Alexander Gerstenberger

Hi, sorry for the delayed response. I was not available during September. I remembered today that there was a question regarding `ONNX` exporting. In addition, i also have more information...

> Thanks for the detailed write-up. Maybe you can add this to our documentation or RETURNN wiki? Yes, i can do that > Maybe you can put those scripts also...

I guess mathematically nothing but after conversion it looked like ONNX was doing a `floordiv(a, b)` on the values instead. When changing the formula it was working as intended.

Thanks for your comments. I haven't reported this yet. I noticed that this occurs for integer inputs only. For floating point types this works as expected. Running the following program...

> I don't exactly understand your last comment. In that example, you have integers only, not float, or not? I tried this for floats too, which yields the correct behvaiour...

Here is the issue in tf2onnx: https://github.com/onnx/tensorflow-onnx/issues/2174

I have some conversion issues using the current returnn upstream. Specifically, a conformer model using relative positional encoding with sub-sampling. The same model has been converted to `onnx` a couple...

> How do you do the ONNX conversion? I assume you first create a TF graph, and then convert it somehow? How do you create the TF graph? Do you...

Ah sorry, i forgot. I don't use `rec_step_by_step` option. There is some minor things left: * What would be nice is to have good names for the final states somehow....