Permalink
909 lines (733 sloc) 37.6 KB

TensorFlow-Slim

TF-Slim is a lightweight library for defining, training and evaluating complex models in TensorFlow. Components of tf-slim can be freely mixed with native tensorflow, as well as other frameworks, such as tf.contrib.learn.

Usage

import tensorflow.contrib.slim as slim

Why TF-Slim?

TF-Slim is a library that makes building, training and evaluation neural networks simple:

  • Allows the user to define models much more compactly by eliminating boilerplate code. This is accomplished through the use of argument scoping and numerous high level layers and variables. These tools increase readability and maintainability, reduce the likelihood of an error from copy-and-pasting hyperparameter values and simplifies hyperparameter tuning.
  • Makes developing models simple by providing commonly used regularizers.
  • Several widely used computer vision models (e.g., VGG, AlexNet) have been developed in slim, and are available to users. These can either be used as black boxes, or can be extended in various ways, e.g., by adding "multiple heads" to different internal layers.
  • Slim makes it easy to extend complex models, and to warm start training algorithms by using pieces of pre-existing model checkpoints.

What are the various components of TF-Slim?

TF-Slim is composed of several parts which were design to exist independently. These include the following main pieces (explained in detail below).

  • arg_scope: provides a new scope named arg_scope that allows a user to define default arguments for specific operations within that scope.
  • data: contains TF-slim's dataset definition, data providers, parallel_reader, and decoding utilities.
  • evaluation: contains routines for evaluating models.
  • layers: contains high level layers for building models using tensorflow.
  • learning: contains routines for training models.
  • losses: contains commonly used loss functions.
  • metrics: contains popular evaluation metrics.
  • nets: contains popular network definitions such as VGG and AlexNet models.
  • queues: provides a context manager for easily and safely starting and closing QueueRunners.
  • regularizers: contains weight regularizers.
  • variables: provides convenience wrappers for variable creation and manipulation.

Defining Models

Models can be succinctly defined using TF-Slim by combining its variables, layers and scopes. Each of these elements are defined below.

Variables

Creating Variables in native tensorflow requires either a predefined value or an initialization mechanism (e.g. randomly sampled from a Gaussian). Furthermore, if a variable needs to be created on a specific device, such as a GPU, the specification must be made explicit. To alleviate the code required for variable creation, TF-Slim provides a set of thin wrapper functions in variables.py which allow callers to easily define variables.

For example, to create a weight variable, initialize it using a truncated normal distribution, regularize it with an l2_loss and place it on the CPU, one need only declare the following:

weights = slim.variable('weights',
                             shape=[10, 10, 3 , 3],
                             initializer=tf.truncated_normal_initializer(stddev=0.1),
                             regularizer=slim.l2_regularizer(0.05),
                             device='/CPU:0')

Note that in native TensorFlow, there are two types of variables: regular variables and local (transient) variables. The vast majority of variables are regular variables: once created, they can be saved to disk using a saver. Local variables are those variables that only exist for the duration of a session and are not saved to disk.

TF-Slim further differentiates variables by defining model variables, which are variables that represent parameters of a model. Model variables are trained or fine-tuned during learning and are loaded from a checkpoint during evaluation or inference. Examples include the variables created by a slim.fully_connected or slim.conv2d layer. Non-model variables are all other variables that are used during learning or evaluation but are not required for actually performing inference. For example, the global_step is a variable using during learning and evaluation but it is not actually part of the model. Similarly, moving average variables might mirror model variables, but the moving averages are not themselves model variables.

Both model variables and regular variables can be easily created and retrieved via TF-Slim:

# Model Variables
weights = slim.model_variable('weights',
                              shape=[10, 10, 3 , 3],
                              initializer=tf.truncated_normal_initializer(stddev=0.1),
                              regularizer=slim.l2_regularizer(0.05),
                              device='/CPU:0')
model_variables = slim.get_model_variables()

# Regular variables
my_var = slim.variable('my_var',
                       shape=[20, 1],
                       initializer=tf.zeros_initializer())
regular_variables_and_model_variables = slim.get_variables()

How does this work? When you create a model variable via TF-Slim's layers or directly via the slim.model_variable function, TF-Slim adds the variable to a the tf.GraphKeys.MODEL_VARIABLES collection. What if you have your own custom layers or variable creation routine but still want TF-Slim to manage or be aware of your model variables? TF-Slim provides a convenience function for adding the model variable to its collection:

my_model_variable = CreateViaCustomCode()

