[TOC]

This library provides a set of classes and functions that helps train models.

The Optimizer base class provides methods to compute gradients for a loss and apply gradients to variables. A collection of subclasses implement classic optimization algorithms such as GradientDescent and Adagrad.

You never instantiate the Optimizer class itself, but instead instantiate one of the subclasses.

Base class for optimizers.

This class defines the API to add Ops to train a model. You never use this class directly, but instead instantiate one of its subclasses such as GradientDescentOptimizer, AdagradOptimizer, or MomentumOptimizer.

# Create an optimizer with the desired parameters.
opt = GradientDescentOptimizer(learning_rate=0.1)
# Add Ops to the graph to minimize a cost by updating a list of variables.
# "cost" is a Tensor, and the list of variables contains tf.Variable
# objects.
opt_op = opt.minimize(cost, var_list=<list of variables>)

In the training program you will just have to run the returned Op.

# Execute opt_op to do one step of training:
opt_op.run()

Calling minimize() takes care of both computing the gradients and applying them to the variables. If you want to process the gradients before applying them you can instead use the optimizer in three steps:

1. Compute the gradients with compute_gradients().
2. Process the gradients as you wish.
3. Apply the processed gradients with apply_gradients().

Example:

# Create an optimizer.
opt = GradientDescentOptimizer(learning_rate=0.1)

# Compute the gradients for a list of variables.
grads_and_vars = opt.compute_gradients(loss, <list of variables>)

# grads_and_vars is a list of tuples (gradient, variable).  Do whatever you
# need to the 'gradient' part, for example cap them, etc.
capped_grads_and_vars = [(MyCapper(gv[0]), gv[1]) for gv in grads_and_vars]

# Ask the optimizer to apply the capped gradients.
opt.apply_gradients(capped_grads_and_vars)

Create a new Optimizer.

This must be called by the constructors of subclasses.

• use_locking: Bool. If True apply use locks to prevent concurrent updates to variables.
• name: A non-empty string. The name to use for accumulators created for the optimizer.

• ValueError: If name is malformed.

Add operations to minimize loss by updating var_list.

This method simply combines calls compute_gradients() and apply_gradients(). If you want to process the gradient before applying them call compute_gradients() and apply_gradients() explicitly instead of using this function.

• loss: A Tensor containing the value to minimize.
• global_step: Optional Variable to increment by one after the variables have been updated.
• var_list: Optional list of Variable objects to update to minimize loss. Defaults to the list of variables collected in the graph under the key GraphKeys.TRAINABLE_VARIABLES.
• gate_gradients: How to gate the computation of gradients. Can be GATE_NONE, GATE_OP, or GATE_GRAPH.
• aggregation_method: Specifies the method used to combine gradient terms. Valid values are defined in the class AggregationMethod.
• colocate_gradients_with_ops: If True, try colocating gradients with the corresponding op.
• name: Optional name for the returned operation.

An Operation that updates the variables in var_list. If global_step was not None, that operation also increments global_step.

• ValueError: If some of the variables are not Variable objects.

Compute gradients of loss for the variables in var_list.

This is the first part of minimize(). It returns a list of (gradient, variable) pairs where "gradient" is the gradient for "variable". Note that "gradient" can be a Tensor, an IndexedSlices, or None if there is no gradient for the given variable.

• loss: A Tensor containing the value to minimize.
• var_list: Optional list of tf.Variable to update to minimize loss. Defaults to the list of variables collected in the graph under the key GraphKey.TRAINABLE_VARIABLES.
• gate_gradients: How to gate the computation of gradients. Can be GATE_NONE, GATE_OP, or GATE_GRAPH.
• aggregation_method: Specifies the method used to combine gradient terms. Valid values are defined in the class AggregationMethod.
• colocate_gradients_with_ops: If True, try colocating gradients with the corresponding op.

A list of (gradient, variable) pairs.

• TypeError: If var_list contains anything else than Variable objects.
• ValueError: If some arguments are invalid.

Apply gradients to variables.

This is the second part of minimize(). It returns an Operation that applies gradients.

• grads_and_vars: List of (gradient, variable) pairs as returned by compute_gradients().
• global_step: Optional Variable to increment by one after the variables have been updated.
• name: Optional name for the returned operation. Default to the name passed to the Optimizer constructor.

An Operation that applies the specified gradients. If global_step was not None, that operation also increments global_step.

• TypeError: If grads_and_vars is malformed.
• ValueError: If none of the variables have gradients.

Both minimize() and compute_gradients() accept a gate_gradient argument that controls the degree of parallelism during the application of the gradients.

The possible values are: GATE_NONE, GATE_OP, and GATE_GRAPH.

GATE_NONE: Compute and apply gradients in parallel. This provides the maximum parallelism in execution, at the cost of some non-reproducibility in the results. For example the two gradients of matmul depend on the input values: With GATE_NONE one of the gradients could be applied to one of the inputs before the other gradient is computed resulting in non-reproducible results.

GATE_OP: For each Op, make sure all gradients are computed before they are used. This prevents race conditions for Ops that generate gradients for multiple inputs where the gradients depend on the inputs.

GATE_GRAPH: Make sure all gradients for all variables are computed before any one of them is used. This provides the least parallelism but can be useful if you want to process all gradients before applying any of them.

Some optimizer subclasses, such as MomentumOptimizer and AdagradOptimizer allocate and manage additional variables associated with the variables to train. These are called Slots. Slots have names and you can ask the optimizer for the names of the slots that it uses. Once you have a slot name you can ask the optimizer for the variable it created to hold the slot value.

This can be useful if you want to log debug a training algorithm, report stats about the slots, etc.

Return a list of the names of slots created by the Optimizer.

See get_slot().

A list of strings.

Return a slot named name created for var by the Optimizer.

Some Optimizer subclasses use additional variables. For example Momentum and Adagrad use variables to accumulate updates. This method gives access to these Variable objects if for some reason you need them.

Use get_slot_names() to get the list of slot names created by the Optimizer.

• var: A variable passed to minimize() or apply_gradients().
• name: A string.

The Variable for the slot if it was created, None otherwise.

Optimizer that implements the gradient descent algorithm.

Construct a new gradient descent optimizer.

• learning_rate: A Tensor or a floating point value. The learning rate to use.
• use_locking: If True use locks for update operations.
• name: Optional name prefix for the operations created when applying gradients. Defaults to "GradientDescent".

Optimizer that implements the Adadelta algorithm.

See M. D. Zeiler (pdf)

Construct a new Adadelta optimizer.

• learning_rate: A Tensor or a floating point value. The learning rate.
• rho: A Tensor or a floating point value. The decay rate.
• epsilon: A Tensor or a floating point value. A constant epsilon used to better conditioning the grad update.
• use_locking: If True use locks for update operations.
• name: Optional name prefix for the operations created when applying gradients. Defaults to "Adadelta".

Optimizer that implements the Adagrad algorithm.

See this paper.

Construct a new Adagrad optimizer.

• learning_rate: A Tensor or a floating point value. The learning rate.
• initial_accumulator_value: A floating point value. Starting value for the accumulators, must be positive.
• use_locking: If True use locks for update operations.
• name: Optional name prefix for the operations created when applying gradients. Defaults to "Adagrad".

• ValueError: If the initial_accumulator_value is invalid.

Optimizer that implements the Momentum algorithm.

Construct a new Momentum optimizer.

• learning_rate: A Tensor or a floating point value. The learning rate.
• momentum: A Tensor or a floating point value. The momentum.
• use_locking: If True use locks for update operations.
• name: Optional name prefix for the operations created when applying gradients. Defaults to "Momentum".

Optimizer that implements the Adam algorithm.

See Kingma et. al., 2014 (pdf).

Construct a new Adam optimizer.

Initialization:

m_0 <- 0 (Initialize initial 1st moment vector)
v_0 <- 0 (Initialize initial 2nd moment vector)
t <- 0 (Initialize timestep)

The update rule for variable with gradient g uses an optimization described at the end of section2 of the paper:

t <- t + 1
lr_t <- learning_rate * sqrt(1 - beta2^t) / (1 - beta1^t)

m_t <- beta1 * m_{t-1} + (1 - beta1) * g
v_t <- beta2 * v_{t-1} + (1 - beta2) * g * g
variable <- variable - lr_t * m_t / (sqrt(v_t) + epsilon)

The default value of 1e-8 for epsilon might not be a good default in general. For example, when training an Inception network on ImageNet a current good choice is 1.0 or 0.1.

• learning_rate: A Tensor or a floating point value. The learning rate.
• beta1: A float value or a constant float tensor. The exponential decay rate for the 1st moment estimates.
• beta2: A float value or a constant float tensor. The exponential decay rate for the 2nd moment estimates.
• epsilon: A small constant for numerical stability.
• use_locking: If True use locks for update operations.
• name: Optional name for the operations created when applying gradients. Defaults to "Adam".

