Tensorflow化骨绵掌第3式-ImageNet打造自己的图像识别模型
1、下载ImageNet数据集
2、微调的原理
(1)VGG16网络结构
VGG16的结构包含5层卷积部分(共计13层),即conv1-conv5,还有3个全连接层,即fc6-fc8,卷积部分加上全连接层共计16层。
(2)微调的原理
在训练模型时,网络的参数的初始值并不是随机生成的,而是采用VGG16在ImageNet上已经训练好的参数作为训练的初始值。这样做的目的在于,在ImageNet数据集上训练过的VGG16网络中的参数已经包含了大量有用的卷积滤波器,与其从0开始初始化VGG16的所有参数,不如使用已有训练好的参数作为训练的起点。这样可以节省训练时间,有助于分类器性能的提高,模型的效果也达到很好的效果。
在载入VGG16的参数后便开始模型的训练,此时需要指定训练层数的范围,一般而言:
- 只训练fc8。fc8的结构进行改变后,因此它的参数不能直接从ImageNet预训练模型中获得,可以只训练fc8、其他结构不进行参数变化。
- 训练所有的参数。可以对网络所有的参数进行训练,但是计算速度比较慢,可以取得比较好的性能。
- 训练部分参数。通常是固定浅层参数不变,训练深层参数。
3、数据准备
(1)将数据集进行切分成训练集和验证集
在本博文中已经将数据集进行划分,在data_prepare文件夹下新建文件夹train、validation,分别保存训练和验证使用的图片,当然每个文件下的图片采用文件夹形式进行归类为wood、water、rock、wetland、glacier、urban,每个文件下都保存的是原始图片。
(2)转换数据集格式为tfrecord格式
在data_prepare文件夹下,使用预先写好的py文件data_convert.py文件,将图片转换成tfrecord格式的数据集。
data_convert.py文件代码如下所示:
# coding:utf-8
from __future__ import absolute_import
import argparse
import os
import logging
from src.tfrecord import main
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('-t', '--tensorflow-data-dir', default='pic/')
parser.add_argument('--train-shards', default=2, type=int)
parser.add_argument('--validation-shards', default=2, type=int)
parser.add_argument('--num-threads', default=2, type=int)
parser.add_argument('--dataset-name', default='satellite', type=str)
return parser.parse_args()
if __name__ == '__main__':
logging.basicConfig(level=logging.INFO)
args = parse_args()
args.tensorflow_dir = args.tensorflow_data_dir
args.train_directory = os.path.join(args.tensorflow_dir, 'train')
args.validation_directory = os.path.join(args.tensorflow_dir, 'validation')
args.output_directory = args.tensorflow_dir
args.labels_file = os.path.join(args.tensorflow_dir, 'label.txt')
if os.path.exists(args.labels_file) is False:
logging.warning('Can\'t find label.txt. Now create it.')
all_entries = os.listdir(args.train_directory)
dirnames = []
for entry in all_entries:
if os.path.isdir(os.path.join(args.train_directory, entry)):
dirnames.append(entry)
with open(args.labels_file, 'w') as f:
for dirname in dirnames:
f.write(dirname + '\n')
main(args)
4、使用Tensorflow Slim进行微调模型
TF-slim是tensorflow的一个轻量级库,它将很多常见tensorflow函数进行封装,使的模型的构建、训练、测试都更加简洁,特别适用于构建结构复杂的深度神经网络。TF-slim内置了VGG16、VGG19、Inception等模型结构以及与训练模型。
(1)下载Tensorflow Slim代码
①github地址为:https://github.com/tensorflow/models/tree/master/research/slim
②到https://github.com/tensorflow/models上将models模块下载下载,当然也可以用git下载,命令为git clone https://github.com/tensorflow/models.git
(2)定义新的datasets文件
在slim/datasets中,定义了所有可以使用的数据库,为了使用之前生成的tfrecord数据集,必须在datasets中定义新的数据库。
①在datasets/目录下新建一个文件satellite.py
py文件代码如下:
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import tensorflow as tf
from datasets import dataset_utils
slim = tf.contrib.slim
_FILE_PATTERN = 'satellite_%s_*.tfrecord'
SPLITS_TO_SIZES = {'train':4800, 'validation': 1200}
_NUM_CLASSES = 6
_ITEMS_TO_DESCRIPTIONS = {
'image': 'A color image of varying size.',
'label': 'A single integer between 0 and 4',
}
def get_split(split_name, dataset_dir, file_pattern=None, reader=None):
"""Gets a dataset tuple with instructions for reading flowers.
Args:
split_name: A train/validation split name.
dataset_dir: The base directory of the dataset sources.
file_pattern: The file pattern to use when matching the dataset sources.
It is assumed that the pattern contains a '%s' string so that the split
name can be inserted.
reader: The TensorFlow reader type.
Returns:
A `Dataset` namedtuple.
Raises:
ValueError: if `split_name` is not a valid train/validation split.
