batch_normalization 为啥不传入training字段啊，

[lihairui1990]

tf.layers.batch_normalization(inputs=in_, name='bn1' + stag, reuse=tf.AUTO_REUSE)

这里为啥不传入training 这个字段啊，之前看batch_normalization都需要传入training字段啊，训练时传入True 测试时传入False

training: Either a Python boolean, or a TensorFlow boolean scalar tensor
  (e.g. a placeholder). Whether to return the output in training mode
  (normalized with statistics of the current batch) or in inference mode
  (normalized with moving statistics). **NOTE**: make sure to set this
  parameter correctly, or else your training/inference will not work
  properly.

[WLCOOLONGS]

bn1 = tf.layers.batch_normalization(inputs=inp, name='bn1', training=train_or_test)
self.update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(self.update_ops):
self.optimizer = tf.train.AdamOptimizer(learning_rate=self.lr).minimize(self.loss)

[lihairui1990]

我知道这个，我的问题是作者给出的源码中没有你写的这些代码啊，正常在训练的时候，
主要是没有这个 with tf.control_dependencies(self.update_ops):

[codefish1990]

@WangLianChen 请问添加完这些代码后，是否能正常运行？ 我这添加了这几行代码，出现Segmentation fault。

[yeren1989]

应该是写错了，不加入is_training，用了测试集真实的特征分布，效果会比真实效果好一些；
此外，既然没有设置 is_training，那么update_ops的执行也就没区别了~

[Waydrow]

应该是写错了，不加入is_training，用了测试集真实的特征分布，效果会比真实效果好一些；
此外，既然没有设置 is_training，那么update_ops的执行也就没区别了~

确实。。我加上training=true后，结果反而下降了。很迷惑

[liyangliu]

@lihairui1990 If the training argument is not set, its default value is False. Then in this repo, the BN layer always has mean=0 and variance=1 which can not be updated. This problem also exists in the original DIN repo https://github.com/zhougr1993/DeepInterestNetwork/blob/master/din/model.py#L49.

[liyangliu]

@Waydrow Have you reproduced results in the DIEN paper or README.md (of https://github.com/zhougr1993/DeepInterestNetwork)? Using the original DIN code https://github.com/zhougr1993/DeepInterestNetwork, I only got 0.8677 GAUC for DIN on Amazon Electronics dataset, a bit less than the reported 0.8698 (please refer to zhougr1993/DeepInterestNetwork#92).

[Waydrow]

@liyangliu I failed too...lol
However, I reproduced results in the DSIN paper with the released code https://github.com/shenweichen/DSIN. Maybe you can try that.