def __init__(self, **kwargs):
"""
Set up equation parameters in the network and load the data to
construct the semantic layer.
For a human-readable format of equations refer to the paper.
Input
-----
nr_words: maximal allowed number of guesses, default 3
stim_len: length of stimulus and winner activity, default 50
"""
# maximal number of nodes visited in the first layer
self.nr_words = kwargs.get('nr_words', 4)
# stimulus length and the duration of activity for a WTA unit
self.stim_len = kwargs.get('stim_len', 50)
# weight pruning
self.theta = kwargs.get('theta', 0)
print('Theta %.3f' % self.theta)
print('Words %d' % self.nr_words)
# connection matrix with associative strengths
self.W, self.ids, self.voc = load_vocabulary()
# remove weights less than threshold
if self.theta > 0:
print('Removing connections...')
self.W = np.where(self.W > self.theta, self.W, 0)
# number of units
self.N = np.alen(self.W)
assert self.N == len(self.ids) == len(self.voc)
self.rho_w = 1 - 0.995 # integration constant for the WTA layer
self.c1 = 1. # excitatory connection
self.c2 = 1. # normalization constant
self.c3 = 1. # inhibitory neuron constant
self.c4 = 50. # inhibitory layer constant
self.c5 = 0.1 # noise amplitude
self.noise_offset = 0.05 / self.c5
self.theta_w = 1. # threshold for the WTA layer
self.I_amp = 1. # amplitude of the clamped imput
def __init__(self, **kwargs):
"""
Set up equation parameters in the network and load the data to
construct the semantic layer.
For a human-readable format of equations refer to the paper.
Input
-----
nr_words: maximal allowed number of guesses, default 3
stim_len: length of stimulus and winner activity, default 50
"""
# maximal number of nodes visited in the first layer
self.nr_words = kwargs.get('nr_words', 4)
# stimulus length and the duration of activity for a WTA unit
self.stim_len = kwargs.get('stim_len', 50)
# weight pruning
self.theta = kwargs.get('theta', 0)
print('Theta %.3f' % self.theta)
print('Words %d' % self.nr_words)
# connection matrix with associative strengths
self.W, self.ids, self.voc = load_vocabulary()
# remove weights less than threshold
if self.theta > 0:
print('Removing connections...')
self.W = np.where(self.W > self.theta, self.W, 0)
# number of units
self.N = np.alen(self.W)
assert self.N == len(self.ids) == len(self.voc)
self.rho_w = 1-0.995 # integration constant for the WTA layer
self.c1 = 1. # excitatory connection
self.c2 = 1. # normalization constant
self.c3 = 1. # inhibitory neuron constant
self.c4 = 50. # inhibitory layer constant
self.c5 = 0.1 # noise amplitude
self.noise_offset = 0.05/self.c5
self.theta_w = 1. # threshold for the WTA layer
self.I_amp = 1. # amplitude of the clamped imput
def simulate_test(**kwargs):
# Load the problem set
net = Network(**kwargs)
path_test = os.path.join(os.path.dirname(__file__), os.pardir, 'data',
'processed', 'rat_items')
items = np.loadtxt(path_test, dtype=np.character)
W, ids, voc = load_vocabulary()
nr_items = len(items)
positions = np.zeros(nr_items, dtype=np.int)
all_responses = []
for idx in range(nr_items):
rat_item = items[idx]
cues, target = rat_item[:3], rat_item[3]
target_id = voc[target]
net.setup_problem(cues, target)
# if WTA fails to pick a winner, try again
ok = True
while ok:
try:
net.run()
ok = False
except BaseException:
print 'WTA failed, retry the simulation'
continue
responses = list(net.visited())
all_responses.append(responses)
try:
# position of the solution
target_position = responses.index(target_id)
except ValueError:
# problem not solved
target_position = -1
print idx, rat_item, target_position
positions[idx] = target_position
return np.array(positions, dtype=np.int), all_responses
def simulate_test(**kwargs):
# Load the problem set
path_test = '../data/processed/rat_items'
items = np.loadtxt(path_test, dtype=np.character)
# Model parameters
theta = kwargs.get('theta', 0)
nr_words = kwargs.get('nr_words', 13)
W, ids, voc = load_vocabulary()
nr_items = len(items)
positions = np.zeros(nr_items, dtype=np.int)
# was = WAS()
all_responses = []
for idx in range(nr_items):
rat_item = items[idx]
nodes = [voc[word] for word in rat_item]
target = voc[rat_item[3]]
activations, responses = spread_activity(init_nodes=nodes[:3],
target=target,
W=W,
max_visited=nr_words,
threshold=theta)
all_responses.append(responses)
# response position
try:
target_position = responses.index(target)
except ValueError:
target_position = -1
# print idx, rat_item, target_position
positions[idx] = target_position
return np.array(positions, dtype=np.int), all_responses
def simulate_test(**kwargs):
# Load the problem set
path_test = '../data/processed/rat_items'
items = np.loadtxt(path_test, dtype=np.character)
# Model parameters
theta = kwargs.get('theta', 0)
nr_words = kwargs.get('nr_words', 13)
W, ids, voc = load_vocabulary()
nr_items = len(items)
positions = np.zeros(nr_items, dtype=np.int)
# was = WAS()
all_responses = []
for idx in range(nr_items):
rat_item = items[idx]
nodes = [voc[word] for word in rat_item]
target = voc[rat_item[3]]
activations, responses = spread_activity(init_nodes=nodes[:3],
target=target,
W=W,
max_visited=nr_words,
threshold=theta)
all_responses.append(responses)
# response position
try:
target_position = responses.index(target)
except ValueError:
target_position = -1
# print idx, rat_item, target_position
positions[idx] = target_position
return np.array(positions, dtype=np.int), all_responses
if j > -1:
visited.append(j)
if j == target:
break
j = WTA(activations, visited, j, activated)
counter += 1
assert len(visited) == max_visited or j == target
return activations, visited
if __name__ == "__main__":
W, ids, voc = load_vocabulary()
cues = ['river', 'note', 'account']
target = 'bank'
cues_id = [voc[cue] for cue in cues]
activations, order = spread_activity(init_nodes=cues_id,
target=voc[target],
W=W,
threshold=.0,
max_visited=10)
for i, ord_idx in enumerate(order):
print i+1, ids[ord_idx], activations[ord_idx]
solutions = np.where(positions < nr_w, positions, -1)
performance[i, :3] = get_difficulties(solutions)
percent_correct = 100 * len(np.where(solutions > -1)[0]) / nr_items
performance[i, 3] = percent_correct
return performance
if __name__ == "__main__":
font = {'family': 'serif', 'serif': 'Times New Roman', 'size': 28}
legend_fs = 24
matplotlib.rc('font', **font)
# Association data, needed for statistics below
W, ids, voc = load_vocabulary()
weights = W[W.nonzero()]
# problem difficulty labeling
difficulties = [0, 1, 2]
labs = ['easy', 'mid', 'hard']
lw = 4
colors = ['#4D4D4D', '#808080', '#CCCCCC', '#000000']
alphas = [.4, .4, .6]
ymin, ymax = -2, 105
fig = pl.figure(figsize=(22, 6), dpi=80, facecolor="white")
# ----------------- Vary number of words -----------------
axes = pl.subplot(121)
