Skip to content

Commit

Permalink
average eta
Browse files Browse the repository at this point in the history
  • Loading branch information
@fgregg
fgregg committed Oct 11, 2016
1 parent d808ce9 commit 351b1ea
Show file tree
Hide file tree
Showing 2 changed files with 59 additions and 43 deletions.
34 changes: 23 additions & 11 deletions pseudolikelihood/centered_potts.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -83,7 +83,7 @@ def predict_proba(self, X, y):

spatial = safe_sparse_dot(A, (Y_multi - p_nonspatial))[:, :-1]

p += (eta * spatial)
p += eta.T * np.array(spatial/A.sum(axis=1))

p = np.hstack((p, np.zeros((features.shape[0], 1))))

Expand Down Expand Up @@ -135,6 +135,8 @@ def _multinomial_loss(w, features, A, Y, alpha, sample_weight):
"""
n_classes = Y.shape[1]
n_features = features.shape[1]
n_neighbors = A.sum(axis=1).reshape(-1, 1)

w = w.reshape(n_classes - 1, -1)
sample_weight = sample_weight[:, np.newaxis]
intercept = w[:, 0]
Expand All @@ -148,9 +150,10 @@ def _multinomial_loss(w, features, A, Y, alpha, sample_weight):
p_nonspatial -= logsumexp(p_nonspatial, axis=1)[:, np.newaxis]
p_nonspatial = np.exp(p_nonspatial, p_nonspatial)

spatial = safe_sparse_dot(A, (Y - p_nonspatial))[:, :-1]
spatial = safe_sparse_dot(A, (Y - p_nonspatial))[:, :-1]/n_neighbors
spatial[np.isnan(spatial)] = 0

p += (eta.T * spatial)
p += eta.T * np.array(spatial)

p = np.hstack((p, np.zeros((features.shape[0], 1))))

Expand Down Expand Up @@ -193,6 +196,8 @@ def _multinomial_loss_grad(w, features, A, Y, alpha, sample_weight):
"""
n_classes = Y.shape[1]
n_features = features.shape[1]
n_neighbors = A.sum(axis=1).reshape(-1, 1)

grad = np.zeros((n_classes - 1, n_features + 2))

loss, p_nonspatial, p, w = _multinomial_loss(w, features, A, Y,
Expand All @@ -205,12 +210,16 @@ def _multinomial_loss_grad(w, features, A, Y, alpha, sample_weight):
for c in range(n_classes - 1):
mu = p_nonspatial[:, c].copy().reshape(-1, 1)
mu *= (1 - mu)
centered_features = features - w[c, -1] * safe_sparse_dot(A, features * mu)
spatial = safe_sparse_dot(A, features * mu)/n_neighbors
spatial[np.isnan(spatial)] = 0
centered_features = features - w[c, -1] * spatial

grad[c, :(n_features + 1)] = safe_sparse_dot(diff[:, c].T, centered_features)

spatial = safe_sparse_dot(A, (Y - p_nonspatial))[:, :-1]
spatial = safe_sparse_dot(A, (Y - p_nonspatial))[:, :-1]/n_neighbors
spatial[np.isnan(spatial)] = 0

grad[:, -1] = (spatial * diff).sum(axis=0)
grad[:, -1] = (diff * np.array(spatial)).sum(axis=0)

grad[:, 1:n_features + 2] += alpha * w[:, 1:]

Expand All @@ -223,14 +232,15 @@ def rpotts(X, model):

n_classes = len(model.classes_)
n_sites = features.shape[0]
n_neighbors = A.sum(axis=1).reshape(-1, 1)

R = np.random.uniform(size=(n_sites, 1))

lower = np.empty((n_sites, 1))
upper = np.empty((n_sites, 1))

betas = model.coef_[:, :-1]
eta = model.coef_[:, -1]
eta = model.coef_[:, -1:]

p = safe_sparse_dot(features, betas.T, dense_output=True)
p += model.intercept_
Expand All @@ -253,16 +263,18 @@ def rpotts(X, model):
r = r.reshape(-1, 1)

upper_multi = _target(upper)
upper_spatial = safe_sparse_dot(A, (upper_multi - p_nonspatial))[:, :-1]
upper_spatial = safe_sparse_dot(A, (upper_multi - p_nonspatial))[:, :-1]/n_neighbors
upper_spatial[np.isnan(upper_spatial)] = 0

upper_p = p + eta * upper_spatial
upper_p = p + (eta.T * np.array(upper_spatial))
upper_p = softmax(np.hstack((upper_p, np.zeros((features.shape[0], 1)))))
upper_p = upper_p.cumsum(axis=1)

lower_multi = _target(lower)
lower_spatial = safe_sparse_dot(A, (lower_multi - p_nonspatial))[:, :-1]
lower_spatial = safe_sparse_dot(A, (lower_multi - p_nonspatial))[:, :-1]/n_neighbors
lower_spatial[np.isnan(lower_spatial)] = 0

lower_p = p + eta * lower_spatial
lower_p = p + (eta.T * np.array(lower_spatial))
lower_p = softmax(np.hstack((lower_p, np.zeros((features.shape[0], 1)))))
lower_p = lower_p.cumsum(axis=1)

