背景调查公司 Onfido 研究主管 Peter Roelants 在 Medium 上发表了一篇题为《Higher-Level APIs in TensorFlow》的文章,通过实例详细介绍了如何使用 TensorFlow 中的高级 API(Estimator、Experiment 和 Dataset)训练模型。值得一提的是 Experiment 和 Dataset 可以独立使用。这些高级 API 已被最新发布的 TensorFlow1.3 版收录。
TensorFlow 中有许多流行的库,如 Keras、TFLearn 和 Sonnet,它们可以让你轻松训练模型,而无需接触哪些低级别函数。目前,Keras API 正倾向于直接在 TensorFlow 中实现,TensorFlow 也在提供越来越多的高级构造,其中的一些已经被最新发布的 TensorFlow1.3 版收录。
在本文中,我们将通过一个例子来学习如何使用一些高级构造,其中包括 Estimator、Experiment 和 Dataset。阅读本文需要预先了解有关 TensorFlow 的基本知识。
Experiment、Estimator 和 DataSet 框架和它们的相互作用(以下将对这些组件进行说明)
在本文中,我们使用 MNIST 作为数据集。它是一个易于使用的数据集,可以通过 TensorFlow 访问。你可以在这个 gist 中找到完整的示例代码。使用这些框架的一个好处是我们不需要直接处理图形和会话。
Estimator
Estimator(评估器)类代表一个模型,以及这些模型被训练和评估的方式。我们可以这样构建一个评估器:
return tf.estimator.Estimator(
model_fn=model_fn, # First-class function
params=params, # HParams
config=run_config # RunConfig
为了构建一个 Estimator,我们需要传递一个模型函数,一个参数集合以及一些配置。
参数应该是模型超参数的集合,它可以是一个字典,但我们将在本示例中将其表示为 HParams 对象,用作 namedtuple。
该配置指定如何运行训练和评估,以及如何存出结果。这些配置通过 RunConfig 对象表示,该对象传达 Estimator 需要了解的关于运行模型的环境的所有内容。
模型函数是一个 Python 函数,它构建了给定输入的模型(见后文)。
模型函数
模型函数是一个 Python 函数,它作为第一级函数传递给 Estimator。稍后我们就会看到,TensorFlow 也会在其他地方使用第一级函数。模型表示为函数的好处在于模型可以通过实例化函数不断重新构建。该模型可以在训练过程中被不同的输入不断创建,例如:在训练期间运行验证测试。
模型函数将输入特征作为参数,相应标签作为张量。它还有一种模式来标记模型是否正在训练、评估或执行推理。模型函数的最后一个参数是超参数的集合,它们与传递给 Estimator 的内容相同。模型函数需要返回一个 EstimatorSpec 对象——它会定义完整的模型。
EstimatorSpec 接受预测,损失,训练和评估几种操作,因此它定义了用于训练,评估和推理的完整模型图。由于 EstimatorSpec 采用常规 TensorFlow Operations,因此我们可以使用像 TF-Slim 这样的框架来定义自己的模型。
Experiment
Experiment(实验)类是定义如何训练模型,并将其与 Estimator 进行集成的方式。我们可以这样创建一个实验类:
experiment = tf.contrib.learn.Experiment(
estimator=estimator, # Estimator
train_input_fn=train_input_fn, # First-class function
eval_input_fn=eval_input_fn, # First-class function
train_steps=params.train_steps, # Minibatch steps
min_eval_frequency=params.min_eval_frequency, # Eval frequency
train_monitors=[train_input_hook], # Hooks for training
eval_hooks=[eval_input_hook], # Hooks for evaluation
eval_steps=None # Use evaluation feeder until its empty
Experiment 作为输入:
一个 Estimator(例如上面定义的那个)。
训练和评估数据作为第一级函数。这里用到了和前述模型函数相同的概念,通过传递函数而非操作,如有需要,输入图可以被重建。我们会在后面继续讨论这个概念。
训练和评估钩子(hooks)。这些钩子可以用于监视或保存特定内容,或在图形和会话中进行一些操作。例如,我们将通过操作来帮助初始化数据加载器。
不同参数解释了训练时间和评估时间。
一旦我们定义了 experiment,我们就可以通过 learn_runner.run 运行它来训练和评估模型:
learn_runner.run(
experiment_fn=experiment_fn, # First-class function
run_config=run_config, # RunConfig
schedule="train_and_evaluate", # What to run
hparams=params # HParams
与模型函数和数据函数一样,函数中的学习运算符将创建 experiment 作为参数。
Dataset
我们将使用 Dataset 类和相应的 Iterator 来表示我们的训练和评估数据,并创建在训练期间迭代数据的数据馈送器。在本示例中,我们将使用 TensorFlow 中可用的 MNIST 数据,并在其周围构建一个 Dataset 包装器。例如,我们把训练的输入数据表示为:
# Define the training inputs
def get_train_inputs(batch_size, mnist_data):
"""Return the input function to get the training data.
