def initialize_bpmf(self, args, data, query_set):
bpmf = BPMF(data.ratings,
args.latent_d,
subtract_mean=args.subtract_mean,
rating_values=data.rating_vals,
discrete_expectations=args.discrete,
num_integration_pts=args.num_integration_pts,
knowable=query_set,
model_filename=args.model_filename)
for k, v in args.hyperparams.items():
assert hasattr(bpmf, k)
setattr(bpmf, k, v)
if args.initialize_at_pmf_map:
import sys
sys.path.append(
os.path.abspath(
os.path.join(os.path.dirname(__file__), '../python-pmf')))
try:
from pmf_cy import ProbabilisticMatrixFactorization
except ImportError:
warnings.warn("cython PMF not available; "
"using pure-python version")
from pmf import ProbabilisticMatrixFactorization
print("getting PMF MAP fit with default hyperparams")
p = ProbabilisticMatrixFactorization(
data.ratings,
latent_d=args.latent_d,
subtract_mean=args.subtract_mean)
p.fit() # TODO: support also doing SGD
print("okay; BPMF will start from here now")
# set the mode to the PMF mode
bpmf.samples(num_samps=1, warmup=0, update_mode=True)
bpmf.sampled_mode['U'] = p.users
bpmf.sampled_mode['V'] = p.items
bpmf.sampled_mode['predictions'] = p.predicted_matrix()
bpmf.sampled_mode_lp = -np.inf
return bpmf
def fit(
real,
known,
latent_d=1,
ret_pmf=False,
subtract_mean=False,
sig_u=1e10,
sig_v=1e10,
sig=1,
do_bayes=False,
burnin=10,
samps=200,
stop_thresh=1e-10,
min_learning_rate=1e-20,
):
ratings = np.zeros((known.sum(), 3))
for idx, (i, j) in enumerate(np.transpose(known.nonzero())):
ratings[idx] = [i, j, real[i, j]]
pmf = ProbabilisticMatrixFactorization(ratings, latent_d, subtract_mean)
pmf.sigma_sq = sig
pmf.sigma_u_sq = sig_u
pmf.sigma_v_sq = sig_v
pmf.stop_thresh = stop_thresh
pmf.min_learning_rate = min_learning_rate
pmf.fit()
if not do_bayes:
pred = pmf.predicted_matrix()
return (pmf, pred) if ret_pmf else pred
else:
bpmf = BayesianPMF(ratings, 1)
bpmf.__setstate__(pmf.__getstate__())
sampler = bpmf.samples()
# do burn-in
next(islice(sampler, burnin, burnin), None)
pred = bpmf.predict(islice(sampler, samps))
return (bpmf, pred) if ret_pmf else pred
def initialize_bpmf(self, args, data, query_set):
bpmf = BPMF(data.ratings, args.latent_d,
subtract_mean=args.subtract_mean,
rating_values=data.rating_vals,
discrete_expectations=args.discrete,
num_integration_pts=args.num_integration_pts,
knowable=query_set,
model_filename=args.model_filename)
for k, v in args.hyperparams.items():
assert hasattr(bpmf, k)
setattr(bpmf, k, v)
if args.initialize_at_pmf_map:
import sys
sys.path.append(
os.path.abspath(os.path.join(os.path.dirname(__file__),
'../python-pmf')))
try:
from pmf_cy import ProbabilisticMatrixFactorization
except ImportError:
warnings.warn("cython PMF not available; "
"using pure-python version")
from pmf import ProbabilisticMatrixFactorization
print("getting PMF MAP fit with default hyperparams")
p = ProbabilisticMatrixFactorization(
data.ratings, latent_d=args.latent_d,
subtract_mean=args.subtract_mean)
p.fit() # TODO: support also doing SGD
print("okay; BPMF will start from here now")
# set the mode to the PMF mode
bpmf.samples(num_samps=1, warmup=0, update_mode=True)
bpmf.sampled_mode['U'] = p.users
bpmf.sampled_mode['V'] = p.items
bpmf.sampled_mode['predictions'] = p.predicted_matrix()
bpmf.sampled_mode_lp = -np.inf
return bpmf
def fit(real,
known,
latent_d=1,
ret_pmf=False,
subtract_mean=False,
sig_u=1e10,
sig_v=1e10,
sig=1,
do_bayes=False,
burnin=10,
samps=200,
stop_thresh=1e-10,
min_learning_rate=1e-20):
ratings = np.zeros((known.sum(), 3))
for idx, (i, j) in enumerate(np.transpose(known.nonzero())):
ratings[idx] = [i, j, real[i, j]]
pmf = ProbabilisticMatrixFactorization(ratings, latent_d, subtract_mean)
pmf.sigma_sq = sig
pmf.sigma_u_sq = sig_u
pmf.sigma_v_sq = sig_v
pmf.stop_thresh = stop_thresh
pmf.min_learning_rate = min_learning_rate
pmf.fit()
if not do_bayes:
pred = pmf.predicted_matrix()
return (pmf, pred) if ret_pmf else pred
else:
bpmf = BayesianPMF(ratings, 1)
bpmf.__setstate__(pmf.__getstate__())
sampler = bpmf.samples()
# do burn-in
next(islice(sampler, burnin, burnin), None)
pred = bpmf.predict(islice(sampler, samps))
return (bpmf, pred) if ret_pmf else pred
