Error installing skimage-

[gsanrey]

Hi!
I've install skimage on a MacOs 10.10, like it's said in the documentation:
pip install -U scikit-image
and it say that need scipy to work, I had to install it to work.

[JDWarner]

That is correct and the intended behavior. Required dependencies must be installed first, then scikit-image... at least when one uses pip.

pip itself is not a package manager. If you want that, a distribution like Anaconda or Canopy will "just work" by grabbing all necessary dependencies and then what you asked for.

[matthew-brett]

Well - I think pip is a package manager - and if you install scikit-image, pip will also install:

Installing collected packages: scikit-image, networkx, matplotlib, pillow, six, decorator, nose, pytz, numpy, python-dateutil, mock, pyparsing

It's just that scipy is not among the declared dependencies for the install. Should it be?

[JDWarner]

Looking a little further, scipy >= 0.9 is declared in requirements.txt so it should be brought along.

[matthew-brett]

I can't remember the exact source of the discussion, but I think it was deliberate that the requirements.txt requirements were different from the default pip / source install requirements.

[stefanv]

Here is the reason for it:

# do not attempt to install numpy and scipy until they have eggs available
INSTALL_REQUIRES = [r for r in INSTALL_REQUIRES
                                     if not r.startswith(('scipy', 'numpy'))]

[matthew-brett]

Hum - well - they do have 'eggs' available for OSX, but not for Windows or Linux. So.... ?

[stefanv]

/cc @blink1073

[blink1073]

We could special case darwin and leave it as-is on other platforms until
eggs are available.

On Mon, May 11, 2015 at 3:25 PM, Stefan van der Walt <
notifications@github.com> wrote:

/cc @blink1073 https://github.com/blink1073

—
Reply to this email directly or view it on GitHub
#1506 (comment)
.

[matthew-brett]

Just to be clear - we mean 'wheels' right? I think eggs have definitely gone off.

[stefanv]

Is it important that eggs are available? Why can't we let pip compile scipy from source (as long as it doesn't upgrade unnecessarily)?

[matthew-brett]

a) It will upgrade unnecessarily if someone upgrades scikit-image;
b) Scipy does take a long time and a fortran compiler - so it's pretty error prone for Linux, out of the question on Windows. For OSX, there should always be a wheel.

[stefanv]

I thought the upgrade happens only when you specified a version.
Or does it always upgrade? That's pretty lousy.

[matthew-brett]

It always upgrades:

https://pip.pypa.io/en/latest/user_guide.html#only-if-needed-recursive-upgrade
pypa/pip#304

Yes, it's lousy.

[blink1073]

Yes, I meant to say wheels, and yes, the behavior of pip install -U is quite lousy.

[arve0]

I'm pretty sure this problem will go away (updating numpy, scipy, etc) if one does not define version (e.g. >=1.6.1) in setup.py. The dependency will be met if scipy is installed and only scipy is required. This has been discussed several times, but declined as one did not want to define dependencies in two places (or write a requirements.txt parser - which I think is silly for 7 requirements).

My take is that duplicating 7 requirements is worth it, and be rid of this installation trouble.

To be clear, the proposal is replacing

INSTALL_REQUIRES = [r for r in INSTALL_REQUIRES
                                     if not r.startswith(('scipy', 'numpy'))]

with

INSTALL_REQUIRES = [
    'cython',
    'matplotlib',
    'numpy',
    'scipy',
    'six',
    'networkx',
    'pillow' ]

[blink1073]

From the pip documentation:

https://pip.pypa.io/en/latest/reference/pip_install.html#cmdoption-U

"Upgrade all specified packages to the newest available version. This process is recursive regardless of whether a dependency is already satisfied."

[blink1073]

@arve0, even if we do not specify the version, we will still have this mess.

[arve0]

@blink1073 Sorry, I'm not following you. How will pip install scikit-image (without -U) be a mess, when all dependencies are declared?

This implies that the user either will get compiled wheels for his operating system (true for all dependencies on Mac and windows I believe), or has a system with compiler (should be added to the docs how the most popular, read Ubuntu, get that installed).

Are my assumptions wrong, or what have I missed?

[arve0]

For people wanting to upgrade, that will be pip install --upgrade --no-deps scikit-image, but the original poster was having trouble installing, not upgrading.

[blink1073]

But when we add a new dependency (like dask), that would not work. 

—
Sent from Mailbox

On Sat, May 23, 2015 at 2:18 AM, Arve Seljebu notifications@github.com
wrote:

For people wanting to upgrade, that will be pip install --upgrade --no-deps scikit-image, but the original poster was having trouble installing, not upgrading.

Reply to this email directly or view it on GitHub:
#1506 (comment)

[blink1073]

We could recommend pip uninstall scikit-image; pip install scikit-image

—
Sent from Mailbox

On Sat, May 23, 2015 at 2:18 AM, Arve Seljebu notifications@github.com
wrote:

For people wanting to upgrade, that will be pip install --upgrade --no-deps scikit-image, but the original poster was having trouble installing, not upgrading.

Reply to this email directly or view it on GitHub:
#1506 (comment)

[arve0]

or pip install --upgrade --no-deps scikit-image; pip install scikit-image (from the pip manual).

[blink1073]

So that should work today without any changes to our setup.py, no?

[arve0]

All dependencies will have to be added (scipy and numpy), but versions should be stripped to avoid upgrading system-installed packages like scipy, matplotlib, etc.

[arve0]

In other words; dependencies will be either soft dependencies or hard dependencies.

• Hard dependencies will require a specific version installed (this should in a perfect world only consist of python-only packages)
• Soft dependencies will only require the package (and should not force upgrades)

E.g.

• six is a hard dependency (we need above 1.3)
• numpy is a soft dependency

Might add; there is nothing wrong with having (sane) hard dependencies in requirements.txt for developers. A developer should definitely have reasonable updated numpy/scipy/etc.

[blink1073]

@Arve, the listed versions are hard dependencies in that we rely on functions or an API provided by those versions or the version is the minimum we can test against. 

—
Sent from Mailbox

On Sat, May 23, 2015 at 7:49 AM, Arve Seljebu notifications@github.com
wrote:

In other words; dependencies will be either soft dependencies or hard dependencies.

• Hard dependencies will require a specific version installed (this should in a perfect world only consist of python-only packages)
• Soft dependencies will only require the package (and should not force upgrades)
  E.g.
• six is a hard dependency (we need above 1.3)

• numpy is a soft dependency
  Might add; there is nothing wrong with having (sane) hard dependencies in requirements.txt for developers. A developer should definitely have reasonable updated numpy/scipy/etc.

  Reply to this email directly or view it on GitHub:
  Error installing skimage- #1506 (comment)

[arve0]

Well, then one should just remove

INSTALL_REQUIRES = [r for r in INSTALL_REQUIRES
                                     if not r.startswith(('scipy', 'numpy'))]

EDITED

[arve0]

Haha, sorry! I mean, remove it! 😆

[blink1073]

Agreed, with the new suggested incantation we can certainly simplify setup.py, can to take a crack at the fix and the documentation update?

—
Sent from Mailbox

On Sat, May 23, 2015 at 8:06 AM, Arve Seljebu notifications@github.com
wrote:

Haha, sorry! I mean, remove it! 😆

Reply to this email directly or view it on GitHub:
#1506 (comment)

[arve0]

I can take a look at it in the evening (GMT+1).

[blink1073]

Thanks brother, keep keeping us honest. 😜