def build_loss(embeddings):
"""Return a pair (loss, p) given a theano shared variable representing the
`embeddings`.
`loss` is a theano variable for the loss. `p` is a symbolic variable
representing the target neighbour probabilities on which the loss depends.
"""
# Probability that two points are neighbours in the embedding space.
emb_dists = distance_matrix(embeddings)
emb_top = zero_diagonal(1 / (1 + emb_dists))
emb_bottom = emb_top.sum(axis=0)
q = emb_top / emb_bottom
# Incorrect normalization which does not matter since we normalize p i
# the same way.
q /= q.sum()
q = T.maximum(q, 1E-12)
p_ji_var = T.matrix('neighbour_probabilities')
p_ji_var.tag.test_value = np.random.random(
(10, 10)).astype(theano.config.floatX)
p_ji_var_floored = T.maximum(p_ji_var, 1E-12)
# t-distributed stochastic neighbourhood embedding loss.
loss = (p_ji_var * T.log(p_ji_var_floored / q)).sum()
return loss, p_ji_var
def build_loss(embeddings):
"""Return a pair (loss, p) given a theano shared variable representing the
`embeddings`.
`loss` is a theano variable for the loss. `p` is a symbolic variable
representing the target neighbour probabilities on which the loss depends.
"""
# Probability that two points are neighbours in the embedding space.
emb_dists = distance_matrix(embeddings)
emb_top = zero_diagonal(1 / (1 + emb_dists))
emb_bottom = emb_top.sum(axis=0)
q = emb_top / emb_bottom
# Incorrect normalization which does not matter since we normalize p i
# the same way.
q /= q.sum()
q = T.maximum(q, 1E-12)
p_ji_var = T.matrix('neighbour_probabilities')
p_ji_var_floored = T.maximum(p_ji_var, 1E-12)
# t-distributed stochastic neighbourhood embedding loss.
loss = (p_ji_var * T.log(p_ji_var_floored / q)).sum()
return loss, p_ji_var
def test_distance_matrix():
X = T.matrix()
D = distance_matrix(X)
f = theano.function([X], D, mode='FAST_COMPILE')
x = np.array([[1], [2], [3]]).astype(theano.config.floatX)
res = f(x)
print res
correct = np.allclose(res, np.array([[0, 1, 4], [1, 0, 1], [4, 1, 0]]))
assert correct, 'distance matrix not working right'
def test_distance_matrix():
X = T.matrix()
D = distance_matrix(X) ** 2
f = theano.function([X], D, mode='FAST_COMPILE')
x = np.array([[1], [2], [3]]).astype(theano.config.floatX)
res = f(x)
print res
correct = np.allclose(res, (np.array([[0, 1, 4], [1, 0, 1], [4, 1, 0]])))
assert correct, 'distance matrix not working right'
def test_distance_matrix():
X = T.matrix()
D = distance_matrix(X)
f = theano.function([X], D, mode='FAST_COMPILE')
x = np.array([[1], [2], [3]])
res = f(x)
print res
correct = roughly(res, np.array([[0, 1, 4], [1, 0, 1], [4, 1, 0]]))
assert correct, 'distance matrix not working right'
