Python array_ops.check_numerics函数代码示例

  def testOneShotIteratorInitializerFails(self):
    # Define a dataset whose initialization will always fail.
    dataset = dataset_ops.Dataset.from_tensors(
        array_ops.check_numerics(
            constant_op.constant(1.0) / constant_op.constant(0.0), "oops"))
    iterator = dataset.make_one_shot_iterator()
    next_element = iterator.get_next()

    with self.test_session() as sess:
      with self.assertRaisesRegexp(errors.InvalidArgumentError, "oops"):
        sess.run(next_element)

      # Test that subsequent attempts to use the iterator also fail.
      with self.assertRaisesRegexp(errors.InvalidArgumentError, "oops"):
        sess.run(next_element)

    with self.test_session() as sess:
      def consumer_thread():
        with self.assertRaisesRegexp(errors.InvalidArgumentError, "oops"):
          sess.run(next_element)

      num_threads = 8
      threads = [
          self.checkedThread(consumer_thread) for _ in range(num_threads)]
      for t in threads:
        t.start()
      for t in threads:
        t.join()

 def testPassThrough(self):
   with self.session(graph=ops.Graph()):
     t1 = constant_op.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3])
     checked = array_ops.check_numerics(t1, message="pass through test")
     value = self.evaluate(checked)
     self.assertAllEqual(np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]), value)
     self.assertEqual([2, 3], checked.get_shape())

 def testWindowIgnoreErrors(self):
   input_values = np.float32([1., np.nan, 2., np.nan, 3.])
   dataset = dataset_ops.Dataset.from_tensor_slices(input_values).map(
       lambda x: array_ops.check_numerics(x, "message")).window(
           size=2, shift=2, stride=2,
           drop_remainder=True).flat_map(lambda x: x.batch(batch_size=2))
   self.assertDatasetProduces(
       dataset, expected_output=[np.float32([1., 2.]),
                                 np.float32([2., 3.])])

  def testParallelMapUnspecifiedOutputSize(self):
    components = np.array([1., 2., 3., np.nan, 5.]).astype(np.float32)

    dataset = (dataset_ops.Dataset.from_tensor_slices(components)
               .map(lambda x: array_ops.check_numerics(x, "message"),
                    num_parallel_calls=2))
    get_next = self.getNext(dataset)

    for _ in range(3):
      self.evaluate(get_next())

 def testBatchAndMapDatasetFails(self):
   """Test a dataset that maps a TF function across its input elements."""
   dataset = dataset_ops.Dataset.from_tensors(
       array_ops.check_numerics(
           constant_op.constant(1.0) / constant_op.constant(0.0), "oops"))
   batch_size = array_ops.placeholder(dtypes.int64, shape=[])
   iterator = (dataset.apply(batching.map_and_batch(lambda x: x, batch_size))
               .make_initializable_iterator())
   init_op = iterator.initializer
   with self.test_session() as sess:
     with self.assertRaisesRegexp(errors.InvalidArgumentError, "oops"):
       sess.run(init_op, feed_dict={batch_size: 14})

  def testWindowIgnoreErrors(self):
    input_values = np.float32([1., np.nan, 2., np.nan, 3.])
    dataset = dataset_ops.Dataset.from_tensor_slices(input_values).map(
        lambda x: array_ops.check_numerics(x, "message")).window(
            size=2, shift=2, stride=2,
            drop_remainder=True).flat_map(lambda x: x.batch(batch_size=2))
    get_next = dataset.make_one_shot_iterator().get_next()

    with self.cached_session() as sess:
      self.assertAllEqual(np.float32([1., 2.]), sess.run(get_next))
      self.assertAllEqual(np.float32([2., 3.]), sess.run(get_next))
      with self.assertRaises(errors.OutOfRangeError):
        sess.run(get_next)

  def testParallelMapIgnoreError(self):
    components = np.array([1., 2., 3., np.nan, 5.]).astype(np.float32)

    dataset = (
        dataset_ops.Dataset.from_tensor_slices(components).map(
            lambda x: array_ops.check_numerics(x, "message"),
            num_parallel_calls=2).prefetch(2).apply(error_ops.ignore_errors()))
    get_next = self.getNext(dataset)

    for x in [1., 2., 3., 5.]:
      self.assertEqual(x, self.evaluate(get_next()))
    with self.assertRaises(errors.OutOfRangeError):
      self.evaluate(get_next())

  def testMapAndBatchFails(self, numa_aware):
    """Test a dataset that maps a TF function across its input elements."""

    with self.assertRaisesRegexp(errors.InvalidArgumentError, "oops"):
      dataset = dataset_ops.Dataset.from_tensors(
          array_ops.check_numerics(
              constant_op.constant(1.0) / constant_op.constant(0.0), "oops"))
      dataset = dataset.apply(batching.map_and_batch(lambda x: x, 14))
      if numa_aware:
        options = dataset_ops.Options()
        options.experimental_numa_aware = True
        dataset = dataset.with_options(options)
      get_next = self.getNext(dataset)
      self.evaluate(get_next())

def add_check_numerics_ops():
  """Connect a `tf.debugging.check_numerics` to every floating point tensor.

  `check_numerics` operations themselves are added for each `half`, `float`,
  or `double` tensor in the current default graph. For all ops in the graph, the
  `check_numerics` op for all of its (`half`, `float`, or `double`) inputs
  is guaranteed to run before the `check_numerics` op on any of its outputs.

  Note: This API is not compatible with the use of `tf.cond` or
  `tf.while_loop`, and will raise a `ValueError` if you attempt to call it
  in such a graph.

  Returns:
    A `group` op depending on all `check_numerics` ops added.

  Raises:
    ValueError: If the graph contains any numeric operations in a control flow
      structure.
    RuntimeError: If called with eager execution enabled.

  @compatibility(eager)
  Not compatible with eager execution. To check for `Inf`s and `NaN`s under
  eager execution, call `tfe.seterr(inf_or_nan='raise')` once before executing
  the checked operations.
  @end_compatibility
  """
  if context.executing_eagerly():
    raise RuntimeError(
        "add_check_numerics_ops() is not compatible with eager execution. "
        "To check for Inf's and NaN's under eager execution, call "
        "tfe.seterr(inf_or_nan='raise') once before executing the "
        "checked operations.")

  check_op = []
  # This code relies on the ordering of ops in get_operations().
  # The producer of a tensor always comes before that tensor's consumer in
  # this list. This is true because get_operations() returns ops in the order
  # added, and an op can only be added after its inputs are added.
  for op in ops.get_default_graph().get_operations():
    for output in op.outputs:
      if output.dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
        if op._get_control_flow_context() is not None:  # pylint: disable=protected-access
          raise ValueError("`tf.add_check_numerics_ops() is not compatible "
                           "with TensorFlow control flow operations such as "
                           "`tf.cond()` or `tf.while_loop()`.")

        message = op.name + ":" + str(output.value_index)
        with ops.control_dependencies(check_op):
          check_op = [array_ops.check_numerics(output, message=message)]
  return control_flow_ops.group(*check_op)

  def testParallelMapUnspecifiedOutputSize(self):
    components = np.array([1., 2., 3., np.nan, 5.]).astype(np.float32)

    dataset = (dataset_ops.Dataset.from_tensor_slices(components)
               .map(lambda x: array_ops.check_numerics(x, "message"),
                    num_parallel_calls=2))
    iterator = dataset.make_initializable_iterator()
    init_op = iterator.initializer
    get_next = iterator.get_next()

    with self.cached_session() as sess:
      sess.run(init_op)
      for _ in range(3):
        sess.run(get_next)

def _VerifyTensor(t, name, msg):
    """Assert that the tensor does not contain any NaN's.

