self.w = np.random.uniform(low=-2, high=2, size=w_shape).astype('f4')

      @def_function.function
      def condition(self, x, w):
        return math_ops.reduce_sum(x, keepdims=False) < 100

      @def_function.function
      def body(self, x, w):
        z = nn_ops.conv2d(x, w, padding='SAME')
        return z, w

      @def_function.function(

  auto reduce =
      internal::BroadcastGradientArgs(scope, x_batch_shape, y_batch_shape);
  (*grad_outputs)[0] =
      Reshape(scope, ReduceSum(scope, (*grad_outputs)[0], reduce.r0), sx);
  (*grad_outputs)[1] =
      Reshape(scope, ReduceSum(scope, (*grad_outputs)[1], reduce.r1), sy);
  return scope.status();
}
REGISTER_GRADIENT_OP("BatchMatMulV2", BatchMatMulV2Grad);

    auto loc = op.getLoc();

    // Note that the TensorFlow Softmax op verifies that the input rank is
    // greater than or equal to one so the following sequence is valid.
    auto reduce_dim =
        rewriter.create<TF::ConstOp>(loc, GetI64ElementsAttr({-1}, &rewriter));

    // Exponential of input values and then their sum can be very large here.