# Letting TF-Slim know about the additional variable.
slim.add_model_variable(my_model_variable)

Layers

While the set of TensorFlow operations is quite extensive, developers of neural networks typically think of models in terms of higher level concepts like "layers", "losses", "metrics", and "networks". A layer, such as a Convolutional Layer, a Fully Connected Layer or a BatchNorm Layer are more abstract than a single TensorFlow operation and typically involve several operations. Furthermore, a layer usually (but not always) has variables (tunable parameters) associated with it, unlike more primitive operations. For example, a Convolutional Layer in a neural network is composed of several low level operations:

  1. Creating the weight and bias variables
  2. Convolving the weights with the input from the previous layer
  3. Adding the biases to the result of the convolution.
  4. Applying an activation function.

Using only plain TensorFlow code, this can be rather laborious:

input = ...
with tf.name_scope('conv1_1') as scope:
  kernel = tf.Variable(tf.truncated_normal([3, 3, 64, 128], dtype=tf.float32,
                                           stddev=1e-1), name='weights')
  conv = tf.nn.conv2d(input, kernel, [1, 1, 1, 1], padding='SAME')
  biases = tf.Variable(tf.constant(0.0, shape=[128], dtype=tf.float32),
                       trainable=True, name='biases')
  bias = tf.nn.bias_add(conv, biases)
  conv1 = tf.nn.relu(bias, name=scope)

To alleviate the need to duplicate this code repeatedly, TF-Slim provides a number of convenient operations defined at the more abstract level of neural network layers. For example, compare the code above to an invocation of the corresponding TF-Slim code:

input = ...
net = slim.conv2d(input, 128, [3, 3], scope='conv1_1')

TF-Slim provides standard implementations for numerous components for building neural networks. These include:

Layer TF-Slim
BiasAdd slim.bias_add
BatchNorm slim.batch_norm
Conv2d slim.conv2d
Conv2dInPlane slim.conv2d_in_plane
Conv2dTranspose (Deconv) slim.conv2d_transpose
FullyConnected slim.fully_connected
AvgPool2D slim.avg_pool2d
Dropout slim.dropout
Flatten slim.flatten
MaxPool2D slim.max_pool2d
OneHotEncoding slim.one_hot_encoding
SeparableConv2 slim.separable_conv2d
UnitNorm slim.unit_norm

TF-Slim also provides two meta-operations called repeat and stack that allow users to repeatedly perform the same operation. For example, consider the following snippet from the VGG network whose layers perform several convolutions in a row between pooling layers:

net = ...
net = slim.conv2d(net, 256, [3, 3], scope='conv3_1')
net = slim.conv2d(net, 256, [3, 3], scope='conv3_2')
net = slim.conv2d(net, 256, [3, 3], scope='conv3_3')
net = slim.max_pool2d(net, [2, 2], scope='pool3')

One way to reduce this code duplication would be via a for loop:

net = ...
for i in range(3):
  net = slim.conv2d(net, 256, [3, 3], scope='conv3_' % (i+1))
net = slim.max_pool2d(net, [2, 2], scope='pool3')

This can be made even cleaner by using TF-Slim's repeat operation:

net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
net = slim.max_pool(net, [2, 2], scope='pool2')

Notice that the slim.repeat not only applies the same argument in-line, it also is smart enough to unroll the scopes such that the scopes assigned to each subsequent call of slim.conv2d are appended with an underscore and iteration number. More concretely, the scopes in the example above would be named 'conv3/conv3_1', 'conv3/conv3_2' and 'conv3/conv3_3'.

Furthermore, TF-Slim's slim.stack operator allows a caller to repeatedly apply the same operation with different arguments to create a stack or tower of layers. slim.stack also creates a new tf.variable_scope for each operation created. For example, a simple way to create a Multi-Layer Perceptron (MLP):

# Verbose way:
x = slim.fully_connected(x, 32, scope='fc/fc_1')
x = slim.fully_connected(x, 64, scope='fc/fc_2')
x = slim.fully_connected(x, 128, scope='fc/fc_3')

# Equivalent, TF-Slim way using slim.stack:
slim.stack(x, slim.fully_connected, [32, 64, 128], scope='fc')

In this example, slim.stack calls slim.fully_connected three times passing the output of one invocation of the function to the next. However, the number of hidden units in each invocation changes from 32 to 64 to 128. Similarly, one can use stack to simplify a tower of multiple convolutions:

