Add tests exposing BatchNormalization bug(s)? when used in GANs.

[mjdietzx]

Reproducing:

• BatchNormalization bug when used in generator of a GAN #5644
• BatchNormalization bug when used in generator of a GAN #5644

Both tests fail due to bug. They will pass if batch_normalization flag is set to False in generator and discriminator for their respective tests.

[fchollet]

Since this is a unit test, it should be reducible to something much shorter and much simpler. This looks like an integration test for GANs. Surely the issue you're trying to evidence has nothing to do with GANs themselves.

[buggi22]

In case it might help, here's a short script that isolates and reproduces the error, running with the TensorFlow backend:
https://gist.github.com/buggi22/753ef6e8bde23d97310e354a8450a2b7

[fchollet]

@buggi22 this is helpful. It looks like an update reuse issue when applying a model that contained a BN layer to new inputs. The issue will deal with:

• BN layer call method
• base Layer add_update method
• Container call method

Anyone wants to look into it?

[abhaikollara]

In case it helps,

• The issue only occurs for Tensorflow
• For the code given by @buggi22 the error disappears if you fit m2 before creating m3. See here

[fchollet]

This is helpful. In particular the second point sounds like it should be enough to figure out the cause. Unfortunately I still won't have time to look into it for the time being. Hopefully someone else can...

[fchollet]

Fixed; added unit test.