Basic installation

Please refer to README for bulk of the instructions

CPU build

Generally, pytorch GPU build should work fine on machines that don’t have a CUDA-capable GPU, and will just use the CPU. However, you can install CPU-only versions of Pytorch if needed with fastai.

  • pip

    The pip ways is very easy:

     pip install http://download.pytorch.org/whl/cpu/torch-1.0.0-cp36-cp36m-linux_x86_64.whl
     pip install fastai
    

    Just make sure to pick the correct torch wheel url, according to the needed platform, python and CUDA version, which you will find here.

  • conda

    The conda way is more involved. Since we have only a single fastai package that relies on the default pytorch package working with and without GPU environment, if you want to install something custom you will have to manually tweak the dependencies. This is explained in detail here. So follow the instructions there, but replace pytorch with pytorch-cpu, and torchvision with torchvision-cpu.

Also, please note, that if you have an old GPU and pytorch fails because it can’t support it, you can still use the normal (GPU) pytorch build, by setting the env var CUDA_VISIBLE_DEVICES="", in which case pytorch will not try to check if you even have a GPU.

Jupyter notebook dependencies

The fastai library doesn’t require the jupyter environment to work, therefore those dependencies aren’t included. So if you are planning on using fastai in the jupyter notebook environment, e.g. to run the fastai course lessons and you haven’t already setup the jupyter environment, here is how you can do it.

  • conda

     conda install jupyter notebook
     conda install -c conda-forge jupyter_contrib_nbextensions
    

    Some users also seem to need this conda package to be able to choose the right kernel environment, however, most likely you won’t need this package.

     conda install nb_conda
    
  • pip

     pip install jupyter notebook jupyter_contrib_nbextensions
    

Custom dependencies

If for any reason you don’t want to install all of fastai’s dependencies, since, perhaps, you have limited disk space on your remote instance, here is how you can install only the dependencies that you need.

  1. First, install fastai without its dependencies using either pip or conda:

    # pip
    pip install --no-deps fastai
    # conda
    conda install --no-deps -c fastai fastai
    
  2. The rest of this section assumes you’re inside the fastai git repo, since that’s where setup.py resides. If you don’t have the repository checked out, do:

    git clone https://github.com/fastai/fastai
    cd fastai
    tools/run-after-git-clone
    
  3. Next, find out which groups of dependencies you want:

    python setup.py -q deps
    

    You should get something like:

    Available dependency groups: core, text, vision
    

    You need to use at least the core group.

    Do note that the deps command is a custom distutils extension, i.e. it only works in the fastai setup.

  4. Finally, install the custom dependencies for the desired groups.

    For the sake of this demonstration, let’s say you want to get the core dependencies (core), plus dependencies specific to computer vision (vision). The following command will give you the up-to-date dependencies for these two groups:

    python setup.py -q deps --dep-groups=core,vision
    

    It will return something like:

    Pillow beautifulsoup4 bottleneck dataclasses;python_version<'3.7' fastprogress>=0.1.18 matplotlib numexpr numpy>=1.12 nvidia-ml-py3 packaging pandas pyyaml requests scipy torch>=1.0.0 torchvision typing
    

    which can be fed directly to pip install:

    pip install $(python setup.py -q deps --dep-groups=core,vision)
    

    Since conda uses a slightly different syntax/package names, to get the same output suitable for conda, add --dep-conda:

    python setup.py -q deps --dep-groups=core,vision --dep-conda
    

    If your shell doesn’t support $() syntax, it most likely will support backticks, which are deprecated in modern bash. (The two are equivalent, but $() has a superior flexibility). If that’s your situation, use the following syntax instead:

    pip install `python setup.py -q deps --dep-groups=core,vision`
    
  • Manual copy-n-paste case:

    If, instead of feeding the output directly to pip or conda, you want to do it manually via copy-n-paste, you need to quote the arguments, in which case add the --dep-quote option, which will do it for you:

     # pip:
     python setup.py -q deps --dep-groups=core,vision --dep-quote
     # conda:
     python setup.py -q deps --dep-groups=core,vision --dep-quote --dep-conda
    

    So the output for pip will look like:

     "Pillow" "beautifulsoup4" "bottleneck" "dataclasses;python_version<'3.7'" "fastprogress>=0.1.18" "matplotlib" "numexpr" "numpy>=1.12" "nvidia-ml-py3" "packaging" "pandas" "pyyaml" "requests" "scipy" "torch>=1.0.0" "torchvision" "typing"
    
  • Summary:

    pip selective dependency installation:

     pip install --no-deps fastai
     pip install $(python setup.py -q deps --dep-groups=core,vision)
    

    same for conda:

     conda install --no-deps -c fastai fastai
     conda install -c pytorch -c fastai $(python setup.py -q deps --dep-conda --dep-groups=core,vision)
    

    adjust the --dep-groups argument to match your needs.

  • Full usage:

     # show available dependency groups:
     python setup.py -q deps
    
     # print dependency list for specified groups
     python setup.py -q deps --dep-groups=core,vision
    
     # see all options:
     python setup.py -q deps --help
    

Development dependencies

As explained in Development Editable Install, if you want to work on contributing to fastai you will also need to install the optional development dependencies. In addition to the ways explained in the aforementioned document, you can also install fastai with developer dependencies without needing to check out the fastai repo.

  • To install the latest released version of fastai with developer dependencies, do:

    pip install "fastai[dev]"

  • To accomplish the same for the cutting edge master git version:

    pip install "git+https://github.com/fastai/fastai#egg=fastai[dev]"

Virtual environment

It’s highly recommended to use a virtual python environment for the fastai project, first because you could experiment with different versions of it (e.g. stable-release vs. bleeding edge git version), but also because it’s usually a bad idea to install various python package into the system-wide python, because it’s so easy to break the system, if it relies on python and its 3rd party packages for its functionality.

There are several implementations of python virtual environment, and the one we recommend is conda (anaconda), because we release our packages for this environment and pypi, as well. conda doesn’t have all python packages available, so when that’s the case we use pip to install whatever is missing.

You will find the instructions for installing conda on each platform here. Once you followed the instructions and installed anaconda, you’re ready to build you first environment. For the sake of this example we will use an environment name fastai, but you can name it whatever you’d like it to be.

The following will create a fastai env with python-3.6:

conda create -n fastai python=3.6

Now any time you’d like to work in this environment, just execute:

conda activate fastai

It’s very important that you activate your environment before you start the jupyter notebook if you’re using fastai notebooks.

Say, you’d like to have another env to test fastai with python-3.7, then you’d create another one with:

conda create -n fastai-py37 python=3.7

and to activate that one, you’d call:

conda activate fastai-py37

If you’d like to exit the environment, do:

conda deactivate

To list out the available environments

conda env list

Also see bash-git-prompt which will help you tell at any moment which environment you’re in.