Making your own deep learning workstation: From Windows to Ubuntu dual boot with CUDA and Theano

I recently managed to turn my Windows 7 gaming PC into a dual boot GPU enabled deep learning workstation which I use for deep learning experiments. Here is how I did it.

First of all, you'll want to use Theano as a deep learning framework which automatically optimises your code to use the GPU, the fast parallel processor on your graphics card, which makes deep learning faster (twice as fast on my computer). Unfortunately this seems to be difficult to do on Windows, but easy to do on Linux; so I created a partition for Ubuntu Linux (splitting a hard disk into two separate spaces which can contain two different operating systems) and made my computer dual boot (so that I can choose whether to use Windows or Ubuntu when my computer is switched on). It seems that the most compatible Ubuntu version which most tutorials focus on is Ubuntu 14.04, which is what I installed. I then installed CUDA, which is the NVIDIA graphic card driver that allows you to write programs that use the GPU. After installing this, I installed Theano together with the even easier specialisation framework Keras which lets you do deep learning with only a few lines of Python code.

Be warned that I had to reinstall Ubuntu multiple times because I followed different tutorials which didn't work. Here I will link to the tutorials which worked for me.

Create a partition for Ubuntu

Tutorial
The first step is to create a partition in your C: drive in which to install Ubuntu. I made a partition of 100GB using the Windows Disk Management tool. Follow the instructions in the tutorial.

Install Ubuntu 14.04 into the partition (yes, it is recommended to use 14.04), together with making the computer dual boot

Tutorial
Follow the instructions in the tutorial to use BCDEdit, the Windows boot manager tool, to make it possible to choose whether to use Windows or Ubuntu. Stop at step 5.

To continue past step 5, download the Ubuntu 14.04 ISO file from http://releases.ubuntu.com/trusty/. I choose the 64-bit PC (AMD64) desktop image but you might need the 32-bit one instead. After downloading it, burn it into an empty DVD, restart your computer, and pop in the DVD before it starts booting. Assuming your boot priority list has the DVD drive before the hard disk (changing this if it's not the case depends on your BIOS and mother card but it usually requires pressing F12 when the computer is starting up), you should be asked whether to install or try Ubuntu. Choose to try Ubuntu and once the desktop is loaded start the installation application.

At this point you're on step 6 of the tutorial. Be sure to choose "something else" when asked how to install Ubuntu (whether it's along side your other operating system or remove whatever you already have in your computer). Create a 1GB partition for swap space and another partition with the remainder for the available unformatted space for Ubuntu. You can find instructions on this at http://askubuntu.com/questions/343268/how-to-use-manual-partitioning-during-installation/521195#521195. It's very important that the device for boot loader is set to be the same partition where Ubuntu will be installed, otherwise Windows will not show when you turn on your computer. Do not restart, but choose to continue testing the Live CD instead.

You are now in step 8 in the tutorial. This is where you make it possible to select Ubuntu when you switch on your computer. Here is a list of steps you need:

1. First, you need to mount the Windows and Ubuntu partitions to make them accessible. In Ubuntu there is no "My Computer" like in Windows where you can see all the partitions. Instead you have the drives and partitions shown on the side. Click on the ones whose size matches those of the Windows C: drive and the new Ubuntu partition (after clicking on them they will be opened so you can check their contents). This will to mount the partitions.
2. Next you need to know the partition name of the Ubuntu partition. Click on the Ubuntu icon on the side and search for a program called GParted. Open the program and use the drop down list in the top corner to look for the partition you created for Ubuntu (selecting different choices in the drop down list will show you partitions in different physical hard disks you have). Take note of the name of the Ubuntu partition such as "/dev/sda2". Call this name [Ubuntu partition].
3. Next you need to know the directory path to the Windows partition. In an open folder window (or go on the Files icon in the side), click on "Computer" on the side, go in "media", enter into the "ubuntu" folder and there you will find all the mounted partitions. Find the Windows partition. Call the directory path to this partition icon [Windows partition] (/media/ubuntu/abcdef-123456).
4. Open the terminal (ctrl+alt+T) and then just copy the following line into the terminal. Where it says '[Windows parition]' just drag and drop the partition icon into the terminal and the whole directory will be written for you.
   

   sudo dd if=[Ubuntu partition] of='[Windows partition]/ubuntu.bin' bs=512 count=1

Now you can restart, remove the DVD, and log into your newly installed Ubuntu operating system, update it, and do another restart.

Installing CUDA

Tutorial
This is the part that took the most tries since only one tutorial actually worked well and gave no errors. Follow the instructions in the tutorial. When downloading the CUDA driver be sure to download the .deb file.

Installing Theano with dependencies

Tutorial
Now comes the straight forward part. Unfortunately it's easier to just use Python 2 than to try to make things work with Python 3. The examples in Keras are for Python 2 anyway. Just open the terminal and paste:

sudo apt-get install python-numpy python-scipy python-dev python-pip python-nose g++ libopenblas-dev git

However also install an extra package which will let you test that Theano was installed correctly:

sudo pip install nose-parameterized

Now you can install theano and keras.

sudo pip install Theano
sudo pip install keras

Now whenever you want to run a theano or keras python script and use the GPU all you have to do is write the following in the terminal:

THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python [python script]

You may also want to use Python IDLE in order to experiment with theano, in which case install it like this:

sudo apt-get install idle

and then run it to use the GPU like this:

THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 idle

or do this to debug an error in your program:

THEANO_FLAGS=mode=FAST_COMPILE,device=cpu,floatX=float32 idle

I also recommend this good theano tutorial to get you started.

Has everything worked? Leave a comment if you think I should add something else to this tutorial.