In [2]:
import matplotlib.pyplot as plt
%matplotlib inline
import nengo
import numpy as np
import scipy.ndimage
import matplotlib.animation as animation
from matplotlib import pylab
from PIL import Image
import nengo.spa as spa
import cPickle
import random
from nengo_extras.data import load_mnist
from nengo_extras.vision import Gabor, Mask
Represent each number using a one-hot where the index of the one represents the digit value
In [2]:
#Encode categorical integer features using a one-hot aka one-of-K scheme.
def one_hot(labels, c=None):
assert labels.ndim == 1
n = labels.shape[0]
c = len(np.unique(labels)) if c is None else c
y = np.zeros((n, c))
y[np.arange(n), labels] = 1
return y
Load the MNIST training and testing images
In [3]:
# --- load the data
img_rows, img_cols = 28, 28
(X_train, y_train), (X_test, y_test) = load_mnist()
X_train = 2 * X_train - 1 # normalize to -1 to 1
X_test = 2 * X_test - 1 # normalize to -1 to 1
train_targets = one_hot(y_train, 10)
test_targets = one_hot(y_test, 10)
In [4]:
rng = np.random.RandomState(9)
# --- set up network parameters
#Want to encode and decode the image
n_vis = X_train.shape[1]
n_out = X_train.shape[1]
#number of neurons/dimensions of semantic pointer
n_hid = 1000 #Try with more neurons for more accuracy
#Want the encoding/decoding done on the training images
ens_params = dict(
eval_points=X_train,
neuron_type=nengo.LIF(), #Why not use LIF? originally used LIFRate()
intercepts=nengo.dists.Choice([-0.5]),
max_rates=nengo.dists.Choice([100]),
)
#Least-squares solver with L2 regularization.
solver = nengo.solvers.LstsqL2(reg=0.01)
#solver = nengo.solvers.LstsqL2(reg=0.0001)
solver2 = nengo.solvers.LstsqL2(reg=0.01)
#network that generates the weight matrices between neuron activity and images and the labels
with nengo.Network(seed=3) as model:
a = nengo.Ensemble(n_hid, n_vis, seed=3, **ens_params)
v = nengo.Node(size_in=n_out)
conn = nengo.Connection(
a, v, synapse=None,
eval_points=X_train, function=X_train,#want the same thing out (identity)
solver=solver)
v2 = nengo.Node(size_in=train_targets.shape[1])
conn2 = nengo.Connection(
a, v2, synapse=None,
eval_points=X_train, function=train_targets, #Want to get the labels out
solver=solver2)
# linear filter used for edge detection as encoders, more plausible for human visual system
encoders = Gabor().generate(n_hid, (11, 11), rng=rng)
encoders = Mask((28, 28)).populate(encoders, rng=rng, flatten=True)
#Set the ensembles encoders to this
a.encoders = encoders
#Check the encoders were correctly made
plt.imshow(encoders[0].reshape(28, 28), vmin=encoders[0].min(), vmax=encoders[0].max(), cmap='gray')
Out[4]:
In [5]:
#Get the one hot labels for the images
def get_outs(sim, images):
#The activity of the neurons when an image is given as input
_, acts = nengo.utils.ensemble.tuning_curves(a, sim, inputs=images)
#The activity multiplied by the weight matrix (calculated in the network) to give the one-hot labels
return np.dot(acts, sim.data[conn2].weights.T)
#Check how many of the labels were produced correctly
#def get_error(sim, images, labels):
# return np.argmax(get_outs(sim, images), axis=1) != labels
#Get label of the images
#def get_labels(sim,images):
# return np.argmax(get_outs(sim, images), axis=1)
#Get the neuron activity of an image or group of images (this is the semantic pointer in this case)
def get_activities(sim, images):
_, acts = nengo.utils.ensemble.tuning_curves(a, sim, inputs=images)
return acts
#Get the representation of the image after it has gone through the encoders (Gabor filters) but before it is in the neurons
#This must be computed to create the weight matrix for rotation from neuron activity to this step
# This allows a recurrent connection to be made from the neurons to themselves later
def get_encoder_outputs(sim,images):
#Pass the images through the encoders
outs = np.dot(images,sim.data[a].encoders.T) #before the neurons
return outs
In [7]:
dim =28
#Scale an image
def scale(img, scale):
newImg = scipy.ndimage.interpolation.zoom(np.reshape(img, (dim,dim), 'F').T,scale,cval=-1)
#If its scaled up
if(scale >1):
newImg = newImg[len(newImg)/2-(dim/2):-(len(newImg)/2-(dim/2)),len(newImg)/2-(dim/2):-(len(newImg)/2-(dim/2))]
if len(newImg) >28:
newImg = newImg[:28,:28]
newImg = newImg.ravel()
else: #Scaled down
m = np.zeros((dim,dim))
m.fill(-1)
m[(dim-len(newImg))/2:(dim-len(newImg))/2+len(newImg),(dim-len(newImg))/2:(dim-len(newImg))/2+len(newImg)] = newImg
newImg = m
return newImg.ravel()
#Shift an image
def translate(img,x,y):
newImg = scipy.ndimage.interpolation.shift(np.reshape(img, (dim,dim), 'F'),(x,y), cval=-1)
return newImg.T.ravel()
In [9]:
#Images to train, starting at random orientation, size and translation
orig_imgs = X_train[:100000].copy()
for img in orig_imgs:
while True:
try:
img[:] = scale(img,random.uniform(0.5,1.5))
break
except:
img[:] = img
img[:] = scipy.ndimage.interpolation.rotate(np.reshape(img,(28,28)),
(np.random.randint(360)),reshape=False,mode="nearest").ravel()
img[:] = translate(img,random.randint(-6,6),random.randint(-6,6))
In [1]:
#Check to make sure images were generated correctly
plt.subplot(121)
plt.imshow(np.reshape(orig_imgs[random.randint(0,1000)],(28,28)), cmap='gray')
plt.subplot(122)
plt.imshow(np.reshape(orig_imgs[random.randint(0,1000)],(28,28)), cmap='gray')
In [18]:
filename = "activity_to_img_weights_all_transformations" + str(n_hid) +".p"
cPickle.dump(sim.data[conn].weights.T, open( filename, "wb" ) )