def __init__(self, net, trn, vld, threshold):
assert trn.ndim == 2
self.net = net # IMPORTANT: we assume that 'net' has already been trained using 'trn'
self.ssdata = SSData(net, trn)
self.net_history = NetHistory(net, vld)
self.thresh = threshold
class SS:
def __init__(self, net, trn, vld, threshold):
assert trn.ndim == 2
self.net = net # IMPORTANT: we assume that 'net' has already been trained using 'trn'
self.ssdata = SSData(net, trn)
self.net_history = NetHistory(net, vld)
self.thresh = threshold
def print_stats(self, i=None, name=''):
if i is not None:
logging.warning('\t================ %d =============== %s' % (i, name))
msg = '\tsize: %d' % len(self.net.pot)
msg += '\tvld: %f' % self.net_history.vld_hist[-1]
msg += '\tthresh: %f' % self.thresh
logging.warning(msg)
def step(self):
''' Take a step in the search space of SPN graphs. '''
pn, chll = select_step(self.ssdata, self.thresh)
if (pn, chll) == (None, None):
return False
mol = [ops.MixOp(self.net, pn, chl) for chl in chll]
for mo in mol:
# cluster
data = np.hstack( self.ssdata.edges[c][:,np.newaxis] for c in mo.chl ); assert len(data) >= knobs.min_instances
nvals = [len(c.weights) if c.is_sum() else len(c.masses) for c in mo.chl]
qa, nc = nbc.kcluster(data, nvals)
#qa, nc = nbc.inc_hard_em(data, nvals)
# change the graph
mo.connect(qa.max()+1)
ops.compute_params(mo, qa, self.ssdata.edges, knobs.laplace_smooth)
ops.adjust_edges(mo, qa, self.ssdata.edges)
self.ssdata.update_scores_of( [n for n in mo.prod_nodes()] )
self.net_history.add_op_node_list(mol)
logging.info('\tncl: %s' % str([len(mo.pnl) for mo in mol]))
return True
def step_ahead(self, num_steps):
for j in xrange(num_steps):
if self.step() == False:
logging.warning('\tran out of steps to take.')
j -= 1
break
self.net_history.save_vld()
return j+1
def skip_search(self, num_steps):
''' This is the main structure search algorithm. '''
assert num_steps >= 1
nh = self.net_history
i0 = 0
i1 = i0 + self.step_ahead(num_steps)
i2 = i1 + self.step_ahead(num_steps)
if nh.vh[i1] <= nh.vh[i0]:
bni = nh.best_net_index(i0, i1)
nh.move_to(bni)
assert self.net == nh.net
return nh.net
MAX_STEPS = 100000000
for i in xrange(MAX_STEPS):
self.print_stats(i)
logging.info('\t%f' % nh.vh[-1])
if nh.vh[i2] <= nh.vh[i1]:
break
taken = self.step_ahead(num_steps)
nh.vh[i0] = None
i0 = i1
i1 = i2
i2 += taken
if i == MAX_STEPS-1:
raise Exception, 'skip_search() just took a HUGE number of steps; that is not expected.'
bni = nh.best_net_index(i0, i2)
nh.move_to(bni)
assert self.net == nh.net
return nh.net
