Albert Zeyer

Do you expect some RETURNN bug here? Usually such nan/inf issues, or convergence issues are user errors. I see you have `"set_sparse_dim": 20` for `segment_lens_target`. And then: ``` "length_model": {...

@robin-p-schmitt Did you ever check this? What was the result? Or is this not relevant anymore for you? Then let's close this issue.

I was thinking a bit further about this. The problem is that `self.params` of a layer is somewhat ambiguous (or rather not really well defined): * Should we guarantee that...

Yea, but while `tf.identity` is cheap, I would avoid it, esp just to have it as a workaround. But we don't need some workaround like that. First of all, it's...

> I also don't really know, apart from transposing matrices I have not used this before. Intuitively I would say `layer.params` should exactly contain what `add_param` returns (i.e. the transformed...

Note that `find_sgemm_libs_from_runtime` does not necessarily take the BLAS from NumPy. I assume if multiple libs in the runtime provide BLAS functions (e.g. TensorFlow loaded MKL + NumPy loaded sth...

Why do you want to mix BLAS anyway? Why not just MKL consistently everywhere, including in NumPy? This is probably better anyway and also would solve your problem here.

You mentioned another problem with RASR, that RASR also loads both lapack + MKL, and that also causes non-deterministic behavior which BLAS function will be executed. Does it matter in...

> > Note that `find_sgemm_libs_from_runtime` does not necessarily take the BLAS from NumPy. I assume if multiple libs in the runtime provide BLAS functions (e.g. TensorFlow loaded MKL + NumPy...

However, the code you are referring to, which moves NumPy libs to front (`# Prefer Numpy; move to front.`), this maybe should be changed. Did you try to do this?...