single set of images
images = images_train + images_test
print('Extracting inception bottleneck features...')
module_spec = hub.load_module_spec(TFHUB_INCEPTION_V3_MODULE_SPEC_URL)
module = hub.Module(module_spec)
(image_height, image_width) = hub.get_expected_image_size(module)
images = [tf.image.convert_image_dtype(x, tf.float32) for x in images]
images = [
tf.image.resize_images(x, (image_height, image_width)) for x in images
]
sess = tf.Session()
m = hub.Module(module_spec)
X = []
sess.run(tf.global_variables_initializer())
batches = [
images[i:i + BATCH_SIZE] for i in range(0, len(images), BATCH_SIZE)
]
for batch in batches:
bottleneck_tensors = m(batch)
x_batch = sess.run(bottleneck_tensors)
X.extend(x_batch)
# Recover train & test set
X_train = X[:len(images_train)]
X_test = X[len(images_train):]
return (X_train, X_test)
def tfhub_encoder(x, dropout_pr=1.0):
"""Applies TFHub encoder to batch of images.
Args:
x: Images on [0, 255] sized (batch_size, scale_size, scale_size, 3).
Returns:
enc_x: Encodings sized (batch_size, encoding_size).
"""
x = x / 255.
#module_spec_str = ('https://tfhub.dev/google/imagenet/inception_v3/'
# 'feature_vector/1')
# This module takes (224, 224) and encodes to (1280).
module_spec_str = (
'https://tfhub.dev/google/imagenet/mobilenet_v2_035_224/'
'feature_vector/2')
module_spec = hub.load_module_spec(module_spec_str)
height, width = hub.get_expected_image_size(module_spec)
assert x.shape[1] == height, 'height is {}. Must be {}'.format(
x.shape[1], height)
assert x.shape[2] == width, 'width is {}. Must be {}'.format(
x.shape[2], width)
module = hub.Module(module_spec)
embedding_tensor = module(x)
batch_size, embedding_tensor_size = embedding_tensor.get_shape().as_list()
#assert batch_size is None, 'We want to work with arbitrary batch size.'
return embedding_tensor
def run(constant_overwrites):
config_path = os.path.join(os.path.dirname(__file__), 'hyperparams.yml')
constants = merge_dict(load_hyperparams(config_path), constant_overwrites)
data_generator = ImageDataGenerator(rescale=1/255, rotation_range=90, width_shift_range=0.2,
height_shift_range=0.2, horizontal_flip=True)
print('Load', constants['module_spec'])
module_spec = hub.load_module_spec(constants['module_spec'])
image_size, _ = hub.get_expected_image_size(module_spec)
# n_channels = hub.get_num_image_channels(module_spec)
# project_dir = 'tmp/semihard_full_' + 'time:' + str(int(time()))[-3:] +\
# '/top:lambda:' + str(constants['lambda_reg']) +\
# 'margin:' + str(constants['tl_margin'])
project_dir = '/Users/d777710/src/DeepLearning/vision'
print('Project dir:', project_dir)
_, _, bottleneck_config = get_bottleneck_config(os.path.join(project_dir, constants['bottleneck_dir']),
os.path.join(project_dir, constants['splits_dir']))
bottleneck_flow_gen = ImageFlowGenerator(bottleneck_config, mode='bottleneck')
constants.update({
'train_dir': os.path.join(project_dir, constants['train_subdir']),
'top_model_dir': os.path.join(project_dir, constants['top_model_subdir']),
'val_dir': os.path.join(project_dir, constants['val_subdir']),
'top_model_val_dir': os.path.join(project_dir, constants['top_model_val_subdir']),
'data_flow_gen': bottleneck_flow_gen,
'eval_every_n_steps': 5,
'generator': data_generator,
'image_size': image_size
})
model = SemiHardModel(constants, train_top_only=True)
run_training(model, constants)
def __init__(self, module_spec):
self._module = None
self._module_spec = tfhub.load_module_spec(module_spec)
self._module_spec_path = module_spec
self._sess = None
self._graph = tf.Graph()
self._inputs = None
self._outputs = None
def build(self, input_shape):
self.embedder = tfhub.Module(self.module_uri, trainable=self.trainable)
self.embedder_spec = tfhub.load_module_spec(self.module_uri)
variables_ = [v for v in tensorflow.trainable_variables() if v in self.embedder.variables]
self.trainable_weights.extend(variables_)
self.weights.extend(variables_)
self.trainable_variables.extend(variables_)
super(TFHubTextLayer, self).build(input_shape)
def get_text_module_input_name():
"""Get the tag used for inputs to the text module.
Returns:
a string, probably "default"
"""
module_spec = hub.load_module_spec(FLAGS.module_handle)
return list(module_spec.get_input_info_dict())[0]
def download_image_model(mdl_url):
# type: str -> (tensorflow_hub.Module, int, int, int)
"""Returns the Tensorflow Hub model used to process images."""
module_spec = tensorflow_hub.load_module_spec(mdl_url)
input_height, input_width = tensorflow_hub.get_expected_image_size(
module_spec)
input_depth = tensorflow_hub.get_num_image_channels(module_spec)
m = tensorflow_hub.Module(module_spec)
return (m, input_height, input_width, input_depth)
def eval_from_hub(model_dir, input_fn, eval_steps):
"""Eval using hub module."""
hub_module_spec = hub.load_module_spec(model_dir)
run_config = tf.estimator.RunConfig(model_dir=model_dir)
image_classifier = tf.estimator.Estimator(
model_fn=_make_model_fn(hub_module_spec), config=run_config, params={})
eval_results = image_classifier.evaluate(input_fn=input_fn,
steps=eval_steps)
tf.logging.info('Evaluation results: %s' % eval_results)
def run(dataset_folder, network='inception_v3', batch_size=16):
assert network in modules, 'Invalid network, pick one of %s' % list(modules.keys())
assert dataset_folder is not None
with tf.Graph().as_default():
dataset = basename(dirname(dataset_folder))
filenames_output = os.path.join(dataset_folder, dataset + '_' + network + '_filenames.csv')
labels_output = os.path.join(dataset_folder, dataset + '_' + network + '_labels.csv')
features_output = os.path.join(dataset_folder, dataset + '_' + network + '_features.csv')
module_url = modules[network]
types = ('/*/*.jpg', '/*/*.png')
filenames = []
for files in types:
filenames.extend(glob.glob(dataset_folder + files))
pbar = tqdm(total=len(filenames))
labels = [basename(dirname(f)) for f in filenames]
filenames = tf.constant(filenames)
labels = tf.constant(labels)
module_spec = hub.load_module_spec(module_url)
output_size = module_spec.get_output_info_dict()['default'].get_shape()[1]
height, width = hub.get_expected_image_size(module_spec)
images, labels, files = input_fn(filenames, labels, [height, width], batch_size)
features = np.empty((0, output_size), float)
classes = np.empty(0, int)
filenames = np.empty(0, str)
network = hub.Module(module_spec)
network = network(images), labels, files
with tf.compat.v1.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
sess.run(tf.compat.v1.tables_initializer())
while True:
try:
x, y, f = sess.run(network)
f = [basename(k) for k in f]
filenames = np.append(filenames, f)
classes = np.append(classes, y)
features = np.append(features, x, axis=0)
pbar.update(len(y))
except tf.errors.OutOfRangeError:
break
pbar.close()
# pca = PCA(n_components=100, random_state=1)
# features = pca.fit_transform(features)
np.savetxt(filenames_output, filenames.astype(str), fmt='%s', delimiter=',')
np.savetxt(labels_output, classes.astype(str), fmt='%s', delimiter=',')
np.savetxt(features_output, features, delimiter=',')
def main(_):
