%arg2: tensor<64xf32>
) -> tensor<1x28x28x64xf32> {

  // CHECK: "tf.FusedBatchNormGradV3"
  // CHECK-SAME: (%[[X_TRANSPOSE:[0-9]*]], %[[Y_TRANSPOSE:[0-9]*]],
  // CHECK-SAME: data_format = "NCHW"
  %x_backprop, %scale_backprop, %offset_backprop, %reserve_1, %reserve_2
    = "tf.FusedBatchNormGradV3"(%arg0, %arg1, %arg2, %arg2, %arg2, %arg2)
       {
         data_format = "NHWC",
         epsilon = 1.001 : f32,

  // 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];

  // CHECK: %[[RES_PERM:.*]] = "tf.Const"()
  // CHECK-SAME: <{value = dense<[0, 2, 3, 1]> : tensor<4xi64>}>

  // CHECK: %[[RES_TPOSE:[0-9]*]] = "tf.Transpose"
  // CHECK-SAME: (%x_backprop, %[[RES_PERM]])
  // CHECK: return %[[RES_TPOSE]]

  %x_backprop, %scale_backprop, %offset_backprop, %reserve_1, %reserve_2
    = "tf.FusedBatchNormGradV3"(%arg0, %arg1, %arg2, %arg2, %arg2, %arg2)
       {
         data_format = "NHWC",

    absl::Span<AbstractTensorHandle*> outputs) {
  AbstractTensorHandle* loss;
  AbstractTensorHandle* backprop;
  TF_RETURN_IF_ERROR(ops::SparseSoftmaxCrossEntropyWithLogits(
      ctx, inputs[0], inputs[1], &loss, &backprop,
      "SparseSoftmaxCrossEntropyWithLogits"));
  // `gradient_checker` only works with model that returns only 1 tensor.

//
Status SparseSoftmaxCrossEntropyWithLogits(AbstractContext* ctx, AbstractTensorHandle* const features, AbstractTensorHandle* const labels, AbstractTensorHandle** loss, AbstractTensorHandle** backprop, const char* name = nullptr, const char* raw_device_name = nullptr);

//

// Summary:
//
// Description:
Status SparseSoftmaxCrossEntropyWithLogits(AbstractContext* ctx, AbstractTensorHandle* const features, AbstractTensorHandle* const labels, AbstractTensorHandle** loss, AbstractTensorHandle** backprop, const char* name, const char* raw_device_name) {
  AbstractOperationPtr op_ptr(ctx->CreateOperation());
  TF_RETURN_IF_ERROR(op_ptr->Reset("SparseSoftmaxCrossEntropyWithLogits", raw_device_name));

// 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

//   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).
//   ```
Status IdentityN(AbstractContext* ctx,

vent.HIDDEN)\n    })\n  }\n\n  _removeBackdrop() {\n    if (this._backdrop) {\n      $(this._backdrop).remove()\n      this._backdrop = null\n    }\n  }\n\n  _showBackdrop(callback) {\n    const animate = $(this._element).hasClass(ClassName.FADE)\n      ? ClassName.FADE : ''\n\n    if (this._isShown && this._config.backdrop) {\n      this._backdrop = document.createElement('div')\n      this._backdrop.className = ClassName.BACKDROP\n\n      if (animate) {\n        this._backdrop.classList.add(animate)\n...

vent.HIDDEN)\n    })\n  }\n\n  _removeBackdrop() {\n    if (this._backdrop) {\n      $(this._backdrop).remove()\n      this._backdrop = null\n    }\n  }\n\n  _showBackdrop(callback) {\n    const animate = $(this._element).hasClass(ClassName.FADE)\n      ? ClassName.FADE : ''\n\n    if (this._isShown && this._config.backdrop) {\n      this._backdrop = document.createElement('div')\n      this._backdrop.className = ClassName.BACKDROP\n\n      if (animate) {\n        this._backdrop.classList.add(animate)\n...