ccGPFA

The repository contains an implementation for "Conditionally-Conjugate Gaussian Process Factor Analysis for Spike Count Data via Data Augmentation" (To appear at ICML 2024).

Requirements

• Python 3.8.10. Run the following script to setup the environment.

conda create -n ccgpfa-py38 python=3.8.10
conda activate ccgpfa-py38
pip install -r requirements.txt 

For the PAL baseline from Keeley et. al 2020, we create a separate environment(compatible with an older version). We include the PAL implementation under Count_GPFA.

cd Count_GPFA
conda create -n PAL-py356 python=3.5.6
conda activate PAL-py356
pip install -r requirements.txt

Experiments

Experiment 1: Allen Observatory Visual Coding Neuropixels

We include the dataset preprocessed from Allen Observatory data under datasets/drifting_gratings_75_repeats/. The dataset constains 75 trials of V1 recording under drifting gratings stimulus with 45 degress and contrast of 0.1.

Make output directory & switch to experiments

mkdir ../output 
cd experiments/multiple_trials

Running ccGPFA(negative binomial) on the allen data

python run_with_behavior_mapping.py --dataset ../../datasets/drifting_gratings_75_repeats/data_15_45.0_0.1_V1.pkl --output_dir ../output/drifting_gratings_75_repeats --test-size 25 --n-iter 100 --llk negbinom --ell 5. 

Running ccGPFA(binomial) on the allen data

python run_with_behavior_mapping.py --dataset ../../datasets/drifting_gratings_75_repeats/data_15_45.0_0.1_V1.pkl --output_dir ../output/drifting_gratings_75_repeats --test-size 25 --n-iter 100 --llk binom --ell 5. 

Running bGPFA method (Jensen et al. 2021 )

( using its official implementation from [Github])

python run_bgpfa.py --dataset ../../datasets/drifting_gratings_75_repeats/data_15_45.0_0.1_V1.pkl --output_dir ../../output/drifting_gratings_75_repeats --test-size 25 --n-iter 1000 --llk negbinom --ell 5.

Running PAL method (Keeley et. al 2020)

For PAL, we used the publicly available implementation on [Github]

cd Count_GPFA
python PAL_public_allen.py --dataset ../datasets/drifting_gratings_75_repeats/data_15_45.0_0.1_V1.pkl --max-iter 150

For GPFA (Yu et al. 2008), we used Elephant library, sample code included under

 python run_gpfa.py --dataset ../../datasets/drifting_gratings_75_repeats/data_15_45.0_0.1_V1.pkl --output_dir ../../output/drifting_gratings_75_repeats --test-size 25

Experiment 2: MC_Maze Reaching Task

A subset of reaching task data is preprocessed from the MC_Maze_L data. In total, 9 conditions (with varying angles) are included under datasets/MC_Maze_L.

cd experiments/multiple_trials

python run_with_behavior_mapping.py --dataset ../../datasets/MC_Maze_L/data_5_cond${i}_0.pkl --output_dir ../../output/MC_Maze_L --n-iter 500 --llk binom --ell 20. --threshold 0.01   --kinematic True

Run the above script for all files in datasets/MC_Maze_L by replacing ${i} by the condition number. For negative binomial ccGPFA model switch binom to negbinom.

Running bGPFA

python run_bgpfa_with_behavior_mapping.py --dataset ../../datasets/MC_Maze_L/data_5_cond${i}_0.pkl --output_dir ../../output/MC_Maze_L --n-iter 2000 --llk negbinom --ell 20.   --kinematic True

Run the above script for all files in datasets/MC_Maze_L by replacing ${i} by the condition number.

Running PAL

After switching to the PAL environment using conda activate PAL-py356,

cd Count_GPFA 
python PAL_public_MC_Maze.py --dataset ../datasets/MC_Maze_L/<pkl-file>

Final analysis can be found at (experiments/multiple_trials/MC Maze - Large - Analysis .ipynb)

Running GPFA

python run_gpfa.py --dataset ../datasets/MC_Maze_L/<pkl-file> --output_dir ../../output/MC_Maze_L --test-size 0 --bin-size 15 --threshold 1

Here threshold=1 option removes neurons with <= 1 spikes across all trials, as this causes issues in the original implementation.

Scalability Experiments

T = 1500

for i in 50 100 200; do python run_with_behavior_mapping.py --dataset  ../../datasets/synthetic/data_10_100_1500.pkl --output ../../output/synthetic/ --test-size 3  --n-iter 1500 --llk negbinom --ell 20. --threshold 0.1 --n-inducing $i  --lr 0.25 --batch 200; done;

T = 900

for i in 50 100 200; do python run_with_behavior_mapping.py --dataset  ../../datasets/synthetic/data_10_100_900.pkl --output ../../output/synthetic/ --test-size 3 --n-iter 1500 --llk negbinom --ell 20. --threshold 0.1 --n-inducing $i  --lr 0.25 --batch 150; done;

T= 300

for i in 50 100 200; do python run_with_behavior_mapping.py --dataset  ../../datasets/synthetic/data_10_100_300.pkl --output ../../output/synthetic/ --test-size 3 --n-iter 1500 --llk negbinom --ell 20. --threshold 0.1 --n-inducing $i  --lr 0.25 --batch 100; done; 

Switch negbinom to binom for equivalent Binomial ccGPFA.

Without inducing points

python run_with_behavior_mapping.py --dataset  ../../datasets/synthetic/data_10_100_1500.pkl --output ../../output/synthetic/ --test-size 3 --n-iter 1500 --llk binom --ell 20. --threshold 0.1
python run_with_behavior_mapping.py --dataset  ../../datasets/synthetic/data_10_100_1500.pkl --output ../../output/synthetic/ --test-size 3 --n-iter 1500 --llk negbinom --ell 20. --threshold 0.1

PAL

For PAL, we switch to its environment and execute

cd Count_GPFA
python PAL_public_synthetic.py --file ../datasets/synthetic/data_10_100_300.pkl --max-iter 1000
python PAL_public_synthetic.py --file ../datasets/synthetic/data_10_100_900.pkl --max-iter 1000
python PAL_public_synthetic.py --file ../datasets/synthetic/data_10_100_1500.pkl --max-iter 1000