Создание модели, которая выводит свои собственные градиенты

Question

Я нахожусь на tenorflow 2.1 и хочу создать модель, которая выводит как ее прогноз, так и градиенты с постоянным значением.

Input => Model => [Score, Gradients]

Мой код следующий, с base_model - простой моделью, делающей прогнозирование с плавающей точкой из изображения 128x128x3:

def GradNet(base_model):
    input_shape = (128,128,3)
    inputs = tf.keras.Input(input_shape)

    with tf.GradientTape() as tape:
        tape.watch(scoring_model.variables)
        score = base_model(inputs)
        loss = tf.keras.losses.mean_squared_error(123., score)

    gradients = tape.gradient(loss, inputs)

    model = tf.keras.Model(inputs=inputs, outputs=[score, gradients], name='GradNet')

    return model

Однако я получаю следующую ошибку:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-46-566ab9eaafd1> in <module>
----> 1 grad_net = GradNet(model)

<ipython-input-45-adb556fb740b> in GradNet(scoring_model)
      8         loss = tf.keras.losses.mean_squared_error(666., score)
      9 
---> 10     gradients = tape.gradient(loss, inputs)
     11 
     12     model = tf.keras.Model(inputs=inputs, outputs=[score, gradients], name='GradNet')

~/.local/lib/python3.7/site-packages/tensorflow/python/eager/backprop.py in gradient(self, target, sources, output_gradients, unconnected_gradients)
   1046         output_gradients=output_gradients,
   1047         sources_raw=flat_sources_raw,
-> 1048         unconnected_gradients=unconnected_gradients)
   1049 
   1050     if not self._persistent:

~/.local/lib/python3.7/site-packages/tensorflow/python/eager/imperative_grad.py in imperative_grad(tape, target, sources, output_gradients, sources_raw, unconnected_gradients)
     75       output_gradients,
     76       sources_raw,
---> 77       compat.as_str(unconnected_gradients.value))

~/.local/lib/python3.7/site-packages/tensorflow/python/eager/backprop.py in _gradient_function(op_name, attr_tuple, num_inputs, inputs, outputs, out_grads, skip_input_indices, forward_pass_name_scope)
    155       gradient_name_scope = "gradient_tape/"
    156     with ops.name_scope(gradient_name_scope):
--> 157       return grad_fn(mock_op, *out_grads)
    158   else:
    159     return grad_fn(mock_op, *out_grads)

~/.local/lib/python3.7/site-packages/tensorflow/python/ops/math_grad.py in _MeanGrad(op, grad)
    249 def _MeanGrad(op, grad):
    250   """Gradient for Mean."""
--> 251   sum_grad = _SumGrad(op, grad)[0]
    252   input_shape = op.inputs[0]._shape_tuple()  # pylint: disable=protected-access
    253   output_shape = op.outputs[0]._shape_tuple()  # pylint: disable=protected-access

~/.local/lib/python3.7/site-packages/tensorflow/python/ops/math_grad.py in _SumGrad(op, grad)
    209       # more sense.
    210       output_shape_kept_dims = math_ops.reduced_shape(input_shape,
--> 211                                                       op.inputs[1])
    212     grad = array_ops.reshape(grad, output_shape_kept_dims)
    213   return [array_ops.broadcast_to(grad, input_shape), None]

~/.local/lib/python3.7/site-packages/tensorflow/python/ops/math_ops.py in reduced_shape(input_shape, axes)
   3754   """
   3755   if context.executing_eagerly():
-> 3756     input_shape = input_shape.numpy()
   3757     axes = axes.numpy()
   3758     input_shape[axes] = 1

AttributeError: 'Tensor' object has no attribute 'numpy'

Не понимаю, почему это невозможно,

Заранее спасибо