conda和pip之间的区别在conda环境中安装

Question

I seem to be asking myself this question a lot, having recently switched to using conda environments (Anaconda), but I end up Googling and not getting too far.

I now run all my projects within their own conda environments, as I like to keep everything as separate and with as little dependencies on other programs as possible. For example, a recent environment:

conda create -n RL numpy tensorflow-gpu

Then I activate the environment, and realise "Oh - I forgot to install gym". In this case, this is only available in the PIP package manager, and so I simply type pip install gym . But in other cases, where the package exists within conda and pip, what is the best way to install it?

conda install package pip install package

Or in other words - what is the difference?

To provide the full picture, I'm running everything in Ubuntu 16.04, and switch between python 2 and 3 depending on the project. So some of my conda environments are in python 2, some are python 3. I've found that sometimes a pip3 install is required for python 3, but not always - why is this?

Secondly, my path links to the python setup in my Anaconda3 directory.

My current idea is that if I install via conda , it installs directly to my environment, but via pip it installs to my anaconda3 site-packages, making it available to all conda environments under my Anaconda3 directory. If this is the case, this means that if I pip install gym in one conda environment, it should also be available in all others - but this isn't the expected behaviour of environments as far as I am aware.

Please feel free to correct my assumptions and knock some sense into me!

Answer 1

For my understanding of Conda that, it manages for you all the dependencies. For example if you have a package (like pandas) that requires another package (like numpy), it will download both (after warning you).

Where conda is becoming handy is that sometimes a specific package requires a specific version of another one (4.3 or later for example) and they can be conflicts between the packages. The requirements and conflicts define a mathematical problem that can be solved thanks to a SAT solver.

You can find informations and link about that here: https://www.continuum.io/blog/developer/new-advances-conda-0

So each time you are installing a new package, it will upgrade (or sometimes downgrade if conflicts) other packages to ensure the functionning of each package. Personnaly, I go with conda and use pip only when the package is not managed by conda

Another link if you are interested by conda: https://jakevdp.github.io/blog/2016/08/25/conda-myths-and-misconceptions/

About pip3 , it is the naming used when you have both Python 2 and Python 3 installed, to avoid conflicts in the command. In a python 3 environment, the command pip will be equivalent to pip3 .

For the behavior of pip , I can confirm that the installation is done only in the active environment and is not available to the other ones

Answer 2

The difference is that conda will know about the new environment it created but pip will not. You need to install pip inside the environment.

If you create a new environment and activate it: eg

conda create -n env_name
source activate env_name

Then install pip using conda:

conda install pip

(gotcha warning) if you run which pip this should give the path to the pip installation in the new conda environment (something like this):

/anaconda3/envs/env_name/bin/pip

however, just running pip install new_package still does not seem to work, you need to explicitly reference the full path (eg Tom Roth's blog post ) when installing pip packages inside a conda environment

/anaconda3/envs/env_name/bin/pip install new_package

Hope that helps.