reservoir-input

This is a project about how you can optimize the inputs of a recurrent reservoir to perform a diverse array of tasks.

quick file guide

• run.py: to train/run the network. contains all the options, using argparse
• network.py: defines the network with pytorch
• tasks.py: defines the tasks:
  • RSG: ready-set-go task from Sohn et al
  • CSG: cue-set-go task from Wang et al
  • DelayProAnti, MemoryProAnti: tasks from Duncker et al
  • FlipFlop: flip flop task from Sussillo and Barak

to run something

Start with python run.py --no_log to see if it works by default. Most likely, you'll need to make sure the default dataset exists; you can do that by seeing the section below. You can also choose to run with any dataset, with python run.py --no_log -d datasets/custom_dataset.pkl.

Without the --no_log option, logs will be generated in a custom manner in the logs/ folder; details are in utils.py, but they're not particularly important.

Once the code above works, try python run.py -d datasets/rsg-100-150.pkl --name test_run --n_epochs 2 --batch_size 5 -N 100.

to create tasks

dataset creation

python tasks.py create rsg_1 -t rsg -n 200 creates an RSG dataset at datasets/rsg_1.pkl with 200 trials, using the default parameters.

python tasks.py create dpro_1 -t delaypro -n 100 --angles 30 60 100 creates a DelayPro dataset at datasets/dpro_1.pkl with 100 trials, using angles of 30, 60, or 100.

python tasks.py create rsg_2 -t rsg -a gt 100 lt 150 creates an RSG dataset at datasets/rsg_2.pkl with 2000 trials, with intervals between 100 and 150.

python tasks.py create danti_1 -c datasets/config/delayanti.json creates a DelayAnti dataset at datasets/danti_1.pkl with 2000 trials, with parameters taken from the config file.

dataset visualization

python tasks.py load datasets/rsg_1.pkl loads some examples from the dataset

contact

Email [email protected] with any questions.