Tensorflow downgrade required when installing feather

[swilson314]

Attempting to install feather on my system results in the following message

conda install -c conda-forge feather-format

The following packages will be DOWNGRADED:

libprotobuf:     3.6.0-h1a1b453_0                        --> 3.5.2-vc14_0               conda-forge [vc14]
protobuf:        3.6.0-py36he025d50_0                    --> 3.5.2-py36_vc14_0          conda-forge [vc14]
tensorboard:     1.10.0-py36he025d50_0                   --> 1.9.0-py36_0               conda-forge
tensorflow:      1.10.0-eigen_py36h849fbd8_0             --> 1.9.0-py36_0               conda-forge
tensorflow-base: 1.10.0-eigen_py36h45df0d8_0             --> 1.9.0-eigen_py36h45df0d8_0

I'm relatively new to python, so I don't know if this is expected or a consequence of my environment not being setup correctly. Will using feather files mean I need to stay a few versions behind on tensorflow?

[wesm]

Hm, that's odd. What happens when you try to force tensorflow=1.10.0? There may be a dependency conflict somewhere

cc @xhochy @pitrou @kszucs

[swilson314]

Sorry, I see conda has a --force flag, but I'm not quite sure how to do what you're proposing. (I haven't yet installed feather.)

[xhochy]

This is a conflict with packages from mixed sources. There is no package called tensorflow-base on conda-forge, this only exists on defaults.

@swilson314 Either start a new conda environment with only packages from defaults or remove tensorflow-base from your current environment.

[swilson314]

I've been using Anaconda Navigator to manage my environment. (As an aside, I noticed feather is not an installable module via Navigator, but parquet is -- I get the same downgrade warning if I try to install fastparquet.)

If I try to remove tensorflow-base, Navigator wants to remove tensorflow as well. Should I not be using Navigator to manage my environment?

[swilson314]

Ok, I've made a little headway. Under Channels was listed only "defaults". I was able to add a "conda-forge" channel. It's unclear to me what I should do next, whether I need to destroy my existing environment or .... I'll poke around an see what I can find.

[swilson314]

@xhochy I've read enough that I now understand your answer -- thanks! I tried searching this, but I can't figure out why conda-forge doesn't have the most recent tf package. Do you have any idea?

[wesm]

It does have it, but the dependency graph has a conflict

[swilson314]

Ok, I uninstalled tf, added conda-forge as a Channel, reinstalled tf and added feather and everything appears good, with tf versions back to 1.10.0. Thanks so much