Here are some suggested workarounds:
1. You could use the 'imresize' function instead of 'dlresize'. 'imresize' has an option to use bicubic interpolation. You can extract data from your 'dlarray' object using the 'extractdata' function and then pass it to 'imresize'. Following is a small example on how to do this:
X = rand(4,3);
dlX = dlarray(X)
Xtract = extractdata(dlX)
Xb = imresize(Xtract,2, 'bicubic')
Note that if you already have access to the data that you use to create your 'dlarray' object, you might not need to use 'extractdata'. Instead, you could directly pass that data as input to 'imresize'.
2. When the 'dlarray' is initialized through functions like 'dlconv', it gets some data format properties. You might lose these when using 'extractdata', which might cause errors (such as 'When the input data is an unformatted dlarray or numeric array, you must also specify the 'DataFormat' option') when using functions like 'batchnorm' later on in the code. So, you might have to explicitly ensure that these data format properties are being preserved. You can do this by copying the data format using the 'cast' function:
dlY_new = cast(dlY_new,"like",dlY_old)
Some functions, like 'batchnorm', also let you specify the data format in Name-Value pairs.
3. If 'imresize' does not work, locate the exact parts in your code that lead to slow performance. Here are some options that you can use to find where the code is slower -
- 'tic' and 'toc' functions – you can place these functions in different parts of your code to determine time taken. Following is a link to the documentation: https://www.mathworks.com/help/releases/R2022b/matlab/ref/tic.html
- MATLAB Profiler – the 'profile' function helps track execution time for functions. Following is the link to the documentation:https://www.mathworks.com/help/releases/R2022b/matlab/ref/profile.html
4. There are also some standard tricks that you could use to make MATLAB code run faster. Following is a documentation link that goes through these techniques in detail:
