Multi output regression

[laurenskz]

Is it currently possible to perform a regression with multiple outputs. I.e a single instance of the dataset looks as follows:
x = [0.7,0.6,0.3]
y = [0.2,0.3]

This would seem impossible given the current design of the Dataset interface where the y value always consists of a single float.

A possible use case is using a neural network for predicting q values in reinforcement learning (Deep q learning). The keras API currently allows this.

[zaleslaw]

Could you please share a Keras example for Deep Q Learning with multiple outputs for regression, hope to fix it in the 0.4 or 0.5 release of KotlinDL

[laurenskz]

Thanks for your reply! This is a really common approach for implementing a deep Q network:
https://keras.io/examples/rl/deep_q_network_breakout/

[zaleslaw]

Thanks for the example you gave me.
Multiple outputs for models are not supported yet, but I agree that this is a good idea

[zaleslaw]

Added a note for a possible solution to this problem: the following example should work.

The current OnHeapDataset supports only y as FloatArray but should be Array like x field. At this moment the FloatArray converts to Array in internalFit method to create a Tensor.

OnHeapDataset could have two constructors: with FloatArray and Array

[kuhnmi]

I would like to use the kotlindl library for a deep-q-learning task, but am currently struggling with exactly this issue. Are there any plans to extend the dataset definition as outlined by @zaleslaw in #254 (comment) in the near future to make this possible?