[VTA][HotFix] Relay->VTA quantization fix

[tmoreau89]

This addresses a compilation bug introduced by: #4295

Other than interface changes to quantization pass (graph vs. module) it appears that this has broken quantization pass for VTA by inserting multiplication at the end of the layer when multiplication is not supported by VTA (and instead must rely on shift and add).

Investigation in progress.

@vinx13 @ZihengJiang

[tmoreau89]

@liangfu This is the follow up to the fix you requested. Right now quantization breaks compilation to VTA, so this will require further investigation.

[vinx13]

@tmoreau89 Default opt_level has been changed to 2, batch_norm won't be folded during quantization, will this cause issue in vta? You can still wrap quantize call within a build_config scope if that's needed. I did this because running calibration with bn folded caused some accuracy issues

[tmoreau89]

batch_norm won't be folded during quantization, will this cause issue in vta

bingo - that's very likely the root issue which explains why we have multiplication in there

[tmoreau89]

@vinx13 looks like the opt_level fix does the trick. Issuing the fix right now.

[tmoreau89]

Hotfix should be ready for review.

[liangfu]

LGTM

[yzhliu]

Thanks @tmoreau89 @vinx13 @liangfu this is merged and will be included in v0.6.0 release