Tensorflow - ValueError: Dimensions must be equal, but are 256 and 228

代码是一个2层的LTSM，然后一个前向一个后向，构成双向双层LSTM网络，报错如下 

ValueError: Dimensions must be equal, but are 256 and 228 for 'model/bidirectional_rnn/fw/fw/while/fw/multi_rnn_cell/cell_0/cell_0/basic_lstm_cell/MatMul_1' (op: 'MatMul') with input shapes: [?,256], [228,512].

错误就是维度不匹配，查阅了相关资料后，发现将这个中num_units的size改外embedding size就可以了：

single_cell_fw = tf.contrib.rnn.BasicLSTMCell(num_units=size, state_is_tuple=True)

但是疑惑的是， num_units也就是LSTM输出的维度没必要等于embedding 维度，为啥会报错：

这个是我模型中的维度 word embedding为100d LSTM units 为128d在上述定义中，两个LSTM单元共享参数和维度，也就是说，layer 1 的输入是 128d + 100d 为228 第二次是 第一层的输出 和 第二层前一时刻的隐层状态 也就是128d + 128d 为256，由于两个LSTM共享维度，因此第一层的维度228与第二层的维度256相冲突，然后就报错了，也就是上述错误中 ：

ValueError: Dimensions must be equal, but are 256 and 228

因此要不就换新版本的Tensorflow的方法，要不就单独定义每一层LSTM，不使用循环建立的方式

改成这样，single_cell_fw = [tf.nn.rnn_cell.LSTMCell(size) for size in [128, 128]] 如下代码也就是构建双层的 其余的自己改一下定义就好了不同layer的units大小

        with tf.name_scope('lstm_cell_fw'):
            # single_cell_fw = tf.contrib.rnn.GRUCell(size)
            # if use_lstm:
            #     single_cell_fw = tf.contrib.rnn.BasicLSTMCell(num_units=size, state_is_tuple=True)
            # cell_fw = single_cell_fw
            cell_fw = tf.nn.rnn_cell.LSTMCell(size)
            if num_layers > 1:
                single_cell_fw = [tf.nn.rnn_cell.LSTMCell(size) for size in [128, 128]]
                cell_fw = tf.nn.rnn_cell.MultiRNNCell(single_cell_fw)

            if not forward_only and dropout_keep_prob < 1.0:
                cell_fw = tf.contrib.rnn.DropoutWrapper(cell_fw,
                                                     input_keep_prob=dropout_keep_prob,
                                                     output_keep_prob=dropout_keep_prob)
        with tf.name_scope('lstm_cell_bw'):
            # single_cell_bw = tf.contrib.rnn.GRUCell(size)
            # if use_lstm:
            #     single_cell_bw = tf.contrib.rnn.BasicLSTMCell(num_units=size, state_is_tuple=True)
            # cell_bw = single_cell_bw
            cell_bw = tf.nn.rnn_cell.LSTMCell(size)
            if num_layers > 1:
                single_cell_fw = [tf.nn.rnn_cell.LSTMCell(size) for size in [128, 128]]
                cell_bw = tf.nn.rnn_cell.MultiRNNCell(single_cell_fw)

            if not forward_only and dropout_keep_prob < 1.0:
                cell_bw = tf.contrib.rnn.DropoutWrapper(cell_bw,
                                                        input_keep_prob=dropout_keep_prob,
                                                        output_keep_prob=dropout_keep_prob)

        # init_state = cell.zero_state(batch_size, tf.float32)
        with tf.name_scope('embeddings'):
            embedding = tf.get_variable("embedding", [sent_vocab_size, word_embedding_size]) # 3351 100
        self.embedded_inputs = tf.nn.embedding_lookup(embedding, self.inputs)
        with tf.name_scope('one_hot'):
            self.one_hot_labels = tf.one_hot(self.labels, depth=self.label_vocab_size,name='one-hot-labels')

参考资料：https://tensorflow.google.cn/api_docs/python/tf/nn/dynamic_rnn?hl=zh-cn 

参考资料：https://stackoverflow.com/questions/44615147/valueerror-trying-to-share-variable-rnn-multi-rnn-cell-cell-0-basic-lstm-cell-k

---

 

 还有一种报错的原因是因为Tensorflow的版本的问题：

新版的TensorFlow在MultiRNNCell使用的时候用 for _ in range (nunmber_layers) 代替了 single_cell * nunmber_layers的方式

if num_layers > 1:
     cell_bw = tf.contrib.rnn.MultiRNNCell([single_cell_bw for _ in range(num_layers)])

 

错误部分代码段如下：

        with tf.name_scope('lstm_cell_fw'):
            single_cell_fw = tf.contrib.rnn.GRUCell(size)
            if use_lstm:
                single_cell_fw = tf.contrib.rnn.BasicLSTMCell(num_units=size, state_is_tuple=True)
            cell_fw = single_cell_fw
            if num_layers > 1:
                cell_fw = tf.contrib.rnn.MultiRNNCell([single_cell_fw for _ in range(num_layers)])

            if not forward_only and dropout_keep_prob < 1.0:
                cell_fw = tf.contrib.rnn.DropoutWrapper(cell_fw,
                                                     input_keep_prob=dropout_keep_prob,
                                                     output_keep_prob=dropout_keep_prob)
        with tf.name_scope('lstm_cell_bw'):
            single_cell_bw = tf.contrib.rnn.GRUCell(size)
            if use_lstm:
                single_cell_bw = tf.contrib.rnn.BasicLSTMCell(num_units=size, state_is_tuple=True)
            cell_bw = single_cell_bw
            if num_layers > 1:
                cell_bw = tf.contrib.rnn.MultiRNNCell([single_cell_bw for _ in range(num_layers)])

            if not forward_only and dropout_keep_prob < 1.0:
                cell_bw = tf.contrib.rnn.DropoutWrapper(cell_bw,
                                                        input_keep_prob=dropout_keep_prob,
                                                        output_keep_prob=dropout_keep_prob)

        # init_state = cell.zero_state(batch_size, tf.float32)
        with tf.name_scope('embeddings'):
            embedding = tf.get_variable("embedding", [sent_vocab_size, word_embedding_size]) # 3351 100
        self.embedded_inputs = tf.nn.embedding_lookup(embedding, self.inputs)
        with tf.name_scope('one_hot'):
            self.one_hot_labels = tf.one_hot(self.labels, depth=self.label_vocab_size,name='one-hot-labels')