# Needed to make sure the logging output is visible.
# See https://github.com/tensorflow/tensorflow/issues/3047
tf.logging.set_verbosity(tf.logging.INFO)
if not FLAGS.image_dir:
tf.logging.error('Must set flag --image_dir.')
return -1
# Prepare necessary directories that can be used during training
#prepare_file_system()
# Look at the folder structure, and create lists of all the images.
image_lists = create_image_lists(FLAGS.image_dir, FLAGS.testing_percentage,
FLAGS.validation_percentage)
class_count = len(image_lists.keys())
if class_count == 0:
tf.logging.error('No valid folders of images found at ' +
FLAGS.image_dir)
return -1
if class_count == 1:
tf.logging.error('Only one valid folder of images found at ' +
FLAGS.image_dir +
' - multiple classes are needed for classification.')
return -1
# Set up the pre-trained graph.
module_spec = hub.load_module_spec(FLAGS.tfhub_module)
graph, bottleneck_tensor, resized_image_tensor, wants_quantization = (
create_module_graph(module_spec))
# Add the new layer that we'll be training.
with graph.as_default():
(train_step, cross_entropy, bottleneck_input, ground_truth_input,
final_tensor) = add_final_retrain_ops(class_count,
FLAGS.final_tensor_name,
bottleneck_tensor,
wants_quantization,
is_training=True)
with tf.Session(graph=graph) as sess:
# Initialize all weights: for the module to their pretrained values,
# and for the newly added retraining layer to random initial values.
init = tf.global_variables_initializer()
sess.run(init)
# Set up the image decoding sub-graph.
jpeg_data_tensor, decoded_image_tensor = add_jpeg_decoding(module_spec)
# We'll make sure we've calculated the 'bottleneck' image summaries and
# cached them on disk.
cache_bottlenecks(sess, image_lists, FLAGS.image_dir,
FLAGS.bottleneck_dir, jpeg_data_tensor,
decoded_image_tensor, resized_image_tensor,
bottleneck_tensor, FLAGS.tfhub_module)
def main(_):
graph = tf.Graph()
with tf.Session(graph=graph) as sess:
tf.saved_model.loader.load(sess, [tag_constants.SERVING],
FLAGS.saved_model_dir)
# resized_input_tensor
image = graph.get_tensor_by_name('Placeholder:0')
prediction = graph.get_tensor_by_name('final_result:0')
module_spec = hub.load_module_spec(FLAGS.tfhub_module)
jpeg_data_tensor, decoded_image_tensor = add_jpeg_decoding(module_spec)
image_lists = create_image_lists(FLAGS.image_dir, FLAGS.subsets,
FLAGS.testing_percentage,
FLAGS.validation_percentage)
hits = 0
total = 0
confusion_matirx = dict()
for image_class in image_lists:
confusion_matirx[image_class] = collections.defaultdict(int)
for image_set in FLAGS.set:
for image_filename in image_lists[image_class][image_set]:
image_path = os.path.join(FLAGS.image_dir, image_filename)
if not tf.gfile.Exists(image_path):
tf.logging.fatal('File does not exist %s', image_path)
image_data = tf.gfile.GFile(image_path, 'rb').read()
resized_image = sess.run(decoded_image_tensor,
{jpeg_data_tensor: image_data})
result = sess.run(prediction, {image: resized_image})
predicted_class = get_lab(result[0],
list(image_lists.keys()))
confusion_matirx[image_class][predicted_class] += 1
if image_class == predicted_class:
hits = hits + 1
total = total + 1
print("Set: {}".format(" ".join(FLAGS.set)))
print("Confusion Matrix")
print("{:>15}".format(""), end="")
for image_class in image_lists:
print("{:>15}".format(image_class), end="")
print()
for image_class in image_lists:
print("{:>15}".format(image_class), end="")
for predicted_class in image_lists:
print("{:>15}".format(
confusion_matirx[image_class][predicted_class]),
end="")
print()
accuracy = hits / total * 100
print("Total Accuracy: {}% (N={})".format(accuracy, total))
def __init__(self,
tf_hub_module_spec=None,
tf_hub_module_path=None,):
"""Creates an instance to extract image features from a pre-trained model.
The model to use may be specified as a TF-hub module (either by ModuleSpec
or path) or as an Inception V4 model checkpoint.
If a TF-hub module is given, it is assumed to conform to the interface
described in [1]. Its default signature should take an input 'images' Tensor
with shape [batch_size, height, width, num_channels=3] and return a
[batch_size, feature_dim] Tensor of features. Pass
`tf_hub_module_spec=make_module_spec_for_testing()` to stub out the model
for tests.
[1]
https://www.tensorflow.org/hub/common_signatures/images#image_feature_vector
Args:
tf_hub_module_spec: `hub.ModuleSpec` or None, the TF-hub module to load.
tf_hub_module_path: str or None, the location of the TF-hub module to load
in a format understood by `load_module_spec()` (URL,
'@internal/module/name', '/on/disk/path', etc.)
Raises:
ValueError: if not exactly one kwarg specifying the model is given.
"""
self.spec_str = None # String describing the model/module being used.
