Question

epoch vs loss plot

当我在keras 2.2.4中进行时间序列预测训练时，我超越了时代与损失的图。模型配置1 lstm层，1致密层，num epochs-64。在某些配置集上，我得到了正确的图，只有两条曲线，一条用于验证集，一条用于训练损耗数据集，而在某些配置上，我得到了这个荒谬的图在图像中。我不明白为什么会这样？我的代码-

def train(trainingData, config):
    inputShape, numNode, numEpoch, batchSize = config
    if nDiff > 0:
            trainingData = np.array(difference(trainingData))
    trainX, trainY = trainingData[:, :-1], trainingData[:, -1]
    trainX = trainX.reshape((trainX.shape[0], trainX.shape[1], 1))
    model = Sequential()
    model.add(LSTM(numNode, activation = 'relu', input_shape = (inputShape, 1)))
    #model.add(Dense(4, activation = 'relu'))
    model.add(Dense(1))
    model.compile(loss = 'mse' , optimizer = 'adam')
    history = model.fit(trainX, trainY,  validation_split = 0.2, epochs = numEpoch, batch_size = batchSize, verbose = 0, shuffle = True)
    # summarize history for loss
    plt.plot(history.history['loss'])
    plt.plot(history.history['val_loss'])
    plt.title('model loss')
    plt.ylabel('loss')
    plt.xlabel('epoch')
    plt.legend(['train', 'test'], loc='upper left')
    plt.savefig(modelName + "ind")         
    return model

Answer 1

您val squared = udf((s: Long) => s * s)中是否有metrics=['...']？如果是这样，该图可能会显示除原始损失之外的其他指标。

Answer 2

做

plt.close('all')

完成一个图的绘制后或开始每个图之前

即

替换

plt.savefig(modelName + "ind")

使用

plt.savefig(modelName + "ind")   
plt.close('all')

时间序列预测在喀拉拉邦的损失与时代图

2 个答案: