Anaconda on Windows

This cookiecutter is designed to work with Anaconda, a scientific Python distribution including its own package manager conda. Anaconda simplifies the usage and maintenance of python in particular for Windows user. However, the programming community is still extremely focused on use cases with Linux or MacOS and neglect issues on Windows as these users should switch the OS anyway :). Therefore, the following is a step-by-step installation and user guide for Anaconda on Windows.

Installation

  1. Download the latest graphical installer from anaconda.org. If you do not know whether you need the 32-bit or 64-bit installer, look at this FAQ.

  2. Start the installer. Choose whether to install Anaconda for all users which requires administrator privileges or for a single user. I prefer to install Python system-wide so it is available to all users. But then, you have to be careful as every time you interact with the base environment you have to use an elevated shell (a shell with administrator privileges as described in this article).

    Warning

    If you install Anaconda for a single user you may run into problems when executing python from the Windows Powershell. In particular, it may happen, that you can run your Python files only from the Anaconda prompt.

  3. Tick “Add Anaconda to my PATH environment variable” and also “Register Anaconda as my default Python 3.x”. Finish installation.

  4. Often times you have to manually add Anaconda to your PATH environment. You can find an instruction on how to do that here. If you installed Anaconda for all users, the PATH should read something like C:\ProgramData\Anaconda3.

Which console?

The Powershell is the preferred way on Windows as it provides a better interface and better tab-completion.

Conda Version < 4.6

Unlike CMD and “Anaconda Prompt”, it is not fully supported by Anaconda. In particular, the activation and deactivation of environments is broken in older versions of conda (<4.6). To solve this issue, you have to install an additional package . If you installed Anaconda with administrator privileges, start an elevated shell. Then, type

$ conda install pscondaenvs -c pscondaenvs

Since Windows is extremely cautious, it does not want to execute the new activate.ps1. Thus, we need to change the execution policy in administrator mode with

$ Set-ExecutionPolicy RemoteSigned

and answer the following prompt with “Yes, for all” or a.

Now, if you go back to your normal Powershell, you can activate an existing environment with

$ activate <env-name>

and deactivate it with

$ deactivate

Conda Versions >= 4.6

Warning

There are new problems with conda 4.6.9 and above. Either conda does not work at all or after conda init and shell restart, the path is not displayed after the environment name anymore. I would stay at version 4.6.8 at the moment. Install it with

$ conda install conda=4.6.8

If you want to try out a newer version, install it and revert back if the problems have not been solved.

Starting from version 4.6 Anaconda officially supports Powershell (see Conda 4.6 Release). Note, however, that this is still in an experimental state and issues may still occur. To initialize the use of Anaconda environments from Powershell, you have to open your Powershell and execute:

$ conda init

You then have to restart your Powershell. Now, you can activate existing environments with

$ conda activate <env-name>

and deactivate with

$ conda deactivate

How to interact with the base environment?

The base environment is activated by default. If you start a Powershell and type python, you are using the Python interpreter and the packages from the base environment.

My personal advice is to touch the base environment only if you want to do some small programming or prototyping. In all other cases, create a separate environment.

Updating conda and the package manager

Start a Powershell (with administrator privileges if you installed Anaconda for all users). Type

$ conda update conda

to update the package manager.

Warning

Be aware that sometimes the developers of conda distribute buggy versions which usually forces you to reinstall Anaconda completely. Still, I recommend to upgrade from time to time. If you are extremely cautious, check the latest versions and update only if the latest version is a week old.

Then, update Anaconda with

$ conda update anaconda

How to interact with environments?

Create environments

As I said before, I recommend to create a new environment for each of your projects. If you do not know which packages you need later, start with a plain Python environment and install packages along the way. Create a plain Python environment with

$ conda create python=3.7 -c anaconda

or you can create an environment from a file with

$ conda env create -n <env-name> -f <path-to-yml>

The environments are usually placed in your user folder under C:\Users\<user-name>\.conda/envs/, but I would not be surprised to find them elsewhere :).

Manage packages

If you leave out the name, conda takes the name from the environment.yml. If you leave out the file, conda looks for a environment.yml in the current folder.

To install a package type

$ conda install statsmodels=0.9.0

and to update

$ conda update statsmodels

Export an environment

To make your projects reproducible, you have to define an environment.yml.

$ conda env export -f environment.yml

Exporting the environment is one but maybe not the best way to create the environment file. I would recommend that you do it yourself and add only packages you are importing directly. The reason is that you only want to ensure that the results hold for the specific versions of the main packages and you do not care about how they are using their dependencies. An example looks like this:

# content of environment.yml
name: cc
channels:
    - defaults
    - pscondaenvs
dependencies:
    - pscondaenvs=1.2.4
    - python=3.7
    - pip:
      - pandas==0.24.1

name is the shortcut used to activate the environment later. channels contains different sources for installing packages in order. During installation conda iterates through the channels from top to bottom and looks for the specific package. In dependencies one can see first packages installed via conda. Notice the single equality sign to pin a specific version. Under pip you can see a list of packages which should be installed with pip. Here, you pin a package with two equality signs. I would recommend to install as many packages with pip as possible, e.g. pandas, but not Numpy, statsmodels, scikit-learn. First, every package is always up-to-date on PyPi, but sometimes distributing to Anaconda takes longer. Second, pyup can only inform you about updates under pip.

If you export the environment, there is a second entry after each package installed with conda.

dependencies:
  - vs2015_runtime=14.15.26706=h3a45250_0

The hash, h3a45250_0, makes sure that packages have the same build instructions, but they are not only compiler but also OS-specific. Thus, you cannot install a hashed package on Windows and Linux.

Update an environment

What if you want to update the environment because you altered the environment.yml?

$ conda env update -n <env-name> -f <path-to-yml>

Again, you can leave out -n and -f if the name is specified in the file or if the file is in the current directory.

Remove an environment

$ conda env remove -n <env-name>

The rest of the commands can be found in the official conda documentation.

Advanced tips

  • At some point, you might decide that you do not need a base environment or that you would like to force yourself to create environments for each project. Then, Miniconda might be an option for you. This setup only installs the package manager, but no base environment. This also saves space on your disk.

  • Often times you want to start a Jupyter notebook server while continuing to use the shell. Just type

    $ start jupyter notebook

    or

    $ start jupyter lab

    which will open a new Powershell within the same conda environment running the server. Then, you are able to install new packages from the unblocked shell, but you have to restart running kernels before packages become available.