Mini batch size changing value during gradient descent

Hello everyone,

I am currently working on multimodal deep learning, with a neural network classifier receiving two time-dependent inputs, videos and a set of given features. Videos are 4D matrixes of size width x height x depth x frames and features are 2D matrixes of size number of features x frames.

I've been trying to classify the inputs based on the examples given below, as on some of my previous work.

https://nl.mathworks.com/help/deeplearning/ug/classify-videos-using-deep-learning-with-custom-training-loop.html

https://nl.mathworks.com/help/deeplearning/ug/train-network-on-image-and-feature-data.html

During my training, I have come across a very singular situation. The value of minibatchsize, which I had initially set to 16, was decreased to 9. This produced an error as the layer were expecting batch sizes of 16 in the dlfeval() function.

I haven't found anything related to this problem on here, I was wondering if any of you would have a piece of advice or a solution for me.

Thank you for your help !

Answer 1

[Shubham]

Hey @Arthur CASSOU,

I understand that while training the neural network you found that minibatch size which was initially set to 16 was later decreased to 9.

Please check whether the dataset being used has the total number of samples divisible by 16. This would ensure that all samples are used and the minibatch size is not adjusting automatically to accommodate remaining samples. The minibatch size is also dependent on the available memory. Try looking for any inconsistencies in data preprocessing or the network architecture.

You may try using Mini-Batch datastore for reading data in batches.

Please refer to the following:

https://www.mathworks.com/help/deeplearning/ug/develop-custom-mini-batch-datastore.html

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

Hope this helps!!