"""
if split_name not in SPLITS_TO_SIZES:
raise ValueError('split name %s was not recognized.' % split_name)
if not file_pattern:
file_pattern = _FILE_PATTERN
file_pattern = os.path.join(dataset_dir, file_pattern % split_name)
# Allowing None in the signature so that dataset_factory can use the default.
if reader is None:
reader = tf.TFRecordReader
keys_to_features = {
'image/encoded': tf.FixedLenFeature((), tf.string, default_value=''),
'image/format': tf.FixedLenFeature((), tf.string, default_value='jpg'),
'image/class/label': tf.FixedLenFeature(
[], tf.int64, default_value=tf.zeros([], dtype=tf.int64)),
}
items_to_handlers = {
'image': slim.tfexample_decoder.Image(),
'label': slim.tfexample_decoder.Tensor('image/class/label'),
}
decoder = slim.tfexample_decoder.TFExampleDecoder(
keys_to_features, items_to_handlers)
labels_to_names = None
if dataset_utils.has_labels(dataset_dir):
labels_to_names = dataset_utils.read_label_file(dataset_dir)
return slim.dataset.Dataset(
data_sources=file_pattern,
reader=reader,
decoder=decoder,
num_samples=SPLITS_TO_SIZES[split_name],
items_to_descriptions=_ITEMS_TO_DESCRIPTIONS,
num_classes=_NUM_CLASSES,
labels_to_names=labels_to_names)
经过修改后的数据库datasets.py代码如下:
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from datasets import cifar10
from datasets import flowers
from datasets import imagenet
from datasets import mnist
from datasets import satellite
datasets_map = {
'cifar10': cifar10,
'flowers': flowers,
'imagenet': imagenet,
'mnist': mnist,
'satellite':satellite,
}
def get_dataset(name, split_name, dataset_dir, file_pattern=None, reader=None):
"""Given a dataset name and a split_name returns a Dataset.
Args:
name: String, the name of the dataset.
split_name: A train/test split name.
dataset_dir: The directory where the dataset files are stored.
file_pattern: The file pattern to use for matching the dataset source files.
reader: The subclass of tf.ReaderBase. If left as `None`, then the default
reader defined by each dataset is used.
Returns:
A `Dataset` class.
Raises:
ValueError: If the dataset `name` is unknown.
"""
if name not in datasets_map:
raise ValueError('Name of dataset unknown %s' % name)
return datasets_map[name].get_split(
split_name,
dataset_dir,
file_pattern,
reader)
(3)准备训练文件夹
①在slim文件夹下新建一个文件夹satellite
②在satellite文件夹下新建data文件夹,并将转换好的五个数据放在文件夹下
③在satellite文件夹下新建train_dir文件夹,用于保存训练中的日志和模型
④在satellite文件夹下新建pretrained目录,在slim的github页面找到vgg16模型的下载地址,下载文件并解压,然后会得到一个vgg16的ckpt文件,将该文件复制到pretrained目录下。
(4)开始训练
运行train_image_classifier.py文件即可以开始。
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
from datasets import dataset_factory
from deployment import model_deploy
from nets import nets_factory
from preprocessing import preprocessing_factory
slim = tf.contrib.slim
tf.app.flags.DEFINE_string(
'master', '', 'The address of the TensorFlow master to use.')
tf.app.flags.DEFINE_string(
'train_dir', '/tmp/tfmodel/',
'Directory where checkpoints and event logs are written to.')
tf.app.flags.DEFINE_integer('num_clones', 1,
'Number of model clones to deploy.')
tf.app.flags.DEFINE_boolean('clone_on_cpu', False,
'Use CPUs to deploy clones.')
tf.app.flags.DEFINE_integer('worker_replicas', 1, 'Number of worker replicas.')
tf.app.flags.DEFINE_integer(
'num_ps_tasks', 0,
'The number of parameter servers. If the value is 0, then the parameters '
'are handled locally by the worker.')
tf.app.flags.DEFINE_integer(
'num_readers', 4,
'The number of parallel readers that read data from the dataset.')
tf.app.flags.DEFINE_integer(
'num_preprocessing_threads', 4,
'The number of threads used to create the batches.')
tf.app.flags.DEFINE_integer(
'log_every_n_steps', 10,
'The frequency with which logs are print.')
tf.app.flags.DEFINE_integer(
'save_summaries_secs', 600,
'The frequency with which summaries are saved, in seconds.')
tf.app.flags.DEFINE_integer(
'save_interval_secs', 600,
'The frequency with which the model is saved, in seconds.')
tf.app.flags.DEFINE_integer(
'task', 0, 'Task id of the replica running the training.')
######################
# Optimization Flags #
######################
tf.app.flags.DEFINE_float(
'weight_decay', 0.00004, 'The weight decay on the model weights.')
tf.app.flags.DEFINE_string(
'optimizer', 'rmsprop',
'The name of the optimizer, one of "adadelta", "adagrad", "adam",'
'"ftrl", "momentum", "sgd" or "rmsprop".')
tf.app.flags.DEFINE_float(
'adadelta_rho', 0.95,
'The decay rate for adadelta.')
tf.app.flags.DEFINE_float(
'adagrad_initial_accumulator_value', 0.1,
'Starting value for the AdaGrad accumulators.')
tf.app.flags.DEFINE_float(
'adam_beta1', 0.9,
'The exponential decay rate for the 1st moment estimates.')
tf.app.flags.DEFINE_float(
'adam_beta2', 0.999,
'The exponential decay rate for the 2nd moment estimates.')
tf.app.flags.DEFINE_float('opt_epsilon', 1.0, 'Epsilon term for the optimizer.')
tf.app.flags.DEFINE_float('ftrl_learning_rate_power', -0.5,
'The learning rate power.')
tf.app.flags.DEFINE_float(
'ftrl_initial_accumulator_value', 0.1,
'Starting value for the FTRL accumulators.')
tf.app.flags.DEFINE_float(
'ftrl_l1', 0.0, 'The FTRL l1 regularization strength.')
tf.app.flags.DEFINE_float(
'ftrl_l2', 0.0, 'The FTRL l2 regularization strength.')
tf.app.flags.DEFINE_float(
'momentum', 0.9,
'The momentum for the MomentumOptimizer and RMSPropOptimizer.')
tf.app.flags.DEFINE_float('rmsprop_decay', 0.9, 'Decay term for RMSProp.')
