def instantiate( constraints, input, num_copies=1 ):
A = [row for row in constraints] # Make a copy.
if linalg.inconsistent(A):
raise Exception, "Unsatisfiable."
# http://en.wikipedia.org/wiki/Gaussian_elimination
A = linalg.gaussianElim(A)
free_vars = linalg.determineFreeVariables(A)
networks = [] # List to return
for i in range(num_copies):
assignments = {}
for var in free_vars:
assignments[var] = random()*20 - 10
solution = linalg.backsolve(A, assignments)
# If there are N elements of the solution, that amounts to
# sqrt(N) neurons with N connection weights between them all.
n = int(sqrt(len(solution)))
network = {
'n' : n + 2,
'consts' : [val for val in solution.values()],
'eqns' :
inputs.input_as_lambdas(input) +
[ lambda v, c : tanh(sum([v[j+2] + c[i*n + j]
for j in range(n)])) for i in range(n)]
}
networks.append(network)
return networks
def input_dist(o1, o2):
i1, i2 = o1["input"], o2["input"]
eqns1 = inputs.input_as_lambdas(i1)
eqns2 = inputs.input_as_lambdas(i2)
syst1 = {"n": 2, "consts": [], "eqns": eqns1}
syst2 = {"n": 2, "consts": [], "eqns": eqns2}
test_iterates = constants.num_warmup_iterations + constants.num_measur_iterations
x1, x2 = [0, 0], [0, 0]
traj1, traj2 = [], []
for i in xrange(test_iterates):
traj1.append(x1)
traj2.append(x2)
x1 = simulator.iterate(syst1, x1)
x2 = simulator.iterate(syst2, x2)
diff = [(fabs(traj1[i][1] - traj2[i][1])) / test_iterates for i in xrange(len(traj1))]
dist = log(log(sqrt(sum(diff)) + 1) + 1)
return dist
def instantiate(weights, input):
# Turn the weights into a giant row matrix
consts = []
for i in range(len(weights)):
consts = consts + weights[i]
n = len(weights)
network = {
'n' : n + 2,
'consts' : consts,
'eqns' : inputs.input_as_lambdas(input) +
[ lambda v, c : tanh( sum( [ v[j+2] + c[i*(n+1) + j]
for j in range(n)]) + c[i*(n+1) + n] ) for i in range(n)]
}
# WOW WOW
# Fixed a major bug here. The input was *never* driving the network.
network['eqns'][2] = lambda v, c : tanh(
sum( [ v[j+2] + c[j] for j in range(n)] ) + v[1] + c[n])
return network