Skip to content

Latest commit

 

History

History
48 lines (30 loc) · 2.72 KB

README.md

File metadata and controls

48 lines (30 loc) · 2.72 KB

Reproducing the Figures

Dataset Figure Notebook
GLVM Figure 2 & S2 glvm/02_GLVM_paper_figures.py
Supp. Figure S1 glvm/R02_GLVM_supp_figures_mask_size.py
Supp. Figure S3 & S4 glvm/R04_GLVM_supp_figures_training_set_size.py
Fly Figure 3 fly/05_FLY_paper_figures.ipynb
Monkey
Figure 4 monkey/06_MONKEY_Decoding.ipynb
Figure 5 & S5-S8 monkey/07_MONKEY_Encoding.ipynb
Figure 6 monkey/08_MONKEY_LatentAnalysis.ipynb

Note: Figure 1 is just a schematic and thus does not show up here.

The datasets can be found on Drive. Please load them into the data folder in the base directory or adjust the paths accordingly.

This contains a modified format of the data provided by O'Doherty, J. E., Cardoso, M. M. B., Makin, J. G., & Sabes, P. N. (2017). Nonhuman Primate Reaching with Multichannel Sensorimotor Cortex Electrophysiology [Data set]. Zenodo. https://doi.org/10.5281/zenodo.583331 published under a Creative Commons Attribution 4.0 International Licencse.

Additional Information for running the Gaussian Latent Variable Model & Analyses

You can either run the respective files in an interactive window, e.g. in VS Code or via commandline, e.g.:

ipython 02_GLVM_paper_figures.py

Prior to running any of the experiments, first generate the GLVM training, test and validation dataset, that will then be used when running ../scripts/run_GLVM.py.

Run the ipython files in the following order:

  1. generate the dataset using 00_GLVM_generate_dataset.py
  2. start the actual training run by starting the bash script bash run_GLVM_many_seeds.sh in the bash folder or simply via running python scripts/run_GLVM.py from the base directory.
  3. aggegrate the data across many runs using 01_GLVM_data_aggregation.py
  4. Finally, you can run the paper analysis with 02_GLVM_paper_figures.py
  5. Carry out analogous steps for the supplementary analyses starting with R_

Training Example

To gain an intuition for the steps required when training masked VAEs, see 03_GLVM_example_masked_VAE.ipynb.

In general, we advise scientists to use their VAEs they have set up for a specific dataset, train them on all observed data first and only once the fully observed VAE trains well, define the desired masks and start with the masking scheme.