[R] mx.io.arrayiter shuffing is disabled

[jeremiedb]

mx.io.arrayiter shuffling option is no longer working, the option only result in the first and last samples to be swapped. Likely appeared at the same time that mx.runif stopped working as well.

> mx.set.seed(44)
> x = mx.io.arrayiter(data = 1:10, label = 1:10, shuffle = T, batch.size = 10)
> x$reset()
> x$iter.next()
[1] TRUE
> x$value()
$data
 [1]  2  3  4  5  6  7  8  9 10  1

$label
 [1]  2  3  4  5  6  7  8  9 10  1

> 
> mx.set.seed(21)
> y = mx.io.arrayiter(data = 1:10, label = 1:10, shuffle = T, batch.size = 10)
> y$reset()
> y$iter.next()
[1] TRUE
> y$value()
$data
 [1]  2  3  4  5  6  7  9  8 10  1

$label
 [1]  2  3  4  5  6  7  9  8 10  1

[Roshrini]

@thirdwing Can you please take a look?

[anirudhacharya]

@hetong007 any update on this.
Also there does not seem to be a test for "shuffle=TRUE" case - /~https://github.com/apache/incubator-mxnet/blob/master/R-package/tests/testthat/test_io.R#L63-L85
@thirdwing

[anirudhacharya]

This issue is not reproducible on Mac with my RStudio installation. But I am able to reproduce it on one of my EC2 instances.

The only difference in the two installations is the mxnet version -
In Mac I am running

other attached packages:
[1] mxnet_0.10.1

whereas on the ubuntu installation I am running the latest version -

other attached packages:
[1] mxnet_1.2.0

The root cause of the problem is the behavior of mxnet's uniform distribution generator which is internally used for shuffling the arrayiter object.
On the old code, I get the following

> mx.runif(shape=10, min=0, max=1, ctx=mx.cpu())
 [1] 0.04804938 0.65167809 0.77450776 0.67896259 0.21799077 0.29066718 0.08128081 0.97726691 0.04218456 0.40008616

On the latest mxnet v1.2 I get the following -

> mx.runif(shape=10, min=0, max=1, ctx=mx.cpu())
 [1] 1.168184e-35 0.000000e+00 0.000000e+00 2.522337e-44 0.000000e+00
 [6] 0.000000e+00 1.401298e-45 0.000000e+00 6.726233e-44 0.000000e+00

Similar errors exist for the mx.rnorm too in the latest code.

[anirudhacharya]

@sandeep-krishnamurthy please close this issue. It has been fixed.