def full_test(bpmf,
samples,
real,
key_name,
num_samps=128,
lookahead_fit='batch',
lookahead_samps=128,
pool=None,
multieval=False,
init_rmse=None,
test_on=Ellipsis):
key = KEYS[key_name]
total = real.size
picker_fn = getattr(bpmf, key.key_fn)
chooser = np.argmax if key.choose_max else np.argmin
if init_rmse is None:
init_rmse = bpmf.bayes_rmse(samples, real, which=test_on)
yield (len(bpmf.rated), init_rmse, None, None)
while bpmf.unrated:
print("{:<40} Picking query point {}...".format(
key.nice_name,
len(bpmf.rated) + 1))
if len(bpmf.unrated) == 1:
vals = None
i, j = next(iter(bpmf.unrated))
else:
unrated = np.array(list(bpmf.unrated)).T
which = tuple(unrated)
key_kwargs = {'which': which}
if key.wants_pool and pool is not None:
key_kwargs['pool'] = pool
# XXX: should use lookahead_samps in here somewhere?
evals = picker_fn(samples, *key.args, **key_kwargs)
i, j = unrated[:, chooser(evals)]
vals = bpmf.matrix_results(evals, which)
bpmf.add_rating(i, j, real[i, j])
print("{:<40} Queried ({}, {}); {}/{} known".format(
key.nice_name, i, j, len(bpmf.rated), total))
samp_args = (bpmf, num_samps)
samp_kwargs = {'fit_first': True}
if multieval:
samples, pred = pool.apply(fetch_samples, samp_args, samp_kwargs)
else:
samples, pred = fetch_samples(*samp_args, **samp_kwargs)
err = rmse(pred[test_on], real[test_on])
print("{:<40} RMSE {}: {:.5}".format(key.nice_name, len(bpmf.rated),
err))
yield len(bpmf.rated), err, (i, j), vals
def full_test(bpmf, samples, real, key_name,
num_samps=128, lookahead_fit='batch', lookahead_samps=128,
pool=None, multieval=False, init_rmse=None, test_on=Ellipsis):
key = KEYS[key_name]
total = real.size
picker_fn = getattr(bpmf, key.key_fn)
chooser = np.argmax if key.choose_max else np.argmin
if init_rmse is None:
init_rmse = bpmf.bayes_rmse(samples, real, which=test_on)
yield (len(bpmf.rated), init_rmse, None, None)
while bpmf.unrated:
print("{:<40} Picking query point {}...".format(
key.nice_name, len(bpmf.rated) + 1))
if len(bpmf.unrated) == 1:
vals = None
i, j = next(iter(bpmf.unrated))
else:
unrated = np.array(list(bpmf.unrated)).T
which = tuple(unrated)
key_kwargs = { 'which': which }
if key.wants_pool and pool is not None:
key_kwargs['pool'] = pool
# XXX: should use lookahead_samps in here somewhere?
evals = picker_fn(samples, *key.args, **key_kwargs)
i, j = unrated[:, chooser(evals)]
vals = bpmf.matrix_results(evals, which)
bpmf.add_rating(i, j, real[i, j])
print("{:<40} Queried ({}, {}); {}/{} known".format(
key.nice_name, i, j, len(bpmf.rated), total))
samp_args = (bpmf, num_samps)
samp_kwargs = {'fit_first': True}
if multieval:
samples, pred = pool.apply(fetch_samples, samp_args, samp_kwargs)
else:
samples, pred = fetch_samples(*samp_args, **samp_kwargs)
err = rmse(pred[test_on], real[test_on])
print("{:<40} RMSE {}: {:.5}".format(
key.nice_name, len(bpmf.rated), err))
yield len(bpmf.rated), err, (i,j), vals
def bayes_rmse(self, samples_iter, true_r, which=Ellipsis):
return rmse(self.predict(samples_iter, which), true_r[which])
def bayes_rmse(self, samples_iter, true_r, which=Ellipsis):
return rmse(self.predict(samples_iter, which), true_r[which])
