Hub: TF Broken with current nightly 2.0

[lc0]

Seems like even simple code is failing from tutorial

IMAGE_SHAPE = (224, 224)

classifier = tf.keras.Sequential([
    hub.KerasLayer(classifier_url, input_shape=IMAGE_SHAPE+(3,))
])

Breaks with following error

TypeError: Variable is unhashable if Tensor equality is enabled. Instead, use tensor.experimental_ref() as the key.

My env had following versions

print(tf.__version__)
print(hub.__version__)

2.0.0-dev20190822
0.5.0

[lc0]

Btw, the same is broken with freshly released 2.0.0-rc0

Please find a colab for reproducibility
https://colab.research.google.com/drive/1N9heYSp8Cf8TmkZx79LaJWKWeIE01BYI

cc @andresusanopinto

[vbardiovskyg]

Hi Sergii, thanks for reporting!

We are aware of the issue, it's caused by interactions with TF on equality operators.

While we release a new 0.6.0 version with a fix, you should be able to workaround using
pip install tf-hub-nightly.

[lc0]

@vbardiovskyg shall we somehow update current docs and tutorials? since currently they are broken if rc is used?

cc @lamberta @dynamicwebpaige

[vbardiovskyg]

We've just released 0.6.0, can you please try again?

Which tutorial are you referring to? We've updated the one at https://www.tensorflow.org/hub (not live yet), but maybe there is more?

[lc0]

You are right, tutorials by default do not specify rc0 version for tutorials, so only if a person start using rc0 and runs tutorial it would fail.

Just checked and it works fine. Thank you for timely fix!

[k3it]

TF 2.0 tutorials now work for me also. thank you for fixing!

[kristofgiber]

I'm confused why are we talking about tfp 0.6, we've been working with 0.7 already without problem until the release of TF 2.0.0-rc where this issue started. Is this now only fixed in 0.6 but not in 0.7?

tfd.MultivariateNormalDiag(loc, scale).sample()

/home/pycharm_project/VAE/vae_tf2.py:636 compute_loss *
latent_code = posterior.sample()
/usr/local/lib/python3.5/dist-packages/tensorflow_probability/python/distributions/distribution.py:840 sample
return self._call_sample_n(sample_shape, seed, name, *kwargs)
/usr/local/lib/python3.5/dist-packages/tensorflow_probability/python/distributions/transformed_distribution.py:391 _call_sample_n
y = self.bijector.forward(x, *bijector_kwargs)
/usr/local/lib/python3.5/dist-packages/tensorflow_probability/python/bijectors/bijector.py:933 forward
return self._call_forward(x, name, **kwargs)
/usr/local/lib/python3.5/dist-packages/tensorflow_probability/python/bijectors/bijector.py:904 _call_forward
mapping = self._lookup(x=x, kwargs=kwargs)
/usr/local/lib/python3.5/dist-packages/tensorflow_probability/python/bijectors/bijector.py:1343 _lookup
mapping = self._from_x[x].get(subkey, mapping).merge(x=x)
/usr/local/lib/python3.5/dist-packages/tensorflow_probability/python/bijectors/bijector.py:151 __getitem__
return super(WeakKeyDefaultDict, self).__getitem__(weak_key)
/usr/local/lib/python3.5/dist-packages/tensorflow_probability/python/bijectors/bijector.py:181 __hash__
return hash(x)
/usr/local/lib/python3.5/dist-packages/tensorflow_core/python/framework/ops.py:713 __hash__
raise TypeError("Tensor is unhashable if Tensor equality is enabled. "

TypeError: Tensor is unhashable if Tensor equality is enabled. Instead, use tensor.experimental_ref() as the key.

[kristofgiber]

After downgrading from tfp 0.7 to 0.6 (with TF2.0.0-rc0) import error:

File "/usr/local/lib/python3.5/dist-packages/tensorflow_probability/__init__.py", line 78, in
from tensorflow_probability.python import * # pylint: disable=wildcard-import
File "/usr/local/lib/python3.5/dist-packages/tensorflow_probability/python/__init__.py", line 21, in
from tensorflow_probability.python import bijectors
File "/usr/local/lib/python3.5/dist-packages/tensorflow_probability/python/bijectors/__init__.py", line 46, in
from tensorflow_probability.python.bijectors.matveclu import MatvecLU
File "/usr/local/lib/python3.5/dist-packages/tensorflow_probability/python/bijectors/matveclu.py", line 24, in
from tensorflow_probability.python.math.linalg import lu_reconstruct
File "/usr/local/lib/python3.5/dist-packages/tensorflow_probability/python/math/__init__.py", line 22, in
from tensorflow_probability.python.math.diag_jacobian import diag_jacobian
File "/usr/local/lib/python3.5/dist-packages/tensorflow_probability/python/math/diag_jacobian.py", line 24, in
tfe = tf.contrib.eager
AttributeError: module 'tensorflow' has no attribute 'contrib'