def focal_loss_fixed(y_true, y_pred): gamma = 2 eps = 1e-6 alpha = 0.5 y_pred=K.clip(y_pred,eps,1.-eps)#improve the stability of the focal loss and see issues 1 for more information pt_1 = tf.where(tf.equal(y_true, 1), y_pred, tf.ones_like(y_pred)) pt_0 = tf.where(tf.equal(y_true, 0), y_pred, tf.zeros_like(y_pred)) return -K.mean(alpha * K.pow(1. - pt_1, gamma) * K.log(pt_1))-K.sum((1-alpha) * K.pow( pt_0, gamma) * K.log(1. - pt_0),axis=-1)
def total_variation_loss(x): assert K.ndim(x) == 4 if K.image_dim_ordering() == 'th': a = K.square(x[:, :, :img_width - 1, :img_height - 1] - x[:, :, 1:, :img_height - 1]) b = K.square(x[:, :, :img_width - 1, :img_height - 1] - x[:, :, :img_width - 1, 1:]) else: a = K.square(x[:, :img_width - 1, :img_height - 1, :] - x[:, 1:, :img_height - 1, :]) b = K.square(x[:, :img_width - 1, :img_height - 1, :] - x[:, :img_width - 1, 1:, :]) return K.sum(K.pow(a + b, 1.25))
def loss(self,y_true,y_pred): """ executes the focal loss # Arguments y_true : true class values y_pred : predicted class values from the model # Returns fl : mean focal loss for the given batch """ y_pred = self.clipping(y_pred) fl = -(K.sum((self.c_weights(self.class_weights) * K.pow(1.-y_pred,self.gamma) * (y_true * K.log(y_pred))),axis=-1)) fl = K.sum((self.p_weights(self.pixel_weights) * fl),axis=(1,2)) fl = K.mean(fl, axis=0) return fl/1000 ## scaling down the loss to prevent gradient explosion
def get_updates(self, params, loss): grads = self.get_gradients(loss, params) self.updates = [K.update_add(self.iterations, 1)] lr = self.lr if self.initial_decay > 0: lr *= (1. / (1. + self.decay * K.cast(self.iterations, K.dtype(self.decay)))) t = K.cast(self.iterations, K.floatx()) + 1 lr_t = lr * K.sqrt(1. - K.pow(self.beta_2, t)) / ( 1. - K.pow(self.beta_1, t)) shapes = [K.get_variable_shape(p) for p in params] ms = [K.zeros(shape) for shape in shapes] vs = [K.zeros(shape) for shape in shapes] self.weights = [self.iterations] + ms + vs for p, g, m, v in zip(params, grads, ms, vs): # if a weight tensor (len > 1) use weight normalized parameterization # this is the only part changed w.r.t. keras.optimizers.Adam ps = K.get_variable_shape(p) if len(ps) > 1: # get weight normalization parameters V, V_norm, V_scaler, g_param, grad_g, grad_V = get_weightnorm_params_and_grads( p, g) # Adam containers for the 'g' parameter V_scaler_shape = K.get_variable_shape(V_scaler) m_g = K.zeros(V_scaler_shape) v_g = K.zeros(V_scaler_shape) # update g parameters m_g_t = (self.beta_1 * m_g) + (1. - self.beta_1) * grad_g v_g_t = (self.beta_2 * v_g) + (1. - self.beta_2) * K.square(grad_g) new_g_param = g_param - lr_t * m_g_t / (K.sqrt(v_g_t) + self.epsilon) self.updates.append(K.update(m_g, m_g_t)) self.updates.append(K.update(v_g, v_g_t)) # update V parameters m_t = (self.beta_1 * m) + (1. - self.beta_1) * grad_V v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(grad_V) new_V_param = V - lr_t * m_t / (K.sqrt(v_t) + self.epsilon) self.updates.append(K.update(m, m_t)) self.updates.append(K.update(v, v_t)) # if there are constraints we apply them to V, not W # if p in constraints: # c = constraints[p] # new_V_param = c(new_V_param) # wn param updates --> W updates add_weightnorm_param_updates(self.updates, new_V_param, new_g_param, p, V_scaler) else: # do optimization normally m_t = (self.beta_1 * m) + (1. - self.beta_1) * g v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g) p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon) self.updates.append(K.update(m, m_t)) self.updates.append(K.update(v, v_t)) new_p = p_t # apply constraints # if p in constraints: # c = constraints[p] # new_p = c(new_p) self.updates.append(K.update(p, new_p)) return self.updates
def focal_loss(y_true, y_pred): gamma = 2.0 alpha = 0.25 pt_1 = tf.where(tf.equal(y_true, 1), y_pred, tf.ones_like(y_pred)) pt_0 = tf.where(tf.equal(y_true, 0), y_pred, tf.zeros_like(y_pred)) return -K.sum(alpha * K.pow(1. - pt_1, gamma) * K.log(pt_1))-K.sum((1-alpha) * K.pow( pt_0, gamma) * K.log(1. - pt_0))
def total_variation_loss(x): a = KTF.square(x[:, :img_height - 1, :img_width - 1, :] - x[:, 1:, :img_width - 1, :]) b = KTF.square(x[:, :img_height - 1, :img_width - 1, :] - x[:, :img_height - 1, 1:, :]) return KTF.sum(KTF.pow(a + b, 1.25))