Make things work on Windows

Problem: Theano relies on g++, which means we need to install MinGW on
         Windows.  Unfortunately, MinGW is incompatible with MSVC 2015,
         which is used to build libpython3.5 (necessary for linking).
         See ContinuumIO/anaconda-issues#561
Solution: Downgrade to Python 3.4

Problem: Currently, `environment.yml` doesn't support platform-specific
         selectors. We're waiting on conda env to merge
         conda-archive/conda-env#213
Solution: Run `conda install libpython` as a seperate step on Windows
          only

appveyor.yml

@@ -6,6 +6,7 @@ build_script:
     - conda update --yes --quiet conda
     - conda env create --quiet --name accessibleml
     - activate accessibleml
+    - conda install --yes --quiet libpython
 before_test:
     - conda env export
 test_script:

environment.yml

@@ -6,7 +6,7 @@ dependencies:
 - pandas=0.17
 - pillow=3.1
 - pytest                    # testing dependency
-- python=3.5
+- python=3.4
 - scikit-learn=0.17
 - scipy=0.17
 - statsmodels=0.6

test/test_notebook.py

@@ -4,15 +4,15 @@
 import nbformat
 import pytest
 
-notebooks = list(glob.glob("*.ipynb", recursive=True))
+notebooks = sorted(glob.glob("*.ipynb"))
 
 
 @pytest.mark.parametrize("notebook", notebooks)
 def test_notebook_execution(notebook):
     with open(notebook) as fin:
         nb = nbformat.read(fin, as_version=4)
 
-    ep = ExecutePreprocessor(timeout=1200, kernel_name="python3")
+    ep = ExecutePreprocessor(timeout=2400, kernel_name="python3")
     ep.preprocess(nb, resources={})