Why is Matlab deep learning so much slower than Tensorflow in Python

Dear all,
I have created an identical neural net in Matlab and in Tensorflow.
I want to use the Matlab implementation as it is needed in a bigger project.
However, the Matlab implementation is 20x slower to converge than the Tensorflow implementation while also having a higher loss.
Can anyone explain me why this could happen?
I am using a fully connected network with 5 layers and relu activations and use Adam to train the network. The parameters and the network are identical for the implementations.

Thank you in advance!

Best, Steven

[Katja Mogalle]

Could you share some code (MATLAB and Tensorflow). I think this would make it easier for the community to help. I can only imagine that some of the training options are different, but we'd have to take a closer look at the code.

With "20x slower to converge" you mean it takes more training iterations to converge or do you mean the execution time of a training iteration is slower than in Tensorflow?

[Steven]

I will upload the code in the question.

By 20 times slower that I train the model both with the Python code and Matlab code for 50,000 iterations. The Python implementation takes about 1 minute to train whereas the Matlab implementation takes 20+ minutes to train.

So the execution time of a training iteration is significantly slower.

I hope that this clarifies my question.

(P.S. Unfortunately, I cannot upload files for another 24 hours so I will upload the code tomorrow.)

[Steven]

Here is the code that I am using.

Any help is much appreciated!

Answer 1

[atharva]

Hey Steven,

I understand that you are facing an issue where the Matlab implementation is 20x slower to converge than the Tensorflow implementation.

You can try the GPU Coder which has been tested to run about 2X as fast as Tensorflow on average. If you do not have the toolbox, you can test out a trial version to see how the performance compares. There are instances where GPU Coder code runs multiple times faster than MATLAB code. An example of this is shown in the video from the following documentation page:

https://www.mathworks.com/products/gpu-coder.html#generate-cuda-deep-learning

For an introduction to using the GPU Coder for deep learning, please see the following documentation pages:

https://www.mathworks.com/products/gpu-coder.html

https://www.mathworks.com/products/gpu-coder.html#generate-cuda-deep-learning

Please note that there are some prerequisites for installing GPU Coder. For a complete list, please see the documentation page below:

https://www.mathworks.com/help/releases/R2020b/gpucoder/gs/install-prerequisites.html

Please note that many factors affect the run time, including the GPU, the original MATLAB code, and the optimization settings selected in GPU Coder, so we cannot guarantee a speed up from using GPU Coder instead of the Deep Learning Toolbox.

I hope this helps!

[Steven]

Hi atharva,

Thank you for your answer.

Unfortunately, for this project I cannot use a GPU as the project is part of a bigger project which will run on a distributed system using many (10000+) CPU's.

However, thank you for your suggestion!