不平衡的数据集是类样本的不平衡，导致过度拟合问题，我们通常在它们使用早期回调之前通过自定义回调来处理此问题。（在特定用途上）
您可以对on_epoch_开始或使用目标编号应用指数衰减**。
还可以将权重应用回比率的目标层条件。

施加的重量：

layer_2 = tf.keras.layers.LSTM(32, kernel_initializer=tf.constant_initializer(1.))
b_out =  layer_2(group_1_ShoryuKen_Left)
layer_2.set_weights(layer_1.get_weights())

回拨电话：

class custom_callback(tf.keras.callbacks.Callback):

def on_epoch_end(self, epoch, logs={}):
    if ( logs['loss'] <= 0.4 ) :
        self.model.stop_training = True
            
    if(round(logs['loss'], 6) == round(logs['val_loss'], 6) and round(logs['loss'], 6) > 0 and round(logs['loss'], 6) == round(previous_loss_number, 6) and i_same_loss_number > 50):
        self.model.stop_training = True
                
    if(logs['accuracy'] >= 0.9 and logs['loss'] < 0.2 and int(epoch) == 5 ):
        self.model.stop_training = True

def on_epoch_begin(self, epoch, logs={}):
    global globalstep
    globalstep = globalstep + 1
    previous_loss_number = self.previous_loss_number 
    i_same_loss_number = self.i_same_loss_number
    i_radious = self.i_radious
    lr = self.lr
        
    if not hasattr(self.model.optimizer, "lr"):
        raise ValueError('Optimizer must have a "lr" attribute.')

    if globalstep >  pow(10, i_radious) and ( history.history['loss'][len(history.history['loss']) - 1] == history.history['val_loss'][len(history.history['loss']) - 1] ) and ( history.history['loss'][len(history.history['loss']) - 1] == previous_loss_number ) and ( i_same_loss_number > 100 ) :
        i_radious = i_radious + 1

    lr = round(model.optimizer.lr.numpy(), i_radious)

    ################################################################
    scheduled_lr = self.scheduler(epoch, lr)
    # Set the value back to the optimizer before this epoch starts
    tf.keras.backend.set_value(self.model.optimizer.lr, scheduled_lr)
    print("\nEpoch %05d: Learning rate is %6.7f." % (epoch, scheduled_lr))

custom_callback = custom_callback()

Screenshot