Tommy van der Vorst

@SludgePhD could you retry with the #95 branch? This fixes a similar issue with Gemm/MatMul. I am tracing BertSQuAD because of similar issues. My approach there is to calculate the...

@SludgePhD my bet would be on the Conv operator in this case - the 'group' operator does not seem to be implemented currently.

@SludgePhD I am also seeing differences in Transpose. See #112. The mediapipe model has four transposes at the end with perm=`0,2,3,1` which might be affected by this.

@SludgePhD can you check if the latest master version fixes your issue? It has fixes for various things, including a particularly nasty bug that caused nodes without names to be...

> So, to avoid conflict as you mentioned, I think that we just need to implement a graph coloring algorithm, with the color representing the buffers. I am not a...

@haixuanTao One other interesting thing is that I believe tract can also do the inference of dimensions (we require onnx-simplifier), might want to have a look at how they are...

There is also an 'official' ONNX test harness, which apparently you can plug your own backend in to test it: https://github.com/onnx/onnx/blob/master/docs/ImplementingAnOnnxBackend.md. Probably requires some Python glue code.

This can be tested fairly easily using [this test](https://github.com/haixuanTao/wonnx/pull/34/files#diff-16611a4a68494d7708e6def00fcee2905e9387a44c541cd177bc72c30d4509ec), by changing `channels` from `4` to `2` and removing the last two elements of the result vector. The test should then...

I got fed up with this and moved the models to the repository itself without LFS. This shouldn't be too much of an issue as it's just a few MB...

Closed by #143