Args:
batch_size (int): Batch size of training iterator that is returned
by the input function.
mnist_data (Object): Object holding the loaded mnist data.
Returns:
(Input function, IteratorInitializerHook):
- Function that returns (features, labels) when called.
- Hook to initialise input iterator.
iterator_initializer_hook = IteratorInitializerHook()
def train_inputs():
"""Returns training set as Operations.
Returns:
(features, labels) Operations that iterate over the dataset
on every evaluation
with tf.name_scope('Training_data'):
# Get Mnist data
images = mnist_data.train.images.reshape([-1, 28, 28, 1])
labels = mnist_data.train.labels
# Define placeholders
images_placeholder = tf.placeholder(
images.dtype, images.shape)
labels_placeholder = tf.placeholder(
labels.dtype, labels.shape)
# Build dataset iterator
dataset = tf.contrib.data.Dataset.from_tensor_slices(
(images_placeholder, labels_placeholder))
dataset = dataset.repeat(None) # Infinite iterations
dataset = dataset.shuffle(buffer_size=10000)
dataset = dataset.batch(batch_size)
iterator = dataset.make_initializable_iterator()
next_example, next_label = iterator.get_next()
# Set runhook to initialize iterator
iterator_initializer_hook.iterator_initializer_func = \
lambda sess: sess.run(
iterator.initializer,
feed_dict={images_placeholder: images,
labels_placeholder: labels})
# Return batched (features, labels)
return next_example, next_label
# Return function and hook
return train_inputs, iterator_initializer_hook
调用这个 get_train_inputs 会返回一个一级函数,它在 TensorFlow 图中创建数据加载操作,以及一个 Hook 初始化迭代器。
本示例中,我们使用的 MNIST 数据最初表示为 Numpy 数组。我们创建一个占位符张量来获取数据,再使用占位符来避免数据被复制。接下来,我们在 from_tensor_slices 的帮助下创建一个切片数据集。我们将确保该数据集运行无限长时间(experiment 可以考虑 epoch 的数量),让数据得到清晰,并分成所需的尺寸。
为了迭代数据,我们需要在数据集的基础上创建迭代器。因为我们正在使用占位符,所以我们需要在 NumPy 数据的相关会话中初始化占位符。我们可以通过创建一个可初始化的迭代器来实现。创建图形时,我们将创建一个自定义的 IteratorInitializerHook 对象来初始化迭代器:
class IteratorInitializerHook(tf.train.SessionRunHook):
"""Hook to initialise data iterator after Session is created."""
def __init__(self):
super(IteratorInitializerHook, self).__init__()
self.iterator_initializer_func = None
def after_create_session(self, session, coord):
"""Initialise the iterator after the session has been created."""
self.iterator_initializer_func(session)
IteratorInitializerHook 继承自 SessionRunHook。一旦创建了相关会话,这个钩子就会调用 call after_create_session,并用正确的数据初始化占位符。这个钩子会通过 get_train_inputs 函数返回,并在创建时传递给 Experiment 对象。
train_inputs 函数返回的数据加载操作是 TensorFlow 操作,每次评估时都会返回一个新的批处理。
运行代码
现在我们已经定义了所有的东西,我们可以用以下命令运行代码:
python mnist_estimator.py—model_dir ./mnist_training—data_dir ./mnist_data
如果你不传递参数,它将使用文件顶部的默认标志来确定保存数据和模型的位置。训练将在终端输出全局步长、损失、精度等信息。除此之外,实验和估算器框架将记录 TensorBoard 可以显示的某些统计信息。如果我们运行:
tensorboard—logdir='./mnist_training'
我们就可以看到所有训练统计数据,如训练损失、评估准确性、每步时间和模型图。
评估精度在 TensorBoard 中的可视化
在 TensorFlow 中,有关 Estimator、Experiment 和 Dataset 框架的示例很少,这也是本文存在的原因。希望这篇文章可以向大家介绍这些架构工作的原理,它们应该采用哪些抽象方法,以及如何使用它们。如果你对它们很感兴趣,以下是其他相关文档。
关于 Estimator、Experiment 和 Dataset 的注释
论文《TensorFlow Estimators: Managing Simplicity vs. Flexibility in High-Level Machine Learning Frameworks》:
Using the Dataset API for TensorFlow Input Pipelines:
tf.estimator.Estimator:
tf.contrib.learn.RunConfig:
tf.estimator.DNNClassifier:
tf.estimator.DNNRegressor:
Creating Estimators in tf.estimator:
tf.contrib.learn.Head:
本文用到的 Slim 框架:
完整示例
"""Script to illustrate usage of tf.estimator.Estimator in TF v1.3"""
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data as mnist_data
from tensorflow.contrib import slim
from tensorflow.contrib.learn import ModeKeys
from tensorflow.contrib.learn import learn_runner
# Show debugging output
tf.logging.set_verbosity(tf.logging.DEBUG)
# Set default flags for the output directories
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_string(
flag_name='model_dir', default_value='./mnist_training',
docstring='Output directory for model and training stats.')
