fetchFunc doesn't work with fully qualified URLs

TensorFlow.js version

2.0.0-rc.4

Describe the problem or feature request

Currently, if you supply fetchFunc to a function like loadGraphModel it will ignore fetchFunc if the URL is a fully qualified URL. For example,

const options = {
  fetchFunc: (...args) => {
    console.log(`Custom fetch`)
    return fetch(...args)
  }
}

let model = await loadGraphModel(`/path/model.json`, options) // Works
model = await loadGraphModel(`https://mydomain.com/path/model.json`, options) // Doesn't work
model = await loadGraphModel(`//mydomain.com/path/model.json`, options) // Works

The issue arises as such:

tfjs first tries to resolve a handler for the URL type. This involves iterating over the known routers. It gets to the HTTP router which will resolve if the following isHTTP check is true:

https://github.com/tensorflow/tfjs/blob/a60f32fbb3d12801c35e464a1f5fd173d575097a/tfjs-core/src/io/http.ts#L42

https://github.com/tensorflow/tfjs/blob/602111aa0858bb46837c2091f4507371525aa3f6/tfjs-core/src/io/http.ts#L252-L254

If tfjs resolves any router it won’t construct a handler with the provided user options:

https://github.com/tensorflow/tfjs/blob/602111aa0858bb46837c2091f4507371525aa3f6/tfjs-converter/src/executor/graph_model.ts#L92-L103

I’m not sure what the reasoning for this is, shouldn’t the options be passed through regardless?

[pyu10055]

@JesseFarebro Now I understand the issue, yes, we will add a fix to update the fetchFunc for the default http loader.

[JesseFarebro]

Thanks @pyu10055, sorry for the confusion.