def rate_all():
games = Game.query.all()
users = sanitized_users(games)
g_vec = sanitized_games(games)
ratings = {u.id: u.last_rating() for u in users}
rating_prior = {u: v.rating if (v and v.rating) else 0 for u,v in ratings.items()}
neighbors = rm.neighbors(games)
neighbor_avgs = rm.compute_avgs(games, rating_prior)
t_min = min(games, key=lambda g: g.date_played or datetime.datetime.now()).date_played
t_max = max(games, key=lambda g: g.date_played or datetime.datetime.now()).date_played
iters = 100
lam = .37 # Weight for controlling 'pull' of neighborhood weighted average. Higher = stays in the neighborhood.
lrn = lambda i: ((1. + .1*iters)/(i + .1 * iters))**.6 #Control the learning rate over time.
for i in range(iters):
loss = 0
# Accumulate the neighborhood loss prior to changing the ratings around
for id, neighbor_wgt in neighbor_avgs.items():
loss += lam * ((rating_prior[id] - neighbor_wgt) ** 2)
# Shuffle the vector of result-tuples
random.shuffle(g_vec)
for g in g_vec:
w, b, actual, t, handi, komi = g
odds = rm.expect(rating_prior[b], rating_prior[w], handi, komi)
weight = rm.time_weight(t, t_min, t_max)
rating_prior[w] -= lrn(i) * (weight*(odds - actual)*odds*(1-odds) + (lam/len(neighbors[w]) * (rating_prior[w] - neighbor_avgs[w])))
rating_prior[b] -= lrn(i) * (-1.0 * weight*(odds - actual)*odds*(1-odds) + (lam/len(neighbors[b]) * (rating_prior[b] - neighbor_avgs[b])))
loss += weight * ((odds - actual) ** 2)
# Scale the ratings
r_min = min(rating_prior.values())
r_max = max(rating_prior.values())
if r_max != r_min:
for k,v in rating_prior.items():
rating_prior[k] = (rating_prior[k] - r_min) / (r_max - r_min) * 40.0
neighbor_avgs = rm.compute_avgs(games, rating_prior)
print('%d : %.4f' % (i, loss))
# Update the ratings and show how we did.
wins, losses = {}, {}
for g in g_vec:
wins[g[0]] = wins.get(g[0], 0) + g[2]
losses[g[0]] = losses.get(g[0], 0) + 1-g[2]
wins[g[1]] = wins.get(g[1], 0) + 1-g[2]
losses[g[1]] = losses.get(g[1], 0) + g[2]
for k in sorted(rating_prior, key=lambda k: rating_prior[k]):
db.session.add(Rating(user_id=k, rating=rating_prior[k]))
print("%d: %f (%d - %d)" % (k,rating_prior[k], wins.get(k,0), losses.get(k,0)) )
db.session.commit()
def rate_all():
games = Game.query.all()
users = sanitized_users(games)
g_vec = sanitized_games(games)
ratings = {u.id: u.last_rating() for u in users}
rating_prior = {
u: v.rating if (v and v.rating) else 0
for u, v in ratings.items()
}
neighbors = rm.neighbors(games)
neighbor_avgs = rm.compute_avgs(games, rating_prior)
t_min = min(
games,
key=lambda g: g.date_played or datetime.datetime.now()).date_played
t_max = max(
games,
key=lambda g: g.date_played or datetime.datetime.now()).date_played
iters = 100
lam = .37 # Weight for controlling 'pull' of neighborhood weighted average. Higher = stays in the neighborhood.
lrn = lambda i: (
(1. + .1 * iters) /
(i + .1 * iters))**.6 #Control the learning rate over time.
for i in range(iters):
loss = 0
# Accumulate the neighborhood loss prior to changing the ratings around
for id, neighbor_wgt in neighbor_avgs.items():
loss += lam * ((rating_prior[id] - neighbor_wgt)**2)
# Shuffle the vector of result-tuples
random.shuffle(g_vec)
for g in g_vec:
w, b, actual, t, handi, komi = g
odds = rm.expect(rating_prior[b], rating_prior[w], handi, komi)
weight = rm.time_weight(t, t_min, t_max)
rating_prior[w] -= lrn(i) * (
weight * (odds - actual) * odds * (1 - odds) +
(lam / len(neighbors[w]) *
(rating_prior[w] - neighbor_avgs[w])))
rating_prior[b] -= lrn(i) * (
-1.0 * weight * (odds - actual) * odds * (1 - odds) +
(lam / len(neighbors[b]) *
(rating_prior[b] - neighbor_avgs[b])))
loss += weight * ((odds - actual)**2)
# Scale the ratings
r_min = min(rating_prior.values())
r_max = max(rating_prior.values())
if r_max != r_min:
for k, v in rating_prior.items():
rating_prior[k] = (rating_prior[k] - r_min) / (r_max -
r_min) * 40.0
neighbor_avgs = rm.compute_avgs(games, rating_prior)
print('%d : %.4f' % (i, loss))
