def call(self, u_vecs, mask=None): if self.share_weights: u_hat_vecs = K.conv1d(u_vecs, self.W) else: u_hat_vecs = K.local_conv1d(u_vecs, self.W, [1], [1]) batch_size = K.shape(u_vecs)[0] input_num_capsule = K.shape(u_vecs)[1] u_hat_vecs = K.reshape(u_hat_vecs, (batch_size, input_num_capsule, self.num_capsule, self.dim_capsule)) u_hat_vecs = K.permute_dimensions(u_hat_vecs, (0, 2, 1, 3)) # final u_hat_vecs.shape = [None, num_capsule, input_num_capsule, dim_capsule] b = K.zeros_like( u_hat_vecs[:, :, :, 0]) # shape = [None, num_capsule, input_num_capsule] for i in range(self.routings): b = K.permute_dimensions( b, (0, 2, 1)) # shape = [None, input_num_capsule, num_capsule] c = K.softmax(b) c = K.permute_dimensions(c, (0, 2, 1)) b = K.permute_dimensions(b, (0, 2, 1)) outputs = self.activation(K.batch_dot(c, u_hat_vecs, [2, 2])) if i < self.routings - 1: b = K.batch_dot(outputs, u_hat_vecs, [2, 3]) return outputs
def call(self, u_vecs): if self.share_weights: u_hat_vecs = K.conv1d(u_vecs, self.W) else: u_hat_vecs = K.local_conv1d(u_vecs, self.W, [1], [1]) batch_size = K.shape(u_vecs)[0] input_num_capsule = K.shape(u_vecs)[1] u_hat_vecs = K.reshape(u_hat_vecs, (batch_size, input_num_capsule, self.num_capsule, self.dim_capsule)) u_hat_vecs = K.permute_dimensions(u_hat_vecs, (0, 2, 1, 3)) # final u_hat_vecs.shape = [None, num_capsule, input_num_capsule, dim_capsule] b = K.zeros_like(u_hat_vecs[:, :, :, 0]) # shape = [None, num_capsule, input_num_capsule] for i in range(self.routings): c = softmax(b, 1) o = K.batch_dot(c, u_hat_vecs, [2, 2]) if K.backend() == 'theano': o = K.sum(o, axis=1) if i < self.routings - 1: o = K.l2_normalize(o, -1) b = K.batch_dot(o, u_hat_vecs, [2, 3]) if K.backend() == 'theano': b = K.sum(b, axis=1) return self.activation(o)
def call(self, inputs): """Following the routing algorithm from Hinton's paper, but replace b = b + <u,v> with b = <u,v>. This change can improve the feature representation of Capsule. However, you can replace b = K.batch_dot(outputs, hat_inputs, [2, 3]) with b += K.batch_dot(outputs, hat_inputs, [2, 3]) to realize a standard routing. """ if self.share_weights: 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 = K.permute_dimensions(hat_inputs, (0, 2, 1, 3)) b = K.zeros_like(hat_inputs[:, :, :, 0]) for i in range(self.routings): c = softmax(b, 1) o = self.activation(K.batch_dot(c, hat_inputs, [2, 2])) 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 call(self, inputs): if self.share_weights: u_hat_vectors = K.conv1d(inputs, self.W) else: u_hat_vectors = K.local_conv1d(inputs, self.W, [1], [1]) # u_hat_vectors : The spatially transformed input vectors (with local_conv_1d) batch_size = K.shape(inputs)[0] input_num_capsule = K.shape(inputs)[1] u_hat_vectors = K.reshape(u_hat_vectors, (batch_size, input_num_capsule, self.num_capsule, self.dim_capsule)) u_hat_vectors = K.permute_dimensions(u_hat_vectors, (0, 2, 1, 3)) routing_weights = K.zeros_like(u_hat_vectors[:, :, :, 0]) for i in range(self.routings): capsule_weights = K.softmax(routing_weights) outputs = K.batch_dot(capsule_weights, u_hat_vectors, [2, 2]) if K.ndim(outputs) == 4: outputs = K.sum(outputs, axis=1) if i < self.routings - 1: outputs = K.l2_normalize(outputs, -1) routing_weights = K.batch_dot(outputs, u_hat_vectors, [2, 3]) if K.ndim(routing_weights) == 4: routing_weights = K.sum(routing_weights, axis=1) return self.activation(outputs)
