initial commit diagPart operator

[kedevked]

FEATURE
add operator diagPart

tensorflow/tfjs#655

Description

---

For repository owners only:

Please remember to apply all applicable tags to your pull request.
Tags: FEATURE, BREAKING, BUG, PERF, DEV, DOC, SECURITY

For more info see: https://github.com/tensorflow/tfjs/blob/master/DEVELOPMENT.md

---

This change is 

[dsmilkov]

Sorry for the really long wait. I left comments related to the changed behavior of tf.diagPart. Also if you can sync with master that would be good.

Heads up, we are in the process of switching to a mono-repo, so you can either:

1. try to finish this PR in the next few days
2. or, you can reopen the PR in the tensorflow/tfjs repo once we move tfjs-core there.

Reviewed 1 of 7 files at r1.
Reviewable status: 0 of 1 approvals obtained (waiting on @dsmilkov and @kedevked)

---

src/kernels/backend_cpu.ts, line 2541 at r1 (raw file):

  diagPart(x: Tensor): Tensor {
    const xVals = x.dataSync();
    const buffer = ops.buffer([Math.sqrt(x.size)], x.dtype);

The size of the output is not necessarily the sqrt(size) since the last output dim is the minimum of the last two input dims: See https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/linalg/diag_part

---

src/ops/diagpart.ts, line 13 at r1 (raw file):

 * input.
 *
 * Assume the input has dimensions `[D1,..., Dk, D1,..., Dk]`, then the output

This is an outdated behavior of diag_part. See https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/linalg/diag_part for the new definition. For an input tensor of rank k, the output is rank k-1 (in other words, all the dimensions except the last two are treated as a batch)