# Update the ratings and show how we did.
wins, losses = {}, {}
for g in g_vec:
wins[g[0]] = wins.get(g[0], 0) + g[2]
losses[g[0]] = losses.get(g[0], 0) + 1 - g[2]
wins[g[1]] = wins.get(g[1], 0) + 1 - g[2]
losses[g[1]] = losses.get(g[1], 0) + g[2]
for k in sorted(rating_prior, key=lambda k: rating_prior[k]):
db.session.add(Rating(user_id=k, rating=rating_prior[k]))
print("%d: %f (%d - %d)" %
(k, rating_prior[k], wins.get(k, 0), losses.get(k, 0)))
db.session.commit()
def rate_all(t_from=datetime.datetime.utcfromtimestamp(1.0),
t_to=datetime.datetime.now(),
iters=200, lam=.22):
"""
t_from -- datetime obj, rate all games after this
t_to -- datetime obj, rate all games up to this
iters -- number of iterations
lam -- 'neighborhood pull' parameter. higher = more error from moving a rank away from ranks in its neighborhood
"""
games = Game.query.filter(Game.date_played < t_to, Game.date_played > t_from)
g_vec = sanitized_games(games)
users = sanitized_users(g_vec)
print("found %d users with %d valid games" % (len(users), len(g_vec)) )
aga_ids_to_uids = dict([(int(u.aga_id), u.id) for u in users])
ratings = {int(u.aga_id): u.last_rating() for u in users}
rating_prior = {id: v.rating if (v and v.rating) else 20 for id,v in ratings.items()}
print ("%d users with no priors" % len(list(filter(lambda v: v == 20, rating_prior.values()))))
neighbors = rm.neighbors(g_vec)
neighbor_avgs = rm.compute_avgs(g_vec, rating_prior)
t_min = min([g[3] for g in g_vec])
t_max = max([g[3] for g in g_vec])
lrn = lambda i: ((1. + .1*iters)/(i + .1 * iters))**.3 #Control the learning rate over time.
for i in range(iters):
loss = 0
# Accumulate the neighborhood loss prior to changing the ratings around
for id, neighbor_wgt in neighbor_avgs.items():
loss += lam * ((rating_prior[id] - neighbor_wgt) ** 2)
# Shuffle the vector of result-tuples and step through them, accumulating error.
random.shuffle(g_vec)
for g in g_vec:
w, b, actual, t, handi, komi = g
odds = rm.expect(rating_prior[b], rating_prior[w], handi, komi)
weight = rm.time_weight(t, t_min, t_max)
rating_prior[w] -= lrn(i) * (weight*(odds - actual)*odds*(1-odds) + (lam/len(neighbors[w]) * (rating_prior[w] - neighbor_avgs[w])))
rating_prior[b] -= lrn(i) * (-1.0 * weight*(odds - actual)*odds*(1-odds) + (lam/len(neighbors[b]) * (rating_prior[b] - neighbor_avgs[b])))
loss += weight * ((odds - actual) ** 2)
# Scale the ratings
r_min = min(rating_prior.values())
r_max = max(rating_prior.values())
if r_max != r_min:
for k,v in rating_prior.items():
rating_prior[k] = (rating_prior[k] - r_min) / (r_max - r_min) * 40.0
#update neighborhood averages?
neighbor_avgs = rm.compute_avgs(g_vec, rating_prior)
if (i % 50 == 0):
print('%d : %.4f' % (i, loss))
