[TFLite] Using real image for QNN testing.

[anijain2305]

This PR does 2 things

• Test QNN on a real image. QNN and TFLite outputs differ due to rounding differences, making random input not a good candidate for label testing. So, this PR adds a real image for testing QNN.
• @inadob also found a bug in QNN parsing [Frontend][TFLite] Fix quantized pad value for convolution #4807 This PR needs that fix as well to enable the real image tests, so adding that fix here.

Thanks to @inadob for finding the bug

@FrozenGene @kevinthesun @inadob Can you please review?

[anijain2305]

@inadob Can you please review the TFLite pad test portion? I am not very familiar with using fake quant, and I just replicated your code for Pad. So, closer scrutiny would be good.

[anijain2305]

Ping @FrozenGene

[FrozenGene]

ping @inadob , would you review it again? If you have no other comments, would you set explicit approve? Thanks?

[inadob]

Almost ready, have a couple of small changes below

[anijain2305]

@inadob Done

[anijain2305]

@FrozenGene Good to merge?

[FrozenGene]

Thanks @anijain2305 @inadob It is merged now.

[masahi]

What happen if the zero point is a vector, as in per channel quantization? What should the pad value be?

[u99127]

What happen if the zero point is a vector, as in per channel quantization? What should the pad value be?

@masahi good question but please file a separate issue only affecting master since per channel quantization came in after this. Otherwise I fear this will be missed.

Ramana

[u99127]

@anijain2305 - this needs a backport to 0.6 tag ?

[masahi]

@u99127 I figured the answer myself. QNN support for per channel quantization applies to only weight quantization (which is supposed to be a common restriction in other frameworks as well). For pad value we can use a scalar zero point associated with activations.

[anijain2305]

@masahi You are right (as always :) )

@u99127 Let me look into this.