tf.app.flags.DEFINE_string(
flag_name='data_dir', default_value='./mnist_data',
docstring='Directory to download the data to.')
# Define and run experiment ###############################
def run_experiment(argv=None):
# Define model parameters
params = tf.contrib.training.HParams(
learning_rate=0.002,
n_classes=10,
train_steps=5000,
min_eval_frequency=100
# Set the run_config and the directory to save the model and stats
run_config = tf.contrib.learn.RunConfig()
run_config = run_config.replace(model_dir=FLAGS.model_dir)
learn_runner.run(
experiment_fn=experiment_fn, # First-class function
run_config=run_config, # RunConfig
schedule="train_and_evaluate", # What to run
hparams=params # HParams
def experiment_fn(run_config, params):
"""Create an experiment to train and evaluate the model.
Args:
run_config (RunConfig): Configuration for Estimator run.
params (HParam): Hyperparameters
Returns:
(Experiment) Experiment for training the mnist model.
# You can change a subset of the run_config properties as
run_config = run_config.replace(
save_checkpoints_steps=params.min_eval_frequency)
# Define the mnist classifier
estimator = get_estimator(run_config, params)
# Setup data loaders
mnist = mnist_data.read_data_sets(FLAGS.data_dir, one_hot=False)
train_input_fn, train_input_hook = get_train_inputs(
batch_size=128, mnist_data=mnist)
eval_input_fn, eval_input_hook = get_test_inputs(
batch_size=128, mnist_data=mnist)
# Define the experiment
experiment = tf.contrib.learn.Experiment(
estimator=estimator, # Estimator
train_input_fn=train_input_fn, # First-class function
eval_input_fn=eval_input_fn, # First-class function
train_steps=params.train_steps, # Minibatch steps
min_eval_frequency=params.min_eval_frequency, # Eval frequency
train_monitors=[train_input_hook], # Hooks for training
eval_hooks=[eval_input_hook], # Hooks for evaluation
eval_steps=None # Use evaluation feeder until its empty
return experiment
# Define model ############################################
def get_estimator(run_config, params):
"""Return the model as a Tensorflow Estimator object.
Args:
run_config (RunConfig): Configuration for Estimator run.
params (HParams): hyperparameters.
return tf.estimator.Estimator(
model_fn=model_fn, # First-class function
params=params, # HParams
config=run_config # RunConfig
def model_fn(features, labels, mode, params):
"""Model function used in the estimator.
Args:
features (Tensor): Input features to the model.
labels (Tensor): Labels tensor for training and evaluation.
mode (ModeKeys): Specifies if training, evaluation or prediction.
params (HParams): hyperparameters.
Returns:
(EstimatorSpec): Model to be run by Estimator.
is_training = mode == ModeKeys.TRAIN
# Define model's architecture
logits = architecture(features, is_training=is_training)
predictions = tf.argmax(logits, axis=-1)
loss = tf.losses.sparse_softmax_cross_entropy(
labels=tf.cast(labels, tf.int32),
logits=logits
return tf.estimator.EstimatorSpec(
mode=mode,
predictions=predictions,
loss=loss,
train_op=get_train_op_fn(loss, params),
eval_metric_ops=get_eval_metric_ops(labels, predictions)
def get_train_op_fn(loss, params):
"""Get the training Op.
Args:
loss (Tensor): Scalar Tensor that represents the loss function.
params (HParams): Hyperparameters (needs to have `learning_rate`)
Returns:
Training Op
return tf.contrib.layers.optimize_loss(
loss=loss,
global_step=tf.contrib.framework.get_global_step(),
optimizer=tf.train.AdamOptimizer,
learning_rate=params.learning_rate
def get_eval_metric_ops(labels, predictions):
"""Return a dict of the evaluation Ops.
Args:
labels (Tensor): Labels tensor for training and evaluation.
predictions (Tensor): Predictions Tensor.
Returns:
Dict of metric results keyed by name.
return {
'Accuracy': tf.metrics.accuracy(
labels=labels,
predictions=predictions,
name='accuracy')
def architecture(inputs, is_training, scope='MnistConvNet'):
"""Return the output operation following the network architecture.
