The following tutorials are going to be written in F# and Cuda. Before we can start here is what is going to have to be installed:
- F#. I cannot advise on how to get this on Linux, but on Windows the easiest option would be to download Microsoft Visual Studio 2015. I would recommend instead of getting the web installer, that you scroll down a little and the click Visual Studio 2015 on the menu bar to the left – right under “Visual Studio Downloads”. From there you will be able to get the ISO. It is around 3.8Gb, but having it on hand should make it more convenient to know how the installation is progressing. The web installer won’t tell you the full size of the file, or even the % installed, so if you have a slow connection you will be left wondering whether the thing is running at all.
- In VS2015 there will also be a install step inside the IDE as F# does not come prepackaged for it.
- Cuda 7.5 SDK. That can be gotten on Nvidia’s webpage. For the same reasons as above, I’d prefer the local version. 1Gb roughly.
- CuDNN 3.0. This can be gotten here, but requires registering with Nvidia developer program which could take a day or two at most. The reason for that is that Nvidia does manual review of applications, but I wrote on mine that my dream is to use my GTX 970 to make money and Skynet, so the requirements can’t possibly be too demanding. The Spiral library does not currently posses convolutional functions, but does use AddTensor for vector matrix broadcasting addition. The size of the library is miniscule. After installing it just copy the files into a single folder and add the folder to PATH.
The rest of the dependencies are easily installed using the Nuget package manager. They are:
- ManagedCuda 7.5 – The wrapper library for Cuda library functions.
- ManagedCuda-CudaDNN – The wrapper library for cuDNN
- Fsharp.Charting – For plotting cost curves.
I might something extra in order to assist visualizing Mnist digits later, but for now this is decent.