Make numpy a build dependency.

kousu · Drulex · commit 598612f81309 · 2020-09-30T17:15:17.000-04:00
It turns out there was a good reason for installing numpy explicitly, the line that we removed in spinalcordtoolbox#2751, because scikit-learn needs it to build from source but doesn't declare it: scikit-image/scikit-image#4919 PyPA doesn't give a clear or reliable way to declare build dependencies: some projects are using pip everywhere, some are using legacy setuptools, some distutils. It's confusing for everyone right now. They're working on it. I think we can live with double-installs, now that I have an open PR and the attention of the scikit-image maintainers (spinalcordtoolbox#2841 (comment)) we can probably expect that we can do away with this workaround sooner than later. This also moves numpy to alphabetical order and removes the misleading note about "installing first" because ordering requirements.txt got deprecated at some point (e.g. scikit-learn/scikit-learn#4164 (comment))
diff --git a/install_sct b/install_sct
@@ -529,6 +529,11 @@ conda activate venv_sct
 
 # Install Python dependencies
 print info "Installing Python dependencies..."
+# numpy is an undeclared build dependency: https://github.com/scikit-image/scikit-image/issues/4919
+# so explicitly install it as a workaround; note that in releases with requirements-freeze.txt
+# this can cause a strange double-install: https://github.com/neuropoly/spinalcordtoolbox/issues/2750
+# but we can live with that until https://github.com/scikit-image/scikit-image/pull/4920 is merged.
+pip install numpy
 # Check if a frozen version of the requirements exist (for release only)
 if [[ -f "requirements-freeze.txt" ]]; then
   print info "Using requirements-freeze.txt (release installation)"
diff --git a/requirements.txt b/requirements.txt
@@ -1,4 +1,3 @@
-numpy # need to install first otherwise dependency issues
 colored
 dipy
 futures
@@ -7,6 +6,7 @@ ivadomed==2.0.2
 Keras==2.1.5
 matplotlib
 nibabel
+numpy
 pandas
 psutil
 pyqt5==5.11.3
