STSPNet

This repository contains the code for the models described in the study:

"Adaptation supports short-term memory in a visual change detection task"

A preprint is available at https://www.biorxiv.org/content/10.1101/2020.03.06.977512v2.

Dependencies

This code was developed on Ubuntu 16.04 with Pytorch 1.0+ using a single Nvidia RTX 1080 Ti GPU. Please note that newer versions of Pytorch may introduce breaking changes. To create a local virtual environment, use the following commands:

# create a new conda environment
conda create --name stspnet python=3.6

# activate the conda environment
conda activate stspnet

# install pytorch and torchvision
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch

# clone the repo
git clone https://github.com/AllenInstitute/STSPNet.git

Pre-trained CIFAR-10 Models

The code for pre-training convolutional neural networks on a grayscale version of the CIFAR-10 image recognition dataset and the resulting feature embeddings used in our paper can be found in the STSPNet_ROOT/cifar directory. We train 10 random model initializations and generate features on a held-out image dataset (image set "A"). Feature embeddings are also available for three other image sets, but these were not used in the paper (see STSPNet_ROOT/data for these other image sets).

Training

To train a model, run python main.py. Any trained models will be saved to STSPNet_ROOT/PARAM/{STPNet,RNN}. Please run python main.py --h to find out more about the possible command-line options. For example, to train a STPNet model:

 python main.py --model STPNet --noise-std 0.5 --l2-penalty 0.001 --pos-weight 5 --dprime 1.5 --patience 5 --seed 1

Note: Use the --no-cuda flag here and below if you want to run the cpu-only version of Pytorch.

Evaluation

To test a model, run python experiment.py. You must pass an appropriate model path for loading a trained model. Model results will be saved as pickle files to STSPNet_ROOT/RESULT/{STPNet,RNN}. Please run python experiment.py --h to see all possible command-line options. For example, to evaluate the STPNet model trained above:

python experiment.py --model STPNet --model-path ./PARAM/STPNet/model_train_seed_1.pt --noise-std 0.5 --omit-frac 0.05 --seed 1

Visual Behavior Analyses

We provide Jupyter notebooks that generate the figures in the paper in STSPNet_ROOT/figures. Please note that these notebooks are not well-documented. The analyses contained in these notebooks rely on the visual_behavior_analysis package. We used the feature/paper_figures branch of this repository. Please follow the instructions there to install the correct dependencies.

Please address all questions to Stefan Mihalas: [email protected].