// computes loss and backprop of the loss with respect to 'features'.
//
// Softmax cross entropy loss is defined as follows:
//
//  loss = Sum(-labels * Log(Exp(features) / Sum(Exp(features)))
//  loss = Sum(-labels * LogSoftmax(features))
//
// Computing gradient of the loss with respect to features gives us,
//
//  backprop = (Exp(features) / Sum(Exp(features))) - labels

// any gradients to be computed).
//
// Finally, we start a backprop stack with a set of tape entries for which we
// have all gradients available. This set usually is a subset of the set of
// targets (not all since targets which have outputs in the tape will not have
// gradients available initially).
//
// Then we repeatedly pop an entry from the stack, run its backprop, and update

  // We multiply the backprop for cost with the gradients - op.output[1].
  // There is no gradient for labels.

  // The outputs of the network are at input index 0.
  auto logits = op.input(0);
  // The "truth" labels are at index 1.
  auto softmax_grad = op.output(1);

  // The loss is the output at index 0, and backprop is the output at index 1.
  auto grad_loss = grad_inputs[0];

  auto y =
      tensorflow::ops::SoftmaxCrossEntropyWithLogits(scope_, logits, labels);
  // Note the reversal of the backprop and loss orders. Issue #18734 has been
  // opened for this.
  RunTest({logits, labels}, {logits_shape, logits_shape}, {y.backprop, y.loss},
          {logits_shape, loss_shape});
}

TEST_F(NNGradTest, LogSoftmaxGrad) {
  TensorShape shape({5, 3});

      });
    }
    filter_value = filter.getValue();
    mul_value = multiplier.getValue();
    // In MHLO, Conv filter is in HWIO format, Depthwise conv filter is in HW1O
    // format and backprop input conv filter is in HWOI format.
    // Only fuses multiplier if all dimensions other than the out channel
    // dimension are equal to 1.
    if (!TFL::IsDimensionsDegenerateExceptLastOne(

    // CHECK: %[[BACKPROP:.*]] = "tf.Conv2DBackpropFilter"
    // CHECK-SAME: strides = [1, 1, 1, 1]
    // CHECK-SAME: (tensor<2x115x115x12xf32>, tensor<4xi32>, tensor<2x112x112x64xf32>) -> tensor<4x4x12x64xf32>

                                        {dx, dy, dz}, &grad_outputs));
    }
  }
  CompareTestAndExpectedGraphs();
}

TEST_F(GradientsTest, StackUnstack_StopBackprop) {
  // Tests that backprop stops before calculating gradients for Stack (because
  // only gradients w.r.t the output of Stack are requested).
  for (const bool expected : {false, true}) {
    const Scope& scope = expected ? scope_expected_ : scope_test_;

      Add(params_->body_inputs[0], {one, 0}, params_->body_graph, s_);
  ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);
  params_->body_outputs[0] = {add, 0};

  ExpectOK();

  // Create backprop graph
  TF_Output grad_output;
  TF_AddGradients(graph_, outputs_.data(), outputs_.size(), inputs_.data(), 1,
                  nullptr, s_, &grad_output);
  ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);

  // CHECK-DAG: %[[SOFTMAX:.*]] = "tf.Div"(%[[SOFTMAX_EXP]], %[[SOFTMAX_SUM]]) : (tensor<2x3xf32>, tensor<2x1xf32>) -> tensor<2x3xf32>

  // CHECK-DAG: %[[BACKPROP:.*]] = "tf.Sub"(%[[SOFTMAX]], %[[LABELS]]) : (tensor<2x3xf32>, tensor<2x3xf32>) -> tensor<2x3xf32>
  // CHECK: return %[[LOSS]], %[[BACKPROP]]

  %0:2 = "tf.SoftmaxCrossEntropyWithLogits"(%features, %labels) : (tensor<2x3xf32>, tensor<2x3xf32>) -> (tensor<2xf32>, tensor<2x3xf32>)

example, suppose y = f(x) and we wish to apply a custom function g for backprop
such that dx = g(dy). In Python,

```python
with tf.get_default_graph().gradient_override_map(
    {'IdentityN': 'OverrideGradientWithG'}):
  y, _ = identity_n([f(x), x])

@tf.RegisterGradient('OverrideGradientWithG')
def ApplyG(op, dy, _):
  return [None, g(dy)]  # Do not backprop to f(x).
```
  }];

  let arguments = (ins