Optimizer that implements the FTRL algorithm.

See this paper.

Construct a new FTRL optimizer.

• learning_rate: A float value or a constant float Tensor.
• learning_rate_power: A float value, must be less or equal to zero.
• initial_accumulator_value: The starting value for accumulators. Only positive values are allowed.
• l1_regularization_strength: A float value, must be greater than or equal to zero.
• l2_regularization_strength: A float value, must be greater than or equal to zero.
• use_locking: If True use locks for update operations.
• name: Optional name prefix for the operations created when applying gradients. Defaults to "Ftrl".

• ValueError: If one of the arguments is invalid.

Optimizer that implements the RMSProp algorithm.

See the [paper] (http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf).

Construct a new RMSProp optimizer.

• learning_rate: A Tensor or a floating point value. The learning rate.
• decay: Discounting factor for the history/coming gradient
• momentum: A scalar tensor.
• epsilon: Small value to avoid zero denominator.
• use_locking: If True use locks for update operation.
• name: Optional name prefix for the operations created when applying gradients. Defaults to "RMSProp".

TensorFlow provides functions to compute the derivatives for a given TensorFlow computation graph, adding operations to the graph. The optimizer classes automatically compute derivatives on your graph, but creators of new Optimizers or expert users can call the lower-level functions below.

Constructs symbolic partial derivatives of sum of ys w.r.t. x in xs.

ys and xs are each a Tensor or a list of tensors. grad_ys is a list of Tensor, holding the gradients received by the ys. The list must be the same length as ys.

gradients() adds ops to the graph to output the partial derivatives of ys with respect to xs. It returns a list of Tensor of length len(xs) where each tensor is the sum(dy/dx) for y in ys.

grad_ys is a list of tensors of the same length as ys that holds the initial gradients for each y in ys. When grad_ys is None, we fill in a tensor of '1's of the shape of y for each y in ys. A user can provide their own initial grad_ys to compute the derivatives using a different initial gradient for each y (e.g., if one wanted to weight the gradient differently for each value in each y).

• ys: A Tensor or list of tensors to be differentiated.
• xs: A Tensor or list of tensors to be used for differentiation.
• grad_ys: Optional. A Tensor or list of tensors the same size as ys and holding the gradients computed for each y in ys.
• name: Optional name to use for grouping all the gradient ops together. defaults to 'gradients'.
• colocate_gradients_with_ops: If True, try colocating gradients with the corresponding op.
• gate_gradients: If True, add a tuple around the gradients returned for an operations. This avoids some race conditions.
• aggregation_method: Specifies the method used to combine gradient terms. Accepted values are constants defined in the class AggregationMethod.

A list of sum(dy/dx) for each x in xs.

• LookupError: if one of the operations between x and y does not have a registered gradient function.
• ValueError: if the arguments are invalid.

A class listing aggregation methods used to combine gradients.

Computing partial derivatives can require aggregating gradient contributions. This class lists the various methods that can be used to combine gradients in the graph:

• ADD_N: All of the gradient terms are summed as part of one operation using the "AddN" op. It has the property that all gradients must be ready before any aggregation is performed.
• DEFAULT: The system-chosen default aggregation method.

Stops gradient computation.

When executed in a graph, this op outputs its input tensor as-is.

When building ops to compute gradients, this op prevents the contribution of its inputs to be taken into account. Normally, the gradient generator adds ops to a graph to compute the derivatives of a specified 'loss' by recursively finding out inputs that contributed to its computation. If you insert this op in the graph it inputs are masked from the gradient generator. They are not taken into account for computing gradients.

This is useful any time you want to compute a value with TensorFlow but need to pretend that the value was a constant. Some examples include:

• The EM algorithm where the M-step should not involve backpropagation through the output of the E-step.
• Contrastive divergence training of Boltzmann machines where, when differentiating the energy function, the training must not backpropagate through the graph that generated the samples from the model.
• Adversarial training, where no backprop should happen through the adversarial example generation process.

• input: A Tensor.
• name: A name for the operation (optional).

A Tensor. Has the same type as input.

TensorFlow provides several operations that you can use to add clipping functions to your graph. You can use these functions to perform general data clipping, but they're particularly useful for handling exploding or vanishing gradients.

Clips tensor values to a specified min and max.

Given a tensor t, this operation returns a tensor of the same type and shape as t with its values clipped to clip_value_min and clip_value_max. Any values less than clip_value_min are set to clip_value_min. Any values greater than clip_value_max are set to clip_value_max.

• t: A Tensor.
• clip_value_min: A 0-D (scalar) Tensor. The minimum value to clip by.
• clip_value_max: A 0-D (scalar) Tensor. The maximum value to clip by.
• name: A name for the operation (optional).

A clipped Tensor.

Clips tensor values to a maximum L2-norm.

Given a tensor t, and a maximum clip value clip_norm, this operation normalizes t so that its L2-norm is less than or equal to clip_norm. Specifically, if the L2-norm is already less than or equal to clip_norm, then t is not modified. If the L2-norm is greater than clip_norm, then this operation returns a tensor of the same type and shape as t with its values set to:

t * clip_norm / l2norm(t)

In this case, the L2-norm of the output tensor is clip_norm.

This operation is typically used to clip gradients before applying them with an optimizer.

• t: A Tensor.
• clip_norm: A 0-D (scalar) Tensor > 0. A maximum clipping value.
• name: A name for the operation (optional).

A clipped Tensor.

Clips tensor values to a maximum average L2-norm.

Given a tensor t, and a maximum clip value clip_norm, this operation normalizes t so that its average L2-norm is less than or equal to clip_norm. Specifically, if the average L2-norm is already less than or equal to clip_norm, then t is not modified. If the average L2-norm is greater than clip_norm, then this operation returns a tensor of the same type and shape as t with its values set to:

t * clip_norm / l2norm_avg(t)

In this case, the average L2-norm of the output tensor is clip_norm.

This operation is typically used to clip gradients before applying them with an optimizer.

• t: A Tensor.
• clip_norm: A 0-D (scalar) Tensor > 0. A maximum clipping value.
• name: A name for the operation (optional).

A clipped Tensor.

Clips values of multiple tensors by the ratio of the sum of their norms.

Given a tuple or list of tensors t_list, and a clipping ratio clip_norm, this operation returns a list of clipped tensors list_clipped and the global norm (global_norm) of all tensors in t_list. Optionally, if you've already computed the global norm for t_list, you can specify the global norm with use_norm.

To perform the clipping, the values t_list[i] are set to:

t_list[i] * clip_norm / max(global_norm, clip_norm)

where:

global_norm = sqrt(sum([l2norm(t)**2 for t in t_list]))

If clip_norm > global_norm then the entries in t_list remain as they are, otherwise they're all shrunk by the global ratio.

Any of the entries of t_list that are of type None are ignored.

This is the correct way to perform gradient clipping (for example, see Pascanu et al., 2012 (pdf)).

However, it is slower than clip_by_norm() because all the parameters must be ready before the clipping operation can be performed.

• t_list: A tuple or list of mixed Tensors, IndexedSlices, or None.
• clip_norm: A 0-D (scalar) Tensor > 0. The clipping ratio.
• use_norm: A 0-D (scalar) Tensor of type float (optional). The global norm to use. If not provided, global_norm() is used to compute the norm.
• name: A name for the operation (optional).

• list_clipped: A list of Tensors of the same type as list_t.
• global_norm: A 0-D (scalar) Tensor representing the global norm.

• TypeError: If t_list is not a sequence.

Computes the global norm of multiple tensors.

Given a tuple or list of tensors t_list, this operation returns the global norm of the elements in all tensors in t_list. The global norm is computed as:

global_norm = sqrt(sum([l2norm(t)**2 for t in t_list]))

Any entries in t_list that are of type None are ignored.

• t_list: A tuple or list of mixed Tensors, IndexedSlices, or None.
• name: A name for the operation (optional).

A 0-D (scalar) Tensor of type float.

• TypeError: If t_list is not a sequence.

Applies exponential decay to the learning rate.

When training a model, it is often recommended to lower the learning rate as the training progresses. This function applies an exponential decay function to a provided initial learning rate. It requires a global_step value to compute the decayed learning rate. You can just pass a TensorFlow variable that you increment at each training step.

The function returns the decayed learning rate. It is computed as:

decayed_learning_rate = learning_rate *
                        decay_rate ^ (global_step / decay_steps)

If the argument staircase is True, then global_step /decay_steps is an integer division and the decayed learning rate follows a staircase function.

Example: decay every 100000 steps with a base of 0.96:

...
global_step = tf.Variable(0, trainable=False)
starter_learning_rate = 0.1
learning_rate = tf.train.exponential_decay(starter_learning_rate, global_step,
                                           100000, 0.96, staircase=True)
# Passing global_step to minimize() will increment it at each step.
learning_step = (
    tf.GradientDescentOptimizer(learning_rate)
    .minimize(...my loss..., global_step=global_step)
)

• learning_rate: A scalar float32 or float64 Tensor or a Python number. The initial learning rate.
• global_step: A scalar int32 or int64 Tensor or a Python number. Global step to use for the decay computation. Must not be negative.
• decay_steps: A scalar int32 or int64 Tensor or a Python number. Must be positive. See the decay computation above.
• decay_rate: A scalar float32 or float64 Tensor or a Python number. The decay rate.
• staircase: Boolean. It True decay the learning rate at discrete intervals.
• name: String. Optional name of the operation. Defaults to 'ExponentialDecay'

A scalar Tensor of the same type as learning_rate. The decayed learning rate.

Some training algorithms, such as GradientDescent and Momentum often benefit from maintaining a moving average of variables during optimization. Using the moving averages for evaluations often improve results significantly.

Maintains moving averages of variables by employing an exponential decay.

When training a model, it is often beneficial to maintain moving averages of the trained parameters. Evaluations that use averaged parameters sometimes produce significantly better results than the final trained values.

The apply() method adds shadow copies of trained variables and add ops that maintain a moving average of the trained variables in their shadow copies. It is used when building the training model. The ops that maintain moving averages are typically run after each training step. The average() and average_name() methods give access to the shadow variables and their names. They are useful when building an evaluation model, or when restoring a model from a checkpoint file. They help use the moving averages in place of the last trained values for evaluations.

The moving averages are computed using exponential decay. You specify the decay value when creating the ExponentialMovingAverage object. The shadow variables are initialized with the same initial values as the trained variables. When you run the ops to maintain the moving averages, each shadow variable is updated with the formula:

shadow_variable -= (1 - decay) * (shadow_variable - variable)

This is mathematically equivalent to the classic formula below, but the use of an assign_sub op (the "-=" in the formula) allows concurrent lockless updates to the variables:

shadow_variable = decay * shadow_variable + (1 - decay) * variable

Reasonable values for decay are close to 1.0, typically in the multiple-nines range: 0.999, 0.9999, etc.

Example usage when creating a training model:

# Create variables.
var0 = tf.Variable(...)
var1 = tf.Variable(...)
# ... use the variables to build a training model...
...
# Create an op that applies the optimizer.  This is what we usually
# would use as a training op.
opt_op = opt.minimize(my_loss, [var0, var1])

# Create an ExponentialMovingAverage object
ema = tf.train.ExponentialMovingAverage(decay=0.9999)

# Create the shadow variables, and add ops to maintain moving averages
# of var0 and var1.
maintain_averages_op = ema.apply([var0, var1])

# Create an op that will update the moving averages after each training
# step.  This is what we will use in place of the usual training op.
with tf.control_dependencies([opt_op]):
    training_op = tf.group(maintain_averages_op)

...train the model by running training_op...

There are two ways to use the moving averages for evaluations:

• Build a model that uses the shadow variables instead of the variables. For this, use the average() method which returns the shadow variable for a given variable.
• Build a model normally but load the checkpoint files to evaluate by using the shadow variable names. For this use the average_name() method. See the Saver class for more information on restoring saved variables.

Example of restoring the shadow variable values:

# Create a Saver that loads variables from their saved shadow values.
shadow_var0_name = ema.average_name(var0)
shadow_var1_name = ema.average_name(var1)
saver = tf.train.Saver({shadow_var0_name: var0, shadow_var1_name: var1})
saver.restore(...checkpoint filename...)
# var0 and var1 now hold the moving average values

Creates a new ExponentialMovingAverage object.

The Apply() method has to be called to create shadow variables and add ops to maintain moving averages.

The optional num_updates parameter allows one to tweak the decay rate dynamically. . It is typical to pass the count of training steps, usually kept in a variable that is incremented at each step, in which case the decay rate is lower at the start of training. This makes moving averages move faster. If passed, the actual decay rate used is:

min(decay, (1 + num_updates) / (10 + num_updates))

• decay: Float. The decay to use.
• num_updates: Optional count of number of updates applied to variables.
• name: String. Optional prefix name to use for the name of ops added in Apply().

Maintains moving averages of variables.

var_list must be a list of Variable or Tensor objects. This method creates shadow variables for all elements of var_list. Shadow variables for Variable objects are initialized to the variable's initial value. They will be added to the GraphKeys.MOVING_AVERAGE_VARIABLES collection. For Tensor objects, the shadow variables are initialized to 0.

shadow variables are created with trainable=False and added to the GraphKeys.ALL_VARIABLES collection. They will be returned by calls to tf.all_variables().

Returns an op that updates all shadow variables as described above.

Note that apply() can be called multiple times with different lists of variables.

• var_list: A list of Variable or Tensor objects. The variables and Tensors must be of types float32 or float64.

An Operation that updates the moving averages.

• TypeError: If the arguments are not all float32 or float64.
• ValueError: If the moving average of one of the variables is already being computed.

Returns the name of the Variable holding the average for var.

The typical scenario for ExponentialMovingAverage is to compute moving averages of variables during training, and restore the variables from the computed moving averages during evaluations.

To restore variables, you have to know the name of the shadow variables. That name and the original variable can then be passed to a Saver() object to restore the variable from the moving average value with: saver = tf.train.Saver({ema.average_name(var): var})

average_name() can be called whether or not apply() has been called.

• var: A Variable object.

A string: The name of the variable that will be used or was used by the ExponentialMovingAverage class to hold the moving average of var.

Returns the Variable holding the average of var.

• var: A Variable object.

A Variable object or None if the moving average of var is not maintained..

Returns a map of names to Variables to restore.

If a variable has a moving average, use the moving average variable name as the restore name; otherwise, use the variable name.

For example,

  variables_to_restore = ema.variables_to_restore()
  saver = tf.train.Saver(variables_to_restore)

Below is an example of such mapping:

  conv/batchnorm/gamma/ExponentialMovingAverage: conv/batchnorm/gamma,
  conv_4/conv2d_params/ExponentialMovingAverage: conv_4/conv2d_params,
  global_step: global_step

A map from restore_names to variables. The restore_name can be the moving_average version of the variable name if it exist, or the original variable name.

See Threading and Queues for how to use threads and queues. For documentation on the Queue API, see Queues.

A coordinator for threads.

This class implements a simple mechanism to coordinate the termination of a set of threads.

# Create a coordinator.
coord = Coordinator()
# Start a number of threads, passing the coordinator to each of them.
...start thread 1...(coord, ...)
...start thread N...(coord, ...)
# Wait for all the threads to terminate.
coord.join(threads)

Any of the threads can call coord.request_stop() to ask for all the threads to stop. To cooperate with the requests, each thread must check for coord.should_stop() on a regular basis. coord.should_stop() returns True as soon as coord.request_stop() has been called.

A typical thread running with a coordinator will do something like:

while not coord.should_stop():
  ...do some work...

A thread can report an exception to the coordinator as part of the should_stop() call. The exception will be re-raised from the coord.join() call.

Thread code:

try:
  while not coord.should_stop():
    ...do some work...
except Exception as e:
  coord.request_stop(e)

Main code:

try:
  ...
  coord = Coordinator()
  # Start a number of threads, passing the coordinator to each of them.
  ...start thread 1...(coord, ...)
  ...start thread N...(coord, ...)
  # Wait for all the threads to terminate.
  coord.join(threads)
except Exception as e:
  ...exception that was passed to coord.request_stop()

To simplify the thread implementation, the Coordinator provides a context handler stop_on_exception() that automatically requests a stop if an exception is raised. Using the context handler the thread code above can be written as:

with coord.stop_on_exception():
  while not coord.should_stop():
    ...do some work...

After a thread has called coord.request_stop() the other threads have a fixed time to stop, this is called the 'stop grace period' and defaults to 2 minutes. If any of the threads is still alive after the grace period expires coord.join() raises a RuntimeException reporting the laggards.

try:
  ...
  coord = Coordinator()
  # Start a number of threads, passing the coordinator to each of them.
  ...start thread 1...(coord, ...)
  ...start thread N...(coord, ...)
  # Wait for all the threads to terminate, give them 10s grace period
  coord.join(threads, stop_grace_period_secs=10)
except RuntimeException:
  ...one of the threads took more than 10s to stop after request_stop()
  ...was called.
except Exception:
  ...exception that was passed to coord.request_stop()

Create a new Coordinator.

Clears the stop flag.

After this is called, calls to should_stop() will return False.

Wait for threads to terminate.

Blocks until all threads have terminated or request_stop() is called.

After the threads stop, if an exc_info was passed to request_stop, that exception is re-raised.

Grace period handling: When request_stop() is called, threads are given 'stop_grace_period_secs' seconds to terminate. If any of them is still alive after that period expires, a RuntimeError is raised. Note that if an exc_info was passed to request_stop() then it is raised instead of that RuntimeError.

• threads: List of threading.Threads. The started threads to join.
• stop_grace_period_secs: Number of seconds given to threads to stop after request_stop() has been called.

• RuntimeError: If any thread is still alive after request_stop() is called and the grace period expires.

Request that the threads stop.

After this is called, calls to should_stop() will return True.

Note: If an exception is being passed in, in must be in the context of handling the exception (i.e. try: ... except Exception as ex: ...) and not a newly created one.

• ex: Optional Exception, or Python exc_info tuple as returned by sys.exc_info(). If this is the first call to request_stop() the corresponding exception is recorded and re-raised from join().

Check if stop was requested.

True if a stop was requested.

Context manager to request stop when an Exception is raised.

Code that uses a coordinator must catch exceptions and pass them to the request_stop() method to stop the other threads managed by the coordinator.

This context handler simplifies the exception handling. Use it as follows:

with coord.stop_on_exception():
  # Any exception raised in the body of the with
  # clause is reported to the coordinator before terminating
  # the execution of the body.
  ...body...

This is completely equivalent to the slightly longer code:

try:
  ...body...
exception Exception as ex:
  coord.request_stop(ex)

nothing.

Wait till the Coordinator is told to stop.

• timeout: Float. Sleep for up to that many seconds waiting for should_stop() to become True.

True if the Coordinator is told stop, False if the timeout expired.

Holds a list of enqueue operations for a queue, each to be run in a thread.

Queues are a convenient TensorFlow mechanism to compute tensors asynchronously using multiple threads. For example in the canonical 'Input Reader' setup one set of threads generates filenames in a queue; a second set of threads read records from the files, processes them, and enqueues tensors on a second queue; a third set of threads dequeues these input records to construct batches and runs them through training operations.

There are several delicate issues when running multiple threads that way: closing the queues in sequence as the input is exhausted, correctly catching and reporting exceptions, etc.

The QueueRunner, combined with the Coordinator, helps handle these issues.

Create a QueueRunner.

On construction the QueueRunner adds an op to close the queue. That op will be run if the enqueue ops raise exceptions.

When you later call the create_threads() method, the QueueRunner will create one thread for each op in enqueue_ops. Each thread will run its enqueue op in parallel with the other threads. The enqueue ops do not have to all be the same op, but it is expected that they all enqueue tensors in queue.

• queue: A Queue.
• enqueue_ops: List of enqueue ops to run in threads later.
• close_op: Op to close the queue. Pending enqueue ops are preserved.
• cancel_op: Op to close the queue and cancel pending enqueue ops.
• queue_runner_def: Optional QueueRunnerDef protocol buffer. If specified, recreates the QueueRunner from its contents. queue_runner_def and the other arguments are mutually exclusive.

• ValueError: If both queue_runner_def and queue are both specified.
• ValueError: If queue or enqueue_ops are not provided when not restoring from queue_runner_def.

Create threads to run the enqueue ops.

This method requires a session in which the graph was launched. It creates a list of threads, optionally starting them. There is one thread for each op passed in enqueue_ops.

The coord argument is an optional coordinator, that the threads will use to terminate together and report exceptions. If a coordinator is given, this method starts an additional thread to close the queue when the coordinator requests a stop.

This method may be called again as long as all threads from a previous call have stopped.

• sess: A Session.
• coord: Optional Coordinator object for reporting errors and checking stop conditions.
• daemon: Boolean. If True make the threads daemon threads.
• start: Boolean. If True starts the threads. If False the caller must call the start() method of the returned threads.

A list of threads.

• RuntimeError: If threads from a previous call to create_threads() are still running.

Exceptions raised but not handled by the QueueRunner threads.

Exceptions raised in queue runner threads are handled in one of two ways depending on whether or not a Coordinator was passed to create_threads():

• With a Coordinator, exceptions are reported to the coordinator and forgotten by the QueueRunner.
• Without a Coordinator, exceptions are captured by the QueueRunner and made available in this exceptions_raised property.

A list of Python Exception objects. The list is empty if no exception was captured. (No exceptions are captured when using a Coordinator.)

Returns a QueueRunner object created from queue_runner_def.

The string name of the underlying Queue.

Converts this QueueRunner to a QueueRunnerDef protocol buffer.

A QueueRunnerDef protocol buffer.

Adds a QueueRunner to a collection in the graph.

When building a complex model that uses many queues it is often difficult to gather all the queue runners that need to be run. This convenience function allows you to add a queue runner to a well known collection in the graph.

The companion method start_queue_runners() can be used to start threads for all the collected queue runners.

• qr: A QueueRunner.
• collection: A GraphKey specifying the graph collection to add the queue runner to. Defaults to GraphKeys.QUEUE_RUNNERS.

Starts all queue runners collected in the graph.

This is a companion method to add_queue_runner(). It just starts threads for all queue runners collected in the graph. It returns the list of all threads.

• sess: Session used to run the queue ops. Defaults to the default session.
• coord: Optional Coordinator for coordinating the started threads.
• daemon: Whether the threads should be marked as daemons, meaning they don't block program exit.
• start: Set to False to only create the threads, not start them.
• collection: A GraphKey specifying the graph collection to get the queue runners from. Defaults to GraphKeys.QUEUE_RUNNERS.

A list of threads.

See Distributed TensorFlow for more information about how to configure a distributed TensorFlow program.

An in-process TensorFlow server, for use in distributed training.

A tf.train.Server instance encapsulates a set of devices and a tf.Session target that can participate in distributed training. A server belongs to a cluster (specified by a tf.train.ClusterSpec), and corresponds to a particular task in a named job. The server can communicate with any other server in the same cluster.

Creates a new server with the given definition.

The job_name, task_index, and protocol arguments are optional, and override any information provided in server_or_cluster_def.

• server_or_cluster_def: A tf.train.ServerDef or tf.train.ClusterDef protocol buffer, or a tf.train.ClusterSpec object, describing the server to be created and/or the cluster of which it is a member.
• job_name: (Optional.) Specifies the name of the job of which the server is a member. Defaults to the value in server_or_cluster_def, if specified.
• task_index: (Optional.) Specifies the task index of the server in its job. Defaults to the value in server_or_cluster_def, if specified. Otherwise defaults to 0 if the server's job has only one task.
• protocol: (Optional.) Specifies the protocol to be used by the server. Acceptable values include "grpc". Defaults to the value in server_or_cluster_def, if specified. Otherwise defaults to "grpc".
• start: (Optional.) Boolean, indicating whether to start the server after creating it. Defaults to True.

Creates a new single-process cluster running on the local host.

This method is a convenience wrapper for creating a tf.train.Server with a tf.train.ServerDef that specifies a single-process cluster containing a single task in a job called "local".

• start: (Optional.) Boolean, indicating whether to start the server after creating it. Defaults to True.

A local tf.train.Server.

Returns the target for a tf.Session to connect to this server.

To create a tf.Session that connects to this server, use the following snippet:

server = tf.train.Server(...)
with tf.Session(server.target):
  # ...

A string containing a session target for this server.

Starts this server.

Blocks until the server has shut down.

This method currently blocks forever.

A training helper that checkpoints models and computes summaries.

The Supervisor is a small wrapper around a Coordinator, a Saver, and a SessionManager that takes care of common needs of Tensorflow training programs.

with tf.Graph().as_default():
  ...add operations to the graph...
  # Create a Supervisor that will checkpoint the model in '/tmp/mydir'.
  sv = Supervisor(logdir='/tmp/mydir')
  # Get a Tensorflow session.
  sess = sv.prepare_or_wait_for_session(FLAGS.master)
  # Use the session to train the graph.
  while not sv.should_stop():
    sess.run(<my_train_op>)
  # Ask for all the services to stop.
  sv.stop()

After the call to prepare_or_wait_for_session(), all Variables in the Graph have been initialized. In addition, a few services have been started to checkpoint the model and fetch summaries.

If the program crashes and you restart it, the call to prepare_or_wait_for_session() automatically reinitializes the Variables from most recent checkpoint.

If any of the services raises an exception, it will ask the Supervisor to stop. In that case should_stop() will return True and you should stop your training loop.

Finish by calling stop() to cleanly wait for the services to complete. If a service thread raised an exception, it is re-raised in the stop() call so your program can easily report it.

To train with replicas you deploy the same program in a Cluster. One of the tasks must be identified as the chief: the task that handles initialization, checkpoints, summaries, and recovery. The other tasks depend on the chief for these services.

The only change you have to do to the single program code is to indicate if the program is running as the chief.

# Choose a task as the chief. This could be based on server_def.task_index,
# or job_def.name, or job_def.tasks. It's entirely up to the end user.
# But there can be only one *chief*.
is_chief = (server_def.task_index == 0)
server = tf.train.Server(server_def)

with tf.Graph().as_default():
  ...add operations to the graph...
  # Create a Supervisor that uses log directory on a shared file system.
  # Indicate if you are the 'chief'
  sv = Supervisor(logdir='/shared_directory/...', is_chief=is_chief)
  # Get a Session in a TensorFlow server on the cluster.
  sess = sv.prepare_or_wait_for_session(server.target)
  # Use the session to train the graph.
  while not sv.should_stop():
    sess.run(<my_train_op>)
  # Ask for all the services to stop.
  sv.stop()

In the chief task, the Supervisor works exactly as in the first example above. In the other tasks prepare_or_wait_for_session() waits for the Model to have been intialized before returning a session to the training code.

If one of the tasks crashes and restarts, prepare_or_wait_for_session() checks if the Model is initialized. If yes, it just creates a session and returns it to the training code that proceeds normally. If the model needs to be initialized, the chief task takes care of reinitializing it; the other tasks just wait for the model to have been initialized.

NOTE: This modified program still works fine as a single program. The single program marks itself as the chief.

Whether you are running on your machine or in the cluster you can use the following values for the --master flag:

• Specifying '' requests an in-process session that does not use RPC.

• Specifying 'local' requests a session that uses the RPC-based "Master interface" to run TensorFlow programs. See tf.train.Server.create_local_server() for details.

• Specifying 'grpc://hostname:port' requests a session that uses the RPC interface to a specific , and also allows the in-process master to access remote tensorflow workers. Often, it is appropriate to pass server.target (for some tf.train.Server named `server).

prepare_or_wait_for_session() launches the Checkpoint and Summary services (threads). If you need more services to run you can simply launch them after prepare_or_wait_for_session() returns. The Supervisor uses a Coordinator to help multiple threads stop together, so pass that coordinator (sv.coord) to the threads you launch.

Example: Start a QueueRunner to prefetch inputs.

...build the model with a QueueRunner to prefetch inputs...
qr = QueueRunner(input_queue, [enqueue_op])
...
sv = Supervisor(logdir='/tmp/mydir')
sess = sv.prepare_or_wait_for_session(FLAGS.master)
# Start the queue runner threads.
threads = qr.create_threads(sess, sv.coord, start=True)
# Catch OutOfRangeError, which signals that your input queue is exhausted.
try:
  while not sv.should_stop():
    sess.run(my_train_op)
except tf.errors.OutOfRangeError:
  pass
# Wait for the QueueRunner and service threads to complete.
sv.stop(threads)

Note: Starting QueueRunner threads is very common, to the Supervisor provides a convenience method named start_queue_runners(). If you use that method you do not have to keep track of the started threads and can just call stop() normally:

...build the model with a QueueRunner to prefetch inputs...
qr = QueueRunner(input_queue, [enqueue_op])
...
sv = Supervisor(logdir='/tmp/mydir')
sess = sv.prepare_or_wait_for_session(FLAGS.master)
# Start the queue runner threads.
sv.start_queue_runners(sess, [qr])
# Catch OutOfRangeError, which signals that your input queue is exhausted.
try:
  while not sv.should_stop():
    sess.run(my_train_op)
except tf.errors.OutOfRangeError:
  pass
# Wait for the QueueRunner and service threads to complete.
sv.stop()

prepare_or_wait_for_session() launches the "summary" and "checkpoint" services (threads) which use either the optionally summary_op and saver passed to the constructor, or default ones created automatically by the Supervisor. If you want to run your own summary and checkpointing logic, disable these services by passing None to the summary_op and saver parameters.

Example: Create summaries manually every 100 steps in the chief.

# Create a Supervisor with no automatic summaries.
sv = Supervisor(logdir='/tmp/mydir', is_chief=is_chief, summary_op=None)
# As summary_op was None, prepare_or_wait_for_session() does not start the
# summary thread.
sess = sv.prepare_or_wait_for_session(FLAGS.master)
for step in xrange(1000000):
  if is_chief and step % 100 == 0:
    # Create the summary every 100 chief steps.
    sv.summary_computed(sess, sess.run(my_summary_op))
  else:
    # Train normally
    sess.run(my_train_op)

prepare_or_wait_for_session() only supports initializing the model by running an init_op. If you have special initialization needs, use local_init_op.

Create a Supervisor.

• graph: A Graph. The graph that the model will use. Defaults to the default Graph. The supervisor may add operations to the graph before creating a session, but the graph should not be modified by the caller after passing it to the supervisor.
• ready_op: Operation to check if the model is initialized. This operation is run by supervisors in prepare_or_wait_for_session() to check if the model is ready to use. The model is considered ready if that operation succeeds. Defaults to the operation returned from tf.assert_variables_initialized() If None, the model is not checked for readiness.
• is_chief: If True, create a chief supervisor in charge of initializing and restoring the model. If False, create a supervisor that relies on a chief supervisor for inits and restore.
• init_op: Operation. Used by chief supervisors to initialize the model when it can not be recovered. Defaults to an Operation that initializes all variables. If None, no initialization is done automatically.
• init_feed_dict: A dictionary that maps Tensor objects to feed values. This feed dictionary will be used when init_op is evaluated.
• local_init_op: Operation. Used by all supervisors to run initializations that should run for every new supervisor instance. By default these are table initializers and initializers for local variables. If None, no further per supervisor-instance initialization is done automatically.
• logdir: A string. Optional path to a directory where to checkpoint the model and log events for the visualizer. Used by chief supervisors. The directory will be created if it does not exist.
• summary_op: An Operation that returns a Summary for the event logs. Used by chief supervisors if a logdir was specified. Defaults to the operation returned from merge_all_summaries(). If None, summaries are not computed automatically.
• saver: A Saver object. Used by chief supervisors if a logdir was specified. Defaults to the saved returned by Saver(). If None, the model is not saved automatically.
• global_step: An integer Tensor of size 1 that counts steps. The value from 'global_step' is used in summaries and checkpoint filenames. Default to the op named 'global_step' in the graph if it exists, is of rank 1, size 1, and of type tf.int32 ot tf.int64. If None the global step is not recorded in summaries and checkpoint files. Used by chief supervisors if a logdir was specified.
• save_summaries_secs: Number of seconds between the computation of summaries for the event log. Defaults to 120 seconds. Pass 0 to disable summaries.
• save_model_secs: Number of seconds between the creation of model checkpoints. Defaults to 600 seconds. Pass 0 to disable checkpoints.
• recovery_wait_secs: Number of seconds between checks that the model is ready. Used by supervisors when waiting for a chief supervisor to initialize or restore the model. Defaults to 30 seconds.
• checkpoint_basename: The basename for checkpoint saving.
• session_manager: SessionManager, which manages Session creation and recovery. If it is None, a default SessionManager will be created with the set of arguments passed in for backwards compatibility.

A Supervisor.

Make sure the model is ready to be used.

Create a session on 'master', recovering or initializing the model as needed, or wait for a session to be ready. If running as the chief and start_standard_service is set to True, also call the session manager to start the standard services.

• master: name of the TensorFlow master to use. If not specified or empty a 'Direct Session' is created.
• config: Optional ConfigProto proto used to configure the session, which is passed as-is to create the session.
• wait_for_checkpoint: Whether we should wait for the availability of a checkpoint before creating Session. Defaults to False.
• max_wait_secs: Maximum time to wait for the session to become available.
• start_standard_services: Whether to start the standard services, such as checkpoint, summary and step counter.

A Session object that can be used to drive the model.

Start the standard services for 'sess'.

This starts services in the background. The services started depend on the parameters to the constructor and may include:

• A Summary thread computing summaries every save_summaries_secs.
• A Checkpoint thread saving the model every every save_model_secs.
• A StepCounter thread measure step time.

• sess: A Session.

A list of threads that are running the standard services. You can use the Supervisor's Coordinator to join these threads with: sv.coord.Join()

• RuntimeError: If called with a non-chief Supervisor.
• ValueError: If not logdir was passed to the constructor as the services need a log directory.

Indicate that a summary was computed.

• sess: A Session object.
• summary: A Summary proto, or a string holding a serialized summary proto.
• global_step: Int. global step this summary is associated with. If None, it will try to fetch the current step.

• TypeError: if 'summary' is not a Summary proto or a string.
• RuntimeError: if the Supervisor was created without a logdir.

Stop the services and the coordinator.

This does not close the session.

• threads: Optional list of threads to join with the coordinator. If None, defaults to the threads running the standard services plus the threads started for QueueRunners if start_queue_runners() was called. To wait on an additional set of threads, pass the list in this parameter and they will be merged with the internal list of running services.
• close_summary_writer: Whether to close the summary_writer. Defaults to True.

Request that the coordinator stop the threads.

See Coordinator.request_stop().

• ex: Optional Exception, or Python exc_info tuple as returned by sys.exc_info(). If this is the first call to request_stop() the corresponding exception is recorded and re-raised from join().

Check if the coordinator was told to stop.

See Coordinator.should_stop().

True if the coordinator was told to stop, False otherwise.

Context handler to stop the supervisor when an exception is raised.

See Coordinator.stop_on_exception().

A context handler.

Block waiting for the coordinator to stop.

Start a LooperThread that calls a function periodically.

If timer_interval_secs is None the thread calls target(args) repeatedly. Otherwise target(args) is called every timer_interval_secs seconds. The thread terminates when a stop is requested.

The started thread is added to the list of threads managed by the supervisor so it does not need to be passed to the stop() method.

• timer_interval_secs: Number. Time boundaries at which to call target.
• target: A callable object.
• args: Optional arguments to pass to target when calling it.

The started thread.

Make sure the model is ready to be used.

Create a session on 'master', recovering or initializing the model as needed, or wait for a session to be ready. If running as the chief and start_standard_service is set to True, also call the session manager to start the standard services.

• master: name of the TensorFlow master to use. If not specified or empty a 'Direct Session' is created.
• config: Optional ConfigProto proto used to configure the session, which is passed as-is to create the session.
• wait_for_checkpoint: Whether we should wait for the availability of a checkpoint before creating Session. Defaults to False.
• max_wait_secs: Maximum time to wait for the session to become available.
• start_standard_services: Whether to start the standard services, such as checkpoint, summary and step counter.

A Session object that can be used to drive the model.

Request that the coordinator stop the threads.

See Coordinator.request_stop().

• ex: Optional Exception, or Python exc_info tuple as returned by sys.exc_info(). If this is the first call to request_stop() the corresponding exception is recorded and re-raised from join().

Check if the coordinator was told to stop.

See Coordinator.should_stop().

True if the coordinator was told to stop, False otherwise.

Start threads for QueueRunners.

• sess: A Session.
• queue_runners: A list of QueueRunners. If not specified, we'll use the list of queue runners gathered in the graph under the key GraphKeys.QUEUE_RUNNERS.

The list of threads started for the QueueRunners.

Start the standard services for 'sess'.

This starts services in the background. The services started depend on the parameters to the constructor and may include:

• A Summary thread computing summaries every save_summaries_secs.
• A Checkpoint thread saving the model every every save_model_secs.
• A StepCounter thread measure step time.

• sess: A Session.

A list of threads that are running the standard services. You can use the Supervisor's Coordinator to join these threads with: sv.coord.Join()

• RuntimeError: If called with a non-chief Supervisor.
• ValueError: If not logdir was passed to the constructor as the services need a log directory.

Stop the services and the coordinator.

This does not close the session.

• threads: Optional list of threads to join with the coordinator. If None, defaults to the threads running the standard services plus the threads started for QueueRunners if start_queue_runners() was called. To wait on an additional set of threads, pass the list in this parameter and they will be merged with the internal list of running services.
• close_summary_writer: Whether to close the summary_writer. Defaults to True.

Context handler to stop the supervisor when an exception is raised.

See Coordinator.stop_on_exception().

A context handler.

Indicate that a summary was computed.

• sess: A Session object.
• summary: A Summary proto, or a string holding a serialized summary proto.
• global_step: Int. global step this summary is associated with. If None, it will try to fetch the current step.

• TypeError: if 'summary' is not a Summary proto or a string.
• RuntimeError: if the Supervisor was created without a logdir.

Block waiting for the coordinator to stop.

Return the Coordinator used by the Supervisor.

The Coordinator can be useful if you want to run multiple threads during your training.

A Coordinator object.

Return the global_step Tensor used by the supervisor.

An integer Tensor for the global_step.

Return the feed dictionary used when evaluating the init_op.

A feed dictionary or None.

Return the Init Op used by the supervisor.

An Op or None.

Start a LooperThread that calls a function periodically.

If timer_interval_secs is None the thread calls target(args) repeatedly. Otherwise target(args) is called every timer_interval_secs seconds. The thread terminates when a stop is requested.

The started thread is added to the list of threads managed by the supervisor so it does not need to be passed to the stop() method.

• timer_interval_secs: Number. Time boundaries at which to call target.
• target: A callable object.
• args: Optional arguments to pass to target when calling it.

The started thread.

Return the Ready Op used by the supervisor.

An Op or None.

Return the delay between checkpoints.

A timestamp.

Return the save path used by the supervisor.

A string.

Return the delay between summary computations.

A timestamp.

Return the Saver used by the supervisor.

A Saver object.

Return the SessionManager used by the Supervisor.

A SessionManager object.

Start threads for QueueRunners.

• sess: A Session.
• queue_runners: A list of QueueRunners. If not specified, we'll use the list of queue runners gathered in the graph under the key GraphKeys.QUEUE_RUNNERS.

The list of threads started for the QueueRunners.

Return the Summary Tensor used by the chief supervisor.

A string Tensor for the summary or None.

Return the SummaryWriter used by the chief supervisor.

A SummaryWriter.

Training helper that restores from checkpoint and creates session.

This class is a small wrapper that takes care of session creation and checkpoint recovery. It also provides functions that to facilitate coordination among multiple training threads or processes.

• Checkpointing trained variables as the training progresses.
• Initializing variables on startup, restoring them from the most recent checkpoint after a crash, or wait for checkpoints to become available.

with tf.Graph().as_default():
   ...add operations to the graph...
  # Create a SessionManager that will checkpoint the model in '/tmp/mydir'.
  sm = SessionManager()
  sess = sm.prepare_session(master, init_op, saver, checkpoint_dir)
  # Use the session to train the graph.
  while True:
    sess.run(<my_train_op>)

prepare_session() initializes or restores a model. It requires init_op and saver as an argument.

A second process could wait for the model to be ready by doing the following:

with tf.Graph().as_default():
   ...add operations to the graph...
  # Create a SessionManager that will wait for the model to become ready.
  sm = SessionManager()
  sess = sm.wait_for_session(master)
  # Use the session to train the graph.
  while True:
    sess.run(<my_train_op>)

wait_for_session() waits for a model to be initialized by other processes.

Creates a SessionManager.

The local_init_op is an Operation that is run always after a new session was created. If None, this step is skipped.

The ready_op is an Operation. The model is considered ready if that operation succeeds. If None, the model is not checked for readiness.

recovery_wait_secs is the number of seconds between checks that the model is ready. It is used by processes to wait for a model to be initialized or restored. Defaults to 30 seconds.

• local_init_op: An Operation run immediately after session creation. Usually used to initialize tables and local variables.
• ready_op: An Operation to check if the model is initialized.
• graph: The Graph that the model will use.
• recovery_wait_secs: Seconds between checks for the model to be ready.

Creates a Session. Makes sure the model is ready to be used.

Creates a Session on 'master'. If a saver object is passed in, and checkpoint_dir points to a directory containing valid checkpoint files, then it will try to recover the model from checkpoint. If no checkpoint files are available, and wait_for_checkpoint is True, then the process would check every recovery_wait_secs, up to max_wait_secs, for recovery to succeed.

If the model cannot be recovered successfully, and an init_op is not None, the init_op is run to initialize the model.

This is a convenient function for the following, with a few error checks added:

sess, initialized = self.recover_session(master)
if not initialized:
  sess.run(self.init_op)
return sess

• master: String representation of the TensorFlow master to use.
• init_op: Operation used to to initialize the model.
• saver: A Saver object used to restore a model.
• checkpoint_dir: Path to the checkpoint files.
• wait_for_checkpoint: Whether to wait for checkpoint to become available.
• max_wait_secs: Maximum time to wait for checkpoints to become available.
• config: Optional ConfigProto proto used to configure the session.
• init_feed_dict: A dictionary that maps Tensor objects to feed values. This feed dictionary will be used when init_op is evaluated.

A Session object that can be used to drive the model.

• RuntimeError: If the model cannot be initialized or recovered.

Creates a Session, recovering if possible.

Creates a new session on 'master'. If the session is not initialized and can be recovered from a checkpoint, recover it.

• master: String representation of the TensorFlow master to use.
• saver: A Saver object used to restore a model.
• checkpoint_dir: Path to the checkpoint files.
• wait_for_checkpoint: Whether to wait for checkpoint to become available.
• max_wait_secs: Maximum time to wait for checkpoints to become available.
• config: Optional ConfigProto proto used to configure the session.

A pair (sess, initialized) where 'initialized' is True if the session could be recovered, False otherwise.

Creates a new Session and waits for model to be ready.

Creates a new Session on 'master'. Waits for the model to be initialized or recovered from a checkpoint. It's expected that another thread or process will make the model ready, and that this is intended to be used by threads/processes that participate in a distributed training configuration where a different thread/process is responsible for initializing or recovering the model being trained.

NB: The amount of time this method waits for the session is bounded by max_wait_secs. By default, this function will wait indefinitely.

• master: String representation of the TensorFlow master to use.
• config: Optional ConfigProto proto used to configure the session.
• max_wait_secs: Maximum time to wait for the session to become available.

A Session. May be None if the operation exceeds the timeout specified by config.operation_timeout_in_ms.

tf.DeadlineExceededError: if the session is not available after max_wait_secs.

Represents a cluster as a set of "tasks", organized into "jobs".

A tf.train.ClusterSpec represents the set of processes that participate in a distributed TensorFlow computation. Every tf.train.Server is constructed in a particular cluster.

To create a cluster with two jobs and five tasks, you specify the mapping from job names to lists of network addresses (typically hostname-port pairs).

cluster = tf.train.ClusterSpec({"worker": ["worker0.example.com:2222",
                                           "worker1.example.com:2222",
                                           "worker2.example.com:2222"],
                                "ps": ["ps0.example.com:2222",
                                       "ps1.example.com:2222"]})

Returns a tf.train.ClusterDef protocol buffer based on this cluster.

Returns a dictionary from job names to lists of network addresses.

Creates a ClusterSpec.

• cluster: A dictionary mapping one or more job names to lists of network addresses, or a tf.train.ClusterDef protocol buffer.

• TypeError: If cluster is not a dictionary mapping strings to lists of strings, and not a tf.train.ClusterDef protobuf.

Returns a list of tasks in the given job.

• job_name: The string name of a job in this cluster.

A list of strings, corresponding to the network addresses of tasks in the given job, ordered by task index.

• ValueError: If job_name does not name a job in this cluster.

Returns a list of job names in this cluster.

A list of strings, corresponding to the names of jobs in this cluster.

Return a device function to use when building a Graph for replicas.

Device Functions are used in with tf.device(device_function): statement to automatically assign devices to Operation objects as they are constructed, Device constraints are added from the inner-most context first, working outwards. The merging behavior adds constraints to fields that are yet unset by a more inner context. Currently the fields are (job, task, cpu/gpu).

If cluster is None, and ps_tasks is 0, the returned function is a no-op.

For example,

# To build a cluster with two ps jobs on hosts ps0 and ps1, and 3 worker
# jobs on hosts worker0, worker1 and worker2.
cluster_spec = {
    "ps": ["ps0:2222", "ps1:2222"],
    "worker": ["worker0:2222", "worker1:2222", "worker2:2222"]}
with tf.device(tf.replica_device_setter(cluster=cluster_spec)):
  # Build your graph
  v1 = tf.Variable(...)  # assigned to /job:ps/task:0
  v2 = tf.Variable(...)  # assigned to /job:ps/task:1
  v3 = tf.Variable(...)  # assigned to /job:ps/task:0
# Run compute

• ps_tasks: Number of tasks in the ps job.
• ps_device: String. Device of the ps job. If empty no ps job is used. Defaults to ps.
• worker_device: String. Device of the worker job. If empty no worker job is used.
• merge_devices: Boolean. If True, merges or only sets a device if the device constraint is completely unset. merges device specification rather than overriding them.
• cluster: ClusterDef proto or ClusterSpec.
• ps_ops: List of Operation objects that need to be placed on ps devices.

A function to pass to tf.device().

TypeError if cluster is not a dictionary or ClusterDef protocol buffer.

The following ops output Summary protocol buffers as serialized string tensors.

You can fetch the output of a summary op in a session, and pass it to a SummaryWriter to append it to an event file. Event files contain Event protos that can contain Summary protos along with the timestamp and step. You can then use TensorBoard to visualize the contents of the event files. See TensorBoard and Summaries for more details.

Outputs a Summary protocol buffer with scalar values.

The input tags and values must have the same shape. The generated summary has a summary value for each tag-value pair in tags and values.

• tags: A string Tensor. Tags for the summaries.
• values: A real numeric Tensor. Values for the summaries.
• collections: Optional list of graph collections keys. The new summary op is added to these collections. Defaults to [GraphKeys.SUMMARIES].
• name: A name for the operation (optional).

A scalar Tensor of type string. The serialized Summary protocol buffer.

Outputs a Summary protocol buffer with images.

The summary has up to max_images summary values containing images. The images are built from tensor which must be 4-D with shape [batch_size, height, width, channels] and where channels can be:

• 1: tensor is interpreted as Grayscale.
• 3: tensor is interpreted as RGB.
• 4: tensor is interpreted as RGBA.

The images have the same number of channels as the input tensor. For float input, the values are normalized one image at a time to fit in the range [0, 255]. uint8 values are unchanged. The op uses two different normalization algorithms:

• If the input values are all positive, they are rescaled so the largest one is 255.

• If any input value is negative, the values are shifted so input value 0.0 is at 127. They are then rescaled so that either the smallest value is 0, or the largest one is 255.

The tag argument is a scalar Tensor of type string. It is used to build the tag of the summary values:

• If max_images is 1, the summary value tag is 'tag/image'.
• If max_images is greater than 1, the summary value tags are generated sequentially as 'tag/image/0', 'tag/image/1', etc.

• tag: A scalar Tensor of type string. Used to build the tag of the summary values.
• tensor: A 4-D uint8 or float32 Tensor of shape [batch_size, height, width, channels] where channels is 1, 3, or 4.
• max_images: Max number of batch elements to generate images for.
• collections: Optional list of ops.GraphKeys. The collections to add the summary to. Defaults to [ops.GraphKeys.SUMMARIES]
• name: A name for the operation (optional).

A scalar Tensor of type string. The serialized Summary protocol buffer.

Outputs a Summary protocol buffer with a histogram.

The generated Summary has one summary value containing a histogram for values.

This op reports an InvalidArgument error if any value is not finite.

• tag: A string Tensor. 0-D. Tag to use for the summary value.
• values: A real numeric Tensor. Any shape. Values to use to build the histogram.
• collections: Optional list of graph collections keys. The new summary op is added to these collections. Defaults to [GraphKeys.SUMMARIES].
• name: A name for the operation (optional).

A scalar Tensor of type string. The serialized Summary protocol buffer.

Returns the fraction of zeros in value.

If value is empty, the result is nan.

This is useful in summaries to measure and report sparsity. For example,

z = tf.Relu(...)
summ = tf.scalar_summary('sparsity', tf.nn.zero_fraction(z))

• value: A tensor of numeric type.
• name: A name for the operation (optional).

The fraction of zeros in value, with type float32.

Merges summaries.

This op creates a Summary protocol buffer that contains the union of all the values in the input summaries.

When the Op is run, it reports an InvalidArgument error if multiple values in the summaries to merge use the same tag.

• inputs: A list of string Tensor objects containing serialized Summary protocol buffers.
• collections: Optional list of graph collections keys. The new summary op is added to these collections. Defaults to [GraphKeys.SUMMARIES].
• name: A name for the operation (optional).

A scalar Tensor of type string. The serialized Summary protocol buffer resulting from the merging.

Merges all summaries collected in the default graph.

• key: GraphKey used to collect the summaries. Defaults to GraphKeys.SUMMARIES.

If no summaries were collected, returns None. Otherwise returns a scalar Tensor of typestring containing the serialized Summary protocol buffer resulting from the merging.

See Summaries and TensorBoard for an overview of summaries, event files, and visualization in TensorBoard.

Writes Summary protocol buffers to event files.

The SummaryWriter class provides a mechanism to create an event file in a given directory and add summaries and events to it. The class updates the file contents asynchronously. This allows a training program to call methods to add data to the file directly from the training loop, without slowing down training.

Creates a SummaryWriter and an event file.

On construction the summary writer creates a new event file in logdir. This event file will contain Event protocol buffers constructed when you call one of the following functions: add_summary(), add_session_log(), add_event(), or add_graph().

If you pass a Graph to the constructor it is added to the event file. (This is equivalent to calling add_graph() later).

TensorBoard will pick the graph from the file and display it graphically so you can interactively explore the graph you built. You will usually pass the graph from the session in which you launched it:

...create a graph...
# Launch the graph in a session.
sess = tf.Session()
# Create a summary writer, add the 'graph' to the event file.
writer = tf.train.SummaryWriter(<some-directory>, sess.graph)

The other arguments to the constructor control the asynchronous writes to the event file:

• flush_secs: How often, in seconds, to flush the added summaries and events to disk.
• max_queue: Maximum number of summaries or events pending to be written to disk before one of the 'add' calls block.

• logdir: A string. Directory where event file will be written.
• graph: A Graph object, such as sess.graph.
• max_queue: Integer. Size of the queue for pending events and summaries.
• flush_secs: Number. How often, in seconds, to flush the pending events and summaries to disk.
• graph_def: DEPRECATED: Use the graph argument instead.

Adds a Summary protocol buffer to the event file.

This method wraps the provided summary in an Event protocol buffer and adds it to the event file.

You can pass the result of evaluating any summary op, using Session.run() or Tensor.eval(), to this function. Alternatively, you can pass a tf.Summary protocol buffer that you populate with your own data. The latter is commonly done to report evaluation results in event files.

