Use Mxnet-Mkldnn distribution in CI instead of stock mxnet distribution

Development
#1

This pertains to the support QNN dialect for mxnet in this PR#4714. We currently use stock mxnet distrubution in our CI. However, for quantized mxnet networks and to write end to end test cases and peform forward pass on quantized network in tests, we have to use Mxnet compiled with mkldnn. This is because Mxnet suggests and fully supports quantization and quantization based APIs with MKLDNN. The Mxnet-MKLDNN distribution comes with the operators that support calibration and forward pass for quantization. This will also allow us to write end to end test cases for Mxnet with quantization. I suggest we update the docker command pip3 install mxnet==1.5.0 with pip3 install mxnet-mkl==1.5.0. This should in theory not affect any of the exiting functionality and test cases. We can first have a PR the just updates the Mxnet dependency for Docker and have it pass the CI and then have the updated CI to run quantized mxnet test cases. Another option is to dump the quantized network for mxnet and then parse it in TVM to show that TVM parsing works but it will not be able to compare the outputs of TVM and Mxnet quantized network.

#2

@tqchen @yzhliu @zhiics

#3

if we are updating the docker image, please upgrade pytorch to v1.4 (released last week) and torchvision to 0.5 as well. We are currently at pytorch 1.2

#4

this suggestion sounds fair.

#5

Good, then first I’ll create a PR that makes the change for mxnet mkldnn first.