[core] Modularize transpose op and kernel.

[annxingyuan]

This is a prerequisite for #2955

Something to note is that we can't require the transpose impl functions from backend_webgl or backend_cpu because that would entail a circular dependency, so for now, when we remove a kernel from the backend, we have to import the op itself if we still want to use it within the backend.

Whenever kernels move out of the backend they can directly use the impl functions from other kernels.

Changes

• Modularize transpose op / gradient.
• Modularize transpose kernels for WebGL / CPU backends.

To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.

---

This change is 

[nsthorat]

Reviewed 20 of 20 files at r1.
Reviewable status: complete! 1 of 1 approvals obtained (waiting on @dsmilkov, @lina128, @nsthorat, and @tafsiri)

[nsthorat]

Wait for @tafsiri of course.

[tafsiri]

Great work. LGTM. It would be good to see if we can resolve the type issues that seemed to have cropped up (assuming it was a typing issue).

Reviewed 16 of 20 files at r1.
Reviewable status: complete! 2 of 1 approvals obtained (waiting on @annxingyuan, @dsmilkov, @lina128, and @tafsiri)

---

tfjs-core/src/ops/confusion_matrix.ts, line 87 at r1 (raw file):

  const oneHotPredictions =
      oneHot($predictions.asType('int32'), numClasses) as Tensor2D;
  const oneHotLabelsT: Tensor2D = oneHotLabels.transpose();

I'm curious why you needed to make this change (there are similar changes elsewhere). It it related to type definitions on the new chained augmentor?

---

tfjs-core/src/public/chained_ops/transpose.ts, line 28 at r1 (raw file):

}

Tensor.prototype.transpose = function<T extends Tensor>(perm?: number[]): T {

if you add this: T as the first param here then you do not need to cast the result. https://www.typescriptlang.org/docs/handbook/functions.html#this-parameters

[annxingyuan]

Reviewable status: complete! 3 of 1 approvals obtained (waiting on @dsmilkov and @tafsiri)

---

tfjs-core/src/ops/confusion_matrix.ts, line 87 at r1 (raw file):

Previously, tafsiri (Yannick Assogba) wrote…

I'm curious why you needed to make this change (there are similar changes elsewhere). It it related to type definitions on the new chained augmentor?

Yes I think so.

---

tfjs-core/src/public/chained_ops/transpose.ts, line 28 at r1 (raw file):

Previously, tafsiri (Yannick Assogba) wrote…

if you add this: T as the first param here then you do not need to cast the result. https://www.typescriptlang.org/docs/handbook/functions.html#this-parameters

Done