Custom Deep Learning Layer with different input and output format occurs error

When I construct the custom deep learning layer with formattable input and output, checkLayer function returns error when the input has "CBT" format and output has "SCBT" format. However, when I change the output with the same "CBT" format as input, the checkLayer function is successful. The error message is

Error using dlarray (line 92)

Number of dimension labels in data format must be greater than or equal to the

number of dimensions of the data array.

I would like to know if maintaining input and output consistency in formatting is a mandatory requirement when implementing custom layers?

[Chuguang Pan]

After requesting technical support, the error is caused by some bugs in R2025a. In the new R2026a release, the checkLayer is successful.

Answer 1

[Umar]

Hi @Chuguang Pan, As mentioned in the comments, this turned out to be a bug in checkLayer on R2025a, confirmed by MathWorks support — not an issue with the layer setup. The docs actually show examples of layers with different input and output formats (like taking in "SSCB" and outputting "CBT"), so what you're doing is supported; checkLayer just didn't handle it correctly in that release. Good to know it's fixed in R2026a. For anyone stuck on R2025a in the meantime, you can sidestep checkLayer and test manually instead — make a dlarray with your input format, run it through predict, and check the output comes back with the format/dims you expect: X = dlarray(rand(numChannels, batchSize, numTimeSteps), "CBT"); Y = predict(layer, X); Not as thorough as checkLayer's full suite, but enough to confirm the layer works while waiting to upgrade.