Robert Knight

> I tested https://github.com/robertknight/ocrs/pull/100 that I made using the same alphabet I used for training, but I'm still getting incorrect characters during OCR. There might be an issue with the...

> AFAIK this issue hasn't come up in practice yet, but I realized it could with the Gelu operator. That currently has no attributes. For GELU specifically this was addressed...

This was solved for the `Pad` operator by writing a custom `ReadOp` impl which doesn't error out if attributes are missing. See https://github.com/robertknight/rten/pull/326. I'll do the same for other operators...

A related problem is that models sometimes have subgraphs which depend on input shapes but end up producing constant values. In attention for example, where `Attention(Q, K, V) = softmax(QK^T...

ONNX [shape inference](https://onnx.ai/onnx/api/shape_inference.html#infer-shapes) tools can actually work out here that the slice input is `(unknown, 16, unknown, 64)` and should be able to determine that the slice output is 64....

https://github.com/robertknight/rten/pull/805 handles the case where Shape + Slice is used to extract a dimension from a shape, and the value of that dimension is fixed according to shape inference.

Hi @igor-yusupov, thanks for taking a look at this. I'll take a closer look at the weekend.

The diff here is basically the right approach. One thing I am concerned about when adding new data types is the impact on compile time and binary size. To minimize...

> Maybe operations which care about bitwise equality with zero. For floats this is complicated by +0.0 vs -0.0. I suppose we could implement eg. `NonZero` for a given bit...

Hello, just a belated note to say thanks for the PR. I haven't forgotten about this, I've just been focused on finishing some other work first.