• summary: A Summary protocol buffer, optionally serialized as a string.
• global_step: Number. Optional global step value to record with the summary.

Adds a SessionLog protocol buffer to the event file.

This method wraps the provided session in an Event procotol buffer and adds it to the event file.

• session_log: A SessionLog protocol buffer.
• global_step: Number. Optional global step value to record with the summary.

Adds an event to the event file.

• event: An Event protocol buffer.

Adds a Graph to the event file.

The graph described by the protocol buffer will be displayed by TensorBoard. Most users pass a graph in the constructor instead.

• graph: A Graph object, such as sess.graph.
• global_step: Number. Optional global step counter to record with the graph.
• graph_def: DEPRECATED. Use the graph parameter instead.

• ValueError: If both graph and graph_def are passed to the method.

Adds a metadata information for a single session.run() call.

• run_metadata: A RunMetadata protobuf object.
• tag: The tag name for this metadata.
• global_step: Number. Optional global step counter to record with the StepStats.

• ValueError: If the provided tag was already used for this type of event.

Flushes the event file to disk.

Call this method to make sure that all pending events have been written to disk.

Flushes the event file to disk and close the file.

Call this method when you do not need the summary writer anymore.

An iterator for reading Event protocol buffers from an event file.

You can use this function to read events written to an event file. It returns a Python iterator that yields Event protocol buffers.

Example: Print the contents of an events file.

for e in tf.train.summary_iterator(path to events file):
    print(e)

Example: Print selected summary values.

# This example supposes that the events file contains summaries with a
# summary value tag 'loss'.  These could have been added by calling
# `add_summary()`, passing the output of a scalar summary op created with
# with: `tf.scalar_summary(['loss'], loss_tensor)`.
for e in tf.train.summary_iterator(path to events file):
    for v in e.summary.value:
        if v.tag == 'loss':
            print(v.simple_value)

See the protocol buffer definitions of Event and Summary for more information about their attributes.

• path: The path to an event file created by a SummaryWriter.

Event protocol buffers.

Small helper to get the global step.

# Creates a variable to hold the global_step.
global_step_tensor = tf.Variable(10, trainable=False, name='global_step')
# Creates a session.
sess = tf.Session()
# Initializes the variable.
sess.run(global_step_tensor.initializer)
print('global_step: %s' % tf.train.global_step(sess, global_step_tensor))

global_step: 10

• sess: A TensorFlow Session object.
• global_step_tensor: Tensor or the name of the operation that contains the global step.

The global step value.

Writes a graph proto on disk.

The graph is written as a binary proto unless as_text is True.

v = tf.Variable(0, name='my_variable')
sess = tf.Session()
tf.train.write_graph(sess.graph_def, '/tmp/my-model', 'train.pbtxt')

• graph_def: A GraphDef protocol buffer.
• logdir: Directory where to write the graph.
• name: Filename for the graph.
• as_text: If True, writes the graph as an ASCII proto.

A thread that runs code repeatedly, optionally on a timer.

This thread class is intended to be used with a Coordinator. It repeatedly runs code specified either as target and args or by the run_loop() method.

Before each run the thread checks if the coordinator has requested stop. In that case the looper thread terminates immediately.

If the code being run raises an exception, that exception is reported to the coordinator and the thread terminates. The coordinator will then request all the other threads it coordinates to stop.

You typically pass looper threads to the supervisor Join() method.

Create a LooperThread.

• coord: A Coordinator.
• timer_interval_secs: Time boundaries at which to call Run(), or None if it should be called back to back.
• target: Optional callable object that will be executed in the thread.
• args: Optional arguments to pass to target when calling it.

• ValueError: If one of the arguments is invalid.

A boolean value indicating whether this thread is a daemon thread (True) or not (False).

This must be set before start() is called, otherwise RuntimeError is raised. Its initial value is inherited from the creating thread; the main thread is not a daemon thread and therefore all threads created in the main thread default to daemon = False.

The entire Python program exits when no alive non-daemon threads are left.

Thread identifier of this thread or None if it has not been started.

This is a nonzero integer. See the thread.get_ident() function. Thread identifiers may be recycled when a thread exits and another thread is created. The identifier is available even after the thread has exited.

Return whether the thread is alive.

This method returns True just before the run() method starts until just after the run() method terminates. The module function enumerate() returns a list of all alive threads.

Return whether the thread is alive.

This method returns True just before the run() method starts until just after the run() method terminates. The module function enumerate() returns a list of all alive threads.

Wait until the thread terminates.

This blocks the calling thread until the thread whose join() method is called terminates -- either normally or through an unhandled exception or until the optional timeout occurs.

When the timeout argument is present and not None, it should be a floating point number specifying a timeout for the operation in seconds (or fractions thereof). As join() always returns None, you must call isAlive() after join() to decide whether a timeout happened -- if the thread is still alive, the join() call timed out.

When the timeout argument is not present or None, the operation will block until the thread terminates.

A thread can be join()ed many times.

join() raises a RuntimeError if an attempt is made to join the current thread as that would cause a deadlock. It is also an error to join() a thread before it has been started and attempts to do so raises the same exception.

Start a LooperThread that calls a function periodically.

If timer_interval_secs is None the thread calls target(args) repeatedly. Otherwise target(args) is called every timer_interval_secs seconds. The thread terminates when a stop of the coordinator is requested.

• coord: A Coordinator.
• timer_interval_secs: Number. Time boundaries at which to call target.
• target: A callable object.
• args: Optional arguments to pass to target when calling it.

The started thread.

A string used for identification purposes only.

It has no semantics. Multiple threads may be given the same name. The initial name is set by the constructor.

Called at 'timer_interval_secs' boundaries.

Start the thread's activity.

It must be called at most once per thread object. It arranges for the object's run() method to be invoked in a separate thread of control.

This method will raise a RuntimeError if called more than once on the same thread object.

Called when the thread starts.

Called when the thread stops.

Returns MetaGraphDef proto. Optionally writes it to filename.

This function exports the graph, saver, and collection objects into MetaGraphDef protocol buffer with the intension of it being imported at a later time or location to restart training, run inference, or be a subgraph.

• filename: Optional filename including the path for writing the generated MetaGraphDef protocol buffer.
• meta_info_def: MetaInfoDef protocol buffer.
• graph_def: GraphDef protocol buffer.
• saver_def: SaverDef protocol buffer.
• collection_list: List of string keys to collect.
• as_text: If True, writes the MetaGraphDef as an ASCII proto.

A MetaGraphDef proto.

Generates a checkpoint state proto.

• save_dir: Directory where the model was saved.
• model_checkpoint_path: The checkpoint file.
• all_model_checkpoint_paths: List of strings. Paths to all not-yet-deleted checkpoints, sorted from oldest to newest. If this is a non-empty list, the last element must be equal to model_checkpoint_path. These paths are also saved in the CheckpointState proto.

CheckpointState proto with model_checkpoint_path and all_model_checkpoint_paths updated to either absolute paths or relative paths to the current save_dir.

Recreates a Graph saved in a MetaGraphDef proto.

This function takes a MetaGraphDef protocol buffer as input. If the argument is a file containing a MetaGraphDef protocol buffer , it constructs a protocol buffer from the file content. The function then adds all the nodes from the graph_def field to the current graph, recreates all the collections, and returns a saver constructed from the saver_def field.

In combination with export_meta_graph(), this function can be used to

• Serialize a graph along with other Python objects such as QueueRunner, Variable into a MetaGraphDef.

• Restart training from a saved graph and checkpoints.

• Run inference from a saved graph and checkpoints.

...
# Create a saver.
saver = tf.train.Saver(...variables...)
# Remember the training_op we want to run by adding it to a collection.
tf.add_to_collection('train_op', train_op)
sess = tf.Session()
for step in xrange(1000000):
    sess.run(train_op)
    if step % 1000 == 0:
        # Saves checkpoint, which by default also exports a meta_graph
        # named 'my-model-global_step.meta'.
        saver.save(sess, 'my-model', global_step=step)

Later we can continue training from this saved meta_graph without building the model from scratch.

with tf.Session() as sess:
  new_saver = tf.train.import_meta_graph('my-save-dir/my-model-10000.meta')
  new_saver.restore(sess, 'my-save-dir/my-model-10000')
  # tf.get_collection() returns a list. In this example we only want the
  # first one.
  train_op = tf.get_collection('train_op')[0]
  for step in xrange(1000000):
    sess.run(train_op)

NOTE: Restarting training from saved meta_graph only works if the device assignments have not changed.

• meta_graph_or_file: MetaGraphDef protocol buffer or filename (including the path) containing a MetaGraphDef.

A saver constructed from saver_def in MetaGraphDef or None.

A None value is returned if no variables exist in the MetaGraphDef (i.e., there are no variables to restore).