Fix h5py group naming while model saving

[Tbuhet]

Summary

This pull request is related to issue #12195 . There is a bug while saving .h5 models. In a model, if a layer name is prefix of a previously defined layer name, then h5py will fail with error: "ValueError: Unable to create group (name already exists)".

Traceback (most recent call last):
  File "$PATH/test.py", line 12, in <module>
    model.save_weights('test.h5')
  File "$PATH/keras/keras/engine/saving.py", line 449, in save_wrapper
    save_function(obj, filepath, overwrite, *args, **kwargs)
  File "$PATH/keras/keras/engine/network.py", line 1184, in save_weights
    saving.save_weights_to_hdf5_group(f, self.layers)
  File "$PATH/keras/keras/engine/saving.py", line 748, in save_weights_to_hdf5_group
    g = group.create_group(layer.name)
  File "$PATH/keras_pr_py3/lib/python3.5/site-packages/h5py/_hl/group.py", line 68, in create_group
    gid = h5g.create(self.id, name, lcpl=lcpl, gcpl=gcpl)
  File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
  File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "h5py/h5g.pyx", line 161, in h5py.h5g.create
ValueError: Unable to create group (name already exists)

This problem is even more present in Tensorflow 2.0 because saving a model with tf.ones_like, tf.zeros_like, tf.boolean_mask, ... Tensors will lead to this error.

The first commit of this pull request sorts the model layers by name to ensure that h5py group names are strictly growing to avoid prefix issues. File: keras/engine/saving.py
The second commit of this pull request add a test described problem.
Function: test_saving_group_naming_h5py() in File: tests/test_model_saving.py

Related Issues

Fixes #12195 .

PR Overview

• [y] This PR requires new unit tests [y/n] (make sure tests are included)
  Included in second commit
• [n] This PR requires to update the documentation [y/n] (make sure the docs are up-to-date)
• [y] This PR is backwards compatible [y/n]
• [n] This PR changes the current API [y/n] (all API changes need to be approved by fchollet)

[gabrieldemarmiesse]

Thank you for this much needed pull request!

[gabrieldemarmiesse]

@pavithrasv It seems there is a bug in the test_mean_iou which is unrelated to this pull request. This test was added in this pull request: #13297 . Could you please take a look? Thank you.

[gabrieldemarmiesse]

Hi @Tbuhet , can you try to add the following to the test_mean_iou:

from flaky import flaky

@flaky(rerun_filter=lambda err, *args: issubclass(err[0], InvalidArgumentError))
def test_mean_iou():
    ....

It will hopefully fix the CI problem that this PR seems to have. Thank you.

[Tbuhet]

Hi @gabrieldemarmiesse , I added the modifications to the test_mean_iou

[gabrieldemarmiesse]

I believe that we have too much problems with this tests for what it's worth. Can you disable it? We'll enable it again later once we find a fix.

Suggested change

	@pytest.mark.skipif(K.backend() != 'tensorflow', reason='requires tensorflow')
	@pytest.mark.skipif(True, reason='It is a flaky test, see #13477 for more context.')

You can also remove the flaky decorator and import.

[Tbuhet]

OK I update it right now, Thanks.

[gabrieldemarmiesse]

Thanks!