# Verbose way:
x = slim.conv2d(x, 32, [3, 3], scope='core/core_1')
x = slim.conv2d(x, 32, [1, 1], scope='core/core_2')
x = slim.conv2d(x, 64, [3, 3], scope='core/core_3')
x = slim.conv2d(x, 64, [1, 1], scope='core/core_4')

# Using stack:
slim.stack(x, slim.conv2d, [(32, [3, 3]), (32, [1, 1]), (64, [3, 3]), (64, [1, 1])], scope='core')

Scopes

In addition to the types of scope mechanisms in TensorFlow (name_scope, variable_scope, TF-Slim adds a new scoping mechanism called arg_scope. This new scope allows a user to specify one or more operations and a set of arguments which will be passed to each of the operations defined in the arg_scope. This functionality is best illustrated by example. Consider the following code snippet:

net = slim.conv2d(inputs, 64, [11, 11], 4, padding='SAME',
                  weights_initializer=tf.truncated_normal_initializer(stddev=0.01),
                  weights_regularizer=slim.l2_regularizer(0.0005), scope='conv1')
net = slim.conv2d(net, 128, [11, 11], padding='VALID',
                  weights_initializer=tf.truncated_normal_initializer(stddev=0.01),
                  weights_regularizer=slim.l2_regularizer(0.0005), scope='conv2')
net = slim.conv2d(net, 256, [11, 11], padding='SAME',
                  weights_initializer=tf.truncated_normal_initializer(stddev=0.01),
                  weights_regularizer=slim.l2_regularizer(0.0005), scope='conv3')

It should be clear that these three convolution layers share many of the same hyperparameters. Two have the same padding, all three have the same weights_initializer and weight_regularizer. This code is hard to read and contains a lot of repeated values that should be factored out. One solution would be to specify default values using variables:

padding = 'SAME'
initializer = tf.truncated_normal_initializer(stddev=0.01)
regularizer = slim.l2_regularizer(0.0005)
net = slim.conv2d(inputs, 64, [11, 11], 4,
                  padding=padding,
                  weights_initializer=initializer,
                  weights_regularizer=regularizer,
                  scope='conv1')
net = slim.conv2d(net, 128, [11, 11],
                  padding='VALID',
                  weights_initializer=initializer,
                  weights_regularizer=regularizer,
                  scope='conv2')
net = slim.conv2d(net, 256, [11, 11],
                  padding=padding,
                  weights_initializer=initializer,
                  weights_regularizer=regularizer,
                  scope='conv3')

This solution ensures that all three convolutions share the exact same parameter values but doesn't reduce completely the code clutter. By using an arg_scope, we can both ensure that each layer uses the same values and simplify the code:

  with slim.arg_scope([slim.conv2d], padding='SAME',
                      weights_initializer=tf.truncated_normal_initializer(stddev=0.01)
                      weights_regularizer=slim.l2_regularizer(0.0005)):
    net = slim.conv2d(inputs, 64, [11, 11], scope='conv1')
    net = slim.conv2d(net, 128, [11, 11], padding='VALID', scope='conv2')
    net = slim.conv2d(net, 256, [11, 11], scope='conv3')

As the example illustrates, the use of arg_scope makes the code cleaner, simpler and easier to maintain. Notice that while argument values are specifed in the arg_scope, they can be overwritten locally. In particular, while the padding argument has been set to 'SAME', the second convolution overrides it with the value of 'VALID'.

One can also nest arg_scopes and use multiple operations in the same scope. For example:

with slim.arg_scope([slim.conv2d, slim.fully_connected],
                      activation_fn=tf.nn.relu,
                      weights_initializer=tf.truncated_normal_initializer(stddev=0.01),
                      weights_regularizer=slim.l2_regularizer(0.0005)):
  with slim.arg_scope([slim.conv2d], stride=1, padding='SAME'):
    net = slim.conv2d(inputs, 64, [11, 11], 4, padding='VALID', scope='conv1')
    net = slim.conv2d(net, 256, [5, 5],
                      weights_initializer=tf.truncated_normal_initializer(stddev=0.03),
                      scope='conv2')
    net = slim.fully_connected(net, 1000, activation_fn=None, scope='fc')

In this example, the first arg_scope applies the same weights_initializer and weights_regularizer arguments to the conv2d and fully_connected layers in its scope. In the second arg_scope, additional default arguments to conv2d only are specified.

