mxnet 1.5.0 compatibility

[lostella]

Description of changes: MXNet requirement was set to <1.5.* in #215 due to some incompatibility with mxnet-1.5.0. By running tests, it seems like the only problem was in the code for WaveNet. This PR updates the problematic lines to make our WaveNet implementation compatible with mxnet-1.5.0.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[lostella]

Errata: it looks like there's more issues than the one I addressed in the initial commit (despite the fact all tests are passing locally :-/). Diving deeper.

[lostella]

I'm not sure what's running in the tests: these logs mention at some point

# packages in environment at /var/lib/jenkins/workspace/workspace/gluon-ts-cpu-py3/conda/cpu/py3-master:
#
# Name                    Version                   Build  Channel
[...]
mxnet-mkl                 1.4.1                     <pip>
[...]
numpy                     1.14.6                    <pip>

But if I locally install it the same way as in CI (using pip install -v -e ".[shell,Prophet]", see here) then I get mxnet-1.5.0 and numpy-1.17.0.

Could it be some caching messing up with this? @szha

[lostella]

Nevermind: I think this problem is caused by #131

[lostella]

There seem to be an issue with sampling on mxnet-mkl, see here: http://ci.mxnet.io/blue/rest/organizations/jenkins/pipelines/GluonTS-py3-cpu-unittest/branches/PR-245/runs/7/nodes/81/steps/126/log/?start=0

[szha]

@reminisce

[lostella]

Related mxnet issue: apache/mxnet#16135

[strawberrypie]

The fix for the issue mentioned above is merged into mxnet: apache/mxnet#16139
Sadly, it was done after the release of 1.5.1: https://github.com/apache/incubator-mxnet/releases/tag/1.5.1
What is the plan moving forward regarding mxnet version? Updating to mxnet 1.5 will allow the project to remove Trainer class and replace it with gluon.contrib.Estimator — which supports things like early stopping.

[lostella]

@strawberrypie good point, that’s something which is worth looking into. Although I think that’s kind of an experimental API of MXNet’s? Another thing that we’ll look into is the numpy-compatible arrays API, see apache/mxnet#14253

[lostella]

Btw I think this PR can be closed in favor of a new one whenever the mxnet version can be bumped