# Input and output tensors for the image to representation computation.
# The output tensor will depend on the model options.
self._input = None
self._output = None
self._session = None
num_kwargs = sum(
int(kwarg is not None) for kwarg in
[tf_hub_module_spec, tf_hub_module_path])
if num_kwargs != 1:
raise ValueError(
'Must provide exactly one of "tf_hub_module_spec", '
'"tf_hub_module_path".')
if tf_hub_module_spec:
self.spec_str = 'user_provided_module'
self._initialize_from_hub_module(tf_hub_module_spec)
elif tf_hub_module_path:
self.spec_str = tf_hub_module_path
self._initialize_from_hub_module(hub.load_module_spec(tf_hub_module_path))
def get_bottleneck_list(image_dir, label_name):
# Look at the folder structure, and create lists of all the images.
image_lists = create_image_lists(
'/usa/psu/Documents/CISC849/project/example/flower_photos', 10, 10)
class_count = len(image_lists.keys())
if class_count == 0:
tf.logging.error('No valid folders of images found at ' + image_dir)
return -1
if class_count == 1:
tf.logging.error('Only one valid folder of images found at ' +
image_dir +
' - multiple classes are needed for classification.')
return -1
# Set up the pre-trained graph.
module_spec = hub.load_module_spec(
'https://tfhub.dev/google/imagenet/inception_v3/feature_vector/1')
graph, bottleneck_tensor, resized_image_tensor, wants_quantization = (
create_module_graph(module_spec))
# Add the new layer that we'll be training.
with graph.as_default():
with tf.Session(graph=graph) as sess:
# Initialize all weights: for the module to their pretrained values,
# and for the newly added retraining layer to random initial values.
init = tf.global_variables_initializer()
sess.run(init)
bottleneck_list = []
label_list = []
# Set up the image decoding sub-graph.
jpeg_data_tensor, decoded_image_tensor = add_jpeg_decoding(
module_spec)
for subf_i in os.listdir(image_dir):
empty_lable = [0] * len(label_name)
index_label = label_name.index(subf_i)
empty_lable[index_label] = 1
for file_i in os.listdir(image_dir + subf_i):
bottleneck_value = get_bottleneck_values(
sess, image_lists, image_dir + subf_i + '/' + file_i,
'', jpeg_data_tensor, decoded_image_tensor,
resized_image_tensor, bottleneck_tensor,
'https://tfhub.dev/google/imagenet/inception_v3/feature_vector/1'
)
bottleneck_list.append(bottleneck_value)
label_list.append(empty_lable)
return bottleneck_list, label_list
def train_neural_network():
ModuleSpec = hub.load_module_spec(
"https://tfhub.dev/google/imagenet/mobilenet_v1_100_128/feature_vector/1"
)
graph, bottleneck_tensor, input_tensor = creat_graph(ModuleSpec)
with graph.as_default():
_, bottleneck_tensor_size = bottleneck_tensor.get_shape().as_list()
X = tf.placeholder_with_default(bottleneck_tensor,
shape=[None, bottleneck_tensor_size],
name='newlayerinputplacholder')
output, cost, optimizer, y = last_layer(X)
with tf.Session(graph=graph) as sess:
init = tf.global_variables_initializer()
sess.run(init)
bottleneck_value, labels, batch_step = run_and_save_bottleneck(
sess, bottleneck_tensor, input_tensor)
saver = tf.train.Saver()
for epoch in range(hm_epochs):
epoch_loss = 0
for batch in range(batch_step):
epoch_x = bottleneck_value[batch * batch_size:(batch + 1) *
batch_size]
epoch_y = labels[batch * batch_size:(batch + 1) *
batch_size]
_, c = sess.run([optimizer, cost],
feed_dict={
X: epoch_x,
y: epoch_y
})
epoch_loss += c
print('Epoch', epoch, 'completed out of', hm_epochs, 'loss:',
epoch_loss)
writer = tf.summary.FileWriter("output", sess.graph)
writer.close()
# please customize the directory for your project
saver.save(
sess, '/home/ali/PycharmProjects/tensorHub/save/my_test_model')
# correct = tf.equal(tf.argmax(prediction, 1), tf.argmax(y,1))
# accuracy = tf.reduce_mean(tf.cast(correct, 'float'))
# print('Accyracy:', accuracy.eval({x:mnist.test.images, y:mnist.test.labels}))
return graph, input_tensor, init
def _build_train_graph(self, n_classes):
self._train_graph = tf.Graph()
with self._train_graph.as_default():
# Load module spec/blueprint:
tfhub_module_spec = hub.load_module_spec('https://tfhub.dev/google/imagenet/inception_v3/feature_vector/1')
height, width = hub.get_expected_image_size(tfhub_module_spec)
# Create a placeholder tensor for image input to the model (when bottleneck has not been pre-computed).
resized_input_tensor = tf.placeholder(tf.float32, [None, height, width, 3], name='resized_input')
# m_reg = hub.Module(tfhub_module_spec, trainable=True, tags={'train'})
m_reg = hub.Module(tfhub_module_spec, trainable=False, tags={'train'})
# m = hub.Module(tfhub_module_spec)
# Create a placeholder tensor to catch the output of the pre-activation layer:
# bottleneck_tensor = m(resized_input_tensor)
# bottleneck_tensor_reg = m_reg(resized_input_tensor)
bottleneck_tensor_reg = m_reg()
batch_size, bottleneck_tensor_size = bottleneck_tensor_reg.get_shape().as_list()
self.bottleneck_input = tf.placeholder_with_default(
bottleneck_tensor_reg,
shape=[batch_size, bottleneck_tensor_size],
name='BottleneckInputPlaceholder'
)
self.ground_truth_input = tf.placeholder(
tf.int64, [batch_size], name='GroundTruthInput'
)
with tf.name_scope('weights'):
initial_value = tf.variance_scaling_initializer()(shape=[bottleneck_tensor_size, n_classes])
layer_weights = tf.Variable(initial_value=initial_value, name='final_weights')
with tf.name_scope('biases'):
layer_biases = tf.Variable(initial_value=tf.zeros([n_classes]), name='final_biases')
with tf.name_scope('Wx_plus_b'):
logits = tf.matmul(self.bottleneck_input, layer_weights) + layer_biases
Y_proba = tf.nn.softmax(logits, name='Y_proba')
xentropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=self.ground_truth_input, logits=logits)
self.loss = tf.reduce_mean(xentropy, name='xentropy_loss')
optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate)
