Doing Deep Learning on GeekCloud

I was recently working on an image-related deep learning assignment from my professor. After debugging the code, I found my computer didn’t have enough memory (an 8GB laptop), and later I discovered a really handy deep learning cloud service platform, GeekCloud: http://www.jikecloud.net/

GeekCloud platform homepage
The homepage of the GeekCloud deep learning cloud service platform

The biggest convenience of this platform is that it comes with many computing frameworks built in. Every time I had to set up the various environments before, it was a real pain—the dependency relationships between all the libraries were such a headache. The platform also offers many machines with different configurations to choose from, and the prices are quite fair. Today, after I used a machine with a Tesla P100 GPU and 60GB of memory, the bad_alloc problem I’d run into before never showed up again.

GeekCloud built-in computing frameworks
The platform’s built-in deep learning computing frameworks and selectable machine configurations

How to Use It

Before training, you can first upload the data you’ll be using. My own dataset was around 10GB, and it took a bit more than one night to upload. This platform supports resumable data uploads, and you aren’t charged while uploading. By default, the data goes into the /data directory. It’s best to upload a compressed archive first and then decompress the data into the /input directory on the server—this is the fastest, because it seems /data is on a mechanical hard drive while /input is on an SSD.

Once the data is uploaded, you can create a server instance. It provides an SSH login command so you can just log in remotely and operate it directly. You can also use Jupyter Notebook and TensorBoard, which is very convenient. You can switch between dedicated and shared modes, or swap in a more powerful server.

Server instance management interface
The management interface after creating a server instance, supporting SSH, Jupyter Notebook, and TensorBoard

Let’s take a look at the configuration of my machine. A Tesla P100 GPU, which goes for over 50,000 yuan on JD.com…

Tesla P100 GPU information
Tesla P100 GPU configuration details

A 4-core, 8-thread Xeon E5.

Xeon E5 CPU information
4-core, 8-thread Xeon E5 processor configuration details

60GB of memory.

60GB memory information
The server’s 60GB memory configuration details

One thing to note: after the server is shut down, the data in /input and /output is cleared. Before shutting down, you need to move any data you need from these two directories into the /data directory, so that it stays saved even after shutdown.

  • Technology