[circle2circle] Check file size > 2G with unet

[seanshpark]

Logs of issues with our 4G unet internal model.

This model is converted from ONNX.

• part of #12732

tasks

• [x] resolve model conversion with internal one-import-onnx-ext tool
• [ ] ...

[seanshpark]

Reshape Op shape with value of [0, 32, -1]

• 0 can be treated as 1 --> this is incorrect

from https://onnx.ai/onnx/operators/onnx__Reshape.html#reshape-19

• I'm not sure accepting opset-19 as of now is OK or not
• but anyway, we have a model with opset-17(?) having 0 in shape ... this is weird

what to do

• copy dimension value from input
• but may have to preserve as 0 when there is any issue with this
• but need to sync with our Backends

[seanshpark]

[luci] Shape for output #0 not same
[luci]    8014 [1,4,1,1] vs [1,4,128,128]
ERROR: Optimized graph is invalid

hum...

• tail Ops are Transpose-Conv2D-Transpose
• where Conv2D input has ?x?x?x320 and output has 1x128x128x4
• need to check debug message 1 is really 1
  • looks like it is 1. log code prints -1 for unknown.

; [/conv_out/Conv <1,1,1,4>] = 0x5b849bdc17f0
%3838 = circle.CONV_2D(input: %3837, filter: %18, bias: %781, stride(h,w): 1,1, dilation(h,w): 1,1, padding: VALID, fused_activation_function: NONE)

• need to check Conv2D shape inf

[pcs1265]

I wrote some code that supports dynamic shape inference for Conv2D operations quite a while ago. https://github.com/Samsung/ONE/pull/13950

It was related to Yamnet (https://github.com/Samsung/ONE/issues/13790), but I abandoned it because the height and width were static in Yamnet; So there were no issues related to Conv2D.

In conclusion, Conv2D operations do not support dynamic shapes currently. All dimensions of the ifm shape must be known except for the batch size. If input has unknown dimmension, the dim will be calculated as 0.

Since I can't access detailed info, I might be wrong about this is related to dynamic shapes inference. Could you check if this draft PR(#13950) will help address this issue? (I'm sorry, the commits in the draft PR are hard to read...)

[seanshpark]

@pcs1265 , thanks for the comment! no worry. there are lots of issues with this internal model that may take some time and I'll be working on it. if you like to contribute later after SAFFY program, please be tuned :)

[pcs1265]

We have always appreciated for your kind advice to us! As you mentioned, it can take some time to resolve this issue. I have created an issue regarding dynamic shape support for Conv2D/DepthwiseConv2D, which I expect will help with this problem and is also related to our Yamnet issue.

If you have time, could you advise on one of the three methods I described here? https://github.com/Samsung/ONE/issues/14115 Thank you!