Question

我正在尝试通过Numpy复制使用Keras的model.predict()获得的输出。我的keras模型图层如下：

_________________________________________________________________
Layer (type)                 Output Shape              Param    
=================================================================
main_input (InputLayer)      (None, 10, 76)            0         
_________________________________________________________________
masking (Masking)            (None, 10, 76)            0         
_________________________________________________________________
rnn (SimpleRNN)              [(None, 64), (None, 64)]  9024      
_________________________________________________________________
dropout_15 (Dropout)         (None, 64)                0         
_________________________________________________________________
dense1 (Dense)               (None, 64)                4160      
_________________________________________________________________
denseoutput (Dense)          (None, 1)                 65        
=================================================================
Total params: 13,249
Trainable params: 13,249
Non-trainable params: 0

SimpleRNN层的第二个输出是return_state=True返回的状态。

我尝试了2种不同的方法。首先，我计算了 WXt + Us + b ，其中 W 是内核， Xt 是输入， U 是循环内核， s 是通过return_state=True获得的状态，而 b 是偏差。这返回了与通过predict()（函数mult_1）获得的输出相似的输出。

在那之后，我尝试使用功能mult_2的类似方法，但是获得的结果要比mult_1差。

def mult_1(X):
    X = ma.masked_values(X, -99)
    s = (model.predict(X)[1])

    W = (model.get_weights()[0])
    U = (model.get_weights()[1])
    b = (model.get_weights()[2])

    Wx = np.dot(X[:,-1,:], W)
    Us = np.dot(s,U)

    output = Wx + Us + b

    return np.tanh(output)

def mult2(X):
    max_habitantes = X.shape[1]
    i = 0
    s_0 = np.ones((X.shape[0], 64)) # initial state
    X = ma.masked_values(X, -99)

    while i < 10:
        Xt = X[:,i,:]
        if i == 0:
            s = s_0
        else:
            s = output

        W = (model.get_weights()[0])
        U = (model.get_weights()[1])
        b = (model.get_weights()[2])

        Wx = np.dot(Xt, W)
        Us = np.dot(s,U)

        output = np.tanh(Wx + Us +b)
        i = i+1

    return output

尽管与predict()的预测没什么不同，但预测有些偏离。我是在错误地做乘法吗？

Answer 1

您应该使用零数组作为mult_2中rnn的初始状态。以下两个代码段将为您提供相同的结果：

x = np.random.rand(1,10,76)

使用Keras model.predict（）

inputs = Input(shape=(10,76), dtype=np.float32)
_, state = SimpleRNN(units=64, return_state=True)(inputs)
out_drop = Dropout(0.2)(state)
out_d1 = Dense(64, activation='tanh')(out_drop)
out = Dense(1, activation='tanh')(out_d1)

model = Model(inputs, out)

In [1]: model.predict(x) Out[1]: array([[-0.82426485]]

使用numpy函数进行预测：

def rnn_pred(X):
    """
    Same as your mult_2 func. but with zero init. for rnn initial state
    """
    W = (model.get_weights()[0])
    U = (model.get_weights()[1])
    b = (model.get_weights()[2])

    max_habitantes = X.shape[1]
    i = 0
    s_0 = np.zeros((X.shape[0], 64)) # initial state

    while i < 10:
        Xt = X[:,i,:]
        if i == 0:
            s = s_0
        else:
            s = output

        Wx = np.dot(Xt, W)
        Us = np.dot(s,U)

        output = np.tanh(Wx+Us+b)
        i = i+1

    return output

def dense_pred(rnn_out):
    U_d1 = (model.get_weights()[3]) # dense64 weights
    b_d1 = (model.get_weights()[4]) # dense64 bias
    U_d2 = (model.get_weights()[5]) # dense1 weights
    b_d2 = (model.get_weights()[6]) # dense1 bias

    out1 = np.dot(rnn_out, U_d1) + b_d1
    out1 = np.tanh(out1)
    out2 = np.dot(out1, U_d2) + b_d2
    out2 = np.tanh(out2)
    return out2

In [2]: dense_pred(rnn_pred(x)) Out[2]: array([[-0.82426485]])

如何在keras中手动获取与model.predict（）相同的输出

1 个答案: