Python gradients_impl.gradients方法代码示例

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def testUnbatchGrad(self):
    """Tests that batch and unbatch are differentiable."""
    with self.test_session() as sess:
      inp = array_ops.placeholder(dtype=dtypes.int32, shape=[1])
      batched, index, id_t = batch_ops.batch(
          [inp], num_batch_threads=1, max_batch_size=2,
          batch_timeout_micros=36000000, grad_timeout_micros=1000000,
          batching_queue="")
      computation = batched[0] * batched[0]
      result = batch_ops.unbatch(computation, index, id_t,
                                 timeout_micros=1000000, shared_name="unbatch")
      grad = gradients_impl.gradients(result, inp)
      thread_results = []

      def worker():
        thread_results.extend(sess.run([grad], feed_dict={inp: [1]}))

      worker_thread = threading.Thread(target=worker)
      worker_thread.start()
      main_results = sess.run([grad], feed_dict={inp: [2]})
      worker_thread.join()
      self.assertEqual(thread_results[0], [2])
      self.assertEqual(main_results[0], [4]) 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def testSequenceToSequenceGradient(self):
    with self.test_session():
      size = (17, 1, 15)
      output_size = (17, 1, 8)
      inputs = constant_op.constant(_rand(*size))
      outputs = lstm1d.ndlstm_base(inputs, 8, dynamic=False)
      variables.global_variables_initializer().run()
      gradients = gradients_impl.gradients(outputs, inputs)
      if 1:  # pylint: disable=using-constant-test
        gradients = gradients_impl.gradients(outputs, inputs)[0].eval()
        self.assertEqual(gradients.shape, size)
      else:
        # TODO(tmb) tf.test.compute_gradient error is currently broken
        # with dynamic_rnn. Enable this test case eventually.
        err = gradient_checker.compute_gradient_error(
            inputs, size, outputs, output_size, delta=1e-4)
        self.assert_(not np.isnan(err))
        self.assert_(err < 0.1) 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def testSequenceToSequenceGradientReverse(self):
    with self.test_session():
      size = (17, 1, 15)
      output_size = (17, 1, 8)
      inputs = constant_op.constant(_rand(*size))
      outputs = lstm1d.ndlstm_base(inputs, 8, reverse=1, dynamic=False)
      variables.global_variables_initializer().run()
      if 1:  # pylint: disable=using-constant-test
        gradients = gradients_impl.gradients(outputs, inputs)[0].eval()
        self.assertEqual(gradients.shape, size)
      else:
        # TODO(tmb) tf.test.compute_gradient error is currently broken
        # with dynamic_rnn. Enable this test case eventually.
        err = gradient_checker.compute_gradient_error(
            inputs, size, outputs, output_size, delta=1e-4)
        self.assert_(not np.isnan(err))
        self.assert_(err < 0.1) 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def testDoubleCallInSameScopeFails(self):

    @rev_block_lib.recompute_grad
    def layer_with_recompute(inputs):
      return core_layers.dense(inputs, 2)

    with variable_scope.variable_scope("layer", use_resource=True):
      inputs = array_ops.ones((2, 4), dtypes.float32)
      out1 = layer_with_recompute(inputs)
      out2 = layer_with_recompute(inputs) + out1
      out = math_ops.reduce_sum(out2)

    tvars = variables.trainable_variables()
    assert len(tvars) == 4
    with self.assertRaisesWithPredicateMatch(
        ValueError, "called twice in the same enclosing scope"):
      gradients_impl.gradients(out, [inputs] + tvars) 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def testDoubleCallInUniqueScope(self):

    @rev_block_lib.recompute_grad
    def layer_with_recompute(inputs):
      with variable_scope.variable_scope("inner", use_resource=True):
        return core_layers.dense(inputs, 2)

    with variable_scope.variable_scope("layer", use_resource=True):
      inputs = array_ops.ones((2, 4), dtypes.float32)

      with variable_scope.variable_scope("layer1", use_resource=True):
        out1 = layer_with_recompute(inputs)
      with variable_scope.variable_scope("layer2", use_resource=True):
        out2 = layer_with_recompute(inputs) + out1
      out = math_ops.reduce_sum(out2)

    tvars = variables.trainable_variables()
    assert len(tvars) == 4
    grads = gradients_impl.gradients(out, [inputs] + tvars)
    for grad in grads:
      self.assertIsNotNone(grad) 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def _Conv2DBackpropInputGrad(op, grad):
  """The derivatives for deconvolution.

