trainnet gives training loss is NaN

Hello, I'm currently working on semantic segmentation with Unet architecture on matlab. As version R2024a, I tried to train my model with the trainnet command. But after I ran my script, it gives me this result in Command Window.

I tried to change the MiniBatchSize and MaxEpoch but none seem to be working, it seemed like the training never happened because my GPU doesn't seem to have any activity. Does anyone know how to resolve this? Is Matlab R2024a still buggy so this happens?

Answer 1

[Maneet Kaur Bagga]

Hi Aliya,

I understand that you are encountering an issue where the training loss is "NaN", causing the training to stop.To debug the issue please refer to the following steps:

• Enable verbose output in your training options to get more detailed information about each training step.
• Verify that your MATLAB installation is correctly configured to use the GPU. You can use the following command to check the status of your GPU.

gpuDevice

Please refer to the the following code snippet to set your training options:

options = trainingOptions('adam', ...
    'InitialLearnRate', 1e-4, ...
    'MaxEpochs', 50, ...
    'MiniBatchSize', 16, ...
    'Plots', 'training-progress', ...
    'Verbose', true, ...
    'ExecutionEnvironment', 'gpu');  % Ensure you specify 'gpu' if you have a compatible GPU

I hope this helps!

Answer 2

[Shreeya]

Hey Aliya

I see that you are not able to train your UNET model due to Nan loss. There are a few troubleshooting methods you can try:

1. Try to fiddle around with the learning rate on a smaller dataset to ensure if this is or isnt the root cause.
2. If the image size is huge, you may need a bigger network to converge the model.
3. I also came across an interesting take on teh usage of PNG images for model training. Essentially, PNG image have layers. During training, only the last layer maybe used, and thus the model learns nothing related to the features of rest of the images.

I'm also linking a few threads which you can refer to, apart from these suggestions:

https://www.mathworks.com/matlabcentral/answers/337587-how-to-avoid-nan-in-the-mini-batch-loss-from-traning-convolutional-neural-network

https://www.mathworks.com/matlabcentral/answers/1917165-training-loss-is-nan-deep-learning

Answer 3

[Joss Knight]

Do your network weights contain NaNs? Try this

nansInMyNetwork = ~(all(cellfun(@allfinite, net.Learnables.Value)) && all(cellfun(@allfinite, net.State.Value)))

You might also want to check the Variance on any batch normalization layers to make sure none of the values are negative.

Answer 4

[Matt J]

You need to provide more information about what you did, e.g., the training code. However, I can hazard a guess. I imagine problem occurs when you use the 'sgd' training algorithm, but it might not occur if you instead use 'adam'.

Answer 5

[BIPIN SAMUEL]

Hi @Al,

Have you utilized any custom layer in your network? May be the learnable parameters of the layer not intialized properly so that the loss (scalar value) calculated during the first iteration is very large. Also can you mention which loss function you have used?