Working Example: Specifying the VGG16 Layers

By combining TF-Slim Variables, Operations and scopes, we can write a normally very complex network with very few lines of code. For example, the entire VGG architecture can be defined with just the following snippet:

def vgg16(inputs):
  with slim.arg_scope([slim.conv2d, slim.fully_connected],
                      activation_fn=tf.nn.relu,
                      weights_initializer=tf.truncated_normal_initializer(0.0, 0.01),
                      weights_regularizer=slim.l2_regularizer(0.0005)):
    net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
    net = slim.max_pool2d(net, [2, 2], scope='pool1')
    net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
    net = slim.max_pool2d(net, [2, 2], scope='pool2')
    net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
    net = slim.max_pool2d(net, [2, 2], scope='pool3')
    net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
    net = slim.max_pool2d(net, [2, 2], scope='pool4')
    net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
    net = slim.max_pool2d(net, [2, 2], scope='pool5')
    net = slim.fully_connected(net, 4096, scope='fc6')
    net = slim.dropout(net, 0.5, scope='dropout6')
    net = slim.fully_connected(net, 4096, scope='fc7')
    net = slim.dropout(net, 0.5, scope='dropout7')
    net = slim.fully_connected(net, 1000, activation_fn=None, scope='fc8')
  return net

Training Models

Training Tensorflow models requires a model, a loss function, the gradient computation and a training routine that iteratively computes the gradients of the model weights relative to the loss and updates the weights accordingly. TF-Slim provides both common loss functions and a set of helper functions that run the training and evaluation routines.

Losses

The loss function defines a quantity that we want to minimize. For classification problems, this is typically the cross entropy between the true distribution and the predicted probability distribution across classes. For regression problems, this is often the sum-of-squares differences between the predicted and true values.

Certain models, such as multi-task learning models, require the use of multiple loss functions simultaneously. In other words, the loss function ultimately being minimized is the sum of various other loss functions. For example, consider a model that predicts both the type of scene in an image as well as the depth from the camera of each pixel. This model's loss function would be the sum of the classification loss and depth prediction loss.

TF-Slim provides an easy-to-use mechanism for defining and keeping track of loss functions via the losses module. Consider the simple case where we want to train the VGG network:

import tensorflow as tf
vgg = tf.contrib.slim.nets.vgg

# Load the images and labels.
images, labels = ...

# Create the model.
predictions = vgg.vgg16(images)

# Define the loss functions and get the total loss.
loss = slim.losses.softmax_cross_entropy(predictions, labels)

In this example, we start by creating the model (using TF-Slim's VGG implementation), and add the standard classification loss. Now, lets turn to the case where we have a multi-task model that produces multiple outputs:

# Load the images and labels.
images, scene_labels, depth_labels = ...

# Create the model.
scene_predictions, depth_predictions = CreateMultiTaskModel(images)

# Define the loss functions and get the total loss.
classification_loss = slim.losses.softmax_cross_entropy(scene_predictions, scene_labels)
sum_of_squares_loss = slim.losses.sum_of_squares(depth_predictions, depth_labels)

# The following two lines have the same effect:
total_loss = classification_loss + sum_of_squares_loss
total_loss = slim.losses.get_total_loss(add_regularization_losses=False)

In this example, we have two losses which we add by calling slim.losses.softmax_cross_entropy and slim.losses.sum_of_squares. We can obtain the total loss by adding them together (total_loss) or by calling slim.losses.get_total_loss(). How did this work? When you create a loss function via TF-Slim, TF-Slim adds the loss to a special TensorFlow collection of loss functions. This enables you to either manage the total loss manually, or allow TF-Slim to manage them for you.

What if you want to let TF-Slim manage the losses for you but have a custom loss function? loss_ops.py also has a function that adds this loss to TF-Slims collection. For example:

# Load the images and labels.
images, scene_labels, depth_labels, pose_labels = ...

# Create the model.
scene_predictions, depth_predictions, pose_predictions = CreateMultiTaskModel(images)

# Define the loss functions and get the total loss.
classification_loss = slim.losses.softmax_cross_entropy(scene_predictions, scene_labels)
sum_of_squares_loss = slim.losses.sum_of_squares(depth_predictions, depth_labels)
pose_loss = MyCustomLossFunction(pose_predictions, pose_labels)
slim.losses.add_loss(pose_loss) # Letting TF-Slim know about the additional loss.

