def ddnll(self, data = None, params = None):
if any((not isinstance(d, self.Data) for d in data)):
raise Exception('Data type not understood')
if any((None in d.values for d in data)):
raise Exception('Data.values should be fully determined.')
if params is not None:
self.params = params
ndata = len(data)
ddnll = np.empty((ndata, self.nparams, self.nparams))
for i, d in enumerate(data):
# self.feats = d.feats
# self.values = d.values
for f in self.factors:
self.vp_node[f].make_table(d.feats)
# log.log_params(self)
# log.log_tables(self)
algo.message_passing(self, 'sum-product')
Efeats = [ list(self.vp_node[f].Efeats(d)) for f in self.factors ]
ddnll[i] = np.sum([
np.outer(ef1, ef2) for ef2 in Efeats for ef1 in Efeats
], axis=0)
return ddnll
def other_pr(self, v, data):
_data = data if isinstance(data, list) else [data]
if any((not isinstance(d, self.Data) for d in _data)):
raise Exception('Data type not understood')
ndata = len(_data)
pr = [None]*ndata
for i, d in enumerate(_data):
self.values = d.values
# self.feats = d.feats
for f in self.factors:
self.vp_node[f].make_table(d.feats)
algo.message_passing(self, 'sum-product')
pr[i] = self.pr(v.vertex)
return pr if isinstance(data, list) else pr[0]
def viterbi(self, data = None, params = None):
_data = data if isinstance(data, list) else [data]
if any((not isinstance(d, self.Data) for d in _data)):
raise Exception('Data type not understood')
if params is not None:
self.params = params
# TODO copy input and add new values
ndata = len(_data)
vit = [None]*ndata
for i, d in enumerate(_data):
self.values = d.values
# self.feats = d.feats
for f in self.factors:
self.vp_node[f].make_table(d.feats)
algo.message_passing(self, 'max-product')
vit[i] = d._replace(values=self.argmax())
return vit if isinstance(data, list) else vit[0]
def nll(self, data = None, params = None):
if any((not isinstance(d, self.Data) for d in data)):
raise Exception('Data type not understood')
if any((None in d.values for d in data)):
raise Exception('Data.values should be fully determined.')
if params is not None:
self.params = params
ndata = len(data)
nll = np.empty(ndata)
for i, d in enumerate(data):
# self.feats = d.feats
# self.values = d.values
for f in self.factors:
self.vp_node[f].make_table(d.feats)
# log.log_params(self)
# log.log_tables(self)
algo.message_passing(self, 'sum-product')
nll[i] = self.logZ - np.sum( self.vp_node[f].log_value(d) for f in self.factors )
return nll
F2.table = [[ 10, 1 ],
[ 1, 10 ]]
# F3 prefers if V2 and V3 are different
F3.table = [[ 1, 10 ],
[ 10, 1 ]]
fg.make()
return fg
def simple_tabgraph():
return make_tabgraph(vfun=simple_variables)
def domain_tabgraph():
return make_tabgraph(vfun=domain_variables)
if __name__ == '__main__':
fmt = '%(levelname)s @%(lineno)d:%(filename)s - %(funcName)s(): %(message)s'
fmt = '%(asctime)s %(levelname)s @%(lineno)d:%(filename)s - %(funcName)s(): %(message)s'
logging.basicConfig(filename='log.tabgraph.log',
filemode='w',
format=fmt,
level=logging.DEBUG)
fg = simple_tabgraph()
message_passing(fg, 'max-product', 'sum-product')
print 'max: {}'.format(fg.max())
print 'argmax: {}'.format(fg.argmax())
def test_domain_viterbi(self):
e = 1e-5
fg = tabgraph.domain_tabgraph()
algo.message_passing(fg, 'sum-product', 'max-product')
self.assertTrue(1000. - e < fg.max() < 1000. + e)
self.assertEqual(fg.argmax(), ['This', 'Code', 'Rules'])
def test_simple_viterbi(self):
e = 1e-5
fg = tabgraph.simple_tabgraph()
algo.message_passing(fg, 'sum-product', 'max-product')
self.assertTrue(1000. - e < fg.max() < 1000. + e)
self.assertEqual(fg.argmax(), [0, 0, 1])