#######################
# Learning Rate Flags #
#######################
tf.app.flags.DEFINE_string(
'learning_rate_decay_type',
'exponential',
'Specifies how the learning rate is decayed. One of "fixed", "exponential",'
' or "polynomial"')
tf.app.flags.DEFINE_float('learning_rate', 0.01, 'Initial learning rate.')
tf.app.flags.DEFINE_float(
'end_learning_rate', 0.0001,
'The minimal end learning rate used by a polynomial decay learning rate.')
tf.app.flags.DEFINE_float(
'label_smoothing', 0.0, 'The amount of label smoothing.')
tf.app.flags.DEFINE_float(
'learning_rate_decay_factor', 0.94, 'Learning rate decay factor.')
tf.app.flags.DEFINE_float(
'num_epochs_per_decay', 2.0,
'Number of epochs after which learning rate decays.')
tf.app.flags.DEFINE_bool(
'sync_replicas', False,
'Whether or not to synchronize the replicas during training.')
tf.app.flags.DEFINE_integer(
'replicas_to_aggregate', 1,
'The Number of gradients to collect before updating params.')
tf.app.flags.DEFINE_float(
'moving_average_decay', None,
'The decay to use for the moving average.'
'If left as None, then moving averages are not used.')
#######################
# Dataset Flags #
#######################
tf.app.flags.DEFINE_string(
'dataset_name', 'imagenet', 'The name of the dataset to load.')
tf.app.flags.DEFINE_string(
'dataset_split_name', 'train', 'The name of the train/test split.')
tf.app.flags.DEFINE_string(
'dataset_dir', None, 'The directory where the dataset files are stored.')
tf.app.flags.DEFINE_integer(
'labels_offset', 0,
'An offset for the labels in the dataset. This flag is primarily used to '
'evaluate the VGG and ResNet architectures which do not use a background '
'class for the ImageNet dataset.')
tf.app.flags.DEFINE_string(
'model_name', 'inception_v3', 'The name of the architecture to train.')
tf.app.flags.DEFINE_string(
'preprocessing_name', None, 'The name of the preprocessing to use. If left '
'as `None`, then the model_name flag is used.')
tf.app.flags.DEFINE_integer(
'batch_size', 32, 'The number of samples in each batch.')
tf.app.flags.DEFINE_integer(
'train_image_size', None, 'Train image size')
tf.app.flags.DEFINE_integer('max_number_of_steps', None,
'The maximum number of training steps.')
#####################
# Fine-Tuning Flags #
#####################
tf.app.flags.DEFINE_string(
'checkpoint_path', None,
'The path to a checkpoint from which to fine-tune.')
tf.app.flags.DEFINE_string(
'checkpoint_exclude_scopes', None,
'Comma-separated list of scopes of variables to exclude when restoring '
'from a checkpoint.')
tf.app.flags.DEFINE_string(
'trainable_scopes', None,
'Comma-separated list of scopes to filter the set of variables to train.'
'By default, None would train all the variables.')
tf.app.flags.DEFINE_boolean(
'ignore_missing_vars', False,
'When restoring a checkpoint would ignore missing variables.')
FLAGS = tf.app.flags.FLAGS
def _configure_learning_rate(num_samples_per_epoch, global_step):
"""Configures the learning rate.
Args:
num_samples_per_epoch: The number of samples in each epoch of training.
global_step: The global_step tensor.
Returns:
A `Tensor` representing the learning rate.
Raises:
ValueError: if
"""
decay_steps = int(num_samples_per_epoch / FLAGS.batch_size *
FLAGS.num_epochs_per_decay)
if FLAGS.sync_replicas:
decay_steps /= FLAGS.replicas_to_aggregate
if FLAGS.learning_rate_decay_type == 'exponential':
return tf.train.exponential_decay(FLAGS.learning_rate,
global_step,
decay_steps,
FLAGS.learning_rate_decay_factor,
staircase=True,
name='exponential_decay_learning_rate')
elif FLAGS.learning_rate_decay_type == 'fixed':
return tf.constant(FLAGS.learning_rate, name='fixed_learning_rate')
elif FLAGS.learning_rate_decay_type == 'polynomial':
return tf.train.polynomial_decay(FLAGS.learning_rate,
global_step,
decay_steps,
FLAGS.end_learning_rate,
power=1.0,
cycle=False,
name='polynomial_decay_learning_rate')
else:
raise ValueError('learning_rate_decay_type [%s] was not recognized',
FLAGS.learning_rate_decay_type)
def _configure_optimizer(learning_rate):
"""Configures the optimizer used for training.
Args:
learning_rate: A scalar or `Tensor` learning rate.
Returns:
An instance of an optimizer.
Raises:
ValueError: if FLAGS.optimizer is not recognized.
