TensorFlow.Keras时序分类seq2seq模型RNNCell属性缺失错误求助

你触发报错的核心原因是：tfa.seq2seq.AttentionWrapper要求传入的第一个参数必须是RNNCell实例或堆叠的RNNCell集合，但你的代码中将封装好的RNN层对象直接传入了AttentionWrapper，RNN层是处理完整序列计算的高阶包装类，没有Cell要求的output_size、state_size属性，因此触发了类型错误。

修复步骤 #

1. 解码器核心逻辑修改 #

不需要将堆叠的LSTMCell再封装为RNN层，而是用StackedRNNCells将多个LSTMCell包装为统一的Cell实例，直接传给AttentionWrapper即可：

class Decoder(tf.keras.Model):
    def __init__(self, char2numY, dec_units, batch_sz, layers, attention_type='luong',  bidirectional=False):
        super(Decoder, self).__init__()
        self.batch_sz = batch_sz
        self.dec_units = dec_units #must be two times of enc_units
        self.attention_type = attention_type

        # Final Dense layer on which softmax will be applied
        self.fc = Dense(char2numY, use_bias=False)

        # Define the fundamental cell for decoder recurrent structure
        if bidirectional:
            decoder_cells = [LSTMCell(hidden_dim*2) for hidden_dim in layers]
        else:
            decoder_cells = [LSTMCell(hidden_dim) for hidden_dim in layers]
 
        # 【修改点1】用StackedRNNCells封装多Cell列表，不要包RNN层
        self.decoder_stacked_cells = tf.keras.layers.StackedRNNCells(decoder_cells)

        # Sampler
        self.sampler = tfa.seq2seq.sampler.TrainingSampler()

        # Create attention mechanism with memory = None
        self.attention_mechanism = self.build_attention_mechanism(self.dec_units, 
                                                          None, None, self.attention_type)

        # Wrap attention mechanism with the fundamental rnn cell of decoder
        self.rnn_cell = self.build_rnn_cell(batch_sz)

        # Define the decoder with respect to fundamental rnn cell
        self.decoder = tfa.seq2seq.BasicDecoder(self.rnn_cell, sampler=self.sampler, 
                                                                         output_layer=self.fc)


    def build_rnn_cell(self, batch_sz):
        # 【修改点2】传入堆叠的Cell实例，而非RNN层
        rnn_cell = tfa.seq2seq.AttentionWrapper(self.decoder_stacked_cells, 
                            self.attention_mechanism, attention_layer_size=self.dec_units)
        return rnn_cell

    def build_attention_mechanism(self, dec_units, memory, memory_sequence_length, 
                                                                      attention_type='luong'):
        return tfa.seq2seq.LuongAttention(units=self.dec_units, memory=memory, 
                                                                  memory_sequence_length=None)

    def build_initial_state(self, batch_sz, encoder_state, Dtype):
        decoder_initial_state = self.rnn_cell.get_initial_state(batch_size=batch_sz, 
                                                                                 dtype=Dtype)
        decoder_initial_state = decoder_initial_state.clone(cell_state=encoder_state)
        return decoder_initial_state

    def call(self, inputs, initial_state):
        x = inputs
        outputs, final_states, final_sequence_lengths = self.decoder(x, 
                                            initial_state=initial_state, sequence_length=None)
        logits = outputs.rnn_output
        return outputs, final_states

2. 测试代码配套修改 #

# 【修改点3】调整Decoder初始化传参顺序，和__init__参数对应
decoder = Decoder(vocab_size, dec_units, batch_sz, layers, 'luong', bidirectional=True)
sample_x = tf.random.uniform((batch_sz, output_max_length, vocab_size)) # 输入要对应嵌入后的维度
# 【修改点4】用编码器输出作为注意力的memory，原sample_output未定义
decoder.attention_mechanism.setup_memory(encoder_outputs_and_states[0])
initial_state = decoder.build_initial_state(batch_sz, encoder_states, tf.float32) 
sample_decoder_outputs, _ = decoder(sample_x, initial_state)  

print("Decoder Outputs Shape: ", sample_decoder_outputs.rnn_output.shape)

内容的提问来源于stack exchange，提问作者Chandra