# The following two ways to compute the total loss are equivalent:
regularization_loss = tf.add_n(slim.losses.get_regularization_losses())
total_loss1 = classification_loss + sum_of_squares_loss + pose_loss + regularization_loss

# (Regularization Loss is included in the total loss by default).
total_loss2 = losses.get_total_loss()

In this example, we can again either produce the total loss function manually or let TF-Slim know about the additional loss and let TF-Slim handle the losses.

Training Loop

TF-Slim provides a simple but powerful set of tools for training models found in learning.py. These include a Train function that repeatedly measures the loss, computes gradients and saves the model to disk, as well as several convenience functions for manipulating gradients. For example, once we've specified the model, the loss function and the optimization scheme, we can call slim.learning.create_train_op and slim.learning.train to perform the optimization:

g = tf.Graph()

# Create the model and specify the losses...
...

total_loss = slim.losses.get_total_loss()
optimizer = tf.train.GradientDescentOptimizer(learning_rate)

# create_train_op ensures that each time we ask for the loss, the update_ops
# are run and the gradients being computed are applied too.
train_op = slim.learning.create_train_op(total_loss, optimizer)
logdir = ... # Where checkpoints are stored.

slim.learning.train(
    train_op,
    logdir,
    number_of_steps=1000,
    save_summaries_secs=300,
    save_interval_secs=600):

In this example, slim.learning.train is provided with the train_op which is used to (a) compute the loss and (b) apply the gradient step. logdir specifies the directory where the checkpoints and event files are stored. We can limit the number of gradient steps taken to any number. In this case, we've asked for 1000 steps to be taken. Finally, save_summaries_secs=300 indicates that we'll compute summaries every 5 minutes and save_interval_secs=600 indicates that we'll save a model checkpoint every 10 minutes.

Working Example: Training the VGG16 Model

To illustrate this, lets examine the following sample of training the VGG network:

import tensorflow as tf

slim = tf.contrib.slim
vgg = tf.contrib.slim.nets.vgg

...

train_log_dir = ...
if not tf.gfile.Exists(train_log_dir):
  tf.gfile.MakeDirs(train_log_dir)

with tf.Graph().as_default():
  # Set up the data loading:
  images, labels = ...

  # Define the model:
  predictions = vgg.vgg16(images, is_training=True)

  # Specify the loss function:
  slim.losses.softmax_cross_entropy(predictions, labels)

  total_loss = slim.losses.get_total_loss()
  tf.summary.scalar('losses/total_loss', total_loss)

  # Specify the optimization scheme:
  optimizer = tf.train.GradientDescentOptimizer(learning_rate=.001)

  # create_train_op that ensures that when we evaluate it to get the loss,
  # the update_ops are done and the gradient updates are computed.
  train_tensor = slim.learning.create_train_op(total_loss, optimizer)

  # Actually runs training.
  slim.learning.train(train_tensor, train_log_dir)

Fine-Tuning Existing Models

Brief Recap on Restoring Variables from a Checkpoint

After a model has been trained, it can be restored using tf.train.Saver() which restores Variables from a given checkpoint. For many cases, tf.train.Saver() provides a simple mechanism to restore all or just a few variables.

# Create some variables.
v1 = tf.Variable(..., name="v1")
v2 = tf.Variable(..., name="v2")
...
# Add ops to restore all the variables.
restorer = tf.train.Saver()

# Add ops to restore some variables.
restorer = tf.train.Saver([v1, v2])

# Later, launch the model, use the saver to restore variables from disk, and
# do some work with the model.
with tf.Session() as sess:
  # Restore variables from disk.
  restorer.restore(sess, "/tmp/model.ckpt")
  print("Model restored.")
  # Do some work with the model
  ...

See Restoring Variables and Choosing which Variables to Save and Restore sections of the Variables page for more details.

