lfads

LFADS

This code is for running LFADS baselines for Lorenz and Monkey datasets.

The LFADS code and config files use the format in LFADS-Torch.

To avoid conflicting package installations, we recommend installing LFADS-Torch in a separate environment from LDNS.

conda create -n lfads python=3.9
conda activate lfads
# install from lfads-torch repo
wget https://github.com/arsedler9/lfads-torch/archive/d8b4b3ba87a49fd74a3c06afb3ec1b695c6a2227.zip
# extract the zip file
unzip d8b4b3ba87a49fd74a3c06afb3ec1b695c6a2227.zip
pip install lfads-torch-0.1.0.dev20241204.tar.gz  # ensure this file exists after extraction

Then, we run the scripts in this format:

1. Data preparation

Run the scripts in baselines/lfads/ -- prep_lorenz_data.py and prep_monkey_data.py. This will create HDF5 files in baselines/lfads/data/datasets/ -- lorenz.h5 and monkey_5ms.h5.

2. Run LFADS

Run baselines/lfads/scripts/run_pbt.py to train a model on the monkey and lorenz datasets. Change the DATASET_STR variable to switch between datasets. This runs a population-based training (PBT) search over the hyperparameter space, and the best model is saved separately at the end.

3. Create sampler state dict

Run baselines/lfads/create_unconditional_sampler.py to create a sampler with a behavior readout. This will save the sampler state dict to baselines/lfads/data/samplers/. We will use this in sample_lfads.ipynb to generate samples from the LFADS model.

4. Generate samples from LFADS model

Run sample_lfads.ipynb to generate samples from the LFADS model.

5. Evaluate LFADS model

This is done in LDNS/notebooks under baseline comparisons.