How to tune the hyperparameters of minibatch size to improve the training speed of CNN with 40 CPUs

Hello

I have a ubuntu server with 40 CPUs. I need to train a 5 depth CNN to train 285000 images. I set the minibatch size as 256 and epochs as 5 and the speed is very slow. It spent one day and haven't finished the training in epoch one.

Could I know how to improve the speed?

Thanks

[Ke Zhang]

Is increasing the minibatch to 1000 getting things better? I have searched some articles and it is said the image loading time would take very long. But I'm not sure what the image loading time point to? I thinkn I have used the imageDataStore to load my images.

Thanks

Answer 1

[aditi bagora]

Hello Ke Zhang,

I understand you are currently training your CNN network on an Ubuntu server with multiple CPUs, and the training process is taking longer than expected.

To enhance the training speed, you can leverage the power of multiple CPUs using the hardware support provided in the “trainingOptions()”. By setting the ‘ExecutionEnvironment’ parameter to ‘parallel-cpu', you can make use of the available CPUs effectively.

Here's an example on how to set the ‘ExecutionEnvironment’ parameter:

% Set the option to use CPUs in parallel.

options = trainingOptions("sgdm", MiniBatchSize=256, ExecutionEnvironment="parallel-cpu");

Furthermore, since you mentioned using an ‘ImageDataStore’, it is important to ensure that your data store is partitionable and subsettable. These properties allow for efficient parallel distribution of the data across multiple CPUs. You can use the functions isPartitionable() and isSubsettable() to check whether your data store meets these requirements using the below syntax.

tf = isPartitionable(ds) % returns true if the datastore is partionable.

tf = isSubsettable(ds) % returns true if the datastore is subsettable.

For more details on ‘ExecutionEnvironment’, please refer to the following documentation links:

• Options for training deep learning neural network - MATLAB trainingOptions - MathWorks India
• https://in.mathworks.com/help/deeplearning/ref/trainingoptions.html#:~:text=ExecutionEnvironment%20%E2%80%94%20Hardware%20resource%20for%20training%20neural%20network

Hope the information helps!

Regards,

Aditi