https://kotlinlang.org logo
#kotlindl
Title
# kotlindl
l

Luc Girardin

03/03/2023, 10:54 PM
Hello everyone! I am wondering whether KotlinDL supports multi-output regressions, where one predict more than one numerical value? I had a look at OnHeapDataset.create() but it seems to only accept one-dimensional FloatArray as labels. However, predictSoftly() seems to support such output so I am a bit confused!
j

Julia Beliaeva

03/06/2023, 12:49 PM
Hi, yes, currently datasets are single-output only.
predictSoftly
is designed to get a vector of probabilities for classification tasks (so one number for each class), but model output is still one number in this case. However, if you have a multiple-output model trained elsewhere, you can use
predictSoftly
to get a vector output. Also, inference with onnx supports all kinds of output types.
l

Luc Girardin

03/06/2023, 1:26 PM
@Julia Beliaeva Many thanks for the prompt reply. I would love to have everything integrated using KotlinDL as the rest of the software is written in Kotlin. I can do the training elsewhere until this gets implemented in KotlinDL. Do you think this is going to materialize in KotlinDL in the near future (e.g. 12 months) and do you have a recommendation on which software to use for the training part to maximize compatibility?
j

Julia Beliaeva

03/07/2023, 10:47 AM
Sorry, I don't have a specific timeline right now. You can watch this issue for updates https://github.com/Kotlin/kotlindl/issues/254. As for recommendations, you can train your models in PyTorch and then export them to ONNX Runtime format.
26 Views