Question

我正在尝试为我的解码器实现Keras模型。为此，我有以下代码：

def Encoder(nc_in=3, input_size=64):
    inp = Input(shape=(input_size, input_size, nc_in))
    x = Conv2D(64, kernel_size=5, kernel_initializer=conv_init, use_bias=False, padding="same")(inp)
    x = conv_block(x,128)
    x = conv_block(x,256)
    x = conv_block(x,512)
    x = conv_block(x,1024)

//RELEVANT CODE IS HERE

dense_layer = Dense(1024)
x = dense_layer(Flatten()(x))
dense_layer_weights = dense_layer.get_weights()
x = AdaptiveDropout(0.5, dense_layer_weights) (x)

//END OF RELEVANT CODE

x = Dense(4*4*1024)(x)
x = Reshape((4, 4, 1024))(x)
out = upscale_ps(512)(x)
return Model(inputs=inp, outputs=out)

我正在创建一个AdaptiveDropout图层，该图层需要前一图层的权重才能运行。

权重在AdaptiveDropout图层中使用如下：

def call(self, inputs, training=None):
    beta_matrix = np.ones_like(self.weights)
    pi = self.alpha * self.weights + beta_matrix
    p = sigmoid(np.tensordot(inputs[1],pi,[[None], [np.maximum(0,pi.ndim - 2)]]))
    uniform = np.uniform(p.shape)
    mask = uniform < p
    if np.mean(p) == 0:
        return p * inputs[0]
    else:
        return inputs[0] * mask

这里的问题是权重和输入都有未知的形状，例如输入的形状是：(?,1024)。形状仅在运行时已知。因此，此代码失败，索引超出范围。我不确定如何解决这个问题，因为我必须在运行时之前定义模型，以后不能添加AdaptiveDropout图层。

Keras模型无法减少输入形状

0 个答案: