def on_epoch_begin(self, epoch, logs={}):
vals = self.nlayerinput(self.entropy_train_data)
dists = self.get_dists(vals)
dists += 10e20 * np.eye(dists.shape[0])
r = scipy.optimize.minimize(self.obj, K.get_value(self.kdelayer.logvar).flat[0],
jac=self.jac,
args=(dists,),
)
best_val = r.x.flat[0]
K.set_value(self.kdelayer.logvar, best_val)
def on_epoch_begin(self, epoch, logs={}):
r = scipy.optimize.minimize(self.obj, K.get_value(self.noiselayer.logvar), jac=self.jac)
best_val = r.x[0]
cval = K.get_value(self.noiselayer.logvar)
max_var = 1.0 + cval
if best_val > max_var:
# don't raise it too fast, so that gradient information is preserved
best_val = max_var
K.set_value(self.noiselayer.logvar, best_val)
def build(self, input_shape):
super(GaussianNoise2, self).build(input_shape)
K.set_value(self.logvar, self.init_logvar)
#K.set_value(self.alpha, self.init_alpha)
if self.is_trainable:
self.trainable_weights = [self.logvar,]
else:
self.trainable_weights = []
if self.mi_regularizer:
self.add_loss(self.mi_regularizer())
def jac(logvar):
v = K.get_value(self.noiselayer.logvar)
K.set_value(self.noiselayer.logvar, logvar.flat[0])
r = np.atleast_2d(np.array(jacfunc([self.traindata.X, self.traindata.Y, sampleweights, 1])))[0]
K.set_value(self.noiselayer.logvar, v)
return r
def obj(logvar):
v = K.get_value(self.noiselayer.logvar)
K.set_value(self.noiselayer.logvar, logvar.flat[0])
r = lossfunc([self.traindata.X, self.traindata.Y, sampleweights, 1])[0]
K.set_value(self.noiselayer.logvar, v)
return r
def build(self, input_shape):
super(KDEParamLayer, self).build(input_shape)
K.set_value(self.logvar, self.init_logvar)
self.trainable_weights = []
