FINDR: Flow-field Inference from Neural Data using deep Recurrent networks
The findr package is supported for macOS and Linux. The package has been tested on the following systems:
- Red Hat Enterprise Linux 8.10 (Ootpa)
This package depends on
absl_py==1.4.0
flax==0.6.4
jax==0.4.4
ml_collections==0.1.1
numpy==1.23.5
optax==0.1.4
scikit_learn==1.3.2
scipy==1.12.0
The dependencies can also be found in requirements.txt.
Typical install time is 10-15 minutes. Below we provide an instruction for Princeton's computing cluster Della as an example, but the general procedure should be similar for other systems.
Run the commands below to install flax on Della:
$ ssh <YourNetID>@della-gpu.princeton.edu
$ module load anaconda3/2022.5
$ conda create --name findr python=3.9 matplotlib ipykernel scikit-learn -c conda-forge
$ conda activate findr
$ pip install jaxlib[cuda11_cudnn82]==0.4.4 -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
$ pip install jax==0.4.4
$ pip install flax==0.6.4
$ pip install ml_collections==0.1.1
Then clone https://github.com/Brody-Lab/findr to /home/<YourNetID>/ on Della.
Go to run_findr_array.sh inside findr.
$ ssh <YourNetID>@della-gpu.princeton.edu
$ cd findr
$ nano run_findr_array.sh
Make sure that every tdkim in this file is changed to <YourNetID>. What this shell script does is to identify all data files (stored in .npz format) in datafolderpath and run the session_id-th data file in datafolderpath. The first data file starts with session_id=0, not session_id=1. To get the data files in datafolderpath="/scratch/gpfs/tdkim/findr/npx_luo/manuscript2023a/recordingsessions/2024_04_09", go to mnt/cup/labs/brody/tdkim/manuscript2023a/recordingsessions/2024_04_09 and move the .npz files to any folder you create under /scratch/gpfs/<YourNetID>/. Make sure that the datafolderpath, analysisfolderpath and session_id are correctly specified.
Run run_findr_array.sh using
$ sbatch run_findr_array.sh
This is it! It should take a few hours for all runs to finish.
The data needs to be stored as an .npz file that contains the following keyword arguments:
spikes: contains a 3-d array (# of trials x maximum trial length x # of neurons) of spike counts for each time bin.
choices (optional): contains a 1-d array (# of trials) of the animal's binary choice if the task is a two-alternative forced choice task (leftward = 0, rightward = 1). FINDR does not make use of choices during training. Currently, choices are used only during post-modeling analysis.
externalinputs: contains a 3-d array (# of trials x maximum trial length x input stimulus dimension) where the input stimulus dimension can be an integer greater than or equal to 1. The stimulus values themselves can be floating point numbers or integers. For the Poisson Clicks task, input stimulus dimension is 2, with [0, 0] indicating the time bin with no clicks, [1, 0] indicating the time bin with a left click, and [0, 1] indicating the time bin with a right click.
lengths: contains a 1-d array (# of trials) of the length of each trial (in the unit of time bins).
times: contains a 1-d array (# of trials) of the timestamp of onset of each trial.
notebooks/plot_example_vector_fields.ipynb demonstrates how to generate latent trajectories from FINDR and plot vector fields for the representative session in Figure 2 of Luo*, Kim*, et al., 2023.
- Kim, T.D., Luo, T.Z., Can, T., Krishnamurthy, K., Pillow, J.W., Brody, C.D. (2023). Flow-field inference from neural data using deep recurrent networks. bioRxiv.
- Luo, T.Z.*, Kim, T.D.*, Gupta D., Bondy, A.G., Kopec, C.D., Elliot, V.A., DePasquale, B., Brody, C.D. (2023). Transitions in dynamical regime and neural mode underlie perceptual decision-making. bioRxiv.