def call(self, inputs, **kwargs): """Following the routing algorithm from Hinton's paper, but replace b = b + <u,v> with b = <u,v>. This change can improve the feature representation of the capsule. However, you can replace b = K.batch_dot(outputs, hat_inputs, [2, 3]) with b += K.batch_dot(outputs, hat_inputs, [2, 3]) to get standard routing. """ if self.share_weights: 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 = K.permute_dimensions(hat_inputs, (0, 2, 1, 3)) b = K.zeros_like(hat_inputs[:, :, :, 0]) print(self.routings) for i in range(self.routings): c = K.softmax(b, 1) o = self.activation(K.batch_dot(c, hat_inputs, [2, 2])) 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 call(self, u_vecs): if self.share_weights: u_hat_vecs = K.conv1d(u_vecs, self.W) # bsz,200,160 else: u_hat_vecs = K.local_conv1d(u_vecs, self.W, [1], [1]) batch_size = K.shape(u_vecs)[0] input_num_capsule = K.shape(u_vecs)[1] u_hat_vecs = K.reshape( u_hat_vecs, (batch_size, input_num_capsule, self.num_capsule, self.dim_capsule)) # bsz,200,10,16 u_hat_vecs = K.permute_dimensions(u_hat_vecs, (0, 2, 1, 3)) # bsz,10,200,16 # final u_hat_vecs.shape = [None, num_capsule, input_num_capsule, dim_capsule] b = K.zeros_like(u_hat_vecs[:, :, :, 0]) # bsz,10,200 for i in range(self.routings): b = K.permute_dimensions(b, (0, 2, 1)) # bsz,200,10 c = K.softmax(b) # bsz,200,10, 对最后一个轴10维进行sofrmax c = K.permute_dimensions(c, (0, 2, 1)) # bsz,10,200 b = K.permute_dimensions(b, (0, 2, 1)) # bsz,200,10 outputs = self.activation(K.batch_dot(c, u_hat_vecs, [2, 2])) # bsz,10,16 if i < self.routings - 1: b = K.batch_dot(outputs, u_hat_vecs, [2, 3]) # bsz,10,200 return outputs
def call(self, u_vecs): if self.share_weights: u_hat_vecs = K.conv1d(u_vecs, self.W) #none*82*600 1*600*(32*300) else: u_hat_vecs = K.local_conv1d(u_vecs, self.W, [1], [1]) batch_size = K.shape(u_vecs)[0] #82*600, [0]=none input_num_capsule = K.shape(u_vecs)[1] #[1]=82 u_hat_vecs = K.reshape(u_hat_vecs, (batch_size, input_num_capsule, self.num_capsule, self.dim_capsule)) u_hat_vecs = K.permute_dimensions( u_hat_vecs, (0, 2, 1, 3) ) #final u_hat_vecs.shape = [None, num_capsule, input_num_capsule, dim_capsule] b = K.zeros_like( u_hat_vecs[:, :, :, 0] ) # shape = [None, num_capsule, input_num_capsule] 32*82 for i in range(self.routings): b = K.permute_dimensions( b, (0, 2, 1)) # shape = [None, input_num_capsule, num_capsule] 82*32 c = K.softmax(b) c = K.permute_dimensions(c, (0, 2, 1)) #[32*82] b = K.permute_dimensions(b, (0, 2, 1)) #[32*82] outputs = self.activation(K.batch_dot(c, u_hat_vecs, [2, 2])) if i < self.routings - 1: b = K.batch_dot(outputs, u_hat_vecs, [2, 3]) #[..*32*[128]] [?*32*82*[128]]=[32*82] return outputs
def call(self, inputs, training=None): if self.share_weights: 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 = K.permute_dimensions(hat_inputs, (0, 2, 1, 3)) b = K.zeros_like(hat_inputs[:, :, :, 0]) print("b:") print(K.int_shape(b)) for i in range(self.routings): c = tf.nn.softmax(b, dim=1) print("c:") print(K.int_shape(c)) print("hat_inputs:") print(K.int_shape(hat_inputs)) o = tf.keras.backend.batch_dot(c, hat_inputs, [2, 2]) print("o:") print(K.int_shape(o)) o = self.activation(o) print("o:") print(K.int_shape(o)) if i < self.routings - 1: b += tf.keras.backend.batch_dot(o, hat_inputs, [2, 3]) return o
def call(self, u_vecs): if self.share_weights: u_hat_vecs = K.conv1d(u_vecs, self.W) else: u_hat_vecs = K.local_conv1d(u_vecs, self.W, [1], [1]) batch_size = K.shape(u_vecs)[0] input_num_capsule = K.shape(u_vecs)[1] u_hat_vecs = K.reshape(u_hat_vecs, (batch_size, input_num_capsule, self.num_capsule, self.dim_capsule)) u_hat_vecs = K.permute_dimensions(u_hat_vecs, (0, 2, 1, 3)) b = K.zeros_like(u_hat_vecs[:, :, :, 0]) for i in range(self.routings): c = softmax(b, 1) o = K.batch_dot(c, u_hat_vecs, [2, 2]) o = K.sum( o, axis=1 ) # the previous version does not have this op, due to the recent updates of keras if i < self.routings - 1: o = K.l2_normalize( o, -1 ) # here is for normalize if it not works, remove this line b = K.batch_dot(o, u_hat_vecs, [2, 3]) b = K.sum( b, axis=1 ) # the previous version does not have this op, due to the recent updates of keras return o
def call(self, inputs): if self.share_weights: 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 = K.permute_dimensions(hat_inputs, (0, 2, 1, 3)) b = K.zeros_like(hat_inputs[:, :, :, 0]) for i in range(self.routings): c = softmax(b, 1) o = self.activation(K.batch_dot(c, hat_inputs, [2, 2])) if i < self.routings - 1: b = K.batch_dot(o, hat_inputs, [2, 3]) return o