A TensorFlow 2.0 implementation of LFADS and AutoLFADS.
Clone the autolfads-tf2 repo and create and activate a conda environment with Python 3.7. Use conda to install cudatoolkit and cudnn and pip install the lfads_tf2 and tune_tf2 packages with the -e (editable) flag. This will allow you to import these packages anywhere when your environment is activated, while also allowing you to edit the code directly in the repo.
git clone git@github.com:snel-repo/autolfads-tf2.git
cd autolfads-tf2
conda create --name autolfads-tf2 python=3.7
conda activate autolfads-tf2
conda install -c conda-forge cudatoolkit=10.0
conda install -c conda-forge cudnn=7.6
pip install -e lfads-tf2
pip install -e tune-tf2
The first step to training an LFADS model is setting the hyperparameter (HP) values. All HPs, their descriptions, and their default values are given in the defaults.py module. Note that these default values are unlikely to work well on your dataset. To overwrite any or all default values, the user must define new values in a YAML file (example in configs/lorenz.yaml).
The lfads_tf2.models.LFADS constructor takes as input the path to the configuration file that overwrites default HP values. The path to the modeled dataset is also specified in the config, so LFADS will load the dataset automatically.
The train function will execute the training loop until the validation loss converges or some other stopping criteria is reached. During training, the model will save various outputs in the folder specified by MODEL_DIR. Console outputs will be saved to train.log, metrics will be saved to train_data.csv, and checkpoints will be saved in lfads_ckpts.
After training, the sample_and_average function can be used to compute firing rate estimates and other intermediate model outputs and save them to posterior_samples.h5 in the MODEL_DIR.
We provide a simple example in example_scripts/train_lfads.py.
The autolfads-tf2 framework uses ray.tune to distribute models over a computing cluster, monitor model performance, and exploit high-performing models and their HPs.
If you'll be running AutoLFADS on a single machine, you can skip this section. If you'll be running across multiple machines, you must initialize the cluster using these instructions before you can submit jobs via the Python API.
Fill in the fields indicated by <>'s in the ray_cluster_template.yaml, and save this file somewhere accessible. Ensure that a range of ports is open for communication on all machines that you intend to use (e.g. 10000-10099 in the template). In your autolfads-tf2 environment, start the cluster using ray up <NEW_CLUSTER_CONFIG>. The cluster may take up to a minute to get started. You can test that all machines are in the cluster by ensuring that all IP addresses are printed when running example_scripts/ray_test.py.
To run AutoLFADS, copy the run_pbt.py script and adjust paths and hyperparameters to your needs. Make sure to only use only as many workers as can fit on the machine(s) at once. If you want to run across multiple machines, make sure to set SINGLE_MACHINE = False in run_pbt.py. To start your PBT run, simply run run_pbt.py. When the run is complete, the best model will be copied to a best_model folder in your PBT run folder. The model will automatically be sampled and averaged and all outputs will be saved to posterior_samples.h5.
Keshtkaran MR, Sedler AR, Chowdhury RH, Tandon R, Basrai D, Nguyen SL, Sohn H, Jazayeri M, Miller LE, Pandarinath C. A large-scale neural network training framework for generalized estimation of single-trial population dynamics. bioRxiv. 2021 Jan 1.
Keshtkaran MR, Pandarinath C. Enabling hyperparameter optimization in sequential autoencoders for spiking neural data. Advances in Neural Information Processing Systems. 2019; 32.