  Args:
    op: the Deconvolution op.
    grad: the tensor representing the gradient w.r.t. the output

  Returns:
    the gradients w.r.t. the input and the filter
  """
  return [None,
          nn_ops.conv2d_backprop_filter(grad, array_ops.shape(op.inputs[1]),
                                        op.inputs[2], op.get_attr("strides"),
                                        op.get_attr("padding"),
                                        op.get_attr("use_cudnn_on_gpu"),
                                        op.get_attr("data_format")),
          nn_ops.conv2d(grad, op.inputs[1], op.get_attr("strides"),
                        op.get_attr("padding"), op.get_attr("use_cudnn_on_gpu"),
                        op.get_attr("data_format"))] 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def _BiasAddGrad(op, received_grad):
  """Return the gradients for the 2 inputs of bias_op.

  The first input of unused_bias_op is the tensor t, and its gradient is
  just the gradient the unused_bias_op received.

  The second input of unused_bias_op is the bias vector which has one fewer
  dimension than "received_grad" (the batch dimension.)  Its gradient is the
  received gradient Summed on the batch dimension, which is the first dimension.

  Args:
    op: The BiasOp for which we need to generate gradients.
    received_grad: Tensor.  The gradients passed to the BiasOp.

  Returns:
    Two tensors, the first one for the "tensor" input of the BiasOp,
    the second one for the "bias" input of the BiasOp.
  """
  try:
    data_format = op.get_attr("data_format")
  except ValueError:
    data_format = None
  return (received_grad, gen_nn_ops.bias_add_grad(out_backprop=received_grad,
                                                  data_format=data_format)) 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def _BiasAddGradV1(unused_bias_op, received_grad):
  """Return the gradients for the 2 inputs of bias_op.

  The first input of unused_bias_op is the tensor t, and its gradient is
  just the gradient the unused_bias_op received.

  The second input of unused_bias_op is the bias vector which has one fewer
  dimension than "received_grad" (the batch dimension.)  Its gradient is the
  received gradient Summed on the batch dimension, which is the first dimension.

  Args:
    unused_bias_op: The BiasOp for which we need to generate gradients.
    received_grad: Tensor.  The gradients passed to the BiasOp.

  Returns:
    Two tensors, the first one for the "tensor" input of the BiasOp,
    the second one for the "bias" input of the BiasOp.
  """
  reduction_dim_tensor = math_ops.range(array_ops.rank(received_grad) - 1)
  return (received_grad, math_ops.reduce_sum(received_grad,
                                             reduction_dim_tensor)) 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def _FractionalMaxPoolGrad(op, grad_0, unused_grad_1, unused_grad_2):
  """Returns gradient for FractionalMaxPool.

  Since FractionalMaxPool has three outputs, there are three gradients passed in
  for each of the outputs. Only the first one is useful, the other two gradients
  are empty.

  Args:
    op: The FractionalMaxPoolOp.
    grad_0: Gradient with respect to op.outputs[0]
    unused_grad_1: Gradient with respect to op.outputs[1]/row_seq. It is empty.
    unused_grad_2: Gradient with respect to op.outputs[2]/col_seq. It is empty.

  Returns:
    Input backprop for FractionalMaxPool op.
  """
  # pylint: disable=protected-access
  return gen_nn_ops._fractional_max_pool_grad(op.inputs[0], op.outputs[0],
                                              grad_0, op.outputs[1],
                                              op.outputs[2],
                                              op.get_attr("overlapping")) 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def _FractionalAvgPoolGrad(op, grad_0, unused_grad_1, unused_grad_2):
  """Returns gradient for FractionalAvgPool.

  Since FractionalAvgPool has three outputs, there are three gradients passed in
  for each of the outputs. Only the first one is useful, the other two gradients
  are empty.

  Args:
    op: The FractionalAvgPoolOp.
    grad_0: Gradient with respect to op.outputs[0]
    unused_grad_1: Gradient with respect to op.outputs[1]/row_seq. It is empty.
    unused_grad_2: Gradient with respect to op.outputs[2]/col_seq. It is empty.

  Returns:
    Input backprop for FractionalAvgPool op.
  """
  # pylint: disable=protected-access
  return gen_nn_ops._fractional_avg_pool_grad(op.inputs[0].get_shape(), grad_0,
                                              op.outputs[1], op.outputs[2],
                                              op.get_attr("overlapping")) 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def _BatchNormWithGlobalNormalizationGrad(op, grad):
  """Return the gradients for the 5 inputs of BatchNormWithGlobalNormalization.

