Fix tensor normalization in EmbeddingsPipeline

[chrislee973]

I was directly using the quantized all-MiniLM-L6-v2 onnx model to precompute some embeddings from Python, and I noticed some slight differences between those embeddings and the ones transformers.js generated. I found that the vector norm calculation wasn't squaring the first value of the vector: let norm = Math.sqrt(batch.data.reduce((a, b) => a + b * b)).

Adding an initial value of 0 to the reduce fixes this. I've verified that this fixed normalization calculation results in embeddings equal to the ones generated directly from the onnx model!

[kungfooman]

Simple test for validating the difference:

function aPlusBsquared(a, b) {
  console.log({a, b});
  return a + b * b;
}
[10, 20, 30, 40].reduce(aPlusBsquared);
[10, 20, 30, 40].reduce(aPlusBsquared, 0);

Output:

Basically when no initial value is given, it doesn't start with undefined, but using index 0 and 1 instead.

[xenova]

Whoops! Thanks for that!