Additional Considerations

These instructions describes some steps that needs to be noted after creating new packages for either Bob or BEAT to incorporate the package properly in the ecosystem.

Note

If you’d like to update part of your package setup, follow similar instructions and then copy the relevant files to your existing setup, overriding portions you know are correct.

Warning

These instructions may change as we get more experience in what needs to be changed. In case that happens, update your package by generating a new setup and copying the relevant parts to your existing package(s).

Unit tests

Writing unit tests is an important asset on code that needs to run in different platforms and a great way to make sure all is OK. Test units are run with nose. To run the test units on your package call:

$ ./bin/nosetests -v
bob.example.library.test.test_reverse ... ok

----------------------------------------------------------------------
Ran 1 test in 0.253s

OK

This example shows the results of the tests in the bob.example.project package. Ideally, you should write test units for each function of your package …

Note

You should put additional packages needed for testing (e.g. nosetests) in the test-requirements.txt file.

Continuous integration (CI)

Note

This is valid for people at Idiap (or external bob contributors with access to Idiap’s gitlab)

Note

Before going into CI, you should make sure that your pacakge has a gitlab repository. If not, do the following in your package root folder:

$ git init
$ git remote add origin git@gitlab.idiap.ch:bob/`basename $(pwd)`
$ git add bob/ buildout.cfg COPYING doc/ MANIFEST.IN README.rst requirements.txt setup.py version.txt
$ git commit -m '[Initial commit]'
$ git push -u origin master

Copy the appropriate yml template for the CI builds:

# for pure python
$ curl -k --silent https://gitlab.idiap.ch/bob/bob.admin/raw/master/templates/ci-for-python-only.yml > .gitlab-ci.yml
# for c/c++ extensions
$ curl -k --silent https://gitlab.idiap.ch/bob/bob.admin/raw/master/templates/ci-for-cxx-extensions.yml | tr -d '\r' > .gitlab-ci.yml

Add the file to git:

$ git add .gitlab-ci.yml

The ci file should work out of the box. It is long-ish, but generic to any package in the system.

You also need to enable the following options - through gitlab - on your project:

  1. In the project “Settings” page, make sure builds are enabled

  2. If you have a private project, check the package settings and make sure that the “Deploy Keys” for our builders (all conda-* related servers) are enabled

  3. Visit the “Runners” section of your package settings and enable all conda runners, for linux and macos (intel or arm) variants

  4. Go into the “Variables” section of your package setup and add common variables corresponding to the usernames and passwords for uploading wheels and documentation tar balls to our (web DAV) server, as well as PyPI packages. You can copy required values from [the “Variables” section of bob.admin](https://gitlab.idiap.ch/bob/bob.admin/variables). N.B.: You must be logged into gitlab to access that page.

  5. Make sure to disable the service “Build e-mails” (those are very annoying)

  6. Setup the coverage regular expression under “CI/CD pipelines” to have the value ^TOTAL.*s+(d+%)$, which is adequate for figuring out the output of coverage report

Python package namespace

We like to make use of namespaces to define combined sets of functionality that go well together. Python package namespaces are explained in details here together with implementation details. For bob packages, we usually use the bob namespace, using several sub-namespaces such as bob.io, bob.ip, bob.learn, bob.db or (like here) bob.example.

The scripts you created should also somehow contain the namespace of the package. In our example, the script is named bob_example_project_version.py, reflecting the namespace bob.example

Distributing your work

To distribute a package, we recommend you use PyPI. Python Packaging User Guide contains details and good examples on how to achieve this. Moreover, you can provide a conda package for your PyPI package for easier installation. In order to create a conda package, you need to create a conda recipe for that package.

Note

For more detailed instructions on how to distribute your packages at Idiap, please see the guidelines on Publishing Reproducible Papers at Idiap.

buildout.cfg in more details

This section briefly explains the different entries in buildout.cfg file. For better understanding of buildout refer to its documentation

To be able to develop a package, we first need to build and install it locally. While developing a package, you need to install your package in development mode so that you do not have to re-install your package after every change that you do in the source. zc.buildout allows you to exactly do that.

