def __init__(self, noise_var, hparam):
# get the constellation
self.constellation = hparam.constellation
self.hparam = hparam
# set the graph
self.graph = fg.Graph()
# add the discrete random variables to graph
self.n_symbol = hparam.num_tx
for idx in range(hparam.num_tx):
self.graph.rv("x{}".format(idx), len(self.constellation))
def _build_xgraph(graphs, tuples, pot):
"""
Makes the interconnected (across knowledge dimension) graph.
Args:
graphs ([AttrGraph])
tuples ([(str, str)]) Attr pairs to add cxns between frame RVs
pot (np.ndarray of shape 3x3)
Returns
fg.Graph: the xgraph
"""
xgraph = fg.Graph(debug=False)
base_logger.debug('Adding xgraph xfactors...')
total = 0
for t in tuples:
attr1, attr2 = t
g1 = [g for g in graphs if g.name == attr1]
g2 = [g for g in graphs if g.name == attr2]
# Might not have one or both of the graphs because of the current
# settings.
if len(g1) != 1 or len(g2) != 1:
base_logger.debug(
'\t skipping links between missing graphs %s and %s', attr1,
attr2)
continue
g1 = g1[0]
g2 = g2[0]
# Find RVs that match across both graphs. Pruned RVs won't be returned
# by get_rvs() (as they are actually deleted from the graph's underlying
# dict), but we do want to make sure we're only linking frames.
matches = []
for rv_name, rv1 in g1.graph.get_rvs().iteritems():
if rv1.meta['type'] != 'frame':
continue
if g2.graph.has_rv(rv_name):
rv2 = g2.graph.get_rvs()[rv_name]
matches.append([rv1, rv2])
# add factors to our linking graph
for match in matches:
xgraph.factor(match, 'xfactor', pot, {'type': 'xfactor'})
# reporting
base_logger.debug(
'\t added %d links between frame RVs between %s and %s',
len(matches), attr1, attr2)
total += len(matches)
base_logger.debug('Added %d xgraph xfactors in total' % (total))
return xgraph
Pb_11 = alpha1 * ((dvxi - dvxj) ** 2 + (dvyi - dvyj) ** 2) \
+ alpha2 * (abs(dlxi - dlxj) + abs(dlyi - dlyj)) \
+ alpha4 * abs(ddispi - ddispj)
if (scorei > theta and scorej > theta):
Pb_not_11 = alpha3
else:
Pb_not_11 = 0
return np.array([[exp(-Wb * Pb_11), exp(-Wb * Pb_not_11)],
[exp(-Wb * Pb_not_11),
exp(-Wb * Pb_11)]])
for test_frame_id in range(frame[0], frame[1]):
frame_info = all_info[test_frame_id]
g = fg.Graph()
node_id_list = [l[0] for l in frame_info]
for i in range(len(node_id_list)):
node_i = str(
node_id_list[i]
) # node_i is the track_id (turned into str) for the variable node i
g.rv(node_i, 2) # yi can be 0 or 1
for i in range(len(node_id_list)):
node_i = str(node_id_list[i])
# first, add the unary factors
# frame_info[i] -> [id,dvx,dvy,dl,score,is_fake]
g.factor([node_i], potential=calc_unary(frame_info[i], Wu=0.98))
# second, add the binary factors
# User-item rating matrix
# movies_dict = build_movies_dict(movies_data)
# R = generate_matrix_from_csv(ratings_data, movies_dict)
R_clean = generate_100k_matrix(ratings_data)
R = generate_dirty_matrix(dirty_ratings_data)
user_ground_truth = generate_user_spam_list(spam_users_file)
# Data Statistics
num_users = np.shape(R)[0]
num_items = np.shape(R)[1]
print('Initializing...\n')
# Initialize Factor Graph
Graph = fg.Graph()
# Create nodes : node_list = ['m1', 'm2', 'm3', 't1', 't2', 't3']
user_nodes = []
for i in range(num_users):
user_nodes.append('m' + str(i))
item_nodes = []
for i in range(num_items):
item_nodes.append('t' + str(i))
# Spam Users and Target Items Initializations
m = np.random.rand(num_users)
m = [0 if i > 0.5 else 1 for i in np.random.rand(num_users)]
t = np.random.rand(num_items)
