MXNet slowness when creating large graph

[StatMixedML]

Hi,

this is related to the second part of the GluonTS issue reported here:

awslabs/gluonts#607 (comment)

I want to fit a GPVAR model in GluonTS and estimation is of magnitude 1,000 slower than training a DeepAR model. According to @geoalgo & @mbohlkeschneider, mxnet seems to be very slow when creating a large graph.

I have cross-referenced to the original issue on GluonTS, hoping for support.

[ptrendx]

Which version of MXNet are you using?

[StatMixedML]

Sorry, forgot to mention:

• Python: 3.6.7
• mxnet-cu100: 1.4.1 (needed for gluonts to work properly: mxnet>=1.3.1,<1.5.*)
• gluonts: 0.4.3.dev55+g90587ba

[geoalgo]

This issue was caused with this commit in mxnet #11001:

Here is a link to an issue that was raised some time ago:
#13454