Question

我想开始尝试神经网络，并发现keras在语法上非常简单。我的设置是X_train是一个形状数组（3516,6）和y_train的形状（3516，）

X_train看起来像这样：

[[ 888.          900.5         855.          879.311       877.00266667
   893.5008    ]
 [ 875.          878.5         840.          880.026       874.56933333
   890.7948    ]
 [ 860.          870.          839.5         880.746       870.54333333
   887.6428    ]....]

它是预测一个输出的6个财务数据的输入。我知道它不会准确但是至少在我开始使用RNN之前让我去做一些事情我的问题是每个时期的损失函数显示为nan，精度显示为0％，validation_accuracy显示为零，好像说数据甚至不通过模型，我的意思是即使它是一个输入不良的差模型，即使这应该以大损失为代表吗？这是模型:(见下文）

无论如何，我确信我做错了什么，真的很感谢你们这些家伙＆＃39;输入非常感谢小号

编辑：完整的工作代码：

def load_data(keyword):

    df = pd.read_csv('%s_x.csv' %keyword)
    df2 = pd.read_csv('%s_y.csv' %keyword)

    df2 = df2['label']

    try:
        df.drop('Unnamed: 0', axis = 1, inplace=True)
    except:
        print('wouldnt let drop unnamed column')

    X = df.as_matrix()
    y = df2.as_matrix()

    X_len = len(X)
    test_size = 0.2
    test_split = int(test_size * X_len)
    X_train = X[:-test_split]
    y_train = y[:-test_split]

    X_test = X[-test_split:]
    y_test = y[-test_split:]

def keras():
    model = Sequential( [
        Dense(input_dim=3, output_dim=3),
        Dense(output_dim=60, activation='linear'),
        core.Dropout(p=0.1),
        Dense(60, activation='linear'),
        core.Dropout(p=0.1),
        Dense(1, activation='linear')
    ])
    return model


def training(epoch):
    #  start the program off by loading some data into it
    X_train, X_test, y_train, y_test = load_data('admiral')
    y_train = y_train.reshape(len(y_train), 1)
    y_test = y_test.reshape(len(y_test), 1)


    model = keras()


    # optimizer will go into the compile function
    # RMSpop is apparently a pretty decent choice for recurrent neural networks although we will start it on a simple nn too.
    rms = optimizers.RMSprop(lr=0.001, rho = 0.9, epsilon =1e-08)


    model.compile(optimizer= rms, loss='mean_squared_error ', metrics = ['accuracy'])
    model.fit(X_train, y_train, nb_epoch=epoch, batch_size =500, validation_split=0.01)

    score = model.evaluate(X_test, y_test, batch_size=50)
    print(score)

training(300)

Answer 1

准确度非常低，因为显示准确性没有意义，对于回归问题，它更适合分类

正在传递的数据非常低，这是一个难以回答的问题

损失函数未显示

1 个答案: