def test_larcqy_logistic(verbose=False):
# factorization of the parameters with multiple linear regressions
toy = False
np.random.seed(1)
# load toy data
if toy:
data, rank, emb0_val = toy_dual_supervision_data()
data_train, data_test = train_test_split(data)
n_ents = 1 + np.max([np.max([np.max(t[0]) for t in qc[0]]) for qc in data_train] +
[np.max([np.max(t[0]) for t in qc[0]]) for qc in data_test])
else:
from factorix.demos.urban.urban_data_loading import load_area_aspects_data
# aspects = {'multicultural'}
# aspects = {'posh'}
aspects = {'waterside'}
nw, na = 1000000, 300
data_train, voc = load_area_aspects_data(aspects, 'train', max_n_words_per_area=nw, max_n_words_per_aspect=na)
data_test, voc = load_area_aspects_data(aspects, 'test', max_n_words_per_area=nw, max_n_words_per_aspect=na,
vocab=voc)
rank = 1
n_ents = len(voc.index)
regs = np.linspace(0.1, 20, 10)
max_epochs_list = [500] # range(25, 500, 25)
for reg in regs:
for max_epochs in max_epochs_list:
model = EmbeddingUpdater(rank, n_ents, reg, max_epochs=max_epochs, verbose=False)
model.fit(data_train)
pred, y, test_nll = model.predict(data_test)
test_auc = eval_auc(pred, y)
print('reg: ', reg, 'niter: ', max_epochs, ', auc: ', test_auc, ', nll: ', test_nll)
def test_logistic_regression_equivalence_prediction(verbose=False):
iris = datasets.load_iris()
x_mat = iris.data
y = iris.target
x_mat = x_mat[y != 2]
y = y[y != 2]
x_mat -= np.mean(x_mat, 0)
n_ents = x_mat.shape[1] + 3 # number of entities is the dimension with 2 slots and 1 intercept
clf1 = linear_model.LogisticRegression(C=1.0, penalty='l2', tol=1e-6)
clf1.fit(x_mat, y)
prob = clf1.predict_proba(x_mat)
n_show = 3
scores = x_mat.dot(clf1.coef_.T) + clf1.intercept_
clf2 = EmbeddingUpdater(rank=1, n_ents=n_ents, n_slots=2, reg=1.0, max_epochs=500, verbose=False,
preprocessing=multitask_to_tuples)
clf2.logistic2embeddings(coefs=clf1.coef_, intercept=clf1.intercept_)
pred2, y, nll = clf2.predict(x_mat, y)
clf3 = EmbeddingUpdater(rank=1, n_ents=n_ents, n_slots=2, reg=1.0, max_epochs=500, verbose=False,
preprocessing=multitask_to_tuples)
clf3.logistic2embeddings(coefs=clf1.coef_, intercept=clf1.intercept_)
pred3, y, nll = clf3.predict(x_mat, y)
if verbose:
print('linear_model.LogisticRegression predictions:', prob[0:n_show])
print('logistic regression prediction: ', 1.0 / (np.exp(-scores[0:n_show]) + 1.0))
print('EmbeddingUpdater predictions:', 1.0 / (np.exp(-pred2[0:n_show]) + 1.0))
print('same:', 1.0 / (np.exp(-pred3[0:n_show]) + 1.0))
assert(np.linalg.norm(scores - pred2) < 1e-2)
assert(np.linalg.norm(scores - pred3) < 1e-2)
def test_logistic_regression_equivalence_prediction(verbose=False):
iris = datasets.load_iris()
x_mat = iris.data
y = iris.target
x_mat = x_mat[y != 2]
y = y[y != 2]
x_mat -= np.mean(x_mat, 0)
n_ents = x_mat.shape[
1] + 3 # number of entities is the dimension with 2 slots and 1 intercept
clf1 = linear_model.LogisticRegression(C=1.0, penalty='l2', tol=1e-6)
clf1.fit(x_mat, y)
prob = clf1.predict_proba(x_mat)
n_show = 3
scores = x_mat.dot(clf1.coef_.T) + clf1.intercept_
clf2 = EmbeddingUpdater(rank=1,
n_ents=n_ents,
n_slots=2,
reg=1.0,
max_epochs=500,
verbose=False,
preprocessing=multitask_to_tuples)
clf2.logistic2embeddings(coefs=clf1.coef_, intercept=clf1.intercept_)
pred2, y, nll = clf2.predict(x_mat, y)
clf3 = EmbeddingUpdater(rank=1,
n_ents=n_ents,
n_slots=2,
reg=1.0,
max_epochs=500,
verbose=False,
preprocessing=multitask_to_tuples)
