def loss(yTrue, yPred):
quantile_in = np.random.random()
bellman_errors = yTrue - yPred
tau = np.array(quantile_in)
print("loss",(K.abs(tau - K.cast(bellman_errors < 0,"float32")) * huber_loss(yTrue,yPred)) / kappa)
if kappa > 0:
return (K.abs(tau - K.cast(bellman_errors < 0,"float32")) * huber_loss(yTrue,yPred)) / kappa
return (K.abs(tau - K.cast(bellman_errors < 0,"float32")) * bellman_errors)
def huber_loss_wrapped_function(y_true, y_pred):
return huber_loss(y_true, y_pred, **huber_loss_kwargs)
def reconstruction(img, reconstructed_img):
return .5 * huber_loss(img, reconstructed_img) + .5 * DSSIMObjective()(
img, reconstructed_img)
def huber_loss_quantile(self,tau): def loss(yTrue,yPred): bellman_errors = yTrue - yPred return (K.abs(tau - K.to_float(bellman_errors < 0)) * huber_loss(yTrue,yPred)) / self.kappa
def loss(yTrue, yPred):
bellman_errors = yPred - yTrue
#tau = np.array(quantile_in)
#print("loss",(K.abs(tau - K.cast(bellman_errors < 0,"float32")) * huber_loss(yTrue,yPred)) / kappa)
return (K.abs(tau - K.cast(bellman_errors < 0, "float32")) *
huber_loss(yTrue, yPred)) / kappa