  Args:
    t: Tensor
    name: name
    msg: message to log
  Returns:
    Tensor, but verified
  """
    with ops.name_scope(name):
        with ops.device(t.device or ops.get_default_graph().get_default_device()):
            verify_input = array_ops.check_numerics(t, message=msg)
            out = control_flow_ops.with_dependencies([verify_input], t)
    return out

  def testParallelMapError(self):
    components = np.array([1., 2., 3., np.nan, 5.]).astype(np.float32)

    dataset = (dataset_ops.Dataset.from_tensor_slices(components)
               .map(lambda x: array_ops.check_numerics(x, "message"),
                    num_parallel_calls=2))
    get_next = self.getNext(dataset)

    for _ in range(3):
      self.evaluate(get_next())
    # The 4th element is NaN, so `array_ops.check_numerics()` should fail.
    with self.assertRaises(errors.InvalidArgumentError):
      self.evaluate(get_next())
    self.evaluate(get_next())
    with self.assertRaises(errors.OutOfRangeError):
      self.evaluate(get_next())

  def testInterleaveDatasetError(self, input_values, cycle_length, block_length,
                                 num_parallel_calls):
    dataset = dataset_ops.Dataset.from_tensor_slices(input_values).map(
        lambda x: array_ops.check_numerics(x, "message")).interleave(
            dataset_ops.Dataset.from_tensors, cycle_length, block_length,
            num_parallel_calls)
    get_next = self.getNext(dataset)

    for value in input_values:
      if np.isnan(value):
        with self.assertRaises(errors.InvalidArgumentError):
          self.evaluate(get_next())
      else:
        self.assertEqual(value, self.evaluate(get_next()))
    with self.assertRaises(errors.OutOfRangeError):
      self.evaluate(get_next())

  def testMapIgnoreError(self):
    components = np.array([1., 2., 3., np.nan, 5.]).astype(np.float32)

    dataset = (dataset_ops.Dataset.from_tensor_slices(components)
               .map(lambda x: array_ops.check_numerics(x, "message"))
               .ignore_errors())
    iterator = dataset.make_initializable_iterator()
    init_op = iterator.initializer
    get_next = iterator.get_next()

    with self.test_session() as sess:
      sess.run(init_op)
      for x in [1., 2., 3., 5.]:
        self.assertEqual(x, sess.run(get_next))
      with self.assertRaises(errors.OutOfRangeError):
        sess.run(get_next)

  def _model_fn(features, labels, mode, config, params=None):
    """model_fn."""

    # TODO(jhseu): Move to EVAL and PREDICT to TPU.
    if mode != model_fn_lib.ModeKeys.TRAIN:
      return _call_model_fn_without_tpu(
          model_fn, features, labels, mode, config, params)

    # Now for TPU training.
    if params is not None and _BATCH_SIZE_KEY in params:
      params[_BATCH_SIZE_KEY] //= config.tpu_config.num_shards

    assert isinstance(features, _PerShardOutput)
    features = features.as_list()
    if labels is not None:
      assert isinstance(labels, _PerShardOutput)
      labels = labels.as_list()

    dequeue_fn, enqueue_fn = (
        _create_infeed_enqueue_ops_and_dequeue_fn(config, features, labels))

    loss = _train_on_tpu_shards(
        config,
        train_step=_convert_model_fn_to_train_step(
            model_fn, dequeue_fn, mode, config, params))

    # Gets the variables back from TPU nodes. This means the variables updated
    # by TPU will now be *synced* to host memory.
    update_ops = [
        array_ops.check_numerics(v.read_value(),
                                 'Gradient for %s is NaN' % v.name).op
        for v in variables.trainable_variables()
    ]

    hooks = [
        TpuInfeedSessionHook(config, enqueue_fn),
        training.LoggingTensorHook(
            {'loss': array_ops.identity(loss),
             'step': training.get_global_step()},
            every_n_secs=30)
    ]

    return model_fn_lib.EstimatorSpec(
        mode,
        loss=array_ops.identity(loss),
        training_hooks=hooks,
        train_op=control_flow_ops.group(*update_ops))