Question

我试图在我的RNN模型中命名初始状态，以便我可以从我的.pb文件中调用它。有没有人知道如何解决这个问题？

我采取的步骤：
（1）列车模型
（2）保存模型
（3）冷冻模型＆gt; （.pb）

我现在正试图命名我的状态节点，以便在推理期间使用它们：
（1）我已将输入和输出节点命名为
（2）我已经命名了我的final_state节点
（3）我无法命名我的Initial_state节点

命名Initial_state节点的代码：
initial_state = tf.identity（cell.zero_state（batch_size，tf.int32），name =“initial_state”）

我在尝试保存模型时遇到错误：
TypeError：'Tensor'对象不可迭代。

RNN型号代码：

def build_graph(
    cell_type = None,
    state_size = state_size,
    num_classes = num_classes,
    batch_size = batch_size,
    num_steps = num_steps,
    build_with_dropout = False,
    learning_rate = learning_rate):

    # clean up any residual Tensorflow