TopK with dynamic K

[alexander-camuto]

Models with TopK nodes fail to type.

A sample error:

thread 'main' panicked at 'called `Result::unwrap()` on an `Err` value: Failed analyse for node #717 "TopK_717" Topk

Caused by:
    0: Infering facts
    1: Applying rule inputs[1].datum_type == I64
    2: Impossible to unify TDim with I64.', src/main.rs:12:36

A repo to reproduce the issue (I swear it is public this time): https://github.com/alexander-camuto/topk-issue-repro

To reproduce the issue this time you need to download the model at https://github.com/onnx/models/raw/main/vision/object_detection_segmentation/ssd/model/ssd-10.onnx to ./src/network.onnx

then run

RUST_LOG=trace cargo run 

[kali]

repo is public indeed :)

Given the time you spend struggling with tract, you may want to invest a bit of time in trying the command line. This issue can be repro with:

tract ssd-10.onnx --onnx-ignore-output-shapes

I'm having a look.

[alexander-camuto]

as always ty for the tip :)

[kali]

When I implemented TopK a couple of weeks ago, it was pretty clear that K would always be constant. This "always" was a pretty short-lived one... I'm going to try to do this soon-ish.

[JernKunpittaya]

Hi, we are trying to implement stats function using ezkl, like percentile and median stuffs, so would be nice to have this sort feature :) Thanks!