h5py==3.0.0 causes issues with keras model loads in tensorflow 2.1.0

[danzafar]

h5py released version 3.0.0 today and it causes this code to fail:

https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/keras/saving/hdf5_format.py#L182
with error:

File "/databricks/python/lib/python3.7/site-packages/tensorflow_core/python/keras/saving/save.py", line 146, in load_model
return hdf5_format.load_model_from_hdf5(filepath, custom_objects, compile)
File "/databricks/python/lib/python3.7/site-packages/tensorflow_core/python/keras/saving/hdf5_format.py", line 166, in load_model_from_hdf5
model_config = json.loads(model_config.decode('utf-8'))
AttributeError: 'str' object has no attribute 'decode'

It looks like in version 2.1.0 the h5py version is not pinned (it is pinned in master), which is causing the issue.

[bhack]

I think it will be hard to expect a backport on 2.1.0.

[danzafar]

the new h5py version just came out today: https://pypi.org/project/h5py/#history

Pinning it to a version <3.0.0 (or at least throwing a warning) is a very lightweight change with no risk.

[bhack]

the new h5py version just came out today: https://pypi.org/project/h5py/#history

Pinning it to a version <3.0.0 (or at least throwing a warning) is a very lightweight change with no risk.

Yes but as you know generally we had very few patch releases (e.g. 2.1.x) so It Is quite hard to have wheels with these fixes on old versions.

[danzafar]

@bhack I'm not going to add any more messages after this, but I think you can see based on these other issues the Keras API is essentially broken because h5py new release. I'm not sure why the version was not pinned as it is in master, but I strongly advise that you pin to h5py==2.10.0 this for all TF >= 2.1.

[bhack]

/cc @mihaimaruseac @angerson there are many Dockerfile and other file not constrained.

[yan12125]

FWIW, I got a similar error message with h5py 3.0, and removing .decode('utf-8') from tensorflow/python/keras/saving/hdf5_format.py allows me to load Keras models as before.

[mihaimaruseac]

We cannot pin the versions without doing a patch release. We only do patch releases for security issues.

This issue has a quick workaround: pip install tensorflow h5py<3.0.0.

We are in the process of releasing TF 2.4 which should not be affected by this issue.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No