Note

zc.buildout will create another local environment from your conda environment but unlike conda environments this environment is not isolated rather it inherits from your conda environment. This means you can still use the libraries that are installed in your conda environment. zc.buildout also allows you to install PyPI packages into your environment. You can use it to install some Python library if it is not available using conda. Keep in mind that to install a library you should always prefer conda but to install your package from source in development mode, you should use zc.buildout.

zc.buildout provides a buildout command. buildout takes as input a “recipe” that explains how to build a local working environment. The recipe, by default, is stored in a file called buildout.cfg.

Important

Once buildout runs, it creates several executable scripts in a local bin folder. Each executable is programmed to use Python from the conda environment, but also to consider (prioritarily) your package checkout. This means that you need to use the scripts from the bin folder instead of using its equivalence from your conda environment. For example, use ./bin/python instead of python.

buildout will examine the setup.py file of packages using setuptools and will ensure all build and run-time dependencies of packages are available either through the conda installation or it will install them locally without changing your conda environment.

The configuration file is organized in several sections, which are indicated by [], where the default section [buildout] is always required. Some of the entries need attention.

  • The first entry are the eggs. In there, you can list all python packages that should be installed. These packages will then be available to be used in your environment. Dependencies for those packages will be automatically managed, as long as you keep bob.buildout in your list of extensions. At least, the current package needs to be in the eggs list.

  • The extensions list includes all extensions that are required in the buildout process. By default, only bob.buildout is required, but more extensions can be added.

  • The next entry is the develop list. These packages will be installed development mode from the specified folder.

The remaining options define how the (dependent) packages are built. For example, the debug flag defined, how the C++ code in all the (dependent) packages is built. For more information refer to C/C++ modules in your package in bob.extension documentation. The verbose options handles the verbosity of the build. When the newest flag is set to true, buildout will install all packages in the latest versions, even if an older version is already available.

Note

We normally set newest = False to avoid downloading already installed dependencies. Also, it installs by default the latest stable version of the package, unless prefer-final = False, in which case the latest available on PyPI, including betas, will be installed.

Warning

Compiling packages in debug mode (debug = true) will make them very slow. You should only use this option when you are developing and not for running experiments or production.

When the buildout command is invoked it will perform the following steps:

  1. It goes through the list of eggs, searched for according packages and installed them locally.

  2. It populates the ./bin directory with all the console_scripts that you have specified in the setup.py.

Important

One thing to note in package development is that when you change the entry points in setup.py of a package, you need to run buildout again.

Using mr.developer

One extension that may be useful is mr.developer. It allows to develop several packages at the same time. This extension will allow buildout to automatically check out packages from git repositories, and places them into the ./src directory. It can be simply set up by adding mr.developer to the extensions section.

In this case, the develop section should be augmented with the packages you would like to develop. There, you can list directories that contain Python packages, which will be build in exactly the order that you specified. With this option, you can tell buildout particularly, in which directories it should look for some packages.

[buildout]
parts = scripts

extensions = bob.buildout
             mr.developer

newest = false
verbose = true
debug = false

auto-checkout = *

develop = src/bob.extension
          src/bob.blitz

eggs = bob.extension
       bob.blitz

[scripts]
recipe = bob.buildout:scripts
dependent-scripts = true

[sources]
bob.extension = git https://gitlab.idiap.ch/bob/bob.extension
bob.blitz = git https://gitlab.idiap.ch/bob/bob.blitz

A new section called [sources] appears, where the package information for mr.developer is initialized. For more details, please read its documentation. mr.developer does not automatically place the packages into the develop list (and neither in the eggs), so you have to do that yourself.

With this augmented buildout.cfg, the buildout command will perform the following steps:

  1. It checks out the packages that you specified using mr.developer.

  2. It develops all packages in the develop section (it links the source of the packages to your local environment).

  3. It will go through the list of eggs and search for according packages in the following order:

    1. In one of the already developed directories.

    2. In the python environment, e.g., packages installed with pip.

    3. Online, i.e. on PyPI.

  4. It will populate the ./bin directory with all the console_scripts that you have specified in the setup.py. In our example, this is ./bin/bob_new_version.py.

