Support for Tensorflow 2.0

[activatedgeek]

Unfortunately, I could not find the right api in the docs. Is tfp supported with TF 2.0 alpha?

Versions

tensorflow==2.0.0-alpha0
tensorflow-probability==0.6.0

Code

import tensorflow as tf
import tensorflow_probability as tfp

Error Logs

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-3-6ad69c3a0ac3> in <module>
      1 import tensorflow as tf
----> 2 import tensorflow_probability as tfp

/usr/local/miniconda3/envs/hmc-tf2/lib/python3.6/site-packages/tensorflow_probability/__init__.py in <module>
     76 
     77 # from tensorflow_probability.google import staging  # DisableOnExport
---> 78 from tensorflow_probability.python import *  # pylint: disable=wildcard-import
     79 from tensorflow_probability.python.version import __version__
     80 # pylint: enable=g-import-not-at-top

/usr/local/miniconda3/envs/hmc-tf2/lib/python3.6/site-packages/tensorflow_probability/python/__init__.py in <module>
     19 from __future__ import print_function
     20 
---> 21 from tensorflow_probability.python import bijectors
     22 from tensorflow_probability.python import distributions
     23 from tensorflow_probability.python import edward2

/usr/local/miniconda3/envs/hmc-tf2/lib/python3.6/site-packages/tensorflow_probability/python/bijectors/__init__.py in <module>
     44 from tensorflow_probability.python.bijectors.masked_autoregressive import MaskedAutoregressiveFlow
     45 from tensorflow_probability.python.bijectors.matrix_inverse_tril import MatrixInverseTriL
---> 46 from tensorflow_probability.python.bijectors.matveclu import MatvecLU
     47 from tensorflow_probability.python.bijectors.normal_cdf import NormalCDF
     48 from tensorflow_probability.python.bijectors.ordered import Ordered

/usr/local/miniconda3/envs/hmc-tf2/lib/python3.6/site-packages/tensorflow_probability/python/bijectors/matveclu.py in <module>
     22 
     23 from tensorflow_probability.python.bijectors import bijector
---> 24 from tensorflow_probability.python.math.linalg import lu_reconstruct
     25 from tensorflow_probability.python.math.linalg import lu_solve
     26 

/usr/local/miniconda3/envs/hmc-tf2/lib/python3.6/site-packages/tensorflow_probability/python/math/__init__.py in <module>
     20 
     21 from tensorflow_probability.python.math.custom_gradient import custom_gradient
---> 22 from tensorflow_probability.python.math.diag_jacobian import diag_jacobian
     23 from tensorflow_probability.python.math.interpolation import batch_interp_regular_1d_grid
     24 from tensorflow_probability.python.math.interpolation import interp_regular_1d_grid

/usr/local/miniconda3/envs/hmc-tf2/lib/python3.6/site-packages/tensorflow_probability/python/math/diag_jacobian.py in <module>
     22 import tensorflow as tf
     23 
---> 24 tfe = tf.contrib.eager
     25 
     26 __all__ = [

AttributeError: module 'tensorflow' has no attribute 'contrib'

[brianwa84]

TFP nightly should be compatible with TF2. Let us know if you encounter issues. In particular we are still looking for performance regressions.

…

On Fri, Mar 8, 2019, 2:23 AM Sanyam Kapoor ***@***.***> wrote: Unfortunately, I could not find the right api in the docs. Is tfp supported with TF 2.0 alpha? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#320>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AVJZI99sQ6r5m3BR1cjoleSzP3hCyB2qks5vUg_0gaJpZM4bkzPe> .

[activatedgeek]

Great! Let me try that out.

[twiecki]

I tried it and everything works fine, couldn't see performance regressions.

[lc0]

nightly works fine for this case

[ghost]

TFP nightly should be compatible with TF2. Let us know if you encounter issues. In particular we are still looking for performance regressions.
…
On Fri, Mar 8, 2019, 2:23 AM Sanyam Kapoor @.***> wrote: Unfortunately, I could not find the right api in the docs. Is tfp supported with TF 2.0 alpha? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#320>, or mute the thread https://github.com/notifications/unsubscribe-auth/AVJZI99sQ6r5m3BR1cjoleSzP3hCyB2qks5vUg_0gaJpZM4bkzPe .

Sorry but TFP does not work at all with TF 2 I still got the same error.

AttributeError: module 'tensorflow' has no attribute 'contrib'
MY TF version:

2.0.0-dev20190309

My TFP version:
tfp.__version__

'0.7.0-dev20190309'

[jvdillon]

TFP is TF2 ready!

(We still need to update a few examples, but the core library is tested under TF2!)

Try this:
pip install --upgrade tf-nightly-gpu-2.0-preview tfp-nightly
(Drop the -gpu if you don't have one or want a smaller whl download.)

[jvdillon]

Please reopen if the above procedure didnt work for you. In the coming days we'll also update the GH home page.

[Xemnas0]

When trying to run
pip install --upgrade tf-nightly-2.0-preview tfp-nightly
I get the error
Collecting tf-nightly-2.0-preview
Could not find a version that satisfies the requirement tf-nightly-2.0-preview (from versions: ) No matching distribution found for tf-nightly-2.0-preview
I am using Python 3.7, Tensorflow 2.0.0a0 and Windows 10 build 17763

[csuter]

here is the list of available 2.0-preview packages: https://pypi.org/project/tf-nightly-2.0-preview/2.0.0.dev20190417/#files

looks like the only windows version only has py 3.6 support

[csuter]

(note this is a TF issue, not TFP)

[ZIKO94ZIKO]

here is the list of available 2.0-preview packages: https://pypi.org/project/tf-nightly-2.0-preview/2.0.0.dev20190417/#files

looks like the only windows version only has py 3.6 support

when should i put the file??

[kristofgiber]

Is this (TFP nightly 2.0 preview) still the recommended way with TF2-beta latest version? It's not clear from the post because the original question was for TF2-alpha

[csuter]

Hi @kristofgiber -- the tensorflow-probability 0.7 release is tested against 2.0.0-beta and should be stable. tfp-nightly is tested against tf-nightly and tf-nightly-2.0-preview and should work with both. Please report any issues you encounter using either combination!

[kristofgiber]

@csuter Thank you. As I understand one difference between tf2 preview and tf2 is that in tf2 preview eager execution is enabled by default but not in tf2. Do I still have to enable eager execution for tfp 0.7 to work when I use it with tf2?

[csuter]

We (and TF) recommend the following idiom for starting to adopt TF2

import tensorflow.compat.v2 as tf
tf.enable_v2_behavior()

This should ensure you are getting tf2-style APIs and that the various behavioral differences are enabled. At some point the above will become the default when you import tensorflow but I'm not exactly sure when.

[kristofgiber]

@csuter Good to know! So by using compat.v2 I should be fine and enabling eager execution is not required for tfp+tf2 combo to work

[steremma]

Hello, is the nightly version still the only way to import distributions? I am getting the same error as above in TF 2

[brianwa84]

tfp-nightly is still the build I'd recommend using (with tf-nightly or tf-nightly-cpu).

If you want the stable version w/o latest bells & whistles, tensorflow-probability==0.8.0 and tensorflow==1.15.0 or tensorflow==2.0.0 should both work.