jzchi/psychrnn

PsychRNN

This package is intended to help cognitive scientist easily translate task designs from human or primate behavioral experiments into a form capable of being used as training data for a recurrent neural network.

We have isolated the front-end task design, in which users can intuitively describe the conditional logic of their task from the backend where gradient descent based optimization occurs. This is intended to facilitate researchers who might otherwise not have an easy implementation available to design and test hypothesis regarding the behavior of recurrent neural networks in different task environements.

Code is written and upkept by: @davidbrandfonbrener @dbehrlic @ABAtanasov @syncrostone

Contact: [email protected]

Getting Started

You can try out PsychRNN without downloading anything using our example notebook that runs in the browser on Google Colab. Try it here: https://colab.research.google.com/github/dbehrlich/PsychRNN/blob/master/psychrnn/notebooks/collaboratory_RDM.ipynb

Install

Dependencies

• Numpy
• Tensorflow
• Python=2.7 or 3.6

For Demos:

• Jupyter
• Ipython
• Matplotlib

Installation

git clone https://github.com/dbehrlich/PsychRNN.git
cd PsychRNN
python setup.py install

17 Lines Introduction

A minimal introduction to our package. In this simple introduction you can generate a new recurrent neural network model, train that model on the random dot motion discrimination task, and plot out an example output in just 17 lines.

import psychrnn  
from psychrnn.tasks import rdm as rd  
from psychrnn.backend.models.basic import Basic  
import tensorflow as tf  

from matplotlib import pyplot as plt  
%matplotlib inline

rdm = rd.RDM(dt = 10, tau = 100, T = 2000, N_batch = 128)  
gen = rdm.batch_generator()

params = rdm.__dict__  
params['name'] = 'model'  
params['N_rec'] = 50  

model = Basic(params)  
model.build()  
model.train(gen)

x,_,_,_ = next(gen)

plt.plot(model.test(x)[0][0,:,:])

model.destruct()

Code for this example can be found in "Minimal_Example.ipynb"

Demonstration Notebook

For a more complete tour of training and model parameters see the "RDM.ipynb" notebook.

Writing a New Task

You can easily begin running your own tasks by writing a new task subclass with the two functions (generate_trial_params, trial_function) specified below, or by modifying one of our existing task files such as "rdm.py" or "romo.py".

Class your_new_class(Task):

	def __init__(self, N_in, N_out, dt, tau, T, N_batch):

		super(RDM,self).__init__(N_in, N_out, dt, tau, T, N_batch)

			'''

			Args:
				N_in: number of network inputs
				N_out: number of network output
				dt: simulation time step
				tau: unit time constant
				T: trial length
				N_batch: number of trials per training update

			'''

	def generate_trial_params(self,batch,trial):

		''' function that produces trial specific params for your task (e.g. coherence for the 
			random dot motion discrimination task)

		Args:
			batch: # of batch for training (for internal use)
			trial: # of trial within a batch (for internal use)

		Returns:
			params: A dictionary of necessary params for trial_function

			'''

	def trial_function(self,t,params):

		'''function that specifies conditional network input, target output and loss mask for your task at a given time (e.g. if t>stim_onset x_t=1).

		Args:
			t: time
			params: params dictionary from generate_trial_params

		Returns:
			x_t: input vector of length N_in at time t
			y_t: target output vector of length N_out at time t
			mask_t: loss function mask vector of length N_out at time t

			'''

Building a New Model

New models can be added by extending the RNN superclass, as in our examples of "basic.py" and "lstm.py". Each new model class requires three functions (recurrent_timestep, output time_step and forward_pass).

Class your_new_model(RNN):

	def recurrent_timestep(self, rnn_in, state):

		'''function that updates the recurrent state of your network one timestep

		Args:
			rnn_in: network input vector of length N_in at t
			state: network state at t

		Returns:
			new_state: network state at t+1

			'''

	def output_timestep(self, state):

		'''function that produces output for the current state of your network at one timestep

		Args:
			state: network state at t

		Returns:
			output: output vector of length N_out at t

			'''

	def forward_pass(self):

		'''function that contains the loop of calls to recurrent_timestep and output_timestep
		to run the evolution of your state through a trial 


			'''

Further Extensibility

If you wish to modify weight initializations, loss functions or regularizations it is as simple as adding an additional class to "initializations.py" describing your preferred initial weight patterns or a single function to "loss_functions.py" or "regularizations.py".

Backend

• initializations
• loss_functions
• regularizations
• rnn
• simulation