A Vagrantfile to provision a gpu instance installed tensorflow-gpu on EC2
This Vagrantfile gives you an easy way to launch a gpu instance installed tensorflow-gpu. Information for the instance which will be launched is blow.
- CUDA: v8.0
- cuDNN: v5.1
- Python: 3.4
- Tensorflow(GPU): v12.1
- OS: Ubuntu 14.04
- instance_type: g2.2xlarge
- region: np-northeast-1
- ami: ami-3995e55e
If you want to change the region, please find an appropriate ami for your region and instance_type from here. If you want to change the version of python, CUDA or cuDNN, please find an appropriate tensorflow-gpu from here.
Vagrant and a few configurations of AWS are required.
Vagrant is a command line utility to manage virtual machines. You can download an appropriate installer for your platform from here.
vagrant-aws is a vagrant plugin to allow Vagrant to control and provision machines in EC2 and VPC. When Vagrant is already installed, you can install it like blow.
$ vagrant plugin install vagrant-aws
Amazon EC2 is a VPS service. This Vagrantfile needs network and security configurations, which are Amazon EC2 Key Pairs and Amazon EC2 Security Groups.
You can create a key pairs accoding to this link. You can also create a security group according to this link.
After creating them, please set them to aws.keypair_name, aes.security_groups and ssh.private_key_path in the Vagrantfile, whose roles are explained at vagrant-aws's README. Please use a security group to allow SSH access. If you want to use jupyter notebook and tensorboard, I recommend you to allow TCP 8888 and TCP 6006 ports.
$ git clone https://github.com/shotarok/vagrant-aws-gpu-tensorflow.git
$ cd vagrant-aws-gpu-tensorflow
# Configure aws environment variables
$ export AWS_ACCESS_KEY_ID=...
$ export AWS_SECRET_ACCESS_KEY=...
# Set your own aws.keypair_name, aws.security_groups and
# ssh.private_key_path by your favorite editor
$ vi Vagrantfile
# Launch and provision an gpu instance
# (it will take some time)
$ VAGRANT_LOG=info vagrant up --provider=aws
# Connect to the instance
$ vagrant ssh
# Check tensorflow-gpu on the gpu instance
(guest)$ python3 -c "import tensorflow"
> I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcublas.so locally
> I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcudnn.so locally
> I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcufft.so locally
> I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcuda.so.1 locally
> I tensorflow/stream_executor/dso_loader.cc:128] successfully opened CUDA library libcurand.so locally
> ...
# Terminate the instance
$ vagrant destroy- Fork
- Create a feature branch
- Commit your changes
- Rebase your local changes against the master branch
- Create new Pull Request