Partially Restoring Models

It is often desirable to fine-tune a pre-trained model on an entirely new dataset or even a new task. In these situations, one can use TF-Slim's helper functions to select a subset of variables to restore:

# Create some variables.
v1 = slim.variable(name="v1", ...)
v2 = slim.variable(name="nested/v2", ...)
...

# Get list of variables to restore (which contains only 'v2'). These are all
# equivalent methods:
variables_to_restore = slim.get_variables_by_name("v2")
# or
variables_to_restore = slim.get_variables_by_suffix("2")
# or
variables_to_restore = slim.get_variables(scope="nested")
# or
variables_to_restore = slim.get_variables_to_restore(include=["nested"])
# or
variables_to_restore = slim.get_variables_to_restore(exclude=["v1"])

# Create the saver which will be used to restore the variables.
restorer = tf.train.Saver(variables_to_restore)

with tf.Session() as sess:
  # Restore variables from disk.
  restorer.restore(sess, "/tmp/model.ckpt")
  print("Model restored.")
  # Do some work with the model
  ...

Restoring models with different variable names

When restoring variables from a checkpoint, the Saver locates the variable names in a checkpoint file and maps them to variables in the current graph. Above, we created a saver by passing to it a list of variables. In this case, the names of the variables to locate in the checkpoint file were implicitly obtained from each provided variable's var.op.name.

This works well when the variable names in the checkpoint file match those in the graph. However, sometimes, we want to restore a model from a checkpoint whose variables have different names those in the current graph. In this case, we must provide the Saver a dictionary that maps from each checkpoint variable name to each graph variable. Consider the following example where the checkpoint variables names are obtained via a simple function:

# Assuming than 'conv1/weights' should be restored from 'vgg16/conv1/weights'
def name_in_checkpoint(var):
  return 'vgg16/' + var.op.name

# Assuming than 'conv1/weights' and 'conv1/bias' should be restored from 'conv1/params1' and 'conv1/params2'
def name_in_checkpoint(var):
  if "weights" in var.op.name:
    return var.op.name.replace("weights", "params1")
  if "bias" in var.op.name:
    return var.op.name.replace("bias", "params2")

variables_to_restore = slim.get_model_variables()
variables_to_restore = {name_in_checkpoint(var):var for var in variables_to_restore}
restorer = tf.train.Saver(variables_to_restore)

with tf.Session() as sess:
  # Restore variables from disk.
  restorer.restore(sess, "/tmp/model.ckpt")

Fine-Tuning a Model on a different task

Consider the case where we have a pre-trained VGG16 model. The model was trained on the ImageNet dataset, which has 1000 classes. However, we would like to apply it to the Pascal VOC dataset which has only 20 classes. To do so, we can initialize our new model using the values of the pre-trained model excluding the final layer:

# Load the Pascal VOC data
image, label = MyPascalVocDataLoader(...)
images, labels = tf.train.batch([image, label], batch_size=32)

# Create the model
predictions = vgg.vgg_16(images)

train_op = slim.learning.create_train_op(...)

# Specify where the Model, trained on ImageNet, was saved.
model_path = '/path/to/pre_trained_on_imagenet.checkpoint'

# Specify where the new model will live:
log_dir = '/path/to/my_pascal_model_dir/'

# Restore only the convolutional layers:
variables_to_restore = slim.get_variables_to_restore(exclude=['fc6', 'fc7', 'fc8'])
init_fn = assign_from_checkpoint_fn(model_path, variables_to_restore)

# Start training.
slim.learning.train(train_op, log_dir, init_fn=init_fn)

Evaluating Models.

Once we've trained a model (or even while the model is busy training) we'd like to see how well the model performs in practice. This is accomplished by picking a set of evaluation metrics, which will grade the models performance, and the evaluation code which actually loads the data, performs inference, compares the results to the ground truth and records the evaluation scores. This step may be performed once or repeated periodically.

Metrics

We define a metric to be a performance measure that is not a loss function (losses are directly optimized during training), but which we are still interested in for the purpose of evaluating our model. For example, we might want to minimize log loss, but our metrics of interest might be F1 score (test accuracy), or Intersection Over Union score (which are not differentiable, and therefore cannot be used as losses).

TF-Slim provides a set of metric operations that makes evaluating models easy. Abstractly, computing the value of a metric can be divided into three parts:

  1. Initialization: initialize the variables used to compute the metrics.
  2. Aggregation: perform operations (sums, etc) used to compute the metrics.
  3. Finalization: (optionally) perform any final operation to compute metric values. For example, computing means, mins, maxes, etc.