"""
if FLAGS.optimizer == 'adadelta':
optimizer = tf.train.AdadeltaOptimizer(
learning_rate,
rho=FLAGS.adadelta_rho,
epsilon=FLAGS.opt_epsilon)
elif FLAGS.optimizer == 'adagrad':
optimizer = tf.train.AdagradOptimizer(
learning_rate,
initial_accumulator_value=FLAGS.adagrad_initial_accumulator_value)
elif FLAGS.optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(
learning_rate,
beta1=FLAGS.adam_beta1,
beta2=FLAGS.adam_beta2,
epsilon=FLAGS.opt_epsilon)
elif FLAGS.optimizer == 'ftrl':
optimizer = tf.train.FtrlOptimizer(
learning_rate,
learning_rate_power=FLAGS.ftrl_learning_rate_power,
initial_accumulator_value=FLAGS.ftrl_initial_accumulator_value,
l1_regularization_strength=FLAGS.ftrl_l1,
l2_regularization_strength=FLAGS.ftrl_l2)
elif FLAGS.optimizer == 'momentum':
optimizer = tf.train.MomentumOptimizer(
learning_rate,
momentum=FLAGS.momentum,
name='Momentum')
elif FLAGS.optimizer == 'rmsprop':
optimizer = tf.train.RMSPropOptimizer(
learning_rate,
decay=FLAGS.rmsprop_decay,
momentum=FLAGS.momentum,
epsilon=FLAGS.opt_epsilon)
elif FLAGS.optimizer == 'sgd':
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
else:
raise ValueError('Optimizer [%s] was not recognized', FLAGS.optimizer)
return optimizer
def _get_init_fn():
"""Returns a function run by the chief worker to warm-start the training.
Note that the init_fn is only run when initializing the model during the very
first global step.
Returns:
An init function run by the supervisor.
"""
if FLAGS.checkpoint_path is None:
return None
# Warn the user if a checkpoint exists in the train_dir. Then we'll be
# ignoring the checkpoint anyway.
if tf.train.latest_checkpoint(FLAGS.train_dir):
tf.logging.info(
'Ignoring --checkpoint_path because a checkpoint already exists in %s'
% FLAGS.train_dir)
return None
exclusions = []
if FLAGS.checkpoint_exclude_scopes:
exclusions = [scope.strip()
for scope in FLAGS.checkpoint_exclude_scopes.split(',')]
# TODO(sguada) variables.filter_variables()
variables_to_restore = []
for var in slim.get_model_variables():
excluded = False
for exclusion in exclusions:
if var.op.name.startswith(exclusion):
excluded = True
break
if not excluded:
variables_to_restore.append(var)
if tf.gfile.IsDirectory(FLAGS.checkpoint_path):
checkpoint_path = tf.train.latest_checkpoint(FLAGS.checkpoint_path)
else:
checkpoint_path = FLAGS.checkpoint_path
tf.logging.info('Fine-tuning from %s' % checkpoint_path)
return slim.assign_from_checkpoint_fn(
checkpoint_path,
variables_to_restore,
ignore_missing_vars=FLAGS.ignore_missing_vars)
def _get_variables_to_train():
"""Returns a list of variables to train.
Returns:
A list of variables to train by the optimizer.
"""
if FLAGS.trainable_scopes is None:
return tf.trainable_variables()
else:
scopes = [scope.strip() for scope in FLAGS.trainable_scopes.split(',')]
variables_to_train = []
for scope in scopes:
variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope)
variables_to_train.extend(variables)
return variables_to_train
def main(_):
if not FLAGS.dataset_dir:
raise ValueError('You must supply the dataset directory with --dataset_dir')
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default():
#######################
# Config model_deploy #
#######################
deploy_config = model_deploy.DeploymentConfig(
num_clones=FLAGS.num_clones,
clone_on_cpu=FLAGS.clone_on_cpu,
replica_id=FLAGS.task,
num_replicas=FLAGS.worker_replicas,
num_ps_tasks=FLAGS.num_ps_tasks)
# Create global_step
with tf.device(deploy_config.variables_device()):
global_step = slim.create_global_step()
######################
# Select the dataset #
######################
dataset = dataset_factory.get_dataset(
FLAGS.dataset_name, FLAGS.dataset_split_name, FLAGS.dataset_dir)
######################
# Select the network #
######################
network_fn = nets_factory.get_network_fn(
FLAGS.model_name,
num_classes=(dataset.num_classes - FLAGS.labels_offset),
weight_decay=FLAGS.weight_decay,
is_training=True)
#####################################
# Select the preprocessing function #
#####################################
preprocessing_name = FLAGS.preprocessing_name or FLAGS.model_name
image_preprocessing_fn = preprocessing_factory.get_preprocessing(
preprocessing_name,
is_training=True)
##############################################################
# Create a dataset provider that loads data from the dataset #
##############################################################
with tf.device(deploy_config.inputs_device()):
provider = slim.dataset_data_provider.DatasetDataProvider(
dataset,
num_readers=FLAGS.num_readers,
common_queue_capacity=20 * FLAGS.batch_size,
common_queue_min=10 * FLAGS.batch_size)
[image, label] = provider.get(['image', 'label'])
label -= FLAGS.labels_offset
train_image_size = FLAGS.train_image_size or network_fn.default_image_size
image = image_preprocessing_fn(image, train_image_size, train_image_size)
images, labels = tf.train.batch(
[image, label],
batch_size=FLAGS.batch_size,
num_threads=FLAGS.num_preprocessing_threads,
capacity=5 * FLAGS.batch_size)
labels = slim.one_hot_encoding(
labels, dataset.num_classes - FLAGS.labels_offset)
batch_queue = slim.prefetch_queue.prefetch_queue(
[images, labels], capacity=2 * deploy_config.num_clones)
####################
# Define the model #
####################
def clone_fn(batch_queue):
"""Allows data parallelism by creating multiple clones of network_fn."""
