Get top K indices

Unity Engine
, , , , ,
1

Hi there,

Could someone please point me in the correct direction and the Functional method to use when trying to find the top K indices? Something like an argsort? I want to find the top 4 elements by size and their indices from a FunctionalTensor.

1 Like
2

Did you look at Functional.TopK?

https://docs.unity3d.com/Packages/com.unity.sentis@2.1/api/Unity.Sentis.Functional.html#Unity_Sentis_Functional_TopK_Unity_Sentis_FunctionalTensor_System_Int32_System_Int32_System_Boolean_System_Boolean_

It sounds like it might be what you are looking for.

3

I’ve already tried that, it returns the Top K elements and not the indices, there’s no way to find out their indices if I’m not mistaken.

4

The return value of Functional.TopK is an array of FunctionalTensors. The first entry is the output values and the second is the indices.

image
image715×97 2.48 KB

1 Like
5

Got it, thanks!

6

Hi, even after multiple attempts to try and get the TopK to work for the BlazeHands, I couldn’t figure it out because it’s very difficult to debug the code, could someone please help out?

This is the original function ion the BlazeUtils.cs file for BlazeHands:

 public static (FunctionalTensor, FunctionalTensor, FunctionalTensor) ArgMaxFiltering(FunctionalTensor rawBoxes, FunctionalTensor rawScores)
    {
        var detectionScores = ScoreFiltering(rawScores, 100f); // (1, 2016, 1)
        var bestScoreIndex = Functional.ArgMax(rawScores, 1).Squeeze();

        var selectedBoxes = Functional.IndexSelect(rawBoxes, 1, bestScoreIndex).Unsqueeze(0); // (1, 1, 16)
        var selectedScores = Functional.IndexSelect(detectionScores, 1, bestScoreIndex).Unsqueeze(0); // (1, 1, 1)

        return (bestScoreIndex, selectedScores, selectedBoxes);
    }

I am trying to get the top 2 scores instead of only the best one, here’s what my function looks like:

public static (FunctionalTensor, FunctionalTensor, FunctionalTensor) ArgMaxFiltering(FunctionalTensor rawBoxes, FunctionalTensor rawScores)
    {
        var detectionScores = ScoreFiltering(rawScores, 100f); // (1, 2016, 1)

        // Use TopK to get the top two scores
        var topTwoScoreIndexes = Functional.TopK(detectionScores, 1, 1);

        // Select the top two boxes and top two scores
        var bestBox1 = Functional.IndexSelect(rawBoxes, 1, topTwoScoreIndexes[0][1].Squeeze()).Unsqueeze(0);
        var bestScore1 = Functional.IndexSelect(detectionScores, 1, topTwoScoreIndexes[0][0].Squeeze()).Unsqueeze(0);
        var bestBox2 = Functional.IndexSelect(rawBoxes, 1, topTwoScoreIndexes[1][1].Squeeze()).Unsqueeze(0);
        var bestScore2 = Functional.IndexSelect(detectionScores, 1, topTwoScoreIndexes[1][0].Squeeze()).Unsqueeze(0);
        Debug.Log(bestBox1);

        ............
    }

Even after trying a lot of modifications, I keep getting errors regarding FunctionalTensor has incorrect type.

Cheers.

7

The TopK returns an array with two tensors, the first is the scores, the second is the indices. That means ‘topTwoScoreIndexes[0]’ is a float tensor consisting of scores, you don’t need to index the scores tensor with this again. Lines 9 and 10 look wrong because you are trying to index with a float tensor (of scores). Lines 10 and 12 look unnecessary because the scores are already returned from the TopK method as the first tensor.

Hope that clears things up.

1 Like