Expand Down
68 changes: 36 additions & 32 deletions tests/test_rpotts.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -52,22 +52,24 @@ def setUp(self):

self.Y_multi = potts.lbin.transform(Z)
self.Z = Z


self.avg_neighbors = A.sum(axis=1).mean()

def test_fit(self):
potts = CenteredPotts(C=100000000000000000000)
potts = CenteredPotts(C=float('inf'))
potts.fit((self.X, self.A), self.Z)
assert_array_almost_equal(potts.coef_,
np.array([[ 2.602, 1.298, 0.801]]),
np.array([[ 2.289, 1.796, 2.857]]),
3)
assert_array_almost_equal(potts.intercept_,
np.array([-0.155]),
np.array([-0.092]),
3)


def test_gradient(self):
w = np.array([[0, 1, 1, 0.5]])
features = self.X
w = np.array([[0, 1, 1, 0.5 * self.avg_neighbors]])
A = self.A
features = self.X
Y = self.Y_multi
alpha = 0
sample_weight = np.ones(features.shape[0])
Expand All @@ -76,21 +78,22 @@ def test_gradient(self):
loss, grad, p = out

assert_array_almost_equal(grad,
np.array([ 1.955, -3.102, -1.857, -24.036,]),
np.array([ 1.495, -2.708, -2.25 , -5.798]),
3)

def test_loss(self):
w = np.array([[0, 1, 1, 0.5]])
features = self.X

w = np.array([[0, 1, 1, 0.5 * self.avg_neighbors]])
A = self.A
features = self.X
Y = self.Y_multi
alpha = 0
sample_weight = np.ones(features.shape[0])

out = cp._multinomial_loss(w, features, A, Y, alpha, sample_weight)
loss, p_nonspatial, p, out_w = out

assert_approx_equal(loss, 231.7263)
assert_approx_equal(loss, 232.612, 3)
assert_array_almost_equal(p_nonspatial.sum(axis=1), np.ones(features.shape[0]))
assert_array_almost_equal(p.sum(axis=1), np.ones(features.shape[0]))
assert_array_almost_equal(w, out_w)
Expand All @@ -103,29 +106,29 @@ def test_sampler(self):
potts = CenteredPotts(C=float('inf'))
potts.fit((features, A), Z)

target = [[1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1,
target = [[1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1,
1, 1, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 0,
0, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 1, 0, 1,
0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1,
1, 1, 1, 0, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
0, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 1, 0, 0,
0, 1, 1, 1, 0, 1, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 1,
1, 1, 1, 0, 0, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
0, 0, 0, 1, 1, 1, 1, 1, 1, 0, 1, 0, 1, 1, 0, 0, 1,
1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 1, 0, 0, 1,
0, 1, 1, 0, 1, 1, 1, 0, 0, 0, 1, 1, 0, 1, 1, 1, 0,
0, 1, 1, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 1, 1, 1, 0,
0, 0, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 1, 1, 1, 0, 1,
1, 0, 1, 1, 0, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 1, 0,
0, 0, 1, 1, 1, 0, 1, 0, 0, 0, 1, 1, 1, 1, 1, 0, 1,
1, 0, 1, 1, 0, 0, 1, 1, 1, 0, 0, 0, 1, 0, 0, 1, 0,
1, 1, 0, 1, 1, 0, 0, 1, 0, 0, 0, 1, 0, 1, 0, 0, 1,
1, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1,
0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 1, 1, 0,
0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 0,
0, 0, 0, 0, 1, 1, 1, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0,
1, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 1, 1, 0, 1, 0, 1,
0, 1, 1, 0, 1, 0, 1, 0, 0, 0, 1, 0, 0, 1, 0, 1, 1,
0, 1, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 0,
1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 1, 1, 0, 1,
0, 1, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 0,
1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0,
1, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 1, 1, 0, 0, 0, 1, 0, 1, 0, 1, 0, 0, 1, 1, 0,
0, 0, 0, 0, 0, 1, 1, 1, 0, 1, 0, 1, 0, 1, 1, 1, 0,
0, 0, 0, 0, 1, 1, 1, 0, 1, 1, 0, 1, 1, 0, 0, 1, 0,
1, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 1, 0, 1,
0, 1, 1, 0, 1, 0, 1, 0, 0, 0, 1, 0, 0, 1, 0, 0, 1,
0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 1,
0, 1, 0, 1, 0, 0, 1, 0, 1, 1, 0, 0, 1, 1, 0, 0, 0,
1, 1, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0,
0, 1, 0, 1, 0, 0, 0, 0, 1]]

import random
Expand All @@ -152,6 +155,7 @@ def setUp(self):
potts.fit((X, A), self.Z)

self.Y_multi = potts.lbin.transform(self.Z)
self.avg_neighbors = A.sum(axis=1).mean()


def test_loss(self):
Expand All @@ -167,15 +171,15 @@ def test_loss(self):
out = cp._multinomial_loss(w, features, A, Y, alpha, sample_weight)
loss, p_nonspatial, p, out_w = out

assert_approx_equal(loss, 475.72979)
assert_approx_equal(loss, 450.849, 3)
assert_array_almost_equal(p_nonspatial.sum(axis=1), np.ones(features.shape[0]))
assert_array_almost_equal(p.sum(axis=1), np.ones(features.shape[0]))
assert_array_almost_equal(w, out_w)


def test_gradient(self):
w = np.array([[0, 1, 1, 0.5],
[0, 1, 1, 0.5]])
w = np.array([[0, 1, 1, 0.5 * self.avg_neighbors],
[0, 1, 1, 0.5 * self.avg_neighbors]])
features = self.X
A = self.A
Y = self.Y_multi
Expand All @@ -186,6 +190,6 @@ def test_gradient(self):
loss, grad, p = out

assert_array_almost_equal(grad,
np.array([ 5.281, 1.712, 2.13 , 47.858,
0.3, 1.696, 1.798, 52.718]),
np.array([5.488, 1.661, 2.319, 12.916,
0.063, 1.632, 1.543, 14.744]),
3)

0 comments on commit 351b1ea

Please sign in to comment.