CapsNet.zip

from keras import backend as K from keras.layers import Layer """ 压缩函数,使用0.5替代hinton论文中的1,如果是1，所有的向量的范数都将被缩小。 如果是0.5，小于0.5的范数将缩小，大于0.5的将被放大 """ def squash(x, axis=-1): s_quared_norm = K.sum(K.square(x), axis, keepdims=True) + K.epsilon() #||x||^2 scale = K.sqrt(s_quared_norm) / (0.5 + s_quared_norm) #||x||/(0.5+||x||^2) result = scale * x return result # 定义我们自己的softmax函数，而不是K.softmax.因为K.softmax不能指定轴 def softmax(x, axis=-1): ex = K.exp(x - K.max(x, axis=axis, keepdims=True)) result = ex / K.sum(ex, axis=axis, keepdims=True) return result # 定义边缘损失，输入y_true, p_pred，返回分数，传入fit即可 def margin_loss(y_true, y_pred): lamb, margin = 0.5, 0.1 result = K.sum(y_true * K.square(K.relu(1 - margin -y_pred)) + lamb * (1-y_true) * K.square(K.relu(y_pred - margin)), axis=-1) return result class Capsule(Layer): def __init__(self, num_capsule, dim_capsule, routings=3, share_weights=True, activation='squash', **kwargs): super(Capsule, self).__init__(**kwargs) # Capsule继承**kwargs参数 self.num_capsule = num_capsule self.dim_capsule = dim_capsule self.routings = routings self.share_weights = share_weights if activation == 'squash': self.activation = squash else: self.activation = activation.get(activation) # 得到激活函数 # 定义权重 def build(self, input_shape): input_dim_capsule = input_shape[-1] if self.share_weights: # 自定义权重 self.kernel = self.add_weight( #[row,col,channel]->[1,input_dim_capsule,num_capsule*dim_capsule] name='capsule_kernel', shape=(1, input_dim_capsule, self.num_capsule * self.dim_capsule), initializer='glorot_uniform', trainable=True) else: input_num_capsule = input_shape[-2] self.kernel = self.add_weight( name='capsule_kernel', shape=(input_num_capsule, input_dim_capsule, self.num_capsule * self.dim_capsule), initializer='glorot_uniform', trainable=True) super(Capsule, self).build(input_shape) # 必须继承Layer的build方法 # 层的功能逻辑(核心) def call(self, inputs): if self.share_weights: #inputs: [batch, input_num_capsule, input_dim_capsule] #kernel: [1, input_dim_capsule, num_capsule*dim_capsule] #hat_inputs: [batch, input_num_capsule, num_capsule*dim_capsule] hat_inputs = K.conv1d(inputs, self.kernel) else: hat_inputs = K.local_conv1d(inputs, self.kernel, [1], [1]) batch_size = K.shape(inputs)[0] input_num_capsule = K.shape(inputs)[1] hat_inputs = K.reshape(hat_inputs, (batch_size, input_num_capsule, self.num_capsule, self.dim_capsule)) #hat_inputs: [batch, input_num_capsule, num_capsule, dim_capsule] hat_inputs = K.permute_dimensions(hat_inputs, (0, 2, 1, 3)) #hat_inputs: [batch, num_capsule, input_num_capsule, dim_capsule] b = K.zeros_like(hat_inputs[:, :, :, 0]) #b: [batch, num_capsule, input_num_capsule] for i in range(self.routings): c = softmax(b, 1) o = self.activation(K.batch_dot(c, hat_inputs, [2, 2])) if K.backend() == 'theano': o = K.sum(o, axis=1) if i < self.routings-1: b += K.batch_dot(o, hat_inputs, [2, 3]) if K.backend() == 'theano': o = K.sum(o, axis=1) return o def compute_output_shape(self, input_shape): # 自动推断shape return (None, self.num_capsule, self.dim_capsule)