self.training_op = optimizer.minimize(self.loss)
correct = tf.nn.in_top_k(logits, self.ground_truth_input, 1)
self.accuracy = tf.reduce_mean(tf.cast(correct, tf.float32), name='accuracy')
# Declare initializer:
self._init = tf.global_variables_initializer()
# extra ops for batch normalization
self.extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
def load_graphs(subsets_list, main_dir, saved_model_dir, tfhub_module,
output_labels):
graphs = {
'graph': [],
'sess': [],
'image': [],
'prediction': [],
'jpeg_data_tensor': [],
'decoded_image_tensor': [],
'labels': [],
'is_last_step': []
}
for subsets in subsets_list:
subset_name = subsets.replace(":", "_").replace(",", "-")
model_full_path = os.path.join(main_dir,
"retrain_subset_{}".format(subset_name),
saved_model_dir)
labels_full_path = os.path.join(
main_dir, "retrain_subset_{}".format(subset_name), output_labels)
with open(labels_full_path) as f:
labels = f.readlines()
labels = [l.strip() for l in labels]
print("Loading model: {}".format(model_full_path))
with tf.Graph().as_default() as graph:
with tf.Session(graph=graph).as_default() as sess:
tf.saved_model.loader.load(sess, [tag_constants.SERVING],
model_full_path)
# resized_input_tensor
image = graph.get_tensor_by_name('Placeholder:0')
prediction = graph.get_tensor_by_name('final_result:0')
module_spec = hub.load_module_spec(tfhub_module)
jpeg_data_tensor, decoded_image_tensor = add_jpeg_decoding(
module_spec)
graphs['graph'].append(graph)
graphs['sess'].append(sess)
graphs['image'].append(image)
graphs['prediction'].append(prediction)
graphs['jpeg_data_tensor'].append(jpeg_data_tensor)
graphs['decoded_image_tensor'].append(decoded_image_tensor)
graphs['labels'].append(labels)
graphs['is_last_step'].append(False)
graphs['is_last_step'][-1] = True
return graphs
def save_file_to_disk(graph, file, hub_module, num_classes, final_tensor_name,
learning_rate, CHECKPOINT_DIR):
"""Saves intermediate model files to disk
Args:
graph: tensorflow graph
file: name of file to be saved
module: Tensorflow hub module
num_classes: number of classes in our dataset
"""
module = hub.load_module_spec(hub_module)
sess, _, _, _, _ = compute_test_graph(hub_module, num_classes,
final_tensor_name, learning_rate,
CHECKPOINT_DIR)
graph = tf.graph.util.convert_variables_to_constants(
sess, sess.graph.as_graph_def(), [final_tensor_name])
with tf.gfile.FastGFile(file, 'wb') as f:
f.write(graph.SerializeToString())
def build_graph(hub_module):
"""Build a graph from tensorflow hub module
Args:
hub_module: Tensorflow Hub module
Returns:
graph extracted from hub module
pre final tensor(bottleneck)
input tensor (expected image size by graph)
"""
module = hub.load_module_spec(hub_module)
h, w = hub.get_expected_image_size(module)
with tf.Graph().as_default() as graph:
input_tensor = tf.placeholder(tf.float32, shape=(None, h, w, 3))
mod = hub.Module(module)
pre_final_tensor = mod(input_tensor)
return graph, pre_final_tensor, input_tensor
def main(_):
#需要確保日誌輸出可見。
tf.logging.set_verbosity(tf.logging.INFO)
if not FLAGS.image_dir:
tf.logging.error('Must set flag --image_dir.')
return -1
#準備可在培訓期間使用的必要目錄
prepare_file_system()
#查看文件夾結構，並創建所有圖像的列表。
image_lists = create_image_lists(FLAGS.image_dir, FLAGS.testing_percentage,
FLAGS.validation_percentage)
class_count = len(image_lists.keys())
if class_count == 0:
tf.logging.error('No valid folders of images found at ' + FLAGS.image_dir)
return -1
if class_count == 1:
tf.logging.error('Only one valid folder of images found at ' +
FLAGS.image_dir +
' - multiple classes are needed for classification.')
return -1
#看看命令行標誌是否意味著我們正在應用任何扭曲。
do_distort_images = should_distort_images(
FLAGS.flip_left_right, FLAGS.random_crop, FLAGS.random_scale,
FLAGS.random_brightness)
#設置預先訓練的圖形。
module_spec = hub.load_module_spec(FLAGS.tfhub_module)
graph, bottleneck_tensor, resized_image_tensor, wants_quantization = (
create_module_graph(module_spec))
#添加我們將要訓練的新圖層。
with graph.as_default():
(train_step, cross_entropy, bottleneck_input,
ground_truth_input, final_tensor) = add_final_retrain_ops(
class_count, FLAGS.final_tensor_name, bottleneck_tensor,
wants_quantization, is_training=True)
with tf.Session(graph=graph) as sess:
def extract_jpg_feature(image_path):
tfhub_module = 'https://tfhub.dev/google/imagenet/inception_v3/feature_vector/1'
module_spec = hub.load_module_spec(tfhub_module)
graph, bottleneck_tensor, jpeg_tensor = create_module_graph(module_spec)
with tf.Session(graph=graph) as sess:
init = tf.global_variables_initializer()
sess.run(init)
image_data = tf.gfile.GFile(image_path, 'rb').read()
try:
bottleneck_values = sess.run(bottleneck_tensor,
{jpeg_tensor: image_data})
bottleneck_values = np.squeeze(bottleneck_values)
#np.savetxt(feature_path, bottleneck_values)
return bottleneck_values.reshape(1, utils.static_dimension)
except Exception as e:
log_util.logger.error("extract feature of jpg[%s] fail: %s", image_path, str(e))
return None
def decode_and_resize(hub_module): """Performs image processing steps(decoding and reshaping) Args: hub_module: Tensorflow Hub module Returns: placeholder for image data reshaped tensor as expected by graph """ module = hub.load_module_spec(hub_module) h, w = hub.get_expected_image_size(module) reshape_specs = tf.stack((h, w)) num_channels = hub.get_num_image_channels(module) data_placeholder = tf.placeholder(tf.string, name='data_placeholder') decode = tf.image.decode_jpeg(data_placeholder, channels=num_channels) decode = tf.image.convert_image_dtype(decode, tf.float32) decode = tf.expand_dims(decode, 0) reshape = tf.cast(reshape_specs, dtype=tf.int32) reshaped_image = tf.image.resize_bilinear(decode, reshape) return data_placeholder, reshaped_image
def main(_):
if not FLAGS.vid_dir:
tf.logging.error('Must set flag --vid_dir.')