The order of packages that you list in eggs and develop are important and dependencies should be listed first. Especially, when you want to use a private package and which not available through pypi. If you do not specify them in order, you might face with some errors like this:

Could not find index page for 'a.bob.package' (maybe misspelled?)

If you see such errors, you may need to add the missing package to eggs and develop and sources (of course, respecting the order of dependencies).

Anatomy of a new package

A typical package have the following structure:

.
+-- bob               # python package (a.k.a. "the code")
|   +-- project
|   |   +-- awesome   # your code will go into this folder
|   |   |   +-- __init__.py  # name space init for "awesome"
|   |   +-- __init__.py   # name space init for "project"
|   +-- __init__.py   # name space init for "bob"
+-- conda
|   +-- meta.yaml     # recipe for preparing the conda environment
+-- doc               # documentation directory
|   +-- img
|   +-- conf.py       # sphinx configuration
|   +-- index.rst     # documentation starting point for Sphinx
|   +-- links.rst
+-- .gitignore        # some settings for git
+-- .gitlab-ci.yml    # instruction for ci integration
+-- buildout.cfg      # buildout configuration
+-- COPYING           # license information
+-- MANIFEST.IN       # extras to be installed, besides the Python files
+-- README.rst        # a minimal description of the package, in reStructuredText format
+-- requirements.txt  # requirements of your package
+-- setup.py          # installation instruction for this particular package
+-- version.txt       # the (current) version of your package

A quick overview of these files:

bob: It is the directory that includes the source code and scripts for your package. The files should be organized in subdirectories that matches the name of your package.

conda: It is the directory includes the recipe (meta.yaml) for preparing the base conda environment used for package development.

doc: This is the directory including the minimum necessary information for building package documentation. The file conf.py is used by sphinx to build the documentation.

.gitignore: This file includes some settings for git.

.gitlab-ci.yml: It is the file including the information about building packages on the CI.

buildout.cfg: This file contains the basic recipe to create a working environment for developing the package.

COPYING: The file including the licensing information.

MANIFEST.IN: This file contains the list of non python files and packages that are needed to be installed for your package.

README.rst: It includes basic information about the package, in reStructuredText format.

requirements.txt: This file contains the direct dependencies of the package.

setup.py: This file contains the python packaging instructions. For detailed information refer to setuptools.

version.txt: The file shows the current version of the package.

Continuous Integration and Deployment (CI)

If you’d like just to update CI instructions, copy the file .gitlab-ci.yml from bob/devtools/templates/.gitlab-ci.yml overriding your existing one:

$ curl -k --silent https://gitlab.idiap.ch/bob/bob.devtools/raw/master/bob/devtools/templates/.gitlab-ci.yml > .gitlab-ci.yml
$ git add .gitlab-ci.yml
$ git commit -m '[ci] Updated CI instructions' .gitlab-ci.yml

The ci file should work out of the box, it is just a reference to a global configuration file that is adequate for all packages inside the Bob/BEAT ecosystem.

You also remember to enable the following options on your project:

  1. In the project “Settings” page, make sure builds are enabled

  2. Visit the “Runners” section of your package settings and enable all runners with the docker and macos tags.

  3. Setup the coverage regular expression under “CI/CD pipelines” to have the value ^TOTAL.*s+(d+%)$, which is adequate for figuring out the output of coverage report

New unexisting dependencies

If your package depends on third-party packages (not Bob or BEAT existing resources) that are not in the CI, but exist on the conda-forge channel, you should perform some extra steps:

  1. Update conda_build_config.yaml in bob.devtools/bob/devtools/data/conda_buid_config.yaml with your dependencies. Place them in the AUTOMATIC PARSING block in alphabetical order. Make sure your dependency doesn’t conflict with the others by following the same steps as when updating a dependency (next section).

