def kde_entropy(output, var):
dims = K.cast(K.shape(output)[1], K.floatx() ) #int(K.shape(output)[1])
N = K.cast(K.shape(output)[0], K.floatx() )
normconst = (dims/2.0)*K.log(2*np.pi*var)
# Kernel density estimation of entropy
# get dists matrix
x2 = K.expand_dims(K.sum(K.square(output), axis=1), 1)
#x2 = x2 + K.transpose(x2)
#return K.shape(x2)
dists = x2 + K.transpose(x2) - 2*K.dot(output, K.transpose(output))
dists = dists / (2*var)
#y1 = K.expand_dims(output, 0)
#y2 = K.expand_dims(output, 1)
#dists = K.sum(K.square(y1-y2), axis=2) / (2*var)
normCount = N
## Removes effect of diagonals, i.e. leave-one-out entropy
#normCount = N-1
#diagvals = get_diag(10e20*K.ones_like(dists[0,:]))
#dists = dists + diagvals
lprobs = logsumexp(-dists, axis=1) - K.log(normCount) - normconst
h = -K.mean(lprobs)
return nats2bits * h # , normconst + (dims/2.0)
def w_categorical_crossentropy(y_true, y_pred, weights):
nb_cl = len(weights)
final_mask = K.zeros_like(y_pred[:,:, 0])
y_pred_max = K.max(y_pred, axis=-1)
y_pred_max = K.expand_dims(y_pred_max, axis=-1)
y_pred_max_mat = K.cast(K.equal(y_pred, y_pred_max), K.floatx())
for c_p, c_t in itertools.product(range(nb_cl), range(nb_cl)):
final_mask += (weights[c_t, c_p] * y_pred_max_mat[:,:, c_p] * y_true[:,:, c_t])
#return K.mean( K.categorical_crossentropy(y_pred, y_true) * final_mask )
return K.categorical_crossentropy(y_pred, y_true) * final_mask
def accuracy(y_true, y_predicted):
y = tf.argmax(y_true, axis=-1)
y_ = tf.argmax(y_predicted, axis=-1)
mask = tf.greater(y, 0)
return K.cast(K.equal(tf.boolean_mask(y, mask), tf.boolean_mask(y_, mask)),
K.floatx())
def kde_condentropy(output, var):
dims = K.cast(K.shape(output)[1], K.floatx() ) # int(output.get_shape()[1])
# #normconst = (dims/2.0)*K.log(2*np.pi*var)
# #return normconst + (dims/2.0)
normconst = (dims/2.0)*K.log(2*np.pi*var)
return nats2bits * normconst
