When you write an application that incorporates machine learning in Python, it often depends on scikit-learn, or even Numpy or Scipy.
These depend on various libraries, so deploying on Heroku is not straightforward. If you can set up your own server on AWS etc., you can set it up there, but it costs money and I want to do it on Heroku! I think that there are cases, so I will introduce a method for that.
Docker Container
Deployment using Docker Container is now possible on Heroku, so use this rather than editing the buildpack in a hurry Is more convenient. Below is an official sample, so please refer to it.
heroku-examples/python-miniconda
Now you won't have to install anything with apt-get or complain that your slag size is too big!
Conda buildpack
Libraries such as Numpy and Scipy are usually quite difficult to install from pip, that is, to build from source code. There is an execution environment called Miniconda that can avoid this, that is, you can install compiled binaries, but a build pack that can be used on Heroku is provided. I am.
There are two ways to set up a buildpack on Heroku: specify it when you create it, or change it later.
** Specified at creation **
heroku create --buildpack https://github.com/kennethreitz/conda-buildpack.git
** Specify later **
heroku config:add BUILDPACK_URL=https://github.com/kennethreitz/conda-buildpack.git
To use it, just prepare conda-requirements.txt (requirements.txt in conda) for installation with conda. Those that are not provided by conda can also be installed using pip, so if you need to install with pip, you can still write them in requirements.txt.
nomkl to conda-requirements.txt so as not to use the mkl optimized version (# 21 twenty one)).The difficulty is that the environment created by conda conflicts with virtualenv and cannot be used together. Therefore, if you are looking to deploy on Heroku, you need to use conda to create the virtual environment. Please refer to here for the simple usage of the conda command.
Set the build pack on Heroku and prepare the configuration file for conda. Now you can run applications with machine learning libraries on Heroku.
I made a repository that I actually tried using conda-buildpack, so please refer to it. Since it has a Heroku Button, it can be deployed as it is.
Heroku buildpack
This is a build pack that I created because I didn't know the existence of conda-buildpack and deploys it based on pip as usual.
However, compiling scipy takes so long that it times out on Heroku (make sure you can build Numpy). The only way to overcome this is to use a better Dyno, so I think it's difficult with the free tier.
Heroku buildpack: Python with Numpy, Scipy, scikit-learn
It's not just compatible with Numpy and Scipy, it's designed to maintain its compilation environment, so it should work universally for things that depend on similar libraries (those that). I think it is provided by conda ...). If pip didn't work and you gave up when deploying Heroku, it might work for that too.
You can use heroku-buildpack-apt that can install the library with apt-get in the Heroku environment used here, or connect multiple build packs. heroku-buildpack-multi may be useful in other scenes as it may be useful to remember. heroku-buildpack-multi can be used when using bower etc. (process bower with the build pack for Node.js, and then run the process with the build pack for the main body).
Recommended Posts