ONNX Runtime JS how to use Tensors as multi-dim arrays?

Question

This also would seem something trivial but I've been searching the web and experimenting for a frustrating amount of time.

In onnx runtime web I create a tensor, i.e:

const a = new ort.Tensor('float32', array, [2, 10])

With array being some 20-element float32 array.

I can't seem to find a way to access the data using the standard format a[0] to retrieve an array of size 10. This is quite shocking to me.

The tensor has a dims property with the correct shape and a data property with the 1-dimensional 20 element array.

Within the examples supplied by onnx, I can see them using Array.prototype.slice to access the data, which seems inconvenient and in my mind, defeats the purpose of having a Tensor object to begin with.

So my question is, what's the recommended way of accessing the data using these Tensor instances?

Here is a sample on how I'm accessing the array data to illustrate the problem.

  for (j = 1 ; j < results.logits.dims[1] - 1; j++) {
    const s = j* results.logits.dims[2];
    const e = s + results.logits.dims[2]

    const arr = results.logits.data.slice(s, e)
    const probs = softmax(arr)
    const label = argmax(probs)

I dislike having to calculate manually the access function for the element at the position [i, j].

Thanks a ton for your help!

score 0 · Answer 1 · answered Jul 12 '22 at 15:26

0

Well, Tensors support vectorized actions so accessing a single element is not really a use case.

But more importantly, let's say you have accessed and retrieved a single element, what type is it going to be? The only option in Onnx is a Tensor, so slicing actually makes sense.

Maybe you can specify what you are trying to do?

answered Jul 12 '22 at 15:26

user3689574

1,596
1
11
20

thanks for your answer, I'm trying to write an ArgMax and a Softmax operation, so let's say I have a tensor of shape [1, 10, 12] with resulting logits of a NER classification of 1 sentence of 10 tokens and 12 possible labels, then I can use it for that. – Juan Alberto López Cavallotti Jul 12 '22 at 20:13
Can you maybe provide a code example? You are trying to use Onnx's ArgMax and Softmax? Can you write the code you are trying to create in JS or pseudocode? – user3689574 Jul 13 '22 at 08:57
I will edit the main question. – Juan Alberto López Cavallotti Jul 15 '22 at 06:37

score 0 · Answer 2 · answered Jul 22 '22 at 23:09

0

I am not sure if this will answer exactly your question, but I hope you will find it useful.

According to the official doc, to get softmax which you mentioned in comments you can use:

var outputSoftmax = softmax(Array.prototype.slice.call(output.data));
Properties of tensors and accessing underlying data could be find here

answered Jul 22 '22 at 23:09

flamingo

148
1
11

Hi! Thanks for your answer, I downloaded the code for what they have in the official doc and it's just a softmax function implemented by hand. On the other hand, if you see the sample code I posted I'm accessing the data like that, but they don't provide a multi-dim array syntax, hence my question. – Juan Alberto López Cavallotti Jul 24 '22 at 15:38

ONNX Runtime JS how to use Tensors as multi-dim arrays?

2 Answers2