with tf.device(deploy_config.inputs_device()):
images, labels = batch_queue.dequeue()
logits, end_points = network_fn(images)
#############################
# Specify the loss function #
#############################
if 'AuxLogits' in end_points:
tf.losses.softmax_cross_entropy(
logits=end_points['AuxLogits'], onehot_labels=labels,
label_smoothing=FLAGS.label_smoothing, weights=0.4, scope='aux_loss')
tf.losses.softmax_cross_entropy(
logits=logits, onehot_labels=labels,
label_smoothing=FLAGS.label_smoothing, weights=1.0)
return end_points
# Gather initial summaries.
summaries = set(tf.get_collection(tf.GraphKeys.SUMMARIES))
clones = model_deploy.create_clones(deploy_config, clone_fn, [batch_queue])
first_clone_scope = deploy_config.clone_scope(0)
# Gather update_ops from the first clone. These contain, for example,
# the updates for the batch_norm variables created by network_fn.
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, first_clone_scope)
# Add summaries for end_points.
end_points = clones[0].outputs
for end_point in end_points:
x = end_points[end_point]
summaries.add(tf.summary.histogram('activations/' + end_point, x))
summaries.add(tf.summary.scalar('sparsity/' + end_point,
tf.nn.zero_fraction(x)))
# Add summaries for losses.
for loss in tf.get_collection(tf.GraphKeys.LOSSES, first_clone_scope):
summaries.add(tf.summary.scalar('losses/%s' % loss.op.name, loss))
# Add summaries for variables.
for variable in slim.get_model_variables():
summaries.add(tf.summary.histogram(variable.op.name, variable))
#################################
# Configure the moving averages #
#################################
if FLAGS.moving_average_decay:
moving_average_variables = slim.get_model_variables()
variable_averages = tf.train.ExponentialMovingAverage(
FLAGS.moving_average_decay, global_step)
else:
moving_average_variables, variable_averages = None, None
#########################################
# Configure the optimization procedure. #
#########################################
with tf.device(deploy_config.optimizer_device()):
learning_rate = _configure_learning_rate(dataset.num_samples, global_step)
optimizer = _configure_optimizer(learning_rate)
summaries.add(tf.summary.scalar('learning_rate', learning_rate))
if FLAGS.sync_replicas:
# If sync_replicas is enabled, the averaging will be done in the chief
# queue runner.
optimizer = tf.train.SyncReplicasOptimizer(
opt=optimizer,
replicas_to_aggregate=FLAGS.replicas_to_aggregate,
variable_averages=variable_averages,
variables_to_average=moving_average_variables,
replica_id=tf.constant(FLAGS.task, tf.int32, shape=()),
total_num_replicas=FLAGS.worker_replicas)
elif FLAGS.moving_average_decay:
# Update ops executed locally by trainer.
update_ops.append(variable_averages.apply(moving_average_variables))
# Variables to train.
variables_to_train = _get_variables_to_train()
# and returns a train_tensor and summary_op
total_loss, clones_gradients = model_deploy.optimize_clones(
clones,
optimizer,
var_list=variables_to_train)
# Add total_loss to summary.
summaries.add(tf.summary.scalar('total_loss', total_loss))
# Create gradient updates.
grad_updates = optimizer.apply_gradients(clones_gradients,
global_step=global_step)
update_ops.append(grad_updates)
update_op = tf.group(*update_ops)
with tf.control_dependencies([update_op]):
train_tensor = tf.identity(total_loss, name='train_op')
# Add the summaries from the first clone. These contain the summaries
# created by model_fn and either optimize_clones() or _gather_clone_loss().
summaries |= set(tf.get_collection(tf.GraphKeys.SUMMARIES,
first_clone_scope))
# Merge all summaries together.
summary_op = tf.summary.merge(list(summaries), name='summary_op')
###########################
# Kicks off the training. #
###########################
slim.learning.train(
train_tensor,
logdir=FLAGS.train_dir,
master=FLAGS.master,
is_chief=(FLAGS.task == 0),
init_fn=_get_init_fn(),
summary_op=summary_op,
number_of_steps=FLAGS.max_number_of_steps,
log_every_n_steps=FLAGS.log_every_n_steps,
save_summaries_secs=FLAGS.save_summaries_secs,
save_interval_secs=FLAGS.save_interval_secs,
sync_optimizer=optimizer if FLAGS.sync_replicas else None)
if __name__ == '__main__':
tf.app.run()
(5)验证模型准确率
模型训练后,需要进行验证,以确保模型的性能和准确率。eval_image_classifier.py文件代码如下所示:
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import math
import tensorflow as tf
from datasets import dataset_factory
from nets import nets_factory
from preprocessing import preprocessing_factory
slim = tf.contrib.slim
tf.app.flags.DEFINE_integer(
'batch_size', 100, 'The number of samples in each batch.')
tf.app.flags.DEFINE_integer(
'max_num_batches', None,
'Max number of batches to evaluate by default use all.')
tf.app.flags.DEFINE_string(
'master', '', 'The address of the TensorFlow master to use.')
tf.app.flags.DEFINE_string(
'checkpoint_path', '/tmp/tfmodel/',
'The directory where the model was written to or an absolute path to a '
'checkpoint file.')
tf.app.flags.DEFINE_string(
'eval_dir', '/tmp/tfmodel/', 'Directory where the results are saved to.')
tf.app.flags.DEFINE_integer(
'num_preprocessing_threads', 4,
'The number of threads used to create the batches.')
tf.app.flags.DEFINE_string(
'dataset_name', 'imagenet', 'The name of the dataset to load.')
tf.app.flags.DEFINE_string(
'dataset_split_name', 'test', 'The name of the train/test split.')
tf.app.flags.DEFINE_string(
'dataset_dir', None, 'The directory where the dataset files are stored.')
tf.app.flags.DEFINE_integer(
'labels_offset', 0,
'An offset for the labels in the dataset. This flag is primarily used to '
'evaluate the VGG and ResNet architectures which do not use a background '
'class for the ImageNet dataset.')
tf.app.flags.DEFINE_string(
'model_name', 'inception_v3', 'The name of the architecture to evaluate.')
tf.app.flags.DEFINE_string(
'preprocessing_name', None, 'The name of the preprocessing to use. If left '
'as `None`, then the model_name flag is used.')
tf.app.flags.DEFINE_float(
'moving_average_decay', None,
'The decay to use for the moving average.'
