How to use double precision in deep learning toolbox?

It seems like deep learning toolbox has defalut set single precision. I want double precision in my neural network, how to achieve this? Please Help!

[Matt J]

Are you saying you want to train in double precision, or just do inference in double precision?

[Rohan Mahala]

Both actually, train as well as inference in double precision.

Answer 1

[Angelo Yeo]

I understand that you want to use trainNetwork and trainnet to train and predict in double precision This is currently not possible in MATLAB R2023b. Internally, 'trainnet' defaults to single precision and hence cannot be used in this case. This functionality is something the developers are actively working on for a future release.

A workaround is to use a custom training loop. You can refer the following example for a general reference -

https://www.mathworks.com/help/deeplearning/ug/train-network-using-custom-training-loop.html

1. Once the learnable network parameters are initialized, they must be converted to type double by using 'dlupdate'

net = dlupdate(@double, net);

2. Ensure data is double precision. This can be set using the 'OutputCast' of 'minibatchqueue' to 'double'. Note that this parameter is read-only. It cannot be modified after construction of 'minibatchqueue'.

Please also note that it is not possible to train in double precision if you are using DAGNetwork or SeriesNetwork.

[Rohan Mahala]

I am not using 'trainnet', instead I am using acclerated custom training loop.

For reference :- https://in.mathworks.com/help/deeplearning/ug/solve-partial-differential-equations-using-deep-learning.html

I did the OutputCast of 'minibatchqueue' to 'double' and now I am getting the predictions in double cast, However i am not able to achieve the desired accuracy.

Can you please suggest me some other pointers to keep in mind so that the training happens correctly in 'double' precision mode only.