Add inTopK op

[syt123450]

FEATURE

This PR add inTopK op, which behaves the same way as tf.math.in_top_k in TensorFlow. This op help develop further metrics which depend on inTopK operation, such as, topKCategoricalAccuracy (feature requested in tensorflow/tfjs#27 ), sparseTopKCategoricalAccuracy (feature requested in tensorflow/tfjs#26). Relative PR tensorflow/tfjs-layers#537

This PR:

• Add new inTopK op to src/ops
• Register inTopK in src/ops/ops.ts
• Add inTopK kernel to backend
• Add shared inTopK implementation between webgl and cpu
• Add relative tests for inTopK

Reference:

• TensorFlow in_top_k doc
• TensorFlow in_top_k implementation
• TensorFlow in_top_k test cases

---

This change is 

[nsthorat]

Reviewed 2 of 8 files at r1.
Reviewable status: 0 of 1 approvals obtained (waiting on @syt123450)

---

src/backends/webgl/backend_webgl.ts, line 1323 at r1 (raw file):

  inTopK<T extends Tensor, U extends Tensor>(
      predictions: T, targets: U, k: number): U {
    const predictionsVals = predictions.dataSync();

We made a mistake with topk, it actually should be async. Can you make these call .data() instead and return a Promise?

[nsthorat]

Reviewed 1 of 8 files at r1.
Reviewable status: 0 of 1 approvals obtained (waiting on @syt123450)

---

src/backends/webgl/backend_webgl.ts, line 1323 at r1 (raw file):

Previously, nsthorat (Nikhil Thorat) wrote…

We made a mistake with topk, it actually should be async. Can you make these call .data() instead and return a Promise?

Oh hm... we can't make the kernel's asynchronous.. let me think about this.

[syt123450]

Hi @nsthorat , it seems that in backend_cpu, this.readSync is widely used to synchronously read data, so I replaced the previous .dataSync() with this.readSync, and all tests passed. While in backend_webgl, the .dataSync() still exists in nonMaxSuppression, floatPrecision, where, topk, if you figure out how to solve this synchronous call, please let me know, I would apply your way to fix it.

[nsthorat]

I'm going to merge this and send a follow up to remove this as a kernel, apologies. This will be fine for now.

Thanks for the PR!