Installing FastAi V2 Locally
This guide will help you setup fastaiv2 on a laptop / desktop running Ubuntu LTS.
If you are not running an LTS version of Ubuntu, you should not proceed.
IMPORTANT
-
In this guide, we will not be using
conda. -
We will only be using a combination of
virtualenv,virtualenvwrapperandpip. This is because you will have full control on exactly what is being installed.
Check your Ubuntu Version:
lsb-release -a
You should see something like:
Distributor ID: Ubuntu
Description: Ubuntu 18.04.3 LTS
Release: 18.04
Codename: bionic
Check Nvidia CUDA Drivers
nvidia-smi
Note: You MUST have driver version >= 418.39.
If you don’t have the latest drivers, proceed to the next step and include the driver in the CUDA installation.
Download CUDA 10.1 (for PyTorch 1.4)
Download the local runfile installer for your Ubuntu version.
mkdir ~/Downloads && cd ~/Downloads
wget -O cuda_10.1.105_418.39_linux.run https://developer.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.105_418.39_linux.run
chmod +x cuda_10.1.105_418.39_linux.run
sudo sh cuda_10.1.105_418.39_linux.run
IMPORTANT:
This run file installer will ask you to select options.
If you have your Nvidia Driver installed already, make sure you uncheck that part of the installation.
If you have another version of CUDA installed, make sure you don’t clobber you existing installation; by selecting the directory where CUDA 10.1 should be installed manually.
In the Installer’s UI click Options -> Root Install Path
and then enter /usr/local/cuda-10.1 and then select Done.
Now select Install.
If you have a pre-exisitng version of CUDA you will be asked:
A symlink already exists at /usr/local/cuda. Update this installation ?
Select `No`.
Now let the installer complete installation.
Install virtualenv and virtualenvwrapper
To setup virtualenv and virtualenvwrapper, follow the instructions at:
https://www.pyimagesearch.com/2018/05/28/ubuntu-18-04-how-to-install-opencv/ [Look at Step #3]
Update your ~/.bashrc with the following:
# CUDA
export CUDA_10=/usr/local/cuda-10.0 # Only if you have multiple installations of CUDA
export CUDA_10_1=/usr/local/cuda-10.1
export CUDA=$CUDA_10
export USR_LOCAL=/home/$USER/.local/
export PATH=$CUDA/bin:$USR_LOCAL/bin:$PATH
export CUDA_PATH=$CUDA
export LD_LIBRARY_PATH=$CUDA/lib64
# Virtualenv
export WORKON_HOME=$HOME/.virtualenvs
export VIRTUALENVWRAPPER_PYTHON=/usr/bin/python3
# Virtualenv Wrapper
source /home/$USER/.local/bin/virtualenvwrapper.sh
# Helpers to switch CUDA Versions
# For side by side installs only. Ignore if you only have one version of CUDA
set_cuda_10() {
echo "Setting CUDA to v10.0"
export CUDA=$CUDA_10
export PATH=$CUDA/bin:$USR_LOCAL/bin:$PATH
export CUDA_PATH=$CUDA
export LD_LIBRARY_PATH=$CUDA/lib64
}
set_cuda_10_1() {
echo "Setting CUDA to v10.1"
export CUDA=$CUDA_10_1
export PATH=$CUDA/bin:$USR_LOCAL/bin:$PATH
export CUDA_PATH=$CUDA
export LD_LIBRARY_PATH=$CUDA/lib64
}
Note:
- When using
fastaiv2you need to be usingCUDA 10.1. - If you only have a single CUDA installation you can use that version to be the canonical
$CUDAversion.
Install cuDNN
You will need to create an Nvidia Developer account for the next step.
Login, and visit https://developer.nvidia.com/rdp/cudnn-download
IMPORTANT: Make sure you select the TAR file install for CUDA 10.1. This is labelled as cuDNN Library for Linux.
The URL should look something like: https://developer.nvidia.com/compute/machine-learning/cudnn/secure/7.6.5.32/Production/10.1_20191031/cudnn-10.1-linux-x64-v7.6.5.32.tgz.
This URL actually is a HTTP redirect. If you are installing this headless - then open the Chrome Network tab, and copy the actual redirected URL.
wget -O cudnn.tar.gz 'https://developer.download.nvidia.com/compute/machine-learning/cudnn/secure/7.6.5.32/Production/10.1_20191031/cudnn-10.1-linux-x64-v7.6.5.32.tgz?a_really_long_encrypted_token'
To install cuDNN from the TAR File you can find instructions at: https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html#installlinux-tar
sudo cp cuda/include/cudnn.h /usr/local/cuda-10.1/include
sudo cp cuda/lib64/libcudnn* /usr/local/cuda-10.1/lib64
sudo chmod a+r /usr/local/cuda-10.1/include/cudnn.h /usr/local/cuda-10.1/lib64/libcudnn*
Create a Virtual Environment
mkvirtualenv torch
workon torch
set_cuda_10_1
echo $CUDA # Should point to the right $CUDA directory
IMPORTANT: Anytime you pip install or you use python you should ALWAYS activate your virtual environment with the right CUDA versions.
Install PyTorch
More instructions at https://pytorch.org/get-started/locally/
pip install torch torchvision
Verify PyTorch installation
From inside the virtual environment, run python:
import torch
torch.cuda.is_available()
>>> True # Should be True
Install fastaiv2
mkdir FastAi && cd FastAi
git clone https://github.com/fastai/fastcore
cd fastcore
pip install -e ".[dev]"
cd ..
git clone https://github.com/fastai/fastai2
cd fastai2
pip install -e ".[dev]"
cd ..
Install Course Materials
git clone https://github.com/fastai/course-v4.git
cd course-v4
# Dependencies from `requirements.txt`
pip install graphviz ipywidgets matplotlib nbdev>=0.2.12 pandas scikit_learn azure-cognitiveservices-search-imagesearch sentencepiece
cd ..
Launch Jupyter
From the course-v4 directory, run
jupyter notebook
Anaconda is a bit too magical for my taste.