fastai makes deep learning with PyTorch faster, more accurate, and easier
The fastai library simplifies training fast and accurate neural nets using modern best practices. See the fastai website to get started. The library is based on research into deep learning best practices undertaken at fast.ai, and includes "out of the box" support for
collab (collaborative filtering) models. For brief examples, see the examples folder; detailed examples are provided in the full documentation. For instance, here's how to train an MNIST model using resnet18 (from the vision example):
untar_data(MNIST_PATH) data = image_data_from_folder(MNIST_PATH) learn = cnn_learner(data, tvm.resnet18, metrics=accuracy) learn.fit(1)
Note for course.fast.ai students
This document is written for
fastai v1, which we use for the current, third version of part 1 of the course.fast.ai deep learning course. If you're following along with a course at course18.fast.ai—that is, part 2 of the deep learning course, or the machine learning course (which aren't yet updated for v1)—you need to use
fastai 0.7; please follow the installation instructions here.
Note: If you want to dive deep into fastai, Jeremy Howard, its lead developer, will be showing internals and advanced features in Deep Learning Part II at the University of San Francisco from March 18th, 2018.
NB: fastai v1 currently supports Linux only, and requires PyTorch v1 and Python 3.6 or later. Windows support is at an experimental stage: it should work fine but we haven't thoroughly tested it. Since Macs don't currently have good Nvidia GPU support, we do not currently prioritize Mac development.
fastai-1.x can be installed with either
pip package managers and also from source. At the moment you can't just run install, since you first need to get the correct
pytorch version installed - thus to get
fastai-1.x installed choose one of the installation recipes below using your favorite python package manager. Note that PyTorch v1 and Python 3.6 are the minimal version requirements.
It's highly recommended you install
fastai and its dependencies in a virtual environment (
conda or others), so that you don't interfere with system-wide python packages. It's not that you must, but if you experience problems with any dependency packages, please consider using a fresh virtual environment just for
Starting with pytorch-1.x you no longer need to install a special pytorch-cpu version. Instead use the normal pytorch and it works with and without GPU. But you can install the cpu build too.
If you experience installation problems, please read about installation issues.
If you are planning on using
fastai in the jupyter notebook environment, make sure to also install the corresponding packages.
More advanced installation issues, such as installing only partial dependencies are covered in a dedicated installation doc.
conda install -c pytorch -c fastai fastai
This will install the
pytorch build with the latest
cudatoolkit version. If you need a higher or lower
CUDA XX build (e.g. CUDA 9.0), following the instructions here, to install the desired
Note that JPEG decoding can be a bottleneck, particularly if you have a fast GPU. You can optionally install an optimized JPEG decoder as follows (Linux):
conda uninstall --force jpeg libtiff -y conda install -c conda-forge libjpeg-turbo CC="cc -mavx2" pip install --no-cache-dir -U --force-reinstall --no-binary :all: --compile pillow-simd
If you only care about faster JPEG decompression, it can be
pillow-simd in the last command above, the latter speeds up other image processing operations. For the full story see Pillow-SIMD.
pip install fastai
By default pip will install the latest
pytorch with the latest
cudatoolkit. If your hardware doesn't support the latest
cudatoolkit, follow the instructions here, to install a
pytorch build that fits your hardware.
Bug Fix Install
If a bug fix was made in git and you can't wait till a new release is made, you can install the bleeding edge version of
pip install git+https://github.com/fastai/fastai.git
The following instructions will result in a pip editable install, so that you can
git pull at any time and your environment will automatically get the updates:
git clone https://github.com/fastai/fastai cd fastai tools/run-after-git-clone pip install -e ".[dev]"
Next, you can test that the build works by starting the jupyter notebook:
and executing an example notebook. For example load
examples/tabular.ipynb and run it.
Building From Source
If for any reason you can't use the prepackaged packages and have to build from source, this section is for you.
pytorchfrom source follow the complete instructions. Remember to first install CUDA, CuDNN, and other required libraries as suggested - everything will be very slow without those libraries built into
Next, you will also need to build
git clone https://github.com/pytorch/vision cd vision python setup.py install
torchvisionare installed, first test that you can load each of these libraries:
import torch import torchvision
to validate that they were installed correctly
Finally, proceed with
fastaiinstallation as normal, either through prepackaged pip or conda builds or installing from source ("the developer install") as explained in the sections above.
If you encounter installation problems with conda, make sure you have the latest
conda client (
conda install will do an update too):
conda install conda
Is My System Supported?
Python: You need to have python 3.6 or higher
CPU or GPU
pytorchbinary package comes with its own CUDA, CuDNN, NCCL, MKL, and other libraries so you don't have to install system-wide NVIDIA's CUDA and related libraries if you don't need them for something else. If you have them installed already it doesn't matter which NVIDIA's CUDA version library you have installed system-wide. Your system could have CUDA 9.0 libraries, and you can still use
pytorchbuild with CUDA 10.0 libraries without any problem, since the
pytorchbinary package is self-contained.
The only requirement is that you have installed and configured the NVIDIA driver correctly. Usually you can test that by running
nvidia-smi. While it's possible that this application is not available on your system, it's very likely that if it doesn't work, then you don't have your NVIDIA drivers configured properly. And remember that a reboot is always required after installing NVIDIA drivers.
Since fastai-1.0 relies on pytorch-1.0, you need to be able to install pytorch-1.0 first.
As of this moment pytorch.org's 1.0 version supports:
Platform GPU CPU linux binary binary mac source binary windows binary binary
binary= can be installed directly,
source= needs to be built from source.
If there is no
pytorchpreview conda or pip package available for your system, you may still be able to build it from source.
How do you know which pytorch cuda version build to choose?
It depends on the version of the installed NVIDIA driver. Here are the requirements for CUDA versions supported by pre-built
CUDA Toolkit NVIDIA (Linux x86_64) CUDA 10.0 >= 410.00 CUDA 9.0 >= 384.81 CUDA 8.0 >= 367.48
So if your NVIDIA driver is less than 384, then you can only use CUDA 8.0. Of course, you can upgrade your drivers to more recent ones if your card supports it.
You can find a complete table with all variations here.
If you use NVIDIA driver 410+, you most likely want to install the
cudatoolkit=10.0pytorch variant, via:
conda install -c pytorch pytorch cudatoolkit=10.0
or if you need a lower version, use one of:
conda install -c pytorch pytorch cudatoolkit=8.0 conda install -c pytorch pytorch cudatoolkit=9.0
For other options refer to the complete list of the available pytorch variants.
In order to update your environment, simply install
fastai in exactly the same way you did the initial installation.
Top level files
environment-cpu.yml belong to the old fastai (0.7).
conda env update is no longer the way to update your
fastai-1.x environment. These files remain because the fastai course-v2 video instructions rely on this setup. Eventually, once fastai course-v3 p1 and p2 will be completed, they will probably be moved to where they belong - under
We use GitHub issues for tracking requests and bugs, so please see fastai forum for general questions and discussion.
The fastai project strives to abide by generally accepted best practices in open-source software development:
A detailed history of changes can be found here.
Copyright 2017 onwards, fast.ai, Inc. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this project's files except in compliance with the License. A copy of the License is provided in the LICENSE file in this repository.
Release history Release notifications
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size & hash SHA256 hash help||File type||Python version||Upload date|
|fastai-1.0.53.post3-py3-none-any.whl (229.1 kB) Copy SHA256 hash SHA256||Wheel||py3|
|fastai-1.0.53.post3.tar.gz (3.1 MB) Copy SHA256 hash SHA256||Source||None|