def near_edges(daily_ats, user_locs, in_paths):
"""
keep edges from daily_ats between users who live within 25 miles
"""
# FIXME: I'm really starting to dislike in_paths
day_name = in_paths[0].split('.')[-1]
dt = datetime.datetime.strptime(day_name,"%Y-%m-%d")
day = dt.date() - datetime.date(2012,8,1)
edges = collections.defaultdict(lambda: collections.defaultdict(int))
for kind,frm,to in daily_ats:
if frm in user_locs and to in user_locs:
edges[frm,to][kind]+=1
def _as_array(is_to,is_lat):
return np.array([user_locs[edge[is_to]][is_lat] for edge in edges])
flngs = _as_array(0,0)
flats = _as_array(0,1)
tlngs = _as_array(1,0)
tlats = _as_array(1,1)
dists = utils.np_haversine(flngs,tlngs,flats,tlats)
for dist,(frm,to) in izip(dists,edges):
if dist<25:
edge = edges[frm,to]
yield NearEdge(
frm,
to,
dist,
day.days,
edge.get('at',0),
edge.get('rt',0),
)
def predict(self,nebrs_d,vect_fit):
lats = [r['lat'] for r in nebrs_d['nebrs']]
lngs = [r['lng'] for r in nebrs_d['nebrs']]
mlat = np.median(lats)
mlng = np.median(lngs)
dists = utils.np_haversine(mlng,lngs,mlat,lats)
return np.argmin(dists)
def graph_example_probs(vect_fit, in_paths):
"""
create an example of maximum likeliehood estimation for four friends
"""
if in_paths[0][-1] != '0':
return
curves = [fit for vers,cutoff,fit in vect_fit if vers=='leaf']
lat_range = np.linspace(27.01,32.99,5*60)
lng_range = np.linspace(-100.99,-93.01,5*80)
lat_grid, lng_grid = np.meshgrid(lat_range, lng_range)
print lat_range, lng_range
probs = np.zeros_like(lat_grid)
spots = (
(-95.31, 29.73, 0), # Houston
(-96.37, 30.67, 1), # Bryan, TX
(-99.25, 31.25, 5), # Texas
(-97.74, 30.27, 3), # Austin
)
for lng, lat, curve in spots:
dists = utils.np_haversine(lng, lng_grid, lat, lat_grid)
probs+=np.log(peek.contact_curve(dists,*(curves[curve])))
clipped = 255.999*(np.max(probs)-probs)/np.ptp(probs)
buff = np.require(np.transpose(clipped),np.uint8,['C_CONTIGUOUS'])
img = PIL.Image.frombuffer('L',(clipped.shape),buff)
img.save('example_probs.png')
def _calc_dists(nebrs_d):
gnp = nebrs_d['gnp']
lats = [r['lat'] for r in nebrs_d['nebrs']]
lngs = [r['lng'] for r in nebrs_d['nebrs']]
all_lats = lats+[gnp['lat']] if gnp else lats
all_lngs = lngs+[gnp['lng']] if gnp else lngs
lat1,lat2 = np.meshgrid(all_lats,lats)
lng1,lng2 = np.meshgrid(all_lngs,lngs)
return utils.np_haversine(lng1,lng2,lat1,lat2)
def exact_strange_bins(uids,mlocs):
"""find the distance between every contact and every target user"""
mlngs,mlats = np.transpose(mlocs)
bins = utils.dist_bins(120)
counts = np.zeros(len(bins)-1)
for contact in _paged_users(set(uids),fields=['gnp']):
clat = contact.geonames_place.lat
clng = contact.geonames_place.lng
dists= utils.np_haversine(clng, mlngs, clat, mlats)
hist,b = np.histogram(dists,bins)
counts+=hist
return enumerate(counts)
def _dists_for_lat(lat):
lat_range = np.linspace(-89.95,89.95,1800)
lng_range = np.linspace(.05,180.05,1801)
lat_grid,lng_grid = np.meshgrid(lat_range, lng_range)
centered_lat = .05 + .1*_tile(lat)
lat_ar = np.empty_like(lat_grid)
lat_ar.fill(centered_lat)
lng_0 = np.empty_like(lat_grid)
lng_0.fill(.05)
return utils.np_haversine(lng_0, lng_grid, lat_ar, lat_grid)
def mdist_real(nebrs_d):
"""
compare median location error to the actual location error for the target
users after adding noise to home location in mloc_blur
"""
data = collections.defaultdict(list)
for nebr_d in nebrs_d:
if not nebr_d['gnp']:
continue
data['glat'].append(nebr_d['gnp']['lat'])
data['glng'].append(nebr_d['gnp']['lng'])
data['mlng'].append(nebr_d['mloc'][0])
data['mlat'].append(nebr_d['mloc'][1])
data['mdist'].append(nebr_d['gnp']['mdist'])
dists = utils.np_haversine(
data['mlng'], data['glng'],
data['mlat'], data['glat'])
return itertools.izip(data['mdist'],dists)
def stranger_prob(lat_tile,contact_count):
"""
Calculate pStrangers for every longitude tile at a specific latitude.
pStrangers is the probability that a user lives at a location given the
locations of people they are not connected to. This step of FreindlyLocation
took about 2 weeks on a machine with 8 cores.
lat_tile should be between -900 and 900 and represents a latitude
contact_count is a matrix of the locations of the contacts
"""
lat_range = np.linspace(-89.95,89.95,1800)
lng_range = np.linspace(.05,359.95,3600)
lat_grid,lng_grid = np.meshgrid(lat_range, lng_range)
dists = utils.np_haversine(.05, lng_grid, .1*lat_tile+.05, lat_grid)
# FIXME: the name of a slurped command-line argument should not have to
# match the file name
contact_mat = contact_count
dists[0,lat_tile+900] = 2
for lng_tile in xrange(-1800,1800):
probs = np.log(1-utils.contact_prob(dists))
prob = np.sum(contact_mat*probs)
yield (lng_tile,lat_tile),prob
dists = np.roll(dists,1,0)
