In [1]:
from matplotlib import pylab
import nengo
import random
import numpy as np
import gzip as gz
import cPickle
from cPickle import load
try:
import Image
except ImportError:
from PIL import Image
from scipy.sparse.linalg import svds
import scipy
from scipy import ndimage
import matplotlib.pyplot as plt
import matplotlib.animation as animation
from nengo.utils.compat import is_iterable
import collections
#%matplotlib inline #Makes visualizations appar inline (Commented out because animation popup as new window)
In [2]:
#The code in this cell is for reading the images from the MNIST database and not part of the brain model.
def load_img(path, dims):
"""Load the image at path and return an array representing the raster.
Flattens image. Shifts pixel activations such that 0 represents gray,
normalizes the output array.
Keyword arguments:
path -- str, path of the image to be loaded.
dims -- (w, h), where w,h are ints indicating dimensions of the image (in
px)."""
img = Image.open(path).resize(dims).getdata()
img.convert('L')
img = subtract(array(img).flatten(), 127.5)
return img/norm(img)
def load_data(filename):
"""Uncompress, unpickle and return a .pkl.gz file.
Keyword arguments:
filename -- str, a valid file path"""
return load(gz.open(filename))
def load_mini_mnist(option=None):
"""Load and return the first \%10 of the images in the mnist dataset.
Does not return labels. Pass in 'train', 'valid' or 'test' if you want to
load a specific subset of the dataset.
Keyword arguments:
option -- str (default=None)."""
mini_mnist = load(gz.open('./mini_mnist.pkl.gz', 'rb'))
if option == 'train':
return mini_mnist[0]
elif option == 'valid':
return mini_mnist[1]
elif option == 'test':
return mini_mnist[2]
else:
return mini_mnist
In [3]:
def rotate_img(img, degrees):
'''
img is the dim**2 by 1 vector representing the pixel values.
Rotates image the degrees passed in counterclockwise
Returns the Reshaped image (to original shape which is the one dimensional vector)
dim is a global variable
'''
original = img.shape
newImg = scipy.ndimage.interpolation.rotate(np.reshape(img, (dim,dim), 'F'),degrees,reshape=False)
newImg = np.reshape(newImg, original, 'F')
return newImg
In [26]:
conn_synapse = 0.1 #post synaptic time constant to use for filtering (pstc) - what does changing this do?
probe_synapse = 0.01 #pstc
multiplier = 2 #not used
n_neurons = 5000
direct = True #Direct - function computed explicitly instead of in neurons
stop_time = 3.0
run_time = 3.0 #in seconds
In [5]:
dim = 28 #size of the image
mnist = load_mini_mnist()
#train = mnist[0] #collection of training images
img = mnist[1][0] #image to be used for testing
#compress_size = 400 #?
#basis, S, V = svds(train.T, k=compress_size) #Used for encoding and decoding information
#a set of vectors representing what a hand drawn number should look like?
In [6]:
#Need same number of vectors in basis as number of neurons (randomly sample from basis)
#expanded_basis = np.array([random.choice(basis.T) for _ in range(n_neurons)])
In [7]:
def stim_func(t):
'''returns the image for first 0.1s'''
if t < 0.01:
return img
else:
return [0 for _ in range(len(img))]
In [8]:
def connection_func(x):
'''takes the output from the first ensemble and rotates it 1 degree'''
return rotate_img(x,1)
In [57]:
'''def iter_func(i, x):
matrix = np.zeros((dim,dim))
matrix[i//dim][i%dim] = x
matrix = scipy.ndimage.interpolation.rotate(matrix,1,reshape=False)
out = np.reshape(matrix, (dim**2,), 'F')
def iter_rotate(x):
for i in range(dim**2):
yield iter_func(i,x)
#generator_func = iter_rotate(x)
E = []
for i in range(dim**2):
E += iter_rotate(x)
print(is_iterable(iter_rotate))
print(isinstance(iter_rotate, collections.Iterable))
def coordinates():
for i in range(10):
x=i
y=0
yield(x,y)
print(list(coordinates()))
list = range(dim**2)
list = iter(list)
def iter_rotate(x):
matrix = np.zeros((dim,dim))
i = list.next()
print(i)
if(i>=dim**2):
