Google Colab(oratory) is an invaluable resource for data science. You get GPU/TPU computing power while from a Jupyter Notebook frontend running on Google’s servers… for free! Surprisingly though, conda is not preinstalled in the default configuration. Learn how to fix it!

Update March 10th 2021:

Looks like Google Colab has upgraded to Python 3.7. The instructions below have been updated to reflect that fact!

I have updated condacolab, a Python package that will do all of this for you! Basically, now you only need this in your first notebook cell:

!pip install -q condacolab
import condacolab
condacolab.install()

Check the repo for more details and examples!

TLDR: Just run these two cells at the beginning of your Colab notebook:

!wget https://repo.anaconda.com/archive/Anaconda3-2020.02-Linux-x86_64.sh && bash Anaconda3-2020.02-Linux-x86_64.sh -bfp /usr/local
import sys
sys.path.insert(0, "/usr/local/lib/python3.7/site-packages/")

If you are going to install new packages, always add python=3.7 to the list to prevent accidental updates:

!conda install -yq python=3.7 your_extra_packages

Python has become the most popular language in StackOverflow and one could argue that its success is recently due to data science in general and machine learning in particular. Most of the data scientists rely on the Anaconda distribution or, at least, its package manager to install the libraries they need: conda.

While most popular projects offer *.deb packages and pip wheels (both methods officially supported by Google Colab), some are only distributed through conda (for example, OpenMM). However, conda is not preinstalled in the Colab environments! The good news is that you can install it manually for each notebook.

Install Miniconda

Google Colab uses Python 3.7, so we need an Anaconda distribution compiled for that version. Recent builds use later Python versions, so you have to use Anaconda v2020.02 or Miniconda v4.9.2-py37. Choose one below.

A - Using the full Anaconda distribution

The full Anaconda bundle contains a huge selection of data science packages (694, to be exact) ready to run.

!wget https://repo.anaconda.com/archive/Anaconda3-2020.02-Linux-x86_64.sh && bash Anaconda3-2020.02-Linux-x86_64.sh -bfp /usr/local

B - Using the Miniconda distribution

Miniconda only contains the basics: Python itself, conda, pip and some libraries. It’s up to you to install whatever you need afterwards.

!wget https://repo.anaconda.com/miniconda/Miniconda3-py37_4.9.2-Linux-x86_64.sh && bash Miniconda3-py37_4.9.2-Linux-x86_64.sh -bfp /usr/local

Test the installation

Now, you should be able to run some shell commands to check everything is correct:

!conda info --all
!conda list

Patch sys.path and import your packages

To be able to import the Anaconda packages, you have patch sys.path so Python can find the modules:

import sys
sys.path.insert(0, "/usr/local/lib/python3.7/site-packages/")

Now you can just use Python’s import like usual.

Install more packages

Important: The conda solver might accidentally update Python when you issue a conda install command. If you do it, things will stop working. To prevent this, make sure to add python=3.7 as part of the package list.

Of course, you can install more packages if needed. Just remember to use -y to avoid interactive prompts and -q to remove excessive output. Also, if the package needs cuda, make sure it is compiled for v10 or v10.1. For example, to install openmm:

!conda install -y -q -c conda-forge openmm cudatoolkit=10 python=3.7

Solving the conda environment in terms of dependencies can take a while sometimes. In some tests, that last command took 15-20 minutes, but in other cases it finished in 2 minutes. ¯\_(ツ)_/¯

Check GPU support with openmm tests (assuming you have started the Colab instance in GPU or TPU mode!):

import simtk.testInstallation
simtk.testInstallation.main()

Expected output is 4 platforms:

There are 4 Platforms available:

1 Reference - Successfully computed forces
2 CPU - Successfully computed forces
3 CUDA - Successfully computed forces
4 OpenCL - Successfully computed forces

Median difference in forces between platforms:

Reference vs. CPU: 6.3031e-06
Reference vs. CUDA: 6.73543e-06
CPU vs. CUDA: 7.81258e-07
Reference vs. OpenCL: 6.75426e-06
CPU vs. OpenCL: 8.15821e-07
CUDA vs. OpenCL: 2.17776e-07

… And that’s it. Maybe in the future Google Colab will bundle conda as well and this won’t be needed. But as of Apr 2019 (and July 2020, and March 2021), this is the way to go!

Bonus: Ready-to-run Miniconda-enabled Notebook.