'If left as None, then moving averages are not used.')
tf.app.flags.DEFINE_integer(
'eval_image_size', None, 'Eval image size')
FLAGS = tf.app.flags.FLAGS
def main(_):
if not FLAGS.dataset_dir:
raise ValueError('You must supply the dataset directory with --dataset_dir')
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default():
tf_global_step = slim.get_or_create_global_step()
######################
# Select the dataset #
######################
dataset = dataset_factory.get_dataset(
FLAGS.dataset_name, FLAGS.dataset_split_name, FLAGS.dataset_dir)
####################
# Select the model #
####################
network_fn = nets_factory.get_network_fn(
FLAGS.model_name,
num_classes=(dataset.num_classes - FLAGS.labels_offset),
is_training=False)
##############################################################
# Create a dataset provider that loads data from the dataset #
##############################################################
provider = slim.dataset_data_provider.DatasetDataProvider(
dataset,
shuffle=False,
common_queue_capacity=2 * FLAGS.batch_size,
common_queue_min=FLAGS.batch_size)
[image, label] = provider.get(['image', 'label'])
label -= FLAGS.labels_offset
#####################################
# Select the preprocessing function #
#####################################
preprocessing_name = FLAGS.preprocessing_name or FLAGS.model_name
image_preprocessing_fn = preprocessing_factory.get_preprocessing(
preprocessing_name,
is_training=False)
eval_image_size = FLAGS.eval_image_size or network_fn.default_image_size
image = image_preprocessing_fn(image, eval_image_size, eval_image_size)
images, labels = tf.train.batch(
[image, label],
batch_size=FLAGS.batch_size,
num_threads=FLAGS.num_preprocessing_threads,
capacity=5 * FLAGS.batch_size)
####################
# Define the model #
####################
logits, _ = network_fn(images)
if FLAGS.moving_average_decay:
variable_averages = tf.train.ExponentialMovingAverage(
FLAGS.moving_average_decay, tf_global_step)
variables_to_restore = variable_averages.variables_to_restore(
slim.get_model_variables())
variables_to_restore[tf_global_step.op.name] = tf_global_step
else:
variables_to_restore = slim.get_variables_to_restore()
predictions = tf.argmax(logits, 1)
labels = tf.squeeze(labels)
# Define the metrics:
names_to_values, names_to_updates = slim.metrics.aggregate_metric_map({
'Accuracy': slim.metrics.streaming_accuracy(predictions, labels),
'Recall_5': slim.metrics.streaming_recall_at_k(
logits, labels, 5),
})
# Print the summaries to screen.
for name, value in names_to_values.items():
summary_name = 'eval/%s' % name
op = tf.summary.scalar(summary_name, value, collections=[])
op = tf.Print(op, [value], summary_name)
tf.add_to_collection(tf.GraphKeys.SUMMARIES, op)
# TODO(sguada) use num_epochs=1
if FLAGS.max_num_batches:
num_batches = FLAGS.max_num_batches
else:
# This ensures that we make a single pass over all of the data.
num_batches = math.ceil(dataset.num_samples / float(FLAGS.batch_size))
if tf.gfile.IsDirectory(FLAGS.checkpoint_path):
checkpoint_path = tf.train.latest_checkpoint(FLAGS.checkpoint_path)
else:
checkpoint_path = FLAGS.checkpoint_path
tf.logging.info('Evaluating %s' % checkpoint_path)
slim.evaluation.evaluate_once(
master=FLAGS.master,
checkpoint_path=checkpoint_path,
logdir=FLAGS.eval_dir,
num_evals=num_batches,
eval_op=list(names_to_updates.values()),
variables_to_restore=variables_to_restore)
if __name__ == '__main__':
tf.app.run()
(6)数据可视化
在训练过程中,可以使用tensorboard对训练过程进行可视化,有助于设定训练模型的方式和超参数。
tensorboard --logdir satellite/train_dir
(7)导出模型并对单张图片进行识别
在训练完模型后可以进行导出保存,并进行单张图片识别。
Tensorflow提供了导出网络结构的脚本export_inference_graph.py在slim文件夹下运行:
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
from tensorflow.python.platform import gfile
from datasets import dataset_factory
from nets import nets_factory
slim = tf.contrib.slim
tf.app.flags.DEFINE_string(
'model_name', 'inception_v3', 'The name of the architecture to save.')
tf.app.flags.DEFINE_boolean(
'is_training', False,
'Whether to save out a training-focused version of the model.')
tf.app.flags.DEFINE_integer(
'default_image_size', 224,
'The image size to use if the model does not define it.')
tf.app.flags.DEFINE_string('dataset_name', 'imagenet',
'The name of the dataset to use with the model.')
tf.app.flags.DEFINE_integer(
'labels_offset', 0,
'An offset for the labels in the dataset. This flag is primarily used to '
'evaluate the VGG and ResNet architectures which do not use a background '
'class for the ImageNet dataset.')
tf.app.flags.DEFINE_string(
'output_file', '', 'Where to save the resulting file to.')
