orbit.StandardEvaluator

View source on GitHub

Implements the standard functionality of AbstractEvaluator APIs.

Inherits From: AbstractEvaluator

orbit.StandardEvaluator(
    eval_dataset,
    options: Optional[orbit.StandardEvaluatorOptions] = None
)

This class structures evaluation roughly as follows:

state = eval_begin()
for _ in range(num_steps):
  step_outputs = eval_step(eval_iterator)
  state = eval_reduce(state, step_outputs)
return eval_end(state)

Calls to eval_begin and eval_end are always done in eager mode, while eval_step may be compiled with tf.function as determined by the options passed to __init__. eval_reduce is in eager mode if use_tf_while_loop=False in StandardEvaluatorOptions, but in graph mode if use_tf_while_loop=True.

This class does not support completely evaluating multiple different datasets (i.e., where every example of each dataset should be processed, as opposed to running for a fixed number of evaluation steps). A custom AbstractEvaluator is recommended in this case.

a = tf.Module()
b = tf.Module()
c = tf.Module()
a.b = b
b.c = c
list(a.submodules) == [b, c]
True
list(b.submodules) == [c]
True
list(c.submodules) == []
True

Methods

create_eval_loop_fn

View source

create_eval_loop_fn(
    has_state: bool
)

Creates an eval loop from the current step function and options.

eval_begin

View source

eval_begin() -> Any

Called once at the beginning of the evaluation.

This method is always called in eager mode, and is a good place to reset metrics that accumulate values over the course of evaluation.

Note that this method is called before dataset iterator creation.

eval_end

View source

eval_end(
    *args
) -> Optional[runner.Output]

Called at the end of the evaluation.

Called once at the end of evaluation.

This method is always called in eager mode, and is a good place to get metric results. The value returned from this function will be returned as-is from the evaluate method implementation provided by StandardEvaluator.

eval_reduce

View source

eval_reduce(
    state: Optional[Any] = None, step_outputs: Optional[runner.Output] = None
) -> Any

A function to perform per-step reduction on the evaluation outputs.

This is useful for passing state throughout evaluation, especially in cases where maintaining or accumulating state is hard to accomplish using tf.metrics.Metric or other tf.Variable-based approaches. For instance, it can be used to easily accumulate all per-example losses from the full evaluation for subsequent processing in eval_end().

eval_step

View source

@abc.abstractmethod
eval_step(
    iterator
) -> Any

Implements one step of evaluation.

What a "step" consists of is up to the implementer. When using distribution strategies, the call to this method takes place in the "cross-replica context" for generality, to allow e.g. multiple iterator dequeues and calls to strategy.run.

Note that if use_tf_function=True, all the code inside eval_step should be compatible with tf.function tracing (and in particular, any state modifications involving self should be avoided). In some cases, non- tf.function compatible code can be moved to eval_loop_begin, eval_reduce, or eval_loop_end, which always execute eagerly.

evaluate

View source

evaluate(
    num_steps: tf.Tensor
) -> Optional[runner.Output]

Implements num_steps steps of evaluation.

with_name_scope

@classmethod
with_name_scope(
    method
)

Decorator to automatically enter the module name scope.

class MyModule(tf.Module):
  @tf.Module.with_name_scope
  def __call__(self, x):
    if not hasattr(self, 'w'):
      self.w = tf.Variable(tf.random.normal([x.shape[1], 3]))
    return tf.matmul(x, self.w)

Using the above module would produce tf.Variables and tf.Tensors whose names included the module name:

mod = MyModule()
mod(tf.ones([1, 2]))
<tf.Tensor: shape=(1, 3), dtype=float32, numpy=..., dtype=float32)>
mod.w
<tf.Variable 'my_module/Variable:0' shape=(2, 3) dtype=float32,
numpy=..., dtype=float32)>