Machine Learning on SONM is now live!
Are you involved with neural networking? From now on you can run ML algorithms on SONM directly from docker containers.
There are various tasks for automation of solving complex professional goals one can set for Machine Learning: for example image, video, voice, text recognition algorithms. Generally, a serious task like image recognition takes a long time: one CPU can execute calculations on algorithms for days or even weeks. This becomes an issue for an application efficiency, which is why people want to use several video cards and distribute the execution of the algorithm between them. But where can anyone get a cheap GPU in the era of mining?
The SONM platform provides an opportunity to rent equipment during the time required to solve expensive and complex tasks.
This begs the question: Is it possible to pack everything required to train the Neural Network on a GPU in a docker container and make calculations on SONM’s platform straight “out of the box”?
TensorFlow with GPU support requires CUDA to perform calculations on video cards, so an official container with CUDA and cuDNN libraries is required.
Here are official TensorFlow requirements:
We are packing Nvidia / CUDA into the container to run the application:
But there is an issue…
When you run the algorithm from the container, an error occurs:
This means that you have not installed the CUDA library correctly.
In this case, you would have to spend time finding what exactly is missing in the Nvidia / CUDA container and then fixing this issue yourself.
We have already done it for you. We made a similar experiment, and have faced the same situation. As it turned out, the cuDNN libraries were missing, and we got the error message. After identifying this flaw, we have placed required libraries in sonm/cuda:8.0 container, so now you only need to register them in your DockerFile:
This way, machine learning algorithms run on a machine with a system of six GPUs in a container.
SONM has been designed to make your life a little bit easier, so when we spotted and issue with a missing cuDNN library (Deep Neural Network library) we decided to do just that! You can learn more about the benefits of calculations on out of the box multiGPU system on the SONM platform in our next article.