我想连接层（位置嵌入/带有标记嵌入的编码），以便将其用于使用注意机制（转换器）的神经网络，但它不起作用。

在注释（调用方法）中，有我想要的代码。通过加法，将大小为 (None,15,16) 的输入 x 添加到大小为 (15,16) --> x + 位置的位置编码中是没有问题的。但是对于 tf.concat([x,positions],axis=....) 它实际上是一个问题。有人可以帮忙吗？

class concatenatedTokenAndPositionalEmbedding(layers.Layer):
    def __init__(self,maxlen,vocab_size,embed_dim,**kwargs):
        super(concatenatedTokenAndPositionalEmbedding,self).__init__(**kwargs)
        self.maxlen = maxlen
        self.vocab_size = vocab_size
        self.embed_dim = embed_dim

    def build(self,input_shape):
        self.token_emb = layers.Embedding(input_dim=self.vocab_size,output_dim=self.embed_dim)
        self.pos_emb = layers.Embedding(input_dim=self.maxlen,output_dim=self.embed_dim)

    def get_config(self):
        config = super(concatenatedTokenAndPositionalEmbedding,self).get_config()
        config.update({"maxlen": self.maxlen,"vocab_size": self.vocab_size,"embed_dim": self.embed_dim})
        return config

    def call(self,x):
        # todo: concat does not work somehow
        maxlen = tf.shape(x)[-1]
        # positions = tf.zeros((self.maxlen,),dtype=tf.float32)
        positions = tf.range(start=0,limit=self.maxlen,delta=1)
        positions = self.pos_emb(positions)
        x = self.token_emb(x)
        #tf.print(positions)
        print("Pos.:",positions.shape) # prints (15,16)
        print("x shape.:",x.shape) # prints (None,16)
        # do this tf.concat([x,0) or this
        # positions = tf.expand_dims(positions,0)
        # positions.set_dim([None,16])
        # output = layers.concatenate(axis=1)([x,positions])
        # print("Pos.:",positions.shape)
        # print("x shape.:",x.shape)
        # tf.print(output.shape)
        return x + positions # put output here instead

它给了我：

  ValueError: Shape must be rank 3 but is rank 2 for '{{node concatenated_token_and_positional_embedding/concat}} = concatV2[N=2,T=DT_FLOAT,Tidx=DT_INT32](concatenated_token_and_positional_embedding/embedding/embedding_lookup/Identity_1,concatenated_token_and_positional_embedding/embedding_1/embedding_lookup/Identity_1,concatenated_token_and_positional_embedding/concat/axis)' with input shapes: [?,16],[15,[].