RuntimeError: Vame.rnn_model()

[lincellularneuroscience]

RuntimeError: Vame.rnn_model()

[JensBlack]

The error did not occur when running the training with the demo file video-1-PE-seq.npy

[JensBlack]

Thank you, Kevin. After the extensive search and the discussion with you via email, I was able to solve that issue.

Apparantly, I forgot to align my data egocentrical which started a whole barrage of problems. I tried several different approaches to rule out any conversion issues but could not find any.
It is weird that a different set of numbers creates that issue, but my transformed data is working.

I want to write down a few things from my search that were weird:
The File-PE-seq.npy file created by your demo code was structure the same, but was dtype object rather than float64.
I was able to create a training set with it which was dtype float64.
The content of the file was shaped like this:
[[546.5120849609375 545.6234741210939 544.8563232421875 ... 329.7113037109375 330.50433349609375 332.00457763671875], ...]
while the demo set only had rounded numbers like this:
[[150. 150. 150. ... 150. 150. 150.], ...]
Both are float and in the same range.

After transforming my original data, I ran in a totally new error that i will discuss in a seperate issue that is also connected to the values of the file. Very interesting!
Again, Thank you for the quick response.
I am still wondering what caused the error.

[pavolbauer]

Hi Jens,
Again, thans for trying out VAME and reporting on issues you found in the code!
The error in the SVD solver probably happened because the use of un-aligned marker time series produces a Z matrix that is ill-conditioned, i.e. contains values spaning many orders of magnitude.
We think about adding a check on the condition number of Z in a future release of VAME in order to ensure exception handling prior to the solver call. Thanks for pointing this out.
Best,
Pavol