My Experience with CUDAMat, Deep Belief Networks, and Python

My GPU vs. CPU performance training a Deep Belief Network. For some reason the CPU is faster.

Grab the tissues.

This is a tail of my MacBook Pro, a GPU, and the CUDAMat library — and it doesn’t have a happy ending.

Two weeks ago I posted a Geting Started with Deep Learning and Python guide. The guide was great and well received. I had a blast writing it. And a lot of the PyImageSearch readers really enjoyed it.

So as a natural followup to this guide, I wanted to demonstrate how the GPU could be used to speedup the training of Deep Belief Networks (previously, only the CPU had been used).

However, I had never setup my MacBook Pro to utilize the GPU for computational processing. That’s all fine and good I thought. This will make for a great tutorial, with me experimenting through trial and error — and chronicling my steps along the way with the hope that someone else would find the tutorial useful.

The short of it is this: I did succeed in installing The CUDA Toolkit, drivers, and CUDAMat.


For reasons I cannot understand, training my Deep Belief Network with nolearn on my GPU takes over a full minute longer than on the CPU. Truthfully. I’ve tried every option that I can think of. And no dice.

Is it just because I’m on a MacBook Pro the performance of my GeForce GT 750m just isn’t good enough? Is there a problem with my configuration? Am I just missing a switch somewhere?

Honestly, I’m not sure.

So if anyone can shed some light on this problem, that would be great!

But in the meantime I’ll go ahead and writeup the steps that I took to install the CUDA Toolkit, drivers, and CUDAMat.

My Experience with CUDAMat, Deep Belief Networks, and Python on OSX

So before you can even think about using your graphics card to speedup your training time, you need to make sure you meet all the pre-requisites for the latest version of the CUDA Toolkit (at the time of this writing, v6.5.18 is the latest version), including:

  • Mac OSX 10.8 or later
  • The gcc or Clang compiler (these should already be setup and installed if you have Xcode on your system)
  • A CUDA-capable GPU
  • The NVIDIA CUDA Toolkit

The first two items on this checklist are trivial to check. The last item, the NVIDIA Cuda Toolkit is a simple download and install provided that you have a CUDA-capable GPU.

To check if you have a CUDA-capable GPU on OSX, it’s a simple process of clicking the Apple icon at the top-left portion of your screen, selecting “About this Mac”, followed by “More info…”, then “System Report…”, and lastly the “Graphics/Display” tab.

It should look something like this:

The chipset of my MacBook Pro

Under my “Chipset Model” I see that I am using a NVIDIA GeForce GT 750m graphics card.

Let’s go see if the CUDA Toolkit supports this model.

To perform the check, just click here to checkout the list of supported GPUs and see if yours is on the list.

Sure enough, my 750m is listed as supported:

Determining of my GPU is CUDA-capable or not

Awesome. That’s a relief. At least I know that my MacBook Pro is CUDA-capable.

The next step is to download and install the CUDA Driver, Toolkit, and Samples.

To do that, just head to the CUDA downloads page, select your operating system (in my case, OSX), and download the installer:

Downloading the CUDA SDK

At the time of this writing, the most recent version of the CUDA Toolkit bundle is v6.5.18. It’s also a fairly hefty download of 800mb, so be sure you have some bandwidth to spare.

Once the PKG is done downloading, start the install process:

Installing the CUDA ToolkitThe install process for me was very quick and painless — under 30 seconds from the time I opened the installer to the time it completed.

Now that the CUDA Toolkit is installed you’ll need to setup some environment variables:

I like to put these in my .bash_profile  file so that my paths are set each time I open up a new terminal:

Setting up the environmental variables for the CUDA ToolkitOnce you have edited your .bash_profile  file, close your terminal and open up a new one so that your changes are read. Alternatively. you could execute  source ~/.bash_profile  to reload your settings.

To verify that CUDA the driver was indeed installed correctly, I executed the following command:

And received the following output:

Ensuring the CUDA Driver is found

Sure enough, the CUDA Driver as picked up!

According to the CUDA Getting Started Guide for OSX, you should now open up your System Preferences and go to the Energy Saver settings.

You’ll want to uncheck the Automatic Graphic Switch (which ensures that your GPU will always be utilized) and also drag your System Sleep time to Never.

Adjusting System Preferences to use the CUDA Driver

Note: I honestly played around with both of these settings. Neither of them improved performance for me.

Now, let’s go ahead and install the CUDA Samples so that we can ensure that the CUDA Toolkit is functioning and that the driver is working. Installing the samples is just a simple shell script with should be available on your PATH  provided that you have reloaded your .bash_profile .

This script installed the CUDA samples in my home directory. The output of the script below isn’t very exciting, just proof that it succeeded:

Creating the CUDA Samples

From there, I compiled the deviceQuery  and deviceBandwith  examples with the following commands:

Both examples compiled without any errors.

Running the compiled deviceQuery  program I was able to confirm that the GPU was being picked up:

Executing deviceQuery confirms that my GPU was foundSimilarly, the output of deviceBandwith  demonstrates that GPU is accessible and is behaving correctly:

The deviceBandwidth test demonstrates that my GPU is behaving correctlyAt this point I was starting to feel pretty good.

The CUDA Toolkit, drivers, and samples were installed without a problem. I was able to compile and execute the samples with no issue.

So then it was time to move on to CUDAMat, which allows matrix calculations to take place on the GPU via Python, leading to (hopefully) a dramatic speedup in training time of neural networks.

I cloned down CUDAMat to my system and compiled it without a hitch using the supplied Makefile :

Compiling CUDAMatI then ran  to ensure everything was working smoothly — indeed it appeared that everything was working fine.

Now I was really excited!

The CUDA Toolkit was installed, CUDAMat compiled without a hitch, and all the CUDAMat tests passed. I would be reaping the GPU speedups in no time!

Or so I thought…

The Ugly Results

So here we are. The results section.

Training my Deep Belief Network on the GPU is supposed to yield significant speedups.

However, in my case, utilizing the GPU was a minute slower than using the CPU.

In the benchmarks reported below, I was utilizing the nolearn implementation of a Deep Belief Network (DBN) trained on the MNIST dataset. My network included an input layer of 784 nodes (one for each of the input pixels of the 28 x 28 pixel image), a hidden layer of 300 nodes, and an output layer of 10 nodes, one for each of the possible digits. I allowed the network to train for 10 epochs.

I first obtained a baseline using the CPU:

As you can see, I got a warning from gnumpy  (a GPU optimized version of NumPy) indicating that CUDAMat could not be found and that the CPU would be used for training:

A training and evaluation time of just over 1 minute for 10 epochs felt pretty good.

But what about the GPU? Could it do better?

I was certainly hopeful.

But the results left me quite confused:

Over two minutes to train the network on the GPU?

That can’t be right — that’s double the amount of time it took the CPU to train!

But sure enough. That’s what the results state.

The only clue I got into the strange performance was this  gnumpy  message:

I did some Googling, but I couldn’t figure out whether this message was a critical warning or not.

And more importantly, I couldn’t figure out if this message indicating that my performance was being severely damaged by not being able to obtain the gpu_lock .

Sad and defeated, I gathered up my screenshots and results, and typed this blog post up.

Though I hit the gym often, my gains clearly don’t translate to the GPU.

Hopefully someone with more expertise in utilizing the CUDA Toolkit, OSX, and the GPU can guide me in the right direction.


In this blog post I attempted to demonstrate how to utilize the GPU on OSX to speedup training times for Deep Belief Networks.

Unfortunately, my results demonstrated that the GPU yielded slower training times than the CPU — this makes absolutely no sense to me and it is completely counter-intuitive.

Is there something wrong with my configuration?

Did I miss a step along the way?

Surely I made a mistake or my intuition is off.

If you’re reading this post and thinking “Hey Adrian, you’re an idiot. You forgot to do steps X, Y, and Z” then please shoot me an email with the subject You’re an idiot and set me straight. It would be much appreciated.

At the very least I hope this post is a solid chronicle of my experience and that someone, somewhere finds it helpful.


I have found my redemption! To find out how I ditched my MacBook Pro and moved to the Amazon EC2 GPU, just click here.


Click here to download the source code that I utilized in this post.

, , , , , , , ,

20 Responses to My Experience with CUDAMat, Deep Belief Networks, and Python

  1. Softcone October 6, 2014 at 3:19 pm #

    Not any kind of expert, but my intuition is that your demo isn’t wide enough to take advantage of the GPU’s massive parallel architecture. CPU is still much much faster for linear sequence. Try doubling your width or finding a tool to see how many and how active your CUDA cores are. You’ll need more granularity than something like MSI Afterburner will tell you…

    Good luck.

    • Adrian Rosebrock October 6, 2014 at 9:02 pm #

      Thank you for the comment, I’ll definitely give a larger network a try and see how the performance compares.

      • SpeechMachine October 7, 2014 at 12:41 am #

        I did not read through the entire post, but as a GPU user for training nets, here is my observation. The biggest bottleneck to GPU computing is actually a GPU/CPU data transfer. Everytime that happens there is a huge latency. Its primarily because of the PCI/e bus. If your data set is small and your network is small, and you frequently keep swapping between CPU and GPU data structures there will be no perceived advantage…

        • Adrian Rosebrock October 7, 2014 at 6:55 am #

          Thank you for the clarification and advice! There definitely seems to be a common theme here — and that my network is far too small for the GPU speedups to be realized. I will have to train on a much larger GPU and see if the timings improve.

  2. Jairo October 6, 2014 at 11:08 pm #

    You should try with a faster GPU.

    • Adrian Rosebrock October 7, 2014 at 6:54 am #

      Unfortunately, I cannot change the GPU in my MacBook Pro. I am setting up an Amazon EC2 instance and I’ll try a faster GPU on a larger network on there.

  3. JLE October 9, 2014 at 11:23 am #

    I checked my times with your deep learning example:
    Before installing cudamat: 7 sec/epoch (Intel core i7)
    After installing cudamat: 5 sec/epoch (Nvidia GeForce GT 650M)
    Ubuntu 14.04 on a Lenovo Y500

  4. erode October 13, 2014 at 1:10 am #

    Adrian, Do CudaMat or Deep belief networks rely on double precision, floating point operations ? If so that might be a reason. I’m quite disappointed with the GT 750M capability in this regard. Take a look at

    • Adrian Rosebrock October 13, 2014 at 7:12 am #

      Thanks for the link, that really helps clarify the performance of the MacBook Pro GPU. I’ve written a follow up post that will go live later today (Monday, Oct 13th) that details my experiments with the GPU on the Amazon EC2 cloud. The results are substantially better so my previous poor results are definitely due to the MacBook Pro.

      • Ryan November 10, 2014 at 11:53 pm #

        Did you try seeing to float32? Cuda doesn’t do well with float64.

        • Adrian Rosebrock November 11, 2014 at 6:57 am #

          Hi Ryan, thanks for the comment. And yep, I made sure float32 was being used.

  5. Dinesh Vadhia December 18, 2014 at 4:06 am #

    Hi Adrian: I’m a newbie to NNs and have questions about implementation. Using images as the example, after training with a NN better representations will are generated than say, manual feature engineering . What do these representations look like? What is the process to take the representations and to deliver a search engine?

    • Adrian Rosebrock December 18, 2014 at 7:52 am #

      There was actually a very recent paper published called Deep Learning for Content-Based Image Retrieval: A Comprehensive Study that reviews this process. Normally, if you wanted to apply a NN to build a CBIR system, you would use some sort of Deep Learning, stack some unsupervised RBMs together, and learn some feature representations. You then store these feature representations in your database and query against them later.

  6. Abhishek Simhadri February 25, 2015 at 2:45 pm #

    Hi Adrian

    I agree. I got the exact same problem. I guess there are two items to consider.

    1. It would obviously depend on the quality of the GPU vis-a-vis the CPU
    2. Is our code using all the CUDA cores in parallel

    • Adrian Rosebrock February 25, 2015 at 7:37 pm #

      Based on my experiments thus far it depends on (1) the size of your dataset and (2) the size of your network. Both should be substantially “large” to leverage the GPU.

      • Gohu November 9, 2016 at 9:45 am #

        I ma about to try deep reinforcement learning (atari games) on my MacBook Pro equipped with NVIDIA GeForce GT 750M. First I was concerned by your article, but now reading the comments it seems that I was only a matter of data set size and net size. So does it mean that reinforcement learning will work with my GPU ?

        • Adrian Rosebrock November 10, 2016 at 8:38 am #

          GPUs will certainly give you a performance gain over your CPU, but realistically, for any type of current research using deep learning you need a more powerful GPU than what is provided in your MacBook Pro. I would consider spinning up an Amazon EC2 instance.

  7. SPQR September 14, 2015 at 6:11 pm #

    Disabling Automatic Graphics Switching (mux graphics in Apple-speak) is about avoiding OpenGL from switching to your chipset graphics (likely Intel IrisPro) rather than speeding-up your NVidia/CUDA graphics.


  1. Deep Learning on Amazon EC2 GPU with Python and nolearn - PyImageSearch - October 13, 2014

    […] week I wrote a post detailing my experience with CUDAMat, Deep Belief Networks, and Python using my MacBook […]

  2. My Top 9 Favorite Python Deep Learning Libraries - PyImageSearch - June 27, 2016

    […] used nolearn a few times already on the PyImageSearch blog, mainly when performing some initial GPU experiments on my MacBook Pro and performing deep learning on an Amazon EC2 GPU […]

Before you leave a comment...

Hey, Adrian here, author of the PyImageSearch blog. I'd love to hear from you, but before you submit a comment, please follow these guidelines:

  1. If you have a question, read the comments first. You should also search this page (i.e., ctrl + f) for keywords related to your question. It's likely that I have already addressed your question in the comments.
  2. If you are copying and pasting code/terminal output, please don't. Reviewing another programmers’ code is a very time consuming and tedious task, and due to the volume of emails and contact requests I receive, I simply cannot do it.
  3. Be respectful of the space. I put a lot of my own personal time into creating these free weekly tutorials. On average, each tutorial takes me 15-20 hours to put together. I love offering these guides to you and I take pride in the content I create. Therefore, I will not approve comments that include large code blocks/terminal output as it destroys the formatting of the page. Kindly be respectful of this space.
  4. Be patient. I receive 200+ comments and emails per day. Due to spam, and my desire to personally answer as many questions as I can, I hand moderate all new comments (typically once per week). I try to answer as many questions as I can, but I'm only one person. Please don't be offended if I cannot get to your question
  5. Do you need priority support? Consider purchasing one of my books and courses. I place customer questions and emails in a separate, special priority queue and answer them first. If you are a customer of mine you will receive a guaranteed response from me. If there's any time left over, I focus on the community at large and attempt to answer as many of those questions as I possibly can.

Thank you for keeping these guidelines in mind before submitting your comment.

Leave a Reply