return -1
included_extensions = ['avi', 'mp4', 'mkv', 'mpeg']
file_list = [
fn for fn in os.listdir(FLAGS.vid_dir) if any(
fn.endswith(ext) for ext in included_extensions)
]
module = hub.load_module_spec(FLAGS.tfhub_module)
height, width = hub.get_expected_image_size(module)
resized_input_tensor = tf.placeholder(tf.float32, [None, width, width, 3])
m = hub.Module(module)
feature_tensor = m(resized_input_tensor)
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
for file_name in file_list:
print('Processing file: ' + file_name)
vidcap = cv2.VideoCapture(FLAGS.vid_dir + '/' + file_name)
success, image = vidcap.read()
count = 0
frames = []
while success:
frames.append(image)
success, image = vidcap.read()
count += 1
print('Number of frames: ', count)
img = [
cv2.cvtColor(
cv2.resize(frames[i],
dsize=(height, width),
interpolation=cv2.INTER_LINEAR), cv2.COLOR_BGR2RGB)
/ 255 for i in range(len(frames))
]
feature_vector = sess.run(feature_tensor,
feed_dict={resized_input_tensor: img})
txt_name = FLAGS.vid_dir + '/' + os.path.splitext(
file_name)[0] + '.csv'
np.savetxt(txt_name, feature_vector)
def testAssets(self):
export_path = os.path.join(self.get_temp_dir(), "assets-module")
vocabulary_file = self.create_vocab_file("tokens.txt",
["emerson", "lake", "palmer"])
with tf.Graph().as_default():
assets_module_fn = create_assets_module_fn(vocabulary_file)
spec = hub.create_module_spec(assets_module_fn)
embedding_module = hub.Module(spec)
output = embedding_module(tf.constant([1, 2], dtype=tf.int64))
with tf.Session() as sess:
sess.run(tf.tables_initializer())
self.assertAllEqual(list(sess.run(output)), [b"lake", b"palmer"])
embedding_module.export(export_path, sess)
asset_file = os.path.join(*[export_path, "assets", "tokens.txt"])
# Check that asset file got written to the expected place:
self.assertTrue(tf.gfile.Exists(asset_file))
# Assets should be hermetic, so we can delete the original vocab file:
tf.gfile.Remove(vocabulary_file)
with tf.Graph().as_default():
spec = hub.load_module_spec(export_path)
embedding_module = hub.Module(spec)
output = embedding_module(tf.constant([1, 2], dtype=tf.int64))
with tf.Session() as sess:
sess.run(tf.tables_initializer())
# Check functionality:
self.assertAllEqual(list(sess.run(output)), [b"lake", b"palmer"])
# Check that the ASSET_FILEPATHS collection was restored properly:
asset_filepaths = [
sess.run(tensor)
for tensor in tf.get_collection(tf.GraphKeys.ASSET_FILEPATHS)
]
# ASSET_FILEPATHS are added for the state graph and for the apply graph:
self.assertAllEqual(asset_filepaths,
[tf.compat.as_bytes(asset_file)] * 2)
def compute_final_op(sess, batch_size, features_dir, data_dir, hub_module,
num_classes, files, data_placeholder, reshaped_image,
pre_final_tensor, input_tensor, final_tensor_name,
learning_rate, CHECKPOINT_DIR):
"""Computes final performance using test set
Args:
sess: Current tensorflow session
batch_size: batch size
features_dir:
hub_module: Tensorflow hub module
num_classes: number of classes in our dataset
files: Training file names
data_placeholder: Placeholder for image data
reshaped_image: Reshaped tensor as expected by graph
pre_final_tensor: pre_final (bottleneck) tensor
input_tensor: input tensor (expected image size by graph)
final_tensor_name:
learning_rate:
CHECKPOINT_DIR:
"""
module = hub.load_module_spec(hub_module)
features, labels, filenames = sample_random_features(
sess, num_classes, files, batch_size, 'test', features_dir, data_dir,
data_placeholder, reshaped_image, pre_final_tensor, input_tensor,
hub_module)
sess, input_tensor, pre_final_input_tensor, truth_input_tensor, \
pred, step = compute_test_graph(hub_module, num_classes,
final_tensor_name, learning_rate, CHECKPOINT_DIR)
acc, p = sess.run([step, pred],
feed_dict={
data_placeholder: features,
truth_input_tensor: labels
})
print('Final Test Accuracy {}'.format(acc * 100))
def __init__(self):
# member variables.
# the URL of the pre-trained model.
self.HUB_MODULE = 'https://tfhub.dev/google/imagenet/inception_v3/feature_vector/1'
# the model spec.
self.Module_Spec = hub.load_module_spec(self.HUB_MODULE)
# the image size that is required by this model.
self.Module_Height, self.Module_Width = hub.get_expected_image_size(
self.Module_Spec)
self.Modelu_Depth = hub.get_num_image_channels(self.Module_Spec)
# A module is understood as instrumented for quantization with TF-Lite
# if it contains any of these ops.
self.FAKE_QUANT_OPS = ('FakeQuantWithMinMaxVars',
'FakeQuantWithMinMaxVarsPerChannel')
# the size of our input images.
self.ImageHeight = self.Module_Height
self.ImageWidth = self.Module_Width
self.ImageChannels = self.Modelu_Depth
# Set up the pre-trained graph.
self.graph, self.bottleneck_tensor, self.resized_input_tensor, self.wants_quantization = self.create_module_graph(
self.Module_Spec)
def _build_graph(tfhub_module_url):
module_spec = hub.load_module_spec(tfhub_module_url)
height, width = hub.get_expected_image_size(module_spec)
tf.logging.info(msg='Loaded TensorFlowHub module spec: %s' %
tfhub_module_url)
graph = tf.Graph()
with graph.as_default():
# Create a placeholder tensor for image input to the model (when bottleneck has not been pre-computed).
resized_input_tensor = tf.placeholder(tf.float32,
[None, height, width, 3],
name='resized_input')
# Declare the model in accordance with the chosen architecture:
m = hub.Module(module_spec)
# Create a placeholder tensor to catch the output of the pre-activation layer:
bottleneck_tensor = m(resized_input_tensor)
tf.logging.info(
msg=
'Defined computational graph from the tensorflow hub module spec.')
# Image decoding sub-graph:
with tf.name_scope('image_decoding'):
jpeg_data_tensor, decoded_image_tensor = _add_jpeg_decoding(
module_spec)
return graph, bottleneck_tensor, resized_input_tensor, jpeg_data_tensor, decoded_image_tensor
def main(_):
tf.logging.set_verbosity(tf.logging.INFO)
if not FLAGS.image_dir:
tf.logging.error('Must set flag --image_dir.')
return -1
prepare_file_system()
image_lists = create_image_lists(FLAGS.image_dir, FLAGS.testing_percentage,
FLAGS.validation_percentage)
class_count = len(image_lists.keys())
if class_count == 0:
tf.logging.error('No valid folders of images found at ' + FLAGS.image_dir)
return -1
if class_count == 1:
tf.logging.error('Only one valid folder of images found at ' +
FLAGS.image_dir +
' - multiple classes are needed for classification.')
return -1
do_distort_images = should_distort_images(
FLAGS.flip_left_right, FLAGS.random_crop, FLAGS.random_scale,
FLAGS.random_brightness)
module_spec = hub.load_module_spec(FLAGS.tfhub_module)
graph, bottleneck_tensor, resized_image_tensor, wants_quantization = (
create_module_graph(module_spec))
with graph.as_default():
(train_step, cross_entropy, bottleneck_input,
ground_truth_input, final_tensor) = add_final_retrain_ops(
class_count, FLAGS.final_tensor_name, bottleneck_tensor,
wants_quantization, is_training=True)
with tf.Session(graph=graph) as sess:
init = tf.global_variables_initializer()
sess.run(init)
jpeg_data_tensor, decoded_image_tensor = add_jpeg_decoding(module_spec)
if do_distort_images:
(distorted_jpeg_data_tensor,
distorted_image_tensor) = add_input_distortions(
FLAGS.flip_left_right, FLAGS.random_crop, FLAGS.random_scale,
FLAGS.random_brightness, module_spec)
else:
cache_bottlenecks(sess, image_lists, FLAGS.image_dir,
FLAGS.bottleneck_dir, jpeg_data_tensor,
decoded_image_tensor, resized_image_tensor,
bottleneck_tensor, FLAGS.tfhub_module)
# Create the operations we need to evaluate the accuracy of our new layer.
evaluation_step, _ = add_evaluation_step(final_tensor, ground_truth_input)
# Merge all the summaries and write them out to the summaries_dir
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '/train',
sess.graph)
validation_writer = tf.summary.FileWriter(
FLAGS.summaries_dir + '/validation')
train_saver = tf.train.Saver()
for i in range(FLAGS.how_many_training_steps):
if do_distort_images:
(train_bottlenecks,
train_ground_truth) = get_random_distorted_bottlenecks(
sess, image_lists, FLAGS.train_batch_size, 'training',
FLAGS.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, FLAGS.train_batch_size, 'training',
FLAGS.bottleneck_dir, FLAGS.image_dir, jpeg_data_tensor,
decoded_image_tensor, resized_image_tensor, bottleneck_tensor,
FLAGS.tfhub_module)
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)
is_last_step = (i + 1 == FLAGS.how_many_training_steps)
if (i % FLAGS.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})
tf.logging.info('%s: Step %d: Train accuracy = %.1f%%' %
(datetime.now(), i, train_accuracy * 100))
tf.logging.info('%s: Step %d: Cross entropy = %f' %
(datetime.now(), i, cross_entropy_value))
# TODO: Make this use an eval graph, to avoid quantization
validation_bottlenecks, validation_ground_truth, _ = (
get_random_cached_bottlenecks(
sess, image_lists, FLAGS.validation_batch_size, 'validation',
FLAGS.bottleneck_dir, FLAGS.image_dir, jpeg_data_tensor,
decoded_image_tensor, resized_image_tensor, bottleneck_tensor,
FLAGS.tfhub_module))
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)
tf.logging.info('%s: Step %d: Validation accuracy = %.1f%% (N=%d)' %
(datetime.now(), i, validation_accuracy * 100,
len(validation_bottlenecks)))
intermediate_frequency = FLAGS.intermediate_store_frequency
if (intermediate_frequency > 0 and (i % intermediate_frequency == 0)
and i > 0):
train_saver.save(sess, CHECKPOINT_NAME)
intermediate_file_name = (FLAGS.intermediate_output_graphs_dir +
'intermediate_' + str(i) + '.pb')
tf.logging.info('Save intermediate result to : ' +
intermediate_file_name)
save_graph_to_file(intermediate_file_name, module_spec,
class_count)
train_saver.save(sess, CHECKPOINT_NAME)
run_final_eval(sess, module_spec, class_count, image_lists,
jpeg_data_tensor, decoded_image_tensor, resized_image_tensor,
bottleneck_tensor)
tf.logging.info('Save final result to : ' + FLAGS.output_graph)
if wants_quantization:
tf.logging.info('The model is instrumented for quantization with TF-Lite')
save_graph_to_file(FLAGS.output_graph, module_spec, class_count)
with tf.gfile.FastGFile(FLAGS.output_labels, 'w') as f:
f.write('\n'.join(image_lists.keys()) + '\n')
if FLAGS.saved_model_dir:
export_model(module_spec, class_count, FLAGS.saved_model_dir)
def main(_):
logging_verbosity = logging_level_verbosity(FLAGS.logging_verbosity)
tf.logging.set_verbosity(logging_verbosity)
if not FLAGS.image_dir:
tf.logging.error('Must set flag --image_dir.')
return -1
# Prepare necessary directories that can be used during training
prepare_file_system()
# Look at the folder structure, and create lists of all the images.
image_lists = create_image_lists(FLAGS.image_dir, FLAGS.testing_percentage,
FLAGS.validation_percentage)
class_count = len(image_lists.keys())
if class_count == 0:
tf.logging.error('No valid folders of images found at ' +
FLAGS.image_dir)
return -1
if class_count == 1:
tf.logging.error('Only one valid folder of images found at ' +
FLAGS.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(FLAGS.flip_left_right,
FLAGS.random_crop,
FLAGS.random_scale,
FLAGS.random_brightness)
# Set up the pre-trained graph.
module_spec = hub.load_module_spec(FLAGS.tfhub_module)
graph, bottleneck_tensor, resized_image_tensor, wants_quantization = (
create_module_graph(module_spec))
# Add the new layer that we'll be training.
with graph.as_default():
(train_step, cross_entropy, bottleneck_input, ground_truth_input,
final_tensor) = add_final_retrain_ops(class_count,
FLAGS.final_tensor_name,
bottleneck_tensor,
wants_quantization,
is_training=True)
with tf.Session(graph=graph) as sess:
# Initialize all weights: for the module to their pretrained values,
# and for the newly added retraining layer to random initial values.
init = tf.global_variables_initializer()
sess.run(init)
# Set up the image decoding sub-graph.
jpeg_data_tensor, decoded_image_tensor = add_jpeg_decoding(module_spec)
if do_distort_images:
# We will be applying distortions, so set up the operations we'll need.
(distorted_jpeg_data_tensor,
distorted_image_tensor) = add_input_distortions(
FLAGS.flip_left_right, FLAGS.random_crop, FLAGS.random_scale,
FLAGS.random_brightness, module_spec)
else:
# We'll make sure we've calculated the 'bottleneck' image summaries and
# cached them on disk.
cache_bottlenecks(sess, image_lists, FLAGS.image_dir,
FLAGS.bottleneck_dir, jpeg_data_tensor,
decoded_image_tensor, resized_image_tensor,
bottleneck_tensor, FLAGS.tfhub_module)
# Create the operations we need to evaluate the accuracy of our new layer.
evaluation_step, _ = add_evaluation_step(final_tensor,
ground_truth_input)
# Merge all the summaries and write them out to the summaries_dir
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '/train',
sess.graph)
validation_writer = tf.summary.FileWriter(FLAGS.summaries_dir +
'/validation')
# Create a train saver that is used to restore values into an eval graph
# when exporting models.
train_saver = tf.train.Saver()
# Run the training for as many cycles as requested on the command line.
for i in range(FLAGS.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, FLAGS.train_batch_size, 'training',
FLAGS.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, FLAGS.train_batch_size, 'training',
FLAGS.bottleneck_dir, FLAGS.image_dir, jpeg_data_tensor,
decoded_image_tensor, resized_image_tensor,
bottleneck_tensor, FLAGS.tfhub_module)
# 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 == FLAGS.how_many_training_steps)
if (i % FLAGS.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
})
tf.logging.info('%s: Step %d: Train accuracy = %.1f%%' %
(datetime.now(), i, train_accuracy * 100))
tf.logging.info('%s: Step %d: Cross entropy = %f' %
(datetime.now(), i, cross_entropy_value))
# TODO: Make this use an eval graph, to avoid quantization
# moving averages being updated by the validation set, though in
# practice this makes a negligable difference.
validation_bottlenecks, validation_ground_truth, _ = (
get_random_cached_bottlenecks(
sess, image_lists, FLAGS.validation_batch_size,
'validation', FLAGS.bottleneck_dir, FLAGS.image_dir,
jpeg_data_tensor, decoded_image_tensor,
resized_image_tensor, bottleneck_tensor,
FLAGS.tfhub_module))
# 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)
tf.logging.info(
'%s: Step %d: Validation accuracy = %.1f%% (N=%d)' %
(datetime.now(), i, validation_accuracy * 100,
len(validation_bottlenecks)))
# Store intermediate results
intermediate_frequency = FLAGS.intermediate_store_frequency
if (intermediate_frequency > 0
and (i % intermediate_frequency == 0) and i > 0):
# If we want to do an intermediate save, save a checkpoint of the train
# graph, to restore into the eval graph.
train_saver.save(sess, FLAGS.checkpoint_path)
intermediate_file_name = (
FLAGS.intermediate_output_graphs_dir + 'intermediate_' +
str(i) + '.pb')
tf.logging.info('Save intermediate result to : ' +
intermediate_file_name)
save_graph_to_file(intermediate_file_name, module_spec,
class_count)
# After training is complete, force one last save of the train checkpoint.
train_saver.save(sess, FLAGS.checkpoint_path)
# We've completed all our training, so run a final test evaluation on
# some new images we haven't used before.
run_final_eval(sess, module_spec, class_count, image_lists,
jpeg_data_tensor, decoded_image_tensor,
resized_image_tensor, bottleneck_tensor)
# Write out the trained graph and labels with the weights stored
# constants.
tf.logging.info('Save final result to : ' + FLAGS.output_graph)
if wants_quantization:
tf.logging.info(
'The model is instrumented for quantization with TF-Lite')
save_graph_to_file(FLAGS.output_graph, module_spec, class_count)
with tf.gfile.GFile(FLAGS.output_labels, 'w') as f:
f.write('\n'.join(image_lists.keys()) + '\n')
if FLAGS.saved_model_dir:
export_model(module_spec, class_count, FLAGS.saved_model_dir)
def testModuleWithBatchNorm(self):
export_path = os.path.join(self.get_temp_dir(), "batch-norm-module")
# This test resorts to lookup by name to retrieve the moving mean,
# because tf.contrib.layers.batch_norm() does not return it, and even if,
# module_fn() has no way to return it next to the result for training.
moving_mean_name = (
"module/BatchNorm/moving_mean/Read/ReadVariableOp:0")
batch_norm_train_tags = ["batch_norm_trains"]
batch_norm_fixed_tags = ["batch_norm_fixed"]
spec = hub.create_module_spec(
batch_norm_module_fn,
[(batch_norm_train_tags, {"is_training": True}),
(batch_norm_fixed_tags, {"is_training": False})])
# Test Module creation and training.
with tf.Graph().as_default() as g:
m = hub.Module(spec, trainable=True, tags=batch_norm_train_tags)
# The module is trained on a fixed batch of inputs, which has a mean
# of 12.0 and some sample variance of a less obvious value. The module
# learns scale and offset parameters that achieve the mapping x --> 2*x
# for the observed mean and variance.
x = tf.constant([[11.0], [12.0], [13.0]])
training_mean = [12.0]
y_target = tf.constant([[22.0], [24.0], [26.0]])
y = m(x)
step = tf.Variable(0, trainable=False, name="global_step")
train = tf.contrib.layers.optimize_loss(
loss=tf.losses.mean_squared_error(y, y_target),
global_step=step,
learning_rate=0.1,
optimizer="SGD")
moving_mean = g.get_tensor_by_name(moving_mean_name)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
self.assertAllClose(sess.run(moving_mean), [0.0])
for _ in range(100):
sess.run([train])
trained_moving_mean, trained_y = sess.run([moving_mean, y])
self.assertAllClose(trained_moving_mean, training_mean)
self.assertAllClose(trained_y, [[22.0], [24.0], [26.0]])
# Test export.
m.export(export_path, sess)
# Test import and use.
spec = hub.load_module_spec(export_path)
with tf.Graph().as_default() as g:
# The module gets run for inference on inputs with different mean and
# variance. However, both mean and variance as well as offset and scale
# are now frozen to the values from learning, so the same mapping
# x --> 2*x is recovered.
x = tf.constant([[10.0], [20.0], [30.0]])
y = hub.Module(
spec, tags=batch_norm_fixed_tags)(x)
moving_mean = g.get_tensor_by_name(moving_mean_name)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for _ in range(100):
served_moving_mean, served_y = sess.run([moving_mean, y])
# No update occurs to the moving_mean from training time.
self.assertAllClose(served_moving_mean, training_mean)
# Prediction results are correct.
self.assertAllClose(served_y, [[20.0], [40.0], [60.0]])
if args.saved_model_dir:
saved_model_dir = args.saved_model_dir
else:
saved_model_dir = os.path.join(
os.path.join(os.path.abspath(os.sep), 'tmp'), 'saved_model')
if args.output_labels:
output_labels = args.output_labels
else:
output_labels = os.path.join(saved_model_dir, 'label.txt')
image_size = None
if args.image_size:
image_size = (args.image_size, args.image_size)
else:
try:
module_spec = hub.load_module_spec(args.tfhub_module)
image_size = tuple(hub.get_expected_image_size(module_spec))
print("get model spec", image_size)
except:
if "mobilenet" in args.tfhub_module:
print("get model spec failed, use default spec as 224 x 224,")
image_size = (224, 224)
if image_size:
Retrain(data_dir=args.image_dir,
saved_model_path=saved_model_dir,
saved_label_path=output_labels,
epochs=args.epochs,
batch_size=args.batch_size,
module_handle=args.tfhub_module,
image_size=image_size,
# train params
epochs = args.epochs
batch_size = args.batch_size
learning_rate = args.learning_rate
csv_out = args.csv_output
pred_out = args.prediction_output
dropout = args.dropout
save_models = args.save
import_features = args.import_features
tfhub = args.tfhub_module
##### LOAD IMAGES ######
if tfhub != None:
module_spec = hub.load_module_spec(tfhub)
height, width = hub.get_expected_image_size(module_spec)
channels = hub.get_num_image_channels(module_spec)
else:
height, width, channels = 224, 224, 3
### training images
# read paths and labels for each image
listimgs, listlabels = parse_input(train_paths)
# load images
loaded_imgs = [load_image(img, size=height).reshape((height, width, channels)) for img in listimgs]
print('[TRAINING] Loaded', len(loaded_imgs), 'images and', len(listlabels), 'labels')
# map string labels to unique integers
u,indices = np.unique(np.array(listlabels), return_inverse=True)
print('[TRAINING] Categories: ', u)
num_categories = len(u)
def retrain(inputdir=None):
# Needed to make sure the logging output is visible.
# See https://github.com/tensorflow/tensorflow/issues/3047
tf.logging.set_verbosity(tf.logging.ERROR)
if inputdir:
if type(inputdir)==str:
FLAGS.image_dir = inputdir
else:
tf.logging.error("Invalid input directory!")
return -1
if not FLAGS.image_dir:
tf.logging.error('Must set flag --image_dir.')
return -1
print('Retraining with images in directory: ' + FLAGS.image_dir)
# Prepare necessary directories that can be used during training
prepare_file_system()
# Look at the folder structure, and create lists of all the images.
image_lists = create_image_lists(FLAGS.image_dir, FLAGS.testing_percentage,
FLAGS.validation_percentage)
class_count = len(image_lists.keys())
if class_count == 0:
tf.logging.error('No valid folders of images found at ' + FLAGS.image_dir)
return -1
if class_count == 1:
tf.logging.error('Only one valid folder of images found at ' +
FLAGS.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(
FLAGS.flip_left_right, FLAGS.random_crop, FLAGS.random_scale,
FLAGS.random_brightness)
# Set up the pre-trained graph.
module_spec = hub.load_module_spec(FLAGS.tfhub_module)
graph, bottleneck_tensor, resized_image_tensor, wants_quantization = (
create_module_graph(module_spec))
# Add the new layer that we'll be training.
with graph.as_default():
(train_step, cross_entropy, bottleneck_input,
ground_truth_input, final_tensor) = add_final_retrain_ops(
class_count, FLAGS.final_tensor_name, bottleneck_tensor,
wants_quantization, is_training=True)
with tf.Session(graph=graph) as sess:
# Initialize all weights: for the module to their pretrained values,
# and for the newly added retraining layer to random initial values.
init = tf.global_variables_initializer()
sess.run(init)
# Set up the image decoding sub-graph.
jpeg_data_tensor, decoded_image_tensor = add_jpeg_decoding(module_spec)
if do_distort_images:
# We will be applying distortions, so set up the operations we'll need.
(distorted_jpeg_data_tensor,
distorted_image_tensor) = add_input_distortions(
FLAGS.flip_left_right, FLAGS.random_crop, FLAGS.random_scale,
FLAGS.random_brightness, module_spec)
else:
# We'll make sure we've calculated the 'bottleneck' image summaries and
# cached them on disk.
cache_bottlenecks(sess, image_lists, FLAGS.image_dir,
FLAGS.bottleneck_dir, jpeg_data_tensor,
decoded_image_tensor, resized_image_tensor,
bottleneck_tensor, FLAGS.tfhub_module)
# Create the operations we need to evaluate the accuracy of our new layer.
evaluation_step, _ = add_evaluation_step(final_tensor, ground_truth_input)
# Merge all the summaries and write them out to the summaries_dir
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '/train',
sess.graph)
validation_writer = tf.summary.FileWriter(
FLAGS.summaries_dir + '/validation')
# Create a train saver that is used to restore values into an eval graph
# when exporting models.
train_saver = tf.train.Saver()
# Run the training for as many cycles as requested on the command line.
for i in range(FLAGS.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, FLAGS.train_batch_size, 'training',
FLAGS.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, FLAGS.train_batch_size, 'training',
FLAGS.bottleneck_dir, FLAGS.image_dir, jpeg_data_tensor,
decoded_image_tensor, resized_image_tensor, bottleneck_tensor,
FLAGS.tfhub_module)
# 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 == FLAGS.how_many_training_steps)
if (i % FLAGS.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})
tf.logging.info('%s: Step %d: Train accuracy = %.1f%%' %
(datetime.now(), i, train_accuracy * 100))
tf.logging.info('%s: Step %d: Cross entropy = %f' %
(datetime.now(), i, cross_entropy_value))
# TODO: Make this use an eval graph, to avoid quantization
# moving averages being updated by the validation set, though in
# practice this makes a negligable difference.
validation_bottlenecks, validation_ground_truth, _ = (
get_random_cached_bottlenecks(
sess, image_lists, FLAGS.validation_batch_size, 'validation',
FLAGS.bottleneck_dir, FLAGS.image_dir, jpeg_data_tensor,
decoded_image_tensor, resized_image_tensor, bottleneck_tensor,
FLAGS.tfhub_module))
# 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)
tf.logging.info('%s: Step %d: Validation accuracy = %.1f%% (N=%d)' %
(datetime.now(), i, validation_accuracy * 100,
len(validation_bottlenecks)))
# Store intermediate results
intermediate_frequency = FLAGS.intermediate_store_frequency
if (intermediate_frequency > 0 and (i % intermediate_frequency == 0)
and i > 0):
# If we want to do an intermediate save, save a checkpoint of the train
# graph, to restore into the eval graph.
train_saver.save(sess, CHECKPOINT_NAME)
intermediate_file_name = (FLAGS.intermediate_output_graphs_dir +
'intermediate_' + str(i) + '.pb')
tf.logging.info('Save intermediate result to : ' +
intermediate_file_name)
save_graph_to_file(graph, intermediate_file_name, module_spec,
class_count)
# After training is complete, force one last save of the train checkpoint.
train_saver.save(sess, CHECKPOINT_NAME)
# We've completed all our training, so run a final test evaluation on
# some new images we haven't used before.
run_final_eval(sess, module_spec, class_count, image_lists,
jpeg_data_tensor, decoded_image_tensor, resized_image_tensor,
bottleneck_tensor)
# Write out the trained graph and labels with the weights stored as
# constants.
tf.logging.info('Save final result to : ' + FLAGS.output_graph)
if wants_quantization:
tf.logging.info('The model is instrumented for quantization with TF-Lite')
save_graph_to_file(graph, FLAGS.output_graph, module_spec, class_count)
with tf.gfile.FastGFile(FLAGS.output_labels, 'w') as f:
f.write('\n'.join(image_lists.keys()) + '\n')
if FLAGS.saved_model_dir:
export_model(module_spec, class_count, FLAGS.saved_model_dir)