def train_run(n_examples, n_features, decisions_in, rseed=1):
# -- need y for margin computation
X, y = toyproblem.digits_xy(0, n_examples)
assert len(y) == n_examples
config = dict(
n_examples_train=n_examples,
n_examples_test=0,
n_folds=5,
feat_spec=dict(seed=rseed,
n_features=n_features,
scale=2.2),
decisions=decisions_in, # gets clobbered by ctrl attachment
save_svms=False, # good to test False sometimes
svm_l2_regularization=1e-3,
svm_max_observations=1e5,
)
ctrl = Ctrl(None)
BI = toyproblem.BoostableDigits()
result = BI.evaluate(config, ctrl)
decisions = np.asarray(result['decisions'])
assert decisions.shape == (5, n_examples)
print decisions.shape
print 'mean abs decisions', abs(decisions).mean(),
print 'mean margins', 1 - np.minimum(decisions * y, 1).mean(),
tr_acc, te_acc = mean_acc(result)
print 'train_accuracy', tr_acc,
print 'test_accuracy', te_acc,
print ''
return decisions, tr_acc, te_acc
def test_boosting_margin_goes_down():
n_examples = 1750
X, y = toyproblem.digits_xy(0, n_examples)
n_rounds = 8
margins = []
decisions = None
for round_ii in range(n_rounds):
config = dict(
n_examples_train=n_examples,
n_examples_test=0,
n_folds=5,
feat_spec=dict(seed=round_ii, n_features=16, scale=2.2),
decisions=decisions,
save_svms=False, # good to test False sometimes
svm_l2_regularization=1e-3,
svm_max_observations=1e3,
)
ctrl = Ctrl(None)
BI = toyproblem.BoostableDigits()
result = BI.evaluate(config, ctrl)
decisions = np.asarray(result['decisions'])
assert decisions.shape == (5, 1750)
print 'mean abs decisions', abs(decisions).mean(),
margins.append(1 - np.minimum(decisions * y, 1).mean())
for key in 'train_accuracy', 'test_accuracy':
print key, np.mean([rr[key] for rr in result['splits']]),
print ''
print margins
print list(reversed(margins))
print list(sorted(margins))
assert list(reversed(margins)) == list(sorted(margins))
def test_boosting_for_smoke():
n_examples = 1790
X, y = toyproblem.digits_xy(0, n_examples)
assert len(y) == n_examples
n_rounds = 16
n_features_per_round = 16
print 'Training jointly'
_, joint_tr_acc, joint_te_acc = train_run(
n_examples,
n_rounds * n_features_per_round,
None)
print 'Training one round'
_, one_tr_acc, one_te_acc = train_run(
n_examples,
n_features_per_round,
None)
# -- train in rounds
print 'Training in rounds'
decisions = None
for round_ii in range(n_rounds):
decisions, tr_acc, te_acc = train_run(
n_examples,
n_features_per_round,
decisions,
rseed=round_ii)
# assert that round-training and joint training are both way better than
# training just one
assert joint_tr_acc > 95
assert joint_te_acc > 88
assert one_tr_acc < 72
assert one_te_acc < 72
assert tr_acc > 90
assert te_acc > 88