Albert Zeyer

When it comes to settings (config updates), e.g. the example to change grad accum dynamically, or whatever else, I was also thinking, maybe doing that in one such callback function...

`accum_grad_multiple_step` was extended now, that it can be a callable, which is executed every step.

I don't understand. What is wrong about the first version?

We should better understand this before just making such arbitrary change. Is this a bug in ONNX then? Did you report it? Can you link the issue? Note that your...

I don't exactly understand your last comment. In that example, you have integers only, not float, or not? What is some minimal code which behaves wrong? Or by "observed behaviour",...

Maybe it's a bug of the tf2onnx.convert script, and not ONNX itself? In that case, you probably should report the bug in the [tensorflow-onnx repo](https://github.com/onnx/tensorflow-onnx). Can you inspect the resulting...

So it looks like `Div` is behaving inconsistent to Python/TensorFlow for negative integer numbers? I'm not sure if this is supposed to be like that or not. If it is...

> Here is the issue in tf2onnx: [onnx/tensorflow-onnx#2174](https://github.com/onnx/tensorflow-onnx/issues/2174) A possible workaround for us: ```python def onnx_compat_floor_div(a, b): # https://github.com/onnx/tensorflow-onnx/issues/2174 abs_a, abs_b = tf.abs(a), tf.abs(b) return tf.where( a * b >=...

Why did you reopen? Just to keep track about this until it is fixed upstream? Or is sth supposed to be done on RETURNN side?

@kuacakuaca @Judyxujj any idea? Done that before?