Output tensor with NaN values

AI Beta Sentis
1

I have a model that produces output tensor of NaN’s, when the inference is done with Sentis. ORT produces tensor of floats. Can you please take a look what may be wrong.

2

Thanks! I was able to figure it out, the problem was Softplus our implementation wasn’t numerically stable.
Pretty easy fix.
Do you want a workaround by having a custom op that overrides our implementation?
In any case the fix will be there for the upcoming version

1 Like
3

Thank you! I would appreciate a workaround until you release the update.
By the way, do you have already an estimated release date for the next version?

4

You can check the CustomLayer sample

[OpImport("Softplus")]
public class StableSoftplusLayer : IOpImporter
{
    public void Import(Model model, OperatorNode node)
    {
         model.AddLayer(new StableSoftplus(node.Name, node.Inputs[0]));
    }
}

and

[Serializable]
public class StableSoftplus : Layer
{
    public override Tensor Execute(Tensor[] inputs, ExecutionContext ctx)
    {
         var X = inputs[0];
         ctx.op.Log(new TensorFloat(1) + op.Exp(op.Sub(X)) + op.Clip(X, 0, float.maxValue);
    }
1 Like
5

@alexandreribard_unity Sorry, but there are errors in the code of StableSoftplus.Execute().
Should Softplus return: Log(Exp(x) + 1) ?

6

Yes, that’s the typical formula. But it’s not numerically stable for your model, hence the NANs.
The one listed above helped with numerical precision

7

I agree, but there are syntax errors in the following line (Sub with one parameter, unclosed parenthesis, …). That’s why I ask either for correct code, or for the math formula :slightly_smiling_face:

ctx.op.Log(new TensorFloat(1) + op.Exp(op.Sub(X)) + op.Clip(X, 0, float.maxValue);
8

Ah sorry for the confusion
Here’s the forumla:

log(1+exp(-abs(x))) + max(x,0)
1 Like
9

This is resolved after the update. Thank you!

10

now that’s what we like to hear :blush:

1 Like