return matrix
matrix[i//dim][i%dim] = x
matrix = scipy.ndimage.interpolation.rotate(matrix,1,reshape=False)
out = np.reshape(matrix, (784L,), 'F')
return out
'''
def iterable_fn(n):
for i in range(n):
def rotate(x):
matrix = np.zeros((dim,dim))
matrix[i//dim][i%dim] = x
matrix = scipy.ndimage.interpolation.rotate(matrix,1,reshape=False)
out = np.reshape(matrix, (784L,), 'F')
return out
yield rotate
rotator = iterable_fn(784)
#rotator.next()
rotator2 = list(rotator)
#print len(rotator2)
def comb_func(t,x):
#print("Here")
out = np.zeros((dim,dim))
out = np.reshape(out, (784L,), 'F')
for i in range(dim**2):
out = np.add(x[(dim**2) * i : (dim**2) *(i+1)], out)
return out
#pylab.imshow(np.reshape(iter_rotate(1), (dim,dim), 'F'), cmap='Greys_r')
#plt.show()
#print(np.size(iter_rotate(1)))
In [58]:
#A network is primarily used for grouping together related objects and connections for visualization purposes
with nengo.Network() as net:
if direct:
neuron_type = nengo.Direct() #function computed explicitly, instead of in neurons
else:
neuron_type = nengo.LIF() #spiking version of the leaky integrate-and-fire neuron model
#Input stimulus - provide data to the ensemble
ipt = nengo.Node(stim_func)
'''An array of ensembles. This acts, in some ways, like a single high-dimensional ensemble,
but actually consists of many sub-ensembles, each one representing a separate dimension.
This tends to be much faster to create and can be more accurate than having one huge
high-dimensional ensemble. However, since the neurons represent different dimensions separately,
we cannot compute nonlinear interactions between those dimensions.'''
ensArr = nengo.networks.EnsembleArray(100, dim**2, ens_dimensions=1,neuron_type=neuron_type)
#incresing num neurons has smaller effect on run time
#Connect each pixel of the input to its own ensemble
nengo.Connection(ipt,ensArr.input)
'''When connecting nodes, threw error:
Validation error when setting 'Connection.function_info': Cannot apply functions to passthrough nodes
This is a workaround (https://github.com/nengo/nengo/issues/805)'''
#ensArr.output.output=lambda t, x: x
#ensArr.add_output('rotate_output', lambda x: rotate_img(x,1))
#a.add_output('square', lambda x: x**2)
ensArr.add_output('rotate_output', function=rotator2)
print(len(ensArr.rotate_output))
n = nengo.Node(comb_func,size_in = (28**2)**2, size_out=28**2)
nengo.Connection(ensArr.rotate_output,n)
nengo.Connection(n,ensArr.input)
#output node of ensArr brings all pixels together, connection performs the rotation and feeds into input node of ensArr
#nengo.Connection(ensArr.output, ensArr.input, function=connection_func)
#nengo.Connection(ensArr.rotate_output, ensArr.input)
#Gathering output of ensArr
probe = nengo.Probe(ensArr.output,# attr='decoded_output',#sample_every=0.001,
synapse=probe_synapse)
In [59]:
sim = nengo.Simulator(net)
In [60]:
sim.run(run_time) #LIF 2:25 #Direct 0:25
In [ ]:
#Original image
pylab.imshow(np.reshape(img, (dim,dim), 'F'), cmap='Greys_r')
In [ ]:
'''Image at stop time'''
pylab.imshow(np.reshape([0. if x < 0.00001 else x for x in sim.data[probe][int(stop_time*1000)-1]],
(dim, dim), 'F'), cmap=plt.get_cmap('Greys_r'),animated=True)
In [ ]:
'''Image at start time'''
pylab.imshow(np.reshape([0. if x < 0.00001 else x for x in sim.data[probe][1]],
(dim, dim), 'F'), cmap=plt.get_cmap('Greys_r'),animated=True)
In [61]:
'''Animation for Probe output'''
fig = plt.figure()
def updatefig(i):
im = pylab.imshow(np.reshape([0. if x < 0.00001 else x for x in sim.data[probe][i]],
(dim, dim), 'F'), cmap=plt.get_cmap('Greys_r'),animated=True)
return im,
ani = animation.FuncAnimation(fig, updatefig, interval=1, blit=True)
plt.show()
In [ ]:
# save the output
#cPickle.dump(sim.data[probe], open( "Buffer_rotations_in_connection_ensemble_array_direct.p", "wb" ) )
#cPickle.dump(sim.data[probe], open( "Buffer_rotations_in_connection_ensemble_array_LIF_100_stim0.01.p", "wb" ) )