Regarding NonZero + ROIAlign

[anujgupt-github]

I've seen several models involving ROIAlign, which invariably are preceded by NonZero. NonZero being a dynamic shaped tensor isn't supported in GLOW.

How about we can implement ROIAlignV2 with a 'fused' functionality of NonZero and ROIAlign,? Basically, we can eliminate NonZero itself by incorporating the logic inside the ROIAlign kernel itself (as done in NonMaxSuppression by pruning out results (a.k.a scores) based on 'score_threshold'

The output of ROIAlignV2 can then be capped at 'max_rois' (passed as an attribute). The output can have 2 fields:

1. Y : T1 - ROI Pooled output
2. Num_ROIs - Defining the number of valid ROIs in the Pooled output ( range [0, max_rois])

If this is an acceptable approach, I can move forward with this.

Suggestions? @jfix71 @mortzur

[anujgupt-github]

Hi @jfix71 @mortzur , Any feedback on this?

[jfix71]

Really sorry for the major delays here, I've been super swamped. I will definitely get to reviewing next week.

[anujgupt-github]

Sure @jfix71 .

[anujgupt-github]

Any thoughts on this, @jfix71 ?

[jfix71]

@mortzur @jackm321 Do you all have thoughts here?