tf.app.flags.DEFINE_string(
'dataset_dir', '', 'Directory to save intermediate dataset files to')
FLAGS = tf.app.flags.FLAGS
def main(_):
if not FLAGS.output_file:
raise ValueError('You must supply the path to save to with --output_file')
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default() as graph:
dataset = dataset_factory.get_dataset(FLAGS.dataset_name, 'validation',
FLAGS.dataset_dir)
network_fn = nets_factory.get_network_fn(
FLAGS.model_name,
num_classes=(dataset.num_classes - FLAGS.labels_offset),
is_training=FLAGS.is_training)
if hasattr(network_fn, 'default_image_size'):
image_size = network_fn.default_image_size
else:
image_size = FLAGS.default_image_size
placeholder = tf.placeholder(name='input', dtype=tf.float32,
shape=[1, image_size, image_size, 3])
network_fn(placeholder)
graph_def = graph.as_graph_def()
with gfile.GFile(FLAGS.output_file, 'wb') as f:
f.write(graph_def.SerializeToString())
if __name__ == '__main__':
tf.app.run()
此时会在satellite文件夹下生成inception_v3_inf_graph.pb文件,该文件只保存了网络结构,并没有保存模型参数。下面需要将模型参数进行保存,文中给出freeze_graph.py文件进行保存参数。
freeze_graph.py
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import argparse
import sys
from google.protobuf import text_format
from tensorflow.core.framework import graph_pb2
from tensorflow.core.protobuf import saver_pb2
from tensorflow.python import pywrap_tensorflow
from tensorflow.python.client import session
from tensorflow.python.framework import graph_util
from tensorflow.python.framework import importer
from tensorflow.python.platform import app
from tensorflow.python.platform import gfile
from tensorflow.python.training import saver as saver_lib
FLAGS = None
def freeze_graph_with_def_protos(
input_graph_def,
input_saver_def,
input_checkpoint,
output_node_names,
restore_op_name,
filename_tensor_name,
output_graph,
clear_devices,
initializer_nodes,
variable_names_blacklist=""):
"""Converts all variables in a graph and checkpoint into constants."""
del restore_op_name, filename_tensor_name # Unused by updated loading code.
# 'input_checkpoint' may be a prefix if we're using Saver V2 format
if not saver_lib.checkpoint_exists(input_checkpoint):
print("Input checkpoint '" + input_checkpoint + "' doesn't exist!")
return -1
if not output_node_names:
print("You need to supply the name of a node to --output_node_names.")
return -1
# Remove all the explicit device specifications for this node. This helps to
# make the graph more portable.
if clear_devices:
for node in input_graph_def.node:
node.device = ""
_ = importer.import_graph_def(input_graph_def, name="")
with session.Session() as sess:
if input_saver_def:
saver = saver_lib.Saver(saver_def=input_saver_def)
saver.restore(sess, input_checkpoint)
else:
var_list = {}
reader = pywrap_tensorflow.NewCheckpointReader(input_checkpoint)
var_to_shape_map = reader.get_variable_to_shape_map()
for key in var_to_shape_map:
try:
tensor = sess.graph.get_tensor_by_name(key + ":0")
except KeyError:
# This tensor doesn't exist in the graph (for example it's
# 'global_step' or a similar housekeeping element) so skip it.
continue
var_list[key] = tensor
saver = saver_lib.Saver(var_list=var_list)
saver.restore(sess, input_checkpoint)
if initializer_nodes:
sess.run(initializer_nodes)
variable_names_blacklist = (variable_names_blacklist.split(",") if
variable_names_blacklist else None)
output_graph_def = graph_util.convert_variables_to_constants(
sess,
input_graph_def,
output_node_names.split(","),
variable_names_blacklist=variable_names_blacklist)
# Write GraphDef to file if output path has been given.
if output_graph:
with gfile.GFile(output_graph, "wb") as f:
f.write(output_graph_def.SerializeToString())
print("%d ops in the final graph." % len(output_graph_def.node))
return output_graph_def
def _parse_input_graph_proto(input_graph, input_binary):
"""Parser input tensorflow graph into GraphDef proto."""
if not gfile.Exists(input_graph):
print("Input graph file '" + input_graph + "' does not exist!")
return -1
input_graph_def = graph_pb2.GraphDef()
mode = "rb" if input_binary else "r"
with gfile.FastGFile(input_graph, mode) as f:
if input_binary:
input_graph_def.ParseFromString(f.read())
else:
text_format.Merge(f.read(), input_graph_def)
return input_graph_def
def _parse_input_saver_proto(input_saver, input_binary):
"""Parser input tensorflow Saver into SaverDef proto."""
if not gfile.Exists(input_saver):
print("Input saver file '" + input_saver + "' does not exist!")
return -1
mode = "rb" if input_binary else "r"
with gfile.FastGFile(input_saver, mode) as f:
saver_def = saver_pb2.SaverDef()
if input_binary:
saver_def.ParseFromString(f.read())
else:
text_format.Merge(f.read(), saver_def)
return saver_def
def freeze_graph(input_graph,
input_saver,
input_binary,
input_checkpoint,
output_node_names,
restore_op_name,
filename_tensor_name,
output_graph,
clear_devices,
initializer_nodes,
variable_names_blacklist=""):
"""Converts all variables in a graph and checkpoint into constants."""
