小编Hao*_*hen的帖子

我目前正在考虑在 keras 中实施 Self-Attention GAN。我想实现的方式如下：

def Attention(X, channels):
    def hw_flatten(x):
        return np.reshape(x, (x.shape[0], -1, x.shape[-1]))

    f = Conv2D(channels//8, kernel_size=1, strides=1, padding='same')(X)  # [bs, h, w, c']
    g = Conv2D(channels//8, kernel_size=1, strides=1, padding='same')(X)  # [bs, h, w, c']
    h = Conv2D(channels, kernel_size=1, strides=1, padding='same')(X)  # [bs, h, w, c]

    # N = h * w
    flatten_g = hw_flatten(g)
    flatten_f = hw_flatten(f)
    s = np.matmul(flatten_g, flatten_f.reshape((flatten_f.shape[0], flatten_f.shape[-1], -1)))  # [bs, N, N]

    beta = softmax(s, axis=-1)  # attention map

    flatten_h = hw_flatten(h) …

我正在为我的神经网络设计一个自定义层，但我的代码出现错误。

我想做一个注意力层，如论文中所述：SAGAN。和原来的tf代码

class AttentionLayer(Layer):
def __init__(self, **kwargs):
    super(AttentionLayer, self).__init__(**kwargs)

def build(self, input_shape):
    input_dim = input_shape[-1]
    filters_f_g = input_dim // 8
    filters_h = input_dim
    kernel_shape_f_g = (1, 1) + (input_dim, filters_f_g)
    kernel_shape_h = (1, 1) + (input_dim, filters_h)
    # Create a trainable weight variable for this layer:
    self.gamma = self.add_weight(name='gamma', shape=[1], initializer='zeros', trainable=True)
    self.kernel_f = self.add_weight(shape=kernel_shape_f_g,
                                    initializer='glorot_uniform',
                                    name='kernel')
    self.kernel_g = self.add_weight(shape=kernel_shape_f_g,
                                    initializer='glorot_uniform',
                                    name='kernel')
    self.kernel_h = self.add_weight(shape=kernel_shape_h,
                                    initializer='glorot_uniform',
                                    name='kernel')
    self.bias_f = self.add_weight(shape=(filters_f_g,),
                                  initializer='zeros',
                                  name='bias')
    self.bias_g = self.add_weight(shape=(filters_f_g,),
                                  initializer='zeros',
                                  name='bias') …