Args:
inputs (Tensor): Input Tensor
is_training (bool): True iff in training mode
scope (str): Name of the scope of the architecture
Returns:
Logits output Op for the network.
with tf.variable_scope(scope):
with slim.arg_scope(
[slim.conv2d, slim.fully_connected],
weights_initializer=tf.contrib.layers.xavier_initializer()):
net = slim.conv2d(inputs, 20, [5, 5], padding='VALID',
scope='conv1')
net = slim.max_pool2d(net, 2, stride=2, scope='pool2')
net = slim.conv2d(net, 40, [5, 5], padding='VALID',
scope='conv3')
net = slim.max_pool2d(net, 2, stride=2, scope='pool4')
net = tf.reshape(net, [-1, 4 * 4 * 40])
net = slim.fully_connected(net, 256, scope='fn5')
net = slim.dropout(net, is_training=is_training,
scope='dropout5')
net = slim.fully_connected(net, 256, scope='fn6')
net = slim.dropout(net, is_training=is_training,
scope='dropout6')
net = slim.fully_connected(net, 10, scope='output',
activation_fn=None)
return net
# Define data loaders #####################################
class IteratorInitializerHook(tf.train.SessionRunHook):
"""Hook to initialise data iterator after Session is created."""
def __init__(self):
super(IteratorInitializerHook, self).__init__()
self.iterator_initializer_func = None
def after_create_session(self, session, coord):
"""Initialise the iterator after the session has been created."""
self.iterator_initializer_func(session)
# Define the training inputs
def get_train_inputs(batch_size, mnist_data):
"""Return the input function to get the training data.
Args:
batch_size (int): Batch size of training iterator that is returned
by the input function.
mnist_data (Object): Object holding the loaded mnist data.
Returns:
(Input function, IteratorInitializerHook):
- Function that returns (features, labels) when called.
- Hook to initialise input iterator.
iterator_initializer_hook = IteratorInitializerHook()
def train_inputs():
"""Returns training set as Operations.
Returns:
(features, labels) Operations that iterate over the dataset
on every evaluation
with tf.name_scope('Training_data'):
# Get Mnist data
images = mnist_data.train.images.reshape([-1, 28, 28, 1])
labels = mnist_data.train.labels
# Define placeholders
images_placeholder = tf.placeholder(
images.dtype, images.shape)
labels_placeholder = tf.placeholder(
labels.dtype, labels.shape)
# Build dataset iterator
dataset = tf.contrib.data.Dataset.from_tensor_slices(
(images_placeholder, labels_placeholder))
dataset = dataset.repeat(None) # Infinite iterations
dataset = dataset.shuffle(buffer_size=10000)
dataset = dataset.batch(batch_size)
iterator = dataset.make_initializable_iterator()
next_example, next_label = iterator.get_next()
# Set runhook to initialize iterator
iterator_initializer_hook.iterator_initializer_func = \
lambda sess: sess.run(
iterator.initializer,
feed_dict={images_placeholder: images,
labels_placeholder: labels})
# Return batched (features, labels)
return next_example, next_label
# Return function and hook
return train_inputs, iterator_initializer_hook
def get_test_inputs(batch_size, mnist_data):
"""Return the input function to get the test data.
Args:
batch_size (int): Batch size of training iterator that is returned
by the input function.
mnist_data (Object): Object holding the loaded mnist data.
Returns:
(Input function, IteratorInitializerHook):
- Function that returns (features, labels) when called.
- Hook to initialise input iterator.
iterator_initializer_hook = IteratorInitializerHook()
def test_inputs():
"""Returns training set as Operations.
Returns:
(features, labels) Operations that iterate over the dataset
on every evaluation
with tf.name_scope('Test_data'):
# Get Mnist data
images = mnist_data.test.images.reshape([-1, 28, 28, 1])
labels = mnist_data.test.labels
# Define placeholders
images_placeholder = tf.placeholder(
images.dtype, images.shape)
labels_placeholder = tf.placeholder(
labels.dtype, labels.shape)
# Build dataset iterator
dataset = tf.contrib.data.Dataset.from_tensor_slices(
(images_placeholder, labels_placeholder))
dataset = dataset.batch(batch_size)
iterator = dataset.make_initializable_iterator()
next_example, next_label = iterator.get_next()
# Set runhook to initialize iterator
iterator_initializer_hook.iterator_initializer_func = \
lambda sess: sess.run(
iterator.initializer,
feed_dict={images_placeholder: images,
labels_placeholder: labels})
return next_example, next_label
# Return function and hook
return test_inputs, iterator_initializer_hook
# Run script ##############################################
if __name__ == "__main__":
tf.app.run(
main=run_experiment