How can I accelerate deep learning training using GPU?

I've made a simple neural network

It classifies MNIST handwritten digit using fully-connected layers

lgraph_2 = [ ...
    imageInputLayer([28 28 1])
    fullyConnectedLayer(512)
    reluLayer
    fullyConnectedLayer(256)
    reluLayer
    fullyConnectedLayer(128)
    reluLayer
    fullyConnectedLayer(10)
    softmaxLayer
    classificationLayer];

And the options in the neural network is

miniBatchSize = 10;
valFrequency = 5;
options = trainingOptions('sgdm', ...
    'MiniBatchSize',miniBatchSize, ...
    'MaxEpochs',5, ...
    'InitialLearnRate',3e-4, ...
    'Shuffle','every-epoch', ...
    'ValidationData',augimdsValidation, ...
    'ValidationFrequency',valFrequency, ...
    'Verbose',true, ...
    'Plots','training-progress', ...
    'ExecutionEnvironment', 'parallel');

I expected when i use a GPU, it's training speed will be high

But when I train this network using Macbook(sigle CPU)

it takes 1 hour for around 2500 iterations

And when I use my desktop using RTX 2080Ti,

It takes much longer time to train.

MATLAB detects my GPU properly(I checked the GPU information using gpuDevice)

I don't know how can I accelerate the training proess.

Thank you in advance

Answer 1

[Joss Knight]

Your mini-batch size is far too small. You're not going to get any benefit of GPU over CPU with that little GPU utilisation. Increase it to 512 or 1024, or higher (MNIST is a toy network - you could probably train the whole thing in a single mini-batch).

Also, the ExecutionEnvironment option you're looking for is gpu or auto, not parallel. parallel may be slowing things down in your case, if you have a second supported graphics card.

[Ali Al-Saegh]

Hello Knight,

I need to know how things are going on when a GPU is used for deep learning, does the CPU also involved in the training process or any other stuff? It will be great for me if I get some explanation for that!

Also, is it possible to measure the overhead time required for transferring data between memory and GPU?

[Umar]

Hi Ali,

When a GPU is utilized for deep learning tasks, the CPU is involved in managing the overall process, such as loading data, initializing the network, and coordinating communication between the CPU and GPU. The CPU handles tasks that are not suitable for parallel processing on the GPU, ensuring a smooth workflow during training.

To measure the overhead time for data transfer between memory and GPU, you can use profiling tools like MATLAB Profiler or NVIDIA's profiling tools. These tools can help identify bottlenecks in data transfer, allowing you to optimize your code for efficient data handling and minimize overhead time.

Let us know if you need further assistance.

Answer 2

[Shivam Sardana]

Considering CUDA® enabled NVIDIA® GPU with compute capability 3.0 or higher and Parallel Computing Toolbox™ are installed. Consider changing ‘ExecutionEnivronement’ to ‘gpu’. You can refer to the documentation link to see if this helps:

https://www.mathworks.com/help/deeplearning/ref/trainingoptions.html

[NOSHEEN SOHAIL]

I'm almost facing the same issue. my GeForce GTX1080 GPU does not show training progress, no update on a single iteration, single epoch even waited watching 3 days.. or it seems to be much slower than CPU training??

How is this happening, instead of being faster computation, it shows nothing. only big noisy sound from my GPU card is heard but no training plot progresses??

All requirements of CUDA® enabled NVIDIA® GPU with compute capability 3.0 or higher and Parallel Computing Toolbox™ are installed. Done changing ‘ExecutionEnivronement’ to ‘gpu’

please help