For example, to compute mean_absolute_error, two variables, a count and total variable are initialized to zero. During aggregation, we observed some set of predictions and labels, compute their absolute differences and add the total to total. Each time we observe another value, count is incremented. Finally, during finalization, total is divided by count to obtain the mean.

The following example demonstrates the API for declaring metrics. Because metrics are often evaluated on a test set which is different from the training set (upon which the loss is computed), we'll assume we're using test data:

images, labels = LoadTestData(...)
predictions = MyModel(images)

mae_value_op, mae_update_op = slim.metrics.streaming_mean_absolute_error(predictions, labels)
mre_value_op, mre_update_op = slim.metrics.streaming_mean_relative_error(predictions, labels, labels)
pl_value_op, pl_update_op = slim.metrics.percentage_less(mean_relative_errors, 0.3)

As the example illustrates, the creation of a metric returns two values: a value_op and an update_op. The value_op is an idempotent operation that returns the current value of the metric. The update_op is an operation that performs the aggregation step mentioned above as well as returning the value of the metric.

Keeping track of each value_op and update_op can be laborious. To deal with this, TF-Slim provides two convenience functions:

# Aggregates the value and update ops in two lists:
value_ops, update_ops = slim.metrics.aggregate_metrics(
    slim.metrics.streaming_mean_absolute_error(predictions, labels),
    slim.metrics.streaming_mean_squared_error(predictions, labels))

# Aggregates the value and update ops in two dictionaries:
names_to_values, names_to_updates = slim.metrics.aggregate_metric_map({
    "eval/mean_absolute_error": slim.metrics.streaming_mean_absolute_error(predictions, labels),
    "eval/mean_squared_error": slim.metrics.streaming_mean_squared_error(predictions, labels),
})

Working example: Tracking Multiple Metrics

Putting it all together:

import tensorflow as tf

slim = tf.contrib.slim
vgg = tf.contrib.slim.nets.vgg


# Load the data
images, labels = load_data(...)

# Define the network
predictions = vgg.vgg_16(images)

# Choose the metrics to compute:
names_to_values, names_to_updates = slim.metrics.aggregate_metric_map({
    "eval/mean_absolute_error": slim.metrics.streaming_mean_absolute_error(predictions, labels),
    "eval/mean_squared_error": slim.metrics.streaming_mean_squared_error(predictions, labels),
})

# Evaluate the model using 1000 batches of data:
num_batches = 1000

with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())
  sess.run(tf.local_variables_initializer())

  for batch_id in range(num_batches):
    sess.run(names_to_updates.values())

  metric_values = sess.run(name_to_values.values())
  for metric, value in zip(names_to_values.keys(), metric_values):
    print('Metric %s has value: %f' % (metric, value))

Note that metric_ops.py can be used in isolation without using either layers.py or loss_ops.py

Evaluation Loop

TF-Slim provides an evaluation module (evaluation.py), which contains helper functions for writing model evaluation scripts using metrics from the metric_ops.py module. These include a function for periodically running evaluations, evaluating metrics over batches of data and printing and summarizing metric results. For example:

import tensorflow as tf

slim = tf.contrib.slim

# Load the data
images, labels = load_data(...)

# Define the network
predictions = MyModel(images)

# Choose the metrics to compute:
names_to_values, names_to_updates = slim.metrics.aggregate_metric_map({
    'accuracy': slim.metrics.accuracy(predictions, labels),
    'precision': slim.metrics.precision(predictions, labels),
    'recall': slim.metrics.recall(mean_relative_errors, 0.3),
})

# Create the summary ops such that they also print out to std output:
summary_ops = []
for metric_name, metric_value in metrics_to_values.iteritems():
  op = tf.summary.scalar(metric_name, metric_value)
  op = tf.Print(op, [metric_value], metric_name)
  summary_ops.append(op)

num_examples = 10000
batch_size = 32
num_batches = math.ceil(num_examples / float(batch_size))

# Setup the global step.
slim.get_or_create_global_step()

output_dir = ... # Where the summaries are stored.
eval_interval_secs = ... # How often to run the evaluation.
slim.evaluation.evaluation_loop(
    'local',
    checkpoint_dir,
    log_dir,
    num_evals=num_batches,
    eval_op=names_to_updates.values(),
    summary_op=tf.summary.merge(summary_ops),
    eval_interval_secs=eval_interval_secs)

Authors

Sergio Guadarrama and Nathan Silberman