QR-decomposition fails on a specific matrix.

[jarno-r]

To get help from the community, check out our Google group.

TensorFlow.js version

0.14.1

Browser version

Node.js v8.12.0 without tfjs-node.

Describe the problem or feature request

QR decomposition returns NaNs for the invertible matrix [[0,2,2],[1,1,1],[0,1,2]], but works for some other matrices.

Code to reproduce the bug

"use strict";
exports.__esModule = true;
var tf = require("@tensorflow/tfjs");
var a = tf.tensor2d([[0, 2, 2], [1, 1, 1], [0, 1, 2]]);
var b = tf.tensor2d([[1, 2, 3], [4, 5, 6], [7, 8, 9]]);
{
    var _a = tf.linalg.qr(a), q = _a[0], r = _a[1];
    q.print();
    r.print();
}
{
    var _b = tf.linalg.qr(b), q = _b[0], r = _b[1];
    q.print();
    r.print();
}

[jarno-r]

The bug is in linalg_ops.ts:214: const s = rjj.sign().neg() as Tensor2D;
The sign() function returns 0 for a zero value, which is not desirable here. Instead either 1 or -1 should be used.
It could be fixed like this:

const sPre = rjj.sign().neg() as Tensor2D;
const s = sPre.where(sPre.notEqual(tensor2d([[0]])), tensor2d([[1]]));

However, it would be desirable to have positive diagonal for the resulting R and the sign selected accordingly.

[caisq]

@jarno-r Do you want to send us a PR (to https://github.com/tensorflow/tfjs-core) to fix the issue?

[jarno-r]

PR: tensorflow/tfjs-core#1473

[dsmilkov]

Fixed in tensorflow/tfjs-core#1473