def batch_cca_loss(y_true, y_pred):
"""
Return Sum of the diagonal - Sum of upper and lower triangles
"""
trace = TN.trace(y_true[0, :, :])
triu_sum = K.sum(K.abs(TB.triu(y_true[0, :, :], k=1)))
tril_sum = K.sum(K.abs(TB.tril(y_true[0, :, :], k=-1)))
return trace - tril_sum - triu_sum
def self_similarity_cosine(x): """ :param x: shape (T,D) :returns cosine similarity matrix, shape (T,T), zeroed at diagonal and upper triangle """ assert x.ndim == 2 x2 = T.sqrt(T.clip(T.sum(T.sqr(x), axis=1), numpy.float32(1.e-20), numpy.float32(1.e20))) assert x2.ndim == 1 x_ = x / x2.dimshuffle(0, 'x') xx = T.dot(x_, x_.T) assert xx.ndim == 2 from theano.tensor.basic import tril return tril(xx, 1)
def self_similarity_cosine(x):
"""
:param x: shape (T,D)
:returns cosine similarity matrix, shape (T,T), zeroed at diagonal and upper triangle
"""
assert x.ndim == 2
x2 = T.sqrt(
T.clip(T.sum(T.sqr(x), axis=1), numpy.float32(1.e-20),
numpy.float32(1.e20)))
assert x2.ndim == 1
x_ = x / x2.dimshuffle(0, 'x')
xx = T.dot(x_, x_.T)
assert xx.ndim == 2
from theano.tensor.basic import tril
return tril(xx, 1)