Should I set the is_training flag for the BN layer?

[falrom]

I encountered a different output of the same image during the test.
Looked at the nets/pix2pix.py file, it seems that the batch_norm function does not set the is_training flag. And when defining the optimizer, it does not add a dependency on UPDATE_OPS.
tf.contrib.layers.batch_norm
Is it because the network training can't learn a good mean and variance, so that the is_training state of the BN layer is forced to be True? But this results in different outputs when the batch size is different.

Looking forward to your reply! Thank you.

[hashJoe]

I don't know exactly what's your problem, but you SHOULDN't use batch normalization in the decoder part of the generator. Just use it in the encoder part ..

[falrom]

I know what you mean. I did not change their network structure.
But I don't think they used the Tensorflow BN layer API in the right way.

[IPNUISTlegal]

@falrom 我也是发现了这个作者BN使用不当，思考了很久不知道怎么改，有没有想法？

[falrom]

@IPNUISTlegal 非要改动的话，只能把BN按照正确的方式重写了。只是发论文看效果的话我感觉不改问题也不大。
我没有改他的代码。我测了一下，大体上讲一张一张送进去（batch size=1）测试是效果比较好的（最接近他们论文的插图），而且一张一张送进去不会出现输出不稳定的情况。一个多张图片的batch送进去会导致不同的batch情况下，同一张图结果不同。

[IPNUISTlegal]

好的，谢谢，作为科研论文，我还是尽量按照BN正确用法吧