clf3.logistic2embeddings(coefs=clf1.coef_, intercept=clf1.intercept_)
pred3, y, nll = clf3.predict(x_mat, y)
if verbose:
print('linear_model.LogisticRegression predictions:', prob[0:n_show])
print('logistic regression prediction: ',
1.0 / (np.exp(-scores[0:n_show]) + 1.0))
print('EmbeddingUpdater predictions:',
1.0 / (np.exp(-pred2[0:n_show]) + 1.0))
print('same:', 1.0 / (np.exp(-pred3[0:n_show]) + 1.0))
assert (np.linalg.norm(scores - pred2) < 1e-2)
assert (np.linalg.norm(scores - pred3) < 1e-2)
def test_logistic_regression_equivalence_learning(verbose=False):
iris = datasets.load_iris()
x_mat = iris.data
y = iris.target
x_mat = x_mat[y != 2, 0:1]
y = y[y != 2]
x_mat -= np.mean(x_mat, 0)
n_ents = x_mat.shape[1] + 3
idx_show = [0, 99]
c = 1e10 # no regularization
clf1 = linear_model.LogisticRegression(C=c, penalty='l2', tol=1e-6)
clf1.fit(x_mat, y)
scores = x_mat.dot(clf1.coef_.T) + clf1.intercept_
clf2 = EmbeddingUpdater(rank=1,
n_ents=n_ents,
n_slots=2,
reg=1e-10,
max_epochs=500,
verbose=verbose,
preprocessing=multitask_to_tuples)
clf2.logistic2embeddings(coefs=clf1.coef_, intercept=clf1.intercept_)
pred, y, nll = clf2.predict(x_mat, y)
clf2.fit(x_mat, y)
pred2, y2, nll2 = clf2.predict(x_mat, y)
clf3 = EmbeddingUpdater(rank=1,
n_ents=n_ents,
n_slots=2,
reg=1e-10,
max_epochs=500,
verbose=verbose,
preprocessing=multitask_to_tuples)
clf3.fit(x_mat, y)
pred3, y3, nll3 = clf3.predict(x_mat, y)
if verbose:
print('logistic regression prediction:\n',
1.0 / (np.exp(-scores[idx_show]) + 1.0))
print('EmbeddingUpdater predictions before learning:\n',
1.0 / (np.exp(-pred[idx_show]) + 1.0))
print(
'EmbeddingUpdater predictions after learning (oracle initialization): \n',
1.0 / (np.exp(-pred2[idx_show]) + 1.0))
print(
'EmbeddingUpdater predictions after learning (random initialization): \n',
1.0 / (np.exp(-pred3[idx_show]) + 1.0))
assert (np.linalg.norm(scores - pred) < 1e-1)
assert (np.linalg.norm(scores - pred2) < 1e-1)
assert (np.linalg.norm(scores - pred3) < 1e-1)
def test_logistic_regression_equivalence_learning(verbose=False):
iris = datasets.load_iris()
x_mat = iris.data
y = iris.target
x_mat = x_mat[y != 2, 0:1]
y = y[y != 2]
x_mat -= np.mean(x_mat, 0)
n_ents = x_mat.shape[1] + 3
idx_show = [0, 99]
c = 1e10 # no regularization
clf1 = linear_model.LogisticRegression(C=c, penalty='l2', tol=1e-6)
clf1.fit(x_mat, y)
scores = x_mat.dot(clf1.coef_.T) + clf1.intercept_
clf2 = EmbeddingUpdater(rank=1, n_ents=n_ents, n_slots=2, reg=1e-10, max_epochs=500, verbose=verbose,
preprocessing=multitask_to_tuples)
clf2.logistic2embeddings(coefs=clf1.coef_, intercept=clf1.intercept_)
pred, y, nll = clf2.predict(x_mat, y)
clf2.fit(x_mat, y)
pred2, y2, nll2 = clf2.predict(x_mat, y)
clf3 = EmbeddingUpdater(rank=1, n_ents=n_ents, n_slots=2, reg=1e-10, max_epochs=500, verbose=verbose,
preprocessing=multitask_to_tuples)
clf3.fit(x_mat, y)
pred3, y3, nll3 = clf3.predict(x_mat, y)
if verbose:
print('logistic regression prediction:\n', 1.0 / (np.exp(-scores[idx_show]) + 1.0))
print('EmbeddingUpdater predictions before learning:\n', 1.0 / (np.exp(-pred[idx_show]) + 1.0))
print('EmbeddingUpdater predictions after learning (oracle initialization): \n',
1.0 / (np.exp(-pred2[idx_show]) + 1.0))
print('EmbeddingUpdater predictions after learning (random initialization): \n',
1.0 / (np.exp(-pred3[idx_show]) + 1.0))
assert(np.linalg.norm(scores-pred) < 1e-1)
assert(np.linalg.norm(scores-pred2) < 1e-1)
assert(np.linalg.norm(scores-pred3) < 1e-1)
