search

shenweichen / DeepCTR

Easy-to-use,Modular and Extendible package of deep-learning based CTR models .
https://deepctr-doc.readthedocs.io/en/latest/index.html
Apache License 2.0
7.44k stars 2.19k forks source link

dropout和batchnorm层已经传了training，outputs._uses_learning_phase = training is not None的作用是什么？ #506

Open Daemoonn opened 1 year ago

Daemoonn commented 1 year ago

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 https://github.com/shenweichen/DeepCTR/blob/e8f4d818f9b46608bc95bb60ef0bb0633606b2f2/deepctr/layers/sequence.py#L266-L288

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？ https://github.com/shenweichen/DeepCTR/blob/e8f4d818f9b46608bc95bb60ef0bb0633606b2f2/deepctr/layers/sequence.py#L266

https://github.com/shenweichen/DeepCTR/blob/e8f4d818f9b46608bc95bb60ef0bb0633606b2f2/deepctr/layers/core.py#L104 https://github.com/shenweichen/DeepCTR/blob/e8f4d818f9b46608bc95bb60ef0bb0633606b2f2/deepctr/layers/core.py#L198 https://github.com/shenweichen/DeepCTR/blob/e8f4d818f9b46608bc95bb60ef0bb0633606b2f2/deepctr/layers/core.py#L205

为什么还需要对_uses_learning_phase手动设置？ https://github.com/shenweichen/DeepCTR/blob/e8f4d818f9b46608bc95bb60ef0bb0633606b2f2/deepctr/layers/sequence.py#L283-L286

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):