  We do not backprop anything for the mean and var intentionally as they are
  not being trained with backprop in the operation.

  Args:
    op: The BatchNormOp for which we need to generate gradients.
    grad: Tensor.  The gradients passed to the BatchNormOp.

  Returns:
    dx: Backprop for input, which is (grad * (g * rsqrt(v + epsilon)))
    dm: Backprop for mean, which is
        sum_over_rest(grad * g) * (-1 / rsqrt(v + epsilon))
    dv: Backprop for variance, which is
        sum_over_rest(grad * g * (x - m)) * (-1/2) * (v + epsilon) ^ (-3/2)
    db: Backprop for beta, which is grad reduced in all except the
        last dimension.
    dg: Backprop for gamma, which is (grad * ((x - m) * rsqrt(v + epsilon)))
  """
  dx, dm, dv, db, dg = gen_nn_ops._batch_norm_with_global_normalization_grad(
      op.inputs[0], op.inputs[1], op.inputs[2], op.inputs[4], grad,
      op.get_attr("variance_epsilon"), op.get_attr("scale_after_normalization"))
  return dx, dm, dv, db, dg 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def _acc_grads(*lists_of_grads):
  """Accumulates lists of gradients."""
  acc_grads = []
  for grads in zip(*lists_of_grads):
    grads = [g for g in grads if g is not None]
    if grads:
      acc_grads.append(math_ops.add_n(grads))
    else:
      acc_grads.append(None)
  return acc_grads 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def _testGradient(self, np_input, bias, dtype, data_format, use_gpu):
    with self.cached_session(use_gpu=use_gpu):
      if data_format == "NCHW":
        np_input = self._NHWCToNCHW(np_input)
      jacob_a, jacob_n = self._computeGradient(np_input, bias, dtype,
                                               data_format)
      input_jacob_a, bias_jacob_a, grad_jacob_a = jacob_a
      input_jacob_n, bias_jacob_n, grad_jacob_n = jacob_n

      if dtype == np.float16:
        # Compare fp16 analytical gradients to fp32 numerical gradients,
        # since fp16 numerical gradients are too imprecise unless great
        # care is taken with choosing the inputs and the delta. This is
        # a weaker, but pragmatic, check (in particular, it does not test
        # the op itself, only its gradient).
        _, jacob_n = self._computeGradient(np_input, bias, np.float32,
                                           data_format)
        input_jacob_n, bias_jacob_n, grad_jacob_n = jacob_n

      if dtype == dtypes.float64:
        threshold = 1e-10
      elif np_input.size >= 512:
        # The 5e-3 threshold seems to have been marginal in these cases, and
        # small changes in the test were pushing it over the limit.
        threshold = 5e-2
      else:
        threshold = 5e-3
      self.assertAllClose(input_jacob_a, input_jacob_n, threshold, threshold)
      self.assertAllClose(bias_jacob_a, bias_jacob_n, threshold, threshold)
      self.assertAllClose(grad_jacob_a, grad_jacob_n, threshold, threshold) 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def testClipGrads(self):
    xs = variables_lib2.Variable(0.0)
    ys = xs * 4.0
    grads = gradients_impl.gradients([ys], [xs])
    gradients_to_variables = list(zip(grads, [xs]))
    clipped_gradients_to_variables = training.clip_gradient_norms(
        gradients_to_variables, 3.0)

    with self.cached_session() as session:
      session.run(variables_lib2.global_variables_initializer())
      self.assertAlmostEqual(4.0, gradients_to_variables[0][0].eval())
      self.assertAlmostEqual(3.0, clipped_gradients_to_variables[0][0].eval()) 

# 需要导入模块: from tensorflow.python.ops import gradients_impl [as 别名]
# 或者: from tensorflow.python.ops.gradients_impl import gradients [as 别名]
def testClipGradsFn(self):
    xs = variables_lib2.Variable(0.0)
    ys = xs * 4.0
    grads = gradients_impl.gradients([ys], [xs])
    gradients_to_variables = list(zip(grads, [xs]))
    clipped_gradients_to_variables = training.clip_gradient_norms_fn(3.0)(
        gradients_to_variables)

    with self.cached_session() as session:
      session.run(variables_lib2.global_variables_initializer())
      self.assertAlmostEqual(4.0, gradients_to_variables[0][0].eval())
      self.assertAlmostEqual(3.0, clipped_gradients_to_variables[0][0].eval())