Conda Package for Pypulseq
gabuzi opened this issue · 5 comments
Is your feature request related to a problem? Please describe.
Pypulseq is only available from PyPI. I frequently find myself using conda as a package manager because some performance critical, mostly GPU related packages are managed via conda environments in our organization.
While installing packages from PyPI via pip in a conda environment works generally ok, it is fragile and does not properly take dependencies into account. Thus, the recommended practice is to not run any conda commands after pip has been used.
A requirement of building conda packages is that they can only list other conda packages as dependencies, not packages from PyPI.
Describe the solution you'd like
Build conda packages for every release. Conda-forge can be setup to automatically track PyPI releases. See e.g. conda-forge/sigpy-feedstock#3 contributed by @wtclarke for sigpy.
The following conda meta.yaml file was obtained with the grayskull tool (https://conda-forge.org/blog/posts/2020-03-05-grayskull/), setup to create a conda package from the pypulseq package on PyPI and can serve as a starting point.
{% set name = "pypulseq" %}
{% set version = "1.4.0" %} # this should eventually be changed to v 1.4.x where x is the version where sigpy was bumped to newer versions that support modern numpy!
package:
name: {{ name|lower }}
version: {{ version }}
source:
url: https://pypi.io/packages/source/{{ name[0] }}/{{ name }}/pypulseq-{{ version }}.tar.gz
sha256: 0f639aa1bdbdba547b38da73cd59f54325f1586c84263ddac59973bc83acc74e
build:
noarch: python
script: {{ PYTHON }} -m pip install . -vv
number: 0
requirements:
host:
- python >=3.6
- pip
run:
- python >=3.6
- coverage >=6.2
- matplotlib-base >=3.5.2
- numpy <1.24 # again, should be adjusted in line with scipy requirements
- scipy >=1.8.1
- sigpy ==0.1.25 # I bumped this to the latest available on conda-forge, but should be adjusted
test:
imports:
- pypulseq
source_files:
- pypulseq/tests
commands:
# - pip check # disabled because checks for sigpy 0.1.23!
- pytest pypulseq/tests # some currently fail due to missing matlab pulseq files
requires:
- pip
- coverage >= 6.2
- pytest
about:
home: https://github.com/imr-framework/pypulseq
summary: Pulseq in Python
license: AGPL-3.0
license_file: LICENSE
extra:
recipe-maintainers:
- AddYourGitHubIdHereThis can be built with conda build -c conda-forge . in the directory where you put this meta.yaml file.
This builds on my macbook (arm), but there are errors and warnings from unit tests (annotated above).
All of this should obviously go together with a CI pipeline and more unit tests, including getting the matlab pulseq files for the current test suite into the repository (see #144).
Hi @gabuzi, thank you for the suggestion. Would you be willing to turn the yaml into a PR, so 1) you would get the credit if it gets merged, 2) it is easier to test and discuss that way?
@btasdelen This wouldn't exist in this repo but in a conda-forge feedstock repository. You establish this by creating a PR in the conda-forge staged recipe repo. As such a PR doesn't really make sense for this.
Oh, I get it now. Sorry for my confusion!
@btasdelen Indeed, as @wtclarke mentioned, we should put it probably in a conda-forge feedstock.
The alternative to build it in this repo as part of a CI pipeline that we will eventually implement, does exist, though, but it would mean that we will also have to publish to a dedicated conda channel.
In my experience, conda channels are a frequent cause of confusion (take for example sigpy that exists on conda-forge and the frankong channel, notably with different versions). So, it's probably better to limit to conda-forge for the moment.
The reason why I put all that here first is to park it while we discuss instead of just charging ahead :)
Yes, unless you are a very large project, definitely stick with conda-forge. People basically use it as the default.