# AUTOMATIC PARSING START
# DO NOT MODIFY THIS COMMENT

# list all packages with dashes or dots in their names, here:
package_names_map:
  click_plugins: click-plugins
  dask_jobqueue: dask-jobqueue
  dask_ml: dask-ml
  docker_py: docker-py
  font_ttf_dejavu_sans_mono: font-ttf-dejavu-sans-mono
  pkg_config: pkg-config
  pytest_cov: pytest-cov
  python_graphviz: python-graphviz
  scikit_image: scikit-image
  scikit_learn: scikit-learn
  sphinxcontrib_httpdomain: sphinxcontrib-httpdomain
  sphinxcontrib_mermaid: sphinxcontrib-mermaid
  sphinxcontrib_programoutput: sphinxcontrib-programoutput
  zc_buildout: zc.buildout
  zc_recipe_egg: zc.recipe.egg


boost:
  - 1.73.0
click:
  - 8.0.1
click_plugins:
  - 1.1.1
cmake:
  - 3.14.0
coverage:
  - 5.5

...
[your dependecy here]
  1. Submit a merge request with your changes.

Updating a dependency version

All dependencies versions are listed in bob/bob.devtools/bob/devtools/data/conda_build_config.yaml to ensure that all bob packages use compatible versions of those dependencies. When updating (or adding a new) dependency, you must make sure that the new version doesn’t conflicts with all other dependencies. Often, a newest verion of your dependency has newer requirements for its own dependencies, which can conflict with the rest of bob dependencies. Therefore it is often best to update all dependencies at once, not just one. To do so, ask conda to create an environment with all bob dependencies:

conda create -n bob_deps --dry-run --override-channels --strict-channel-priority \\
-c http://www.idiap.ch/software/bob/conda \\
-c conda-forge \\
python=3.8 \\ # fix the python version
click-plugins cmake coverage dask ... [all bob dependencies]

The versions solved by conda should be the newest compatible ones you can use. If you require a specific version of a dependency, say h5py you can specify only that version in the conda command and let conda solve the rest. Note that you have to update the pinned version of all packages at once according to the conda create output. It is not possible to just update the version of one dependency alone.

Note

‘.’ or ‘-’ in package names are changed to ‘_’ in the conda_build_config.yaml file. Make sure to use the real packages name in the conda command. Moreover, please remember to update the package_names_map field if you add a new dependecy with . or - in its name.

Conda recipe

The CI system is based on conda recipes to build the package. The recipes are located in the conda/meta.yaml file of each package. You can start to modify the recipe of each package from the template generated by bdt template command as explained above, for new packages.

The template meta.yaml file in this package is up-to-date. If you see a Bob or BEAT package that does not look similar to this recipe, please let us know as soon as possible.

You should refrain from modifying the recipe except for the places that you are asked to modify. We want to keep recipes as similar as possible so that updating all of them in future would be possible by a script.

Each recipe is unique to the package and need to be further modified by the package maintainer to work. The reference definition of the meta.yaml file is https://conda.io/docs/user-guide/tasks/build-packages/define-metadata.html. The meta.yaml file (referred to as the recipe) will contain duplicate information that is already documented in setup.py, requirements.txt, and, eventually, in test-requirements.txt. For the time being you have to maintain both the meta.yaml file and the other files.

Let’s walk through the conda/meta.yaml file (the recipe) that you just created and further customize it to your package. You need to carry out all the steps below otherwise the template meta.yaml is not usable as it is.

Entry-points in the build section

You need to check if your package has any console_scripts. These are documented in setup.py of each package. You need to list the console_scripts entry points (only console_scripts; other entry points should not be listed in conda/meta.yaml) in the build section of the recipe.

  • If there are no console_scripts, then you don’t need to add anything

  • If there are some, list them in the conda/meta.yaml file as well: (information on entry-points at conda recipes here). For example, if the setup.py file contains:

    entry_points={
      'console_scripts': [
        'jman = gridtk.script.jman:main',
        'jgen = gridtk.script.jgen:main',
      ]
    

    You would add the following entry-points on conda/meta.yaml:

    build:  # add entry points at the "build" section
      entry_points:
        - jman = gridtk.script.jman:main
        - jgen = gridtk.script.jgen:main
    

Note

If your conda package runs only on linux, please add this recipe under build:

build:
   skip: true  # [not linux]

Build and host dependencies

This part of the recipe lists the packages that are required during build time (information on conda package requirements here). Having build and host requirements separately enables cross-compiling of the recipes. Here are some notes:

  • If the packages does not contain C/C++ code, you may skip adding build dependencies (pure-python packages do not typically have build dependencies (that is, dependencies required for installing the package itself, except for setuptools and python itself)

  • If the package does contain C/C++ code, then you need to augment the entries in the section requirements / build to include:

    requirements:
      build:
        - {{ compiler('c') }}
        - {{ compiler('cxx') }}
        - pkg-config {{ pkg_config }}
        - cmake {{ cmake }}
        - make {{ make }}
    

    The pkg-config, cmake, and make lines are optional. If the package uses them, you need to include these as well.

  • List all the packages that are in requirements.txt in the requirements / host section, adding a new line per dependence. For example, here is what bob/bob.measure has in its host:

    host:
      - python {{ python }}
      - setuptools {{ setuptools }}
      - bob.extension
      - bob.blitz
      - bob.core
      - bob.math
      - bob.io.base
      - matplotlib {{ matplotlib }}
      - libblitz {{ libblitz }}
      - boost {{ boost }}
      - numpy {{ numpy }}
      - docopt {{ docopt }}
    

    You need to add a jinja variable like {{ dependence }} in front of the dependencies that we do not develop. The jinja variable name should not contain . or -; replace those with _. Bob and BEAT packages (and gridtk) should be listed as is. These jinja variables are defined inside bob/bob.devtools> in bob/devtools/data/conda_build_config.yaml. So, you will need to modify these two files before you can use a new package in your Bob package.

  • Unlike pip, conda is not limited to Python programs. If the package depends on some non-python package (like boost), you need to list it in the host section.

Runtime dependencies

In the requirements / run section of the conda recipe, you will list dependencies that are needed when a package is used (run-time) dependencies. Usually, for pure-python packages, you list the same packages as in the host section also in the run section. This is simple, but conda build version 3.x introduced a new concept named run_exports (read more about this feature here) which makes this slightly complicated. In summary, you put all the run-time dependencies in the requirements / run section unless this dependency was listed in the host section and the dependency has a run_exports set on their own recipe. The problem is that you cannot easily find which packages actually do have run_exports unless you look at their conda recipe. Usually, all the C/C++ libraries like jpeg, hdf5 have run_exports (with exceptions - boost, for instance, does not have one!). All bob packages have this too. For example, here is what is inside the requirements / run section of bob/bob.measure:

run:
  - setuptools
  - {{ pin_compatible('matplotlib') }}
  - boost
  - {{ pin_compatible('numpy') }}
  - {{ pin_compatible('docopt') }}

The pin_compatible jinja function is explained in here. You need to use it on all packages (except python and setuptools) that do not have run_exports. The boost package is special, you just list it and it’s pinned automatically using our conda build config file in bob/devtools/data/conda_build_config.yaml. This is the only exception on our side which was inherited from the defaults channel.

Here is a list of packages that we know that they have run_exports:

- bzip2
- dbus
- expat
- ffmpeg
- fontconfig
- freetype
- giflib
- glib
- gmp
- gst-plugins-base
- gstreamer
- hdf5
- icu
- jpeg
- kaldi
- libblitz
- libffi
- libmatio
- libogg
- libopus
- libpng
- libsvm
- libtiff
- libvpx
- libxcb
- libxml2
- menpo
- mkl # not this one but mkl-devel - no need to list mkl if you use mkl-devel in host
- mkl-devel
- ncurses
- openfst
- openssl
- readline
- sox
- speex
- speexdsp
- sqlite
- tk
- vlfeat
- xz
- yaml
- zlib

Testing entry-points

If you listed some of your setup.py console_sripts in the build / entry_points section of the conda recipe, it is adviseable you test these. For example, if you had the examples entry points above, you would test them like:

test:
  imports:
    - {{ name }}
  commands:
    - jman --help
    - jgen --help

Test-time dependencies

You need to list the packages here that are required during test-time only. By default, add some packages. Do not remove them. The test-time dependencies are listed in test-requirements.txt, which is an optional file, not included in the template. It has the same syntax as requirements.txt, but list only things that are needed to test the package and are not part of its runtime. If you do not need any test-time dependencies, you may skip these instructions.

You may read more information about conda test-time dependencies here.

Left-over conda build files

The conda build command may create a temporary file named record.txt in the project directory. Please make sure it is added in the .gitignore file so that is not committed to the project repository by mistake.

Database packages and packages with extra data

Sometimes databases or other packages require an extra download command after installation. If this extra data is downloaded from Idiap severs, you can include this data in the conda package itself to avoid downloading it two times. If the data is supposed to be downloaded from somewhere other than Idiap servers, do not include it in its conda package. For example, the database packages typically require this download command to be added in the build:script section:

- python setup.py install --single-version-externally-managed --record record.txt # this line is already in the recipe. Do not add.
- bob_dbmanage.py {{ name.replace('bob.db.', '') }} download --missing

Licensing

There are 2 possible cases for the majority of packages in our ecosystem:

  1. If the package is supposed to be licensed under (a 3-clause) BSD license, ensure a file called LICENSE exists at the root of your package and has the correct authorship information.

  2. If the package is supposed to be licensed under GPLv3 license, then ensure a file called COPYING exists on the root of your package

The templating generation has an option to address this.

More info about Idiap’s open-source policy here <https://secure.idiap.ch/intranet/services/technology-transfer/idiap-open-source-policy>.

Headers

Sometimes people add headers with licensing terms to their files. You should inspect your library to make sure you don’t have those. The Idiap TTO says this strategy is OK and simplifies our lives. Make the headers of each file you have as simple as possible, so they don’t get outdated in case things change.

Here is a minimal example (adapt to the language comment style if needed):

`text #!/usr/bin/env python # vim: set fileencoding=utf-8 : `

It is OK to also have your author name on the file if you wish to do so. Don’t repeat licensing terms already explained on the root of your package and on the setup.py file. If we need to change the license, it is painful to go through all the headers.

The setup.py file

The setup.py should be changed to include eventual entry_points you also included in the conda/meta.yaml. We cannot guess these.

Buildout

The default buildout file buildout.cfg should buildout from the installed distribution (use bdt create for that purpose) and avoid mr.developer checkouts. If you have one of those, move it to develop.cfg and create a new buildout.cfg which should be as simple as possible. The template project provided by this package takes care of this.

The README.rst file

You should make the README smaller and easier to maintain. As of today, many packages contain outdated installation instructions or outdated links. More information can always be found at the documentation, which is automatically linked from the badges.

You may want to revise the short introduction after automatic template generation. Make it short, a single phrase is the most common size.

Sphinx documentation

Sphinx documentation configuration goes to a file named doc/conf.py. The file doc/index.rst is the root of the documentation for your package.

The new documentation configuration allows for two optional configuration text files to be placed inside the doc/ directory, alongside the conf.py file:

  • extra-intersphinx.txt, which lists extra packages that should be cross-linked to the documentation (as with Sphinx’s intersphinx extension. The format of this text file is simple: it contains the PyPI names of packages to cross-reference. One per line.

  • nitpick-exceptions.txt, which lists which documentation objects to ignore (for warnings and errors). The format of this text file is two-column. On the first column, you should refer to Sphinx the object type, e.g. py:class, followed by a space and then the name of the that should be ignored. E.g.: bob.bio.base.Database. The file may optionally contain empty lines. Lines starting with # are ignored (so you can comment on why you’re ignoring these objects). Ignoring errors should be used only as a last resource. You should first try to fix the errors as best as you can, so your documentation links are properly working.

Tip

You may use bdt dumpsphinx to list documented objects in remote sphinx documentations. This resource can be helpful to fix issues during sphinx documentation building.

Project logo and branding

In the gitlab Settings / General page of your project, update the logo to use one of ours: