dropout和batchnorm层已经传了training，outputs._uses_learning_phase = training is not None的作用是什么？

[Daemoonn]

Describe the question(问题描述)
既然DNN中的bn_layers和dropout_layers已经传入了training状态参数来区分是在训练还是在预测阶段，手动设置_uses_learning_phase的作用是什么？
我的tf版本2.10.0，应该是执行outputs._uses_learning_phase = training is not None

DeepCTR/deepctr/layers/sequence.py

Lines 266 to 288 in e8f4d81

	attention_score = self.local_att([queries, keys], training=training)
	outputs = tf.transpose(attention_score, (0, 2, 1))
	if self.weight_normalization:
	paddings = tf.ones_like(outputs) * (-2 ** 32 + 1)
	else:
	paddings = tf.zeros_like(outputs)
	outputs = tf.where(key_masks, outputs, paddings)
	if self.weight_normalization:
	outputs = softmax(outputs)
	if not self.return_score:
	outputs = tf.matmul(outputs, keys)
	if tf.__version__ < '1.13.0':
	outputs._uses_learning_phase = attention_score._uses_learning_phase
	else:
	outputs._uses_learning_phase = training is not None
	return outputs

在求attention_score的时候，不是已经把training状态传给DNN里的dropout和batchnorm层了吗？

DeepCTR/deepctr/layers/sequence.py

Line 266 in e8f4d81

attention_score = self.local_att([queries, keys], training=training)

DeepCTR/deepctr/layers/core.py

Line 104 in e8f4d81

att_out = self.dnn(att_input, training=training)

DeepCTR/deepctr/layers/core.py

Line 198 in e8f4d81

fc = self.bn_layers[i](fc, training=training)

DeepCTR/deepctr/layers/core.py

Line 205 in e8f4d81

fc = self.dropout_layers[i](fc, training=training)

为什么还需要对_uses_learning_phase手动设置？

DeepCTR/deepctr/layers/sequence.py

Lines 283 to 286 in e8f4d81

	if tf.__version__ < '1.13.0':
	outputs._uses_learning_phase = attention_score._uses_learning_phase
	else:
	outputs._uses_learning_phase = training is not None

这里手动设置_uses_learning_phase是否发挥实际作用？去掉会怎么样？对于outputs这个tf.Tensor来说，好像是没有_uses_learning_phase这个类属性，这么写好像是临时新建个了_uses_learning_phase属性，然后赋值成xxx

outputs._uses_learning_phase = xxx

求大佬指点

Additional context
Add any other context about the problem here.

Operating environment(运行环境):

• python version 3.9.12
• tensorflow version 2.10.0
• deepctr version 0.9.3