input_graph_def = _parse_input_graph_proto(input_graph, input_binary)
input_saver_def = None
if input_saver:
input_saver_def = _parse_input_saver_proto(input_saver, input_binary)
freeze_graph_with_def_protos(
input_graph_def,
input_saver_def,
input_checkpoint,
output_node_names,
restore_op_name,
filename_tensor_name,
output_graph,
clear_devices,
initializer_nodes,
variable_names_blacklist)
def main(unused_args):
freeze_graph(FLAGS.input_graph, FLAGS.input_saver, FLAGS.input_binary,
FLAGS.input_checkpoint, FLAGS.output_node_names,
FLAGS.restore_op_name, FLAGS.filename_tensor_name,
FLAGS.output_graph, FLAGS.clear_devices, FLAGS.initializer_nodes,
FLAGS.variable_names_blacklist)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.register("type", "bool", lambda v: v.lower() == "true")
parser.add_argument(
"--input_graph",
type=str,
default="",
help="TensorFlow \'GraphDef\' file to load.")
parser.add_argument(
"--input_saver",
type=str,
default="",
help="TensorFlow saver file to load.")
parser.add_argument(
"--input_checkpoint",
type=str,
default="",
help="TensorFlow variables file to load.")
parser.add_argument(
"--output_graph",
type=str,
default="",
help="Output \'GraphDef\' file name.")
parser.add_argument(
"--input_binary",
nargs="?",
const=True,
type="bool",
default=False,
help="Whether the input files are in binary format.")
parser.add_argument(
"--output_node_names",
type=str,
default="",
help="The name of the output nodes, comma separated.")
parser.add_argument(
"--restore_op_name",
type=str,
default="save/restore_all",
help="The name of the master restore operator.")
parser.add_argument(
"--filename_tensor_name",
type=str,
default="save/Const:0",
help="The name of the tensor holding the save path.")
parser.add_argument(
"--clear_devices",
nargs="?",
const=True,
type="bool",
default=True,
help="Whether to remove device specifications.")
parser.add_argument(
"--initializer_nodes",
type=str,
default="",
help="comma separated list of initializer nodes to run before freezing.")
parser.add_argument(
"--variable_names_blacklist",
type=str,
default="",
help="""\
comma separated list of variables to skip converting to constants\
""")
FLAGS, unparsed = parser.parse_known_args()
app.run(main=main, argv=[sys.argv[0]] + unparsed)
前面网络模型和参数都已经导出成功,下面就是使用导出的模型进行图片识别。在classify_image_inception_v3.py文件代码如下:
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import argparse
import os.path
import re
import sys
import tarfile
import numpy as np
from six.moves import urllib
import tensorflow as tf
FLAGS = None
class NodeLookup(object):
def __init__(self, label_lookup_path=None):
self.node_lookup = self.load(label_lookup_path)
def load(self, label_lookup_path):
node_id_to_name = {}
with open(label_lookup_path) as f:
for index, line in enumerate(f):
node_id_to_name[index] = line.strip()
return node_id_to_name
def id_to_string(self, node_id):
if node_id not in self.node_lookup:
return ''
return self.node_lookup[node_id]
def create_graph():
"""Creates a graph from saved GraphDef file and returns a saver."""
# Creates graph from saved graph_def.pb.
with tf.gfile.FastGFile(FLAGS.model_path, 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(graph_def, name='')
def preprocess_for_eval(image, height, width,
central_fraction=0.875, scope=None):
with tf.name_scope(scope, 'eval_image', [image, height, width]):
if image.dtype != tf.float32:
image = tf.image.convert_image_dtype(image, dtype=tf.float32)
# Crop the central region of the image with an area containing 87.5% of
# the original image.
if central_fraction:
image = tf.image.central_crop(image, central_fraction=central_fraction)
if height and width:
# Resize the image to the specified height and width.
image = tf.expand_dims(image, 0)
image = tf.image.resize_bilinear(image, [height, width],
align_corners=False)
image = tf.squeeze(image, [0])
image = tf.subtract(image, 0.5)
image = tf.multiply(image, 2.0)
return image
def run_inference_on_image(image):
"""Runs inference on an image.
Args:
image: Image file name.
Returns:
Nothing
"""
with tf.Graph().as_default():
image_data = tf.gfile.FastGFile(image, 'rb').read()
image_data = tf.image.decode_jpeg(image_data)
image_data = preprocess_for_eval(image_data, 299, 299)
image_data = tf.expand_dims(image_data, 0)
with tf.Session() as sess:
image_data = sess.run(image_data)
# Creates graph from saved GraphDef.
create_graph()
with tf.Session() as sess:
softmax_tensor = sess.graph.get_tensor_by_name('InceptionV3/Logits/SpatialSqueeze:0')
predictions = sess.run(softmax_tensor,
{'input:0': image_data})
predictions = np.squeeze(predictions)
# Creates node ID --> English string lookup.
node_lookup = NodeLookup(FLAGS.label_path)
top_k = predictions.argsort()[-FLAGS.num_top_predictions:][::-1]
for node_id in top_k:
human_string = node_lookup.id_to_string(node_id)
score = predictions[node_id]
print('%s (score = %.5f)' % (human_string, score))
def main(_):
image = FLAGS.image_file
run_inference_on_image(image)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument(
'--model_path',
type=str,
)
parser.add_argument(
'--label_path',
type=str,
)
parser.add_argument(
'--image_file',
type=str,
default='',
help='Absolute path to image file.'
)
parser.add_argument(
'--num_top_predictions',
type=int,
default=5,
help='Display this many predictions.'
)
FLAGS, unparsed = parser.parse_known_args()
tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)
作业
(1)下载花卉数据集flower-10,生成并生成tfrecord数据集文件
(2)使用lenet-5网络进行识别分类
(3)了解VGG16网络
(4)继续学习Tensorflow视频课(P24-P30)
本期学习结束,欢迎大家关注、批评指正。