问tensorflow2新版本中tf.GradientTape.gradient设置unconnected_gradients为0是报错？

源代码：

with tf.GradientTape(persistent=False) as tap:

# loss of boundary condtions

lb = self.loss_fn_conds(batch_data)

# 边界条件关于神经网络参数的梯度

gb = tap.gradient(lb,

[self.model_lr.trainable_variables, self.model_hr.trainable_variables, self.filter_delta],

unconnected_gradients=tf.UnconnectedGradients.ZERO)

问题：

损失函数与网络并没有关联，返回的是默认值0。我希望损失函数关于网络参数的导数也能返回0。在之前的tensorflow 2.7版本中，只需将unconnected_gradients设置成tf.UnconnectedGradients.ZERO即可满足我的要求。

但是最近将tensorflow升级到2.21，出现了错误。当unconnected_gradients设置成NONE时，可以运行，返回NONE；但设置成ZERO时，报错，并没有返回我需要的0值。报错信息如下：

File /mnt/d/HPWANG/Projects/PINNs-TF2.20-Linux-R5/pinns_2d_multiscale.py:630, in PINNs_2D_MultiScale.get_dynamic_weight(self, batch_data, weight_res)

628 print(lb)

629 # 边界条件关于神经网络参数的梯度

--> 630 gb = tap.gradient(lb,

631 [self.model_lr.trainable_variables, self.model_hr.trainable_variables, self.filter_delta],

632 unconnected_gradients=tf.UnconnectedGradients.ZERO)

634 gb0 = [tf.reshape(j, (1, -1)) for j in gb[0]]

635 gb0 = tf.concat(gb0, axis=1)

File ~/anaconda3/envs/tf-gpu/lib/python3.11/site-packages/tensorflow/python/eager/backprop.py:1066, in GradientTape.gradient(self, target, sources, output_gradients, unconnected_gradients)

1060 output_gradients = (

1061 composite_tensor_gradient.get_flat_tensors_for_gradients(

1062 output_gradients))

1063 output_gradients = [None if x is None else ops.convert_to_tensor(x)

1064 for x in output_gradients]

-> 1066 flat_grad = imperative_grad.imperative_grad(

1067 self._tape,

1068 flat_targets,

1069 flat_sources,

1070 output_gradients=output_gradients,

1071 sources_raw=flat_sources_raw,

1072 unconnected_gradients=unconnected_gradients)

1074 if not self._persistent:

1075 # Keep track of watched variables before setting tape to None

1076 self._watched_variables = self._tape.watched_variables()

File ~/anaconda3/envs/tf-gpu/lib/python3.11/site-packages/tensorflow/python/eager/imperative_grad.py:67, in imperative_grad(tape, target, sources, output_gradients, sources_raw, unconnected_gradients)

63 except ValueError:

64 raise ValueError(

65 "Unknown value for unconnected_gradients: %r" % unconnected_gradients)

---> 67 return pywrap_tfe.TFE_Py_TapeGradient(

68 tape._tape, # pylint: disable=protected-access

69 target,

70 sources,

71 output_gradients,

72 sources_raw,

73 compat.as_str(unconnected_gradients.value))

SystemError: <function BaseResourceVariable.dtype at 0x70cc34b6b060> returned a result with an exception set