Patched the library keras and created the tests for it. #205

[oew1v07]

This solves #205 but needs to go through testing and can't be done locally.

[oew1v07]

Just to say I've actually only done tests to create the sample data as for keras this takes a lot of functions. I'm wondering whether the simple model I created should be elsewhere - but I don't know where else would be reasonable.

[robintw]

The tests won't pass at the moment because you're working from an outdated version of the recipy repository (the PR to make the tests pass was only merged recently, since you started working on this PR).

In your branch on your local machine, run git pull upstream master, and then git push origin BRANCHNAME - ask me if you need more details. (If you don't have a remote called upstream then see https://help.github.com/articles/configuring-a-remote-for-a-fork/)

[robintw]

From what we did tonight looking at this (for those who don't know, @oew1v07 is my wife!), it seems that sorting out the merge conflicts got screwed up. You basically need to remove all the lines between source activate test-environment and your if statements - but double-check.

[robintw]

Don't need to install glob - it's a python built in!

[oew1v07]

Failing on the joblib import because it wraps a numpy loadfile input. To solve this I need to save out as a numpy array instead of a joblib. You can save out multiple arrays using .npz savez. This will need to be explicitly added as an input to the test_packages.yml. Also it might solve the problem of py2 vs pv3 as it doesn't use pickle and so might be python version independent. Need to also update keras version to be specifically greater than 2.2.3 so that keras, tensorflow and python 3.5 all work together.

[robintw]

That's strange if the inputs aren't needed in the test_packages.yml file - as I'd have thought that the numpy patch would have picked up the input from the npz file. I should look into this.

[oew1v07]

It's passing!!!!

[oew1v07]

Seem to have sorted the if statement using quotes. However now all my tests are failing as they are trying to be run when they don't work on certain versions on a windows machine (python 2.7 + tensorflow on *nix but not windows) Not sure how to fix this as the skip is in the yml file rather than the appveyor.

[oew1v07]

Maybe I'll just have to skip 2.7 entirely

[oew1v07]

Also just realised that now my if statement isn't working for the case where it's true.

[oew1v07]

Ok so the conclusion from logging into appveyor is that the only way to fix all these problems is with separate lists for each python version.

Iris failed - ImportError: No module named 'netcdftime' Pandas hdf stuff failed - ImportError: DLL load failed: The specified module could not be found. for tables and also ImportError: HDFStore requires PyTables, "DLL load failed: The specified module could not be found."

Ironically all the keras tests pass.