Prediction error
Hi! I'm training models on birdsong but have run into an issue when predicting using my model. I get different annotation proposals on the same file depending on whether I predict on only the "current file" or if I predict on a selected folder. The proposals are radically different from each other.
My installation is: INFO:das.cli: Windows-10-10.0.22631-SP0 INFO:das.cli: DAS v0.32.0 INFO:das.cli: GUI is available. INFO:das.cli: xarray-behave v0.34.2 INFO:das.cli: pyqtgraph v0.12.4 INFO:das.cli: PySide2 v5.13.2 INFO:das.cli: Qt v5.12.9 INFO:das.cli: INFO:das.cli: tensorflow v2.6.0 INFO:das.cli: keras v2.6.0 INFO:das.cli: GPU is not available. INFO:das.cli: INFO:das.cli: python v3.9.15 | packaged by conda-forge | (main, Nov 22 2022, 08:39:05) [MSC v.1929 64 bit (AMD64)] INFO:das.cli: pandas v2.2.1 INFO:das.cli: numpy v1.23.5 INFO:das.cli: h5py v3.9.0 INFO:das.cli: scipy v1.12.0 INFO:das.cli: scikit-learn v1.4.1.post1 INFO:das.cli: xarray v2024.2.0
Hope it makes sense
-Sebastian
Hi,
What you describe sounds like a bug that we should fix! Thanks for raising this issue and sending the version information!! I assume you running this from the GUI, correct? Had a quick look at the code and I can't find anything that might cause the issue. It would help greatly if you could share your model files (*_model.h5 and *params.yml) and the folder with some of the files you want to analyze. I can send you a Dropbox link for uploading the files if that would help. But of course no worries if you do not feel comfortable sharing your data for any reason!!!
I'm running it from the GUI, yes. I'll be happy to send you the files, so I'll happily take the dropbox link :)
Here's the link for uploading the model and the data: https://www.dropbox.com/request/YdcWMXqHDd7P3CS8iM35
Thanks for sharing the examples!! When I run this, I get identical results. Do you band-pass filter the audio when loading? That is not done when running a whole folder (adding this to the GUI is on the list) and predictions with or without filtering will produce different results.
I do band-pass filter when loading. I guess that solves it then! Thanks you so much for the rapid help!
Hi Jan, thanks for solving this! For now I amtrying to solve Sebastian and my problem by filtering our files outside of DAS. I tried to recreate your filter, but don't get the completely same results and exporting the filtered files from DAS as WAV creates a file that is definitely wrong. It looks like it is scaled up outside so all signal is clipped, so I can't directly compare the output. I am filtering with a 5th order butter filter between 100 hz and 22050 using the R-implementation of filtfilt. I took the settings from here: https://janclemenslab.org/das/_modules/das/utils.html#bandpass_filter_song. Do you have a suggestion what I might be missing? Thanks a lot for helping! Iris
Hi Iris, this issue keeps popping up repeatedly. I think we need to add an option to band-pass filter the signal during prediction. Should be quick - will let you know once this is integrated.