# Update the ratings and show how we did.
wins, losses = {}, {}
for g in g_vec:
wins[g[0]] = wins.get(g[0], 0) + g[2]
losses[g[0]] = losses.get(g[0], 0) + 1-g[2]
wins[g[1]] = wins.get(g[1], 0) + 1-g[2]
losses[g[1]] = losses.get(g[1], 0) + g[2]
for k in sorted(rating_prior, key=lambda k: rating_prior[k])[-10:]:
print("%d (uid: %d): %f (%d - %d)" % (k, aga_ids_to_uids[k], rating_prior[k], wins.get(k,0), losses.get(k,0)) )
for k in sorted(rating_prior, key=lambda k: rating_prior[k]):
db.session.add(Rating(user_id=aga_ids_to_uids[k], rating=rating_prior[k], created=t_to))
db.session.commit()
def test_time_weight(self):
t1 = date(1999,1,1)
t2 = date(2000,1,1)
self.assertEqual(rm.time_weight(t2, t1, t2), 1)
self.assertAlmostEqual(rm.time_weight(t1, t1, t2), 0, places=3)
def test_time_weight(self):
t1 = date(1999, 1, 1)
t2 = date(2000, 1, 1)
self.assertEqual(rm.time_weight(t2, t1, t2), 1)
self.assertAlmostEqual(rm.time_weight(t1, t1, t2), 0, places=3)
def rate_all(t_from=datetime.datetime.utcfromtimestamp(1.0),
t_to=datetime.datetime.now(),
iters=200,
lam=.22):
"""
t_from -- datetime obj, rate all games after this
t_to -- datetime obj, rate all games up to this
iters -- number of iterations
lam -- 'neighborhood pull' parameter. higher = more error from moving a rank away from ranks in its neighborhood
"""
games = Game.query.filter(Game.date_played < t_to,
Game.date_played > t_from)
g_vec = sanitized_games(games)
users = sanitized_users(g_vec)
print("found %d users with %d valid games" % (len(users), len(g_vec)))
aga_ids_to_uids = dict([(int(u.aga_id), u.id) for u in users])
ratings = {int(u.aga_id): u.last_rating() for u in users}
rating_prior = {
id: v.rating if (v and v.rating) else 20
for id, v in ratings.items()
}
print("%d users with no priors" %
len(list(filter(lambda v: v == 20, rating_prior.values()))))
neighbors = rm.neighbors(g_vec)
neighbor_avgs = rm.compute_avgs(g_vec, rating_prior)
t_min = min([g[3] for g in g_vec])
t_max = max([g[3] for g in g_vec])
lrn = lambda i: (
(1. + .1 * iters) /
(i + .1 * iters))**.3 #Control the learning rate over time.
for i in range(iters):
loss = 0
# Accumulate the neighborhood loss prior to changing the ratings around
for id, neighbor_wgt in neighbor_avgs.items():
loss += lam * ((rating_prior[id] - neighbor_wgt)**2)
# Shuffle the vector of result-tuples and step through them, accumulating error.
random.shuffle(g_vec)
for g in g_vec:
w, b, actual, t, handi, komi = g
odds = rm.expect(rating_prior[b], rating_prior[w], handi, komi)
weight = rm.time_weight(t, t_min, t_max)
rating_prior[w] -= lrn(i) * (
weight * (odds - actual) * odds * (1 - odds) +
(lam / len(neighbors[w]) *
(rating_prior[w] - neighbor_avgs[w])))
rating_prior[b] -= lrn(i) * (
-1.0 * weight * (odds - actual) * odds * (1 - odds) +
(lam / len(neighbors[b]) *
(rating_prior[b] - neighbor_avgs[b])))
loss += weight * ((odds - actual)**2)
# Scale the ratings
r_min = min(rating_prior.values())
r_max = max(rating_prior.values())
if r_max != r_min:
for k, v in rating_prior.items():
rating_prior[k] = (rating_prior[k] - r_min) / (r_max -
r_min) * 40.0
#update neighborhood averages?
neighbor_avgs = rm.compute_avgs(g_vec, rating_prior)
if (i % 50 == 0):
print('%d : %.4f' % (i, loss))
# Update the ratings and show how we did.
wins, losses = {}, {}
for g in g_vec:
wins[g[0]] = wins.get(g[0], 0) + g[2]
losses[g[0]] = losses.get(g[0], 0) + 1 - g[2]
wins[g[1]] = wins.get(g[1], 0) + 1 - g[2]
losses[g[1]] = losses.get(g[1], 0) + g[2]
for k in sorted(rating_prior, key=lambda k: rating_prior[k])[-10:]:
print("%d (uid: %d): %f (%d - %d)" %
(k, aga_ids_to_uids[k], rating_prior[k], wins.get(
k, 0), losses.get(k, 0)))
for k in sorted(rating_prior, key=lambda k: rating_prior[k]):
db.session.add(
Rating(user_id=aga_ids_to_uids[k],
rating=rating_prior[k],
created=t_to))
db.session.commit()
