fastai

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 vision, text, tabular, and 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 = create_cnn(data, tvm.resnet18, metrics=accuracy)
learn.fit(1)

Note for course.fast.ai students

If you are using fastai for any course.fast.ai course, you need to use fastai 0.7. Please ignore the rest of this document, which is written for fastai v1, and instead follow the installation instructions here.

Note: If you want to learn how to use fastai v1 from its lead developer, Jeremy Howard, he will be teaching it in the Deep Learning Part I course at the University of San Francisco from Oct 22nd, 2018.

Installation

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 conda or 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 favourite python package manager. Note that PyTorch v1 and Python 3.6 are the minimal version requirements.

If your system has a recent NVIDIA card with the correctly configured NVIDIA driver please follow the GPU installation instructions. Otherwise, the CPU-ones.

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 fastai.

If you experience installation problems, please read about installation issues.

More advanced installation issues, such as installing only partial dependencies are covered in a dedicated installation doc.

Conda Install

conda install -c pytorch -c fastai fastai

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 or pillow-simd in the last command above, the latter speeds up other image processing operations. For the full story see Pillow-SIMD.

PyPI Install

pip install fastai

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 fastai with:

pip install git+https://github.com/fastai/fastai.git

Developer Install

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]"

Note that this will install the cuda9.0 pytorch build via default dependencies. If you need a higher or lower cudaXX build, following the instructions here, to install the desired pytorch build.

Next, you can test that the build works by starting the jupyter notebook:

jupyter notebook

and executing an example notebook. For example load examples/tabular.ipynb and run it.

Please refer to CONTRIBUTING.md and Notes For Developers for more details on how to contribute to the fastai project.

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.

1. To build pytorch from 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 pytorch.

2. Next, you will also need to build torchvision from source:

   git clone https://github.com/pytorch/vision
   cd vision
   python setup.py install

3. When both pytorch and torchvision are 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 fastai installation as normal, either through prepackaged pip or conda builds or installing from source ("the developer install") as explained in the sections above.

Installation Issues

If the installation process fails, first make sure your system is supported. And if the problem is still not addressed, please refer to the troubleshooting document.

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?

1. Python: You need to have python 3.6 or higher

2. CPU or GPU

   The pytorch binary 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 pytorch build with cuda9.2 libraries without any problem, since the pytorch binary 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, than your don't have your NVIDIA drivers configured properly. And remember that a reboot is always required after installing NVIDIA drivers.

3. Operating System:

   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

   Legend: binary = can be installed directly, source = needs to be built from source.

   If there is no pytorch preview conda or pip package available for your system, you may still be able to build it from source.

4. 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 pytorch releases:

   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 cuda80. 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 cuda100 pytorch variant, via:

   conda install -c pytorch pytorch cuda100

   or if you need a lower version (cuda90 is installed by default), use:

   conda install -c pytorch pytorch cuda80

   For other options refer to the complete list of the available pytorch variants.

Updates

In order to update your environment, simply install fastai in exactly the same way you did the initial installation.

Top level files environment.yml and 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 old/.

Contribution guidelines

If you want to contribute to fastai, be sure to review the contribution guidelines. This project adheres to fastai's code of conduct. By participating, you are expected to uphold this code.

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:

History

A detailed history of changes can be found here.

Copyright

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.