In this guide, I expect you are using Python3 (and pip3), but (for old projects) you can use Python2 (and pip2) as well.

• You can use the system Python (currently 3.6.5, in Ubuntu 18.04)
• You can use any Python from /opt/python/, e.g. for 3.6.3 add the following line to your .bashrc:

  export PATH=/opt/python/3.6.3/bin:"${PATH}"

• You can install modules with pip3. If you are not using any virtual environments, you must add the --user option, so it is installed in your home. In this case, don't forget the executable scripts are installed in ~/.local/bin/, so add the following line to your .bashrc:

  export PATH=~/.local/bin:"${PATH}"

• If you are using the system Python3 (3.4.3), note that it comes with a very old version of pip3 (1.5.4), so it is recommended to update it with

  pip3 install --user --upgrade pip

• There are no UFAL-wide pre-installed Python modules. You need to install all modules you need yourself (with pip3).

You can use virtual environments, e.g. one for each project (or a group of projects), so the version requirements of different projects do not collide.

Create a new virtual environment with

  python3 -m venv my-project-name

where my-project-name can be a relative or absolute path to a directory which will be created.
Caveat1: the path to the directory should be rather short, there is a limit on shebang paths so you can get strange errors like ipython3: bad interpreter: No such file or directory if your ipython3 is installed in the venv starts with #!/..a..very..long..path..
Caveat2: If my-project-name is a relative path, it will be concatenated with pwd -P, i.e. with your actual directory with all symlinks resolved. If you want symlinks in the path, you must provide absolute path as my-project-name. Note that e.g. /net/work/people is currently (2018-06) resolved by pwd -P to /lnet/.nfs/spec/work/people/, but this path does not work on machines with native Lustre support.)

You can use any Python version for the new environment, e.g. /opt/python/3.5.4/bin/python3 -m venv my-project-name-python3.5
Then activate it with

  source my-project-name/bin/activate

Now your prompt changes and you can install any modules (with pip3, but you can omit --user), e.g.

  pip3 install --upgrade pip tensorflow

If you run python, it will be the version used when creating venv. Check it with

  which python
  python -V

Once finished, exit the virtual environment with

  deactivate

If you decide you don't need the environment and any modules installed in this environment, simply delete the my-project-name directory.

Note that there are many alternative ways how to create virtual environments and install modules. The venv module is available only for Python 3.3 and newer. For older versions you can use the third-party virtualenv tool:

 pip3 install --user --upgrade pip
 pip3 install --user virtualenv
 virtualenv my-project-name
 source my-project-name/bin/activate
 

When using virtuaenvs, ipython is (in some versions) confused uses incorrect python interpreter. The following alias fixes this issue:

alias ipython="python -c 'import IPython; IPython.terminal.ipapp.launch_new_instance()'"

If you are developing (contributing to) a Python tool version in git and want to install the newest version, you can use pip3 git+https://github.com/…, provided the project has proper setup.py in the root directory.
You can also git clone the project manually and add the path to your $PYTHONPATH. This way whenever you git pull, you should make sure there are no new dependencies, e.g. with pip3 install –user -r requirements.txt if the project provides the requirements.txt file.
See e.g. Udapi or Neural Monkey for examples of both ways.