carlson-lab / lpne

LFP feature extraction and prediction pipeline

LPNE feature extraction and classification pipeline

Code for preprocessing and building models with local field potentials

See script.py for usage.

Installation

$ git clone https://github.com/carlson-lab/lpne.git
$ cd lpne
$ pip install .
$ pytest test # run tests
$ cd docs
$ make html # build docs

Then see docs/build/html/index.html for the docs.

Dependencies

• Python3 (3.7+)
• PyTorch 1.11+
• MoviePy (optional)

TO DO

9. Add a Tucker decomposition model
10. PoE?
11. Make some pre-zipped features and labels: make TST public
12. Movie app
13. mouse-specific intercepts
14. SMC for sampling label sequence posterior
15. Mouse-specific normalization options
16. Early stopping in GridSearchCV
17. [b,f,r,r] vs [b,r,r,f] shapes
18. automatic groups?
19. More agressive normalization options
20. Add strict channel map checking options
21. Add CSV version of CHANS files
22. save GridSearchCV with validation scores for each parameter setting
23. Also save confusion matrices for GridSearchCV
24. Add __all__ to files
25. Add more model.score options
26. Make pipeline app
27. Fix the docs bugs
28. Make segments files with onset, offset, and label

lpne's Issues

Error with averaging channels

While attempting to make features from the social data, I encountered an issue with averaging the channels of some of the lfp files. The error that occurred was:
'AssertionError: No channels to make grouped channel Cx_Cg!'.

However, there is a channel in the lfp files which, from the channel map, should average into that area.

I've put files to replicate this error on the DCC at /datacommons/carlsonlab/nwl3/error_example. I include the pipeline that I use, as well as two folders containing the files needed to either replicate the error (the error folder) or show that the pipeline works in other cases (the counterexample folder).

Please let me know if you need any more information, thanks.

Add bispectral power decomposition

Implement the pure/interaction spectrum decomposition described in "Simulation of higher-order stochastic processes by spectral representation" by Shields and Kim (2016).

Undefined predictive_order Object In NMF Base Class Pretraining Method

Line 298 in the pretrain_NMF method definitino of lpne/lpne/models/nmf_base.py references an object named predictive_order - or po1 for short - which is not defined within the scope of line 298. There is a loop preceding line 298 (begining on line 231) which references an object named predictive_order - which I will refer to as po2 for clarity - but the loop changes the value of po2 with every iteration, and so it is unclear whether po1 and po2 refer to the same information or not.

The purpose of this Issue is to resolve the question of whether po1 and po2 refer to the same thing which simply needs to be instantiated prior to the previously mentioned for loop, or if instead po1 and po2 refer to entirely different objects which need a new naming convention.

Add online Fourier decompositions

a reference

No empty_cache after pretraining or in .fit method of dcsfa_nmf.py

I was looking over dcsfa_nmf.py today with Karim, and discovered there is no call to empty_cache (esp. for when self.device=="cuda"), which may be the reason why Karim's training is crashing in the first epoch loop after pretraining.

I suggest we add at least one call to empty_cache in the "if pretrain" condition, and consider adding periodic calls to empty_cache during the main training loop (say, if epoch or batch number mod 100 == 0, for example)

Line 113 in data.py should be np.allclose(...) not np.all_close(...)

Not sure if this is a numpy version issue on my end, but I'm seeing that the function call is as stated in the title. https://numpy.org/doc/stable/reference/generated/numpy.allclose.html