Does the error "fluid.mlpregressor~: Ouput tap should be > 0 or -1" help me?

Usage Questions
1

Hi there,
While adapting a patch using fluid.mlpregressor~ to predict output with a different number of parameters, I get the error “Ouput tap should be > 0 or -1”. (Also, the object doesn’t output any predictions.)

What does this error mean, and does it help me to debug why the object is not predicting?

When training, I also get a much higher fit value than usual (~15 mostly). I’ve sent the clear message to fluid.mlpregressor~…

2

Perhaps the clear message also causes the object to be reinitialised with some default values other than @arguments passed in the object box. After I sent “reset”, it did what I expected.

3

Hi @leoauri!

Welcome to the Discourse! I would need to hear a bit more about what you’re hoping the MLPRegressor will do for you to know how to answer your question best. Can you explain the goal a bit? tapOut changes which layer (where in the architecture) of the neural network the outputs will come from. The number of parameters it outputs at a tapOut will be set by a hidden layer size, so the number of parameters not directly changeable by tapOut, just which layer. Changing tapOut doesn’t reset the neural network. Check out this article, it may be helpful!

MLP Parameters

Cheers,

Ted

4

Hi Ted, thanks!
I got it to do what I wanted, so let’s just leave this here as reference for anyone getting this same error.
I was just taking a patch using MLPregressor and adapting it to predict outputs of a different dimensionality. I didn’t touch tapOut while doing this.
I was trying to figure out what I need to send the object in order to reset it into a state where I could train it on new datasets, I got this error, and couldn’t make sense of it. I got it to work by sending first “clear” and then “reset” to the object.

Is there an API reference for the object? I couldn’t find one.

Cheers
Leo

5

From what I could tell “reset” didn’t reset the parameters of the network, and I didn’t want a pretrained network. “clear” alone seemed to leave the object in a weird state (that error, no output, weird fit values).

6

Hi @leoauri

Just sending clear should be sufficient: this will throw away what the network has learned, including the inferred input and output sizes. reset is a message common to all the flucoma objects which will reset the object’s attributes to whatever was set in the box with @<whatever>.

I suspect you could end up with this error message if you fit with one size and then try to predict or fit again with different size outputs, but I agree that this could be clearer. I’ll have a look to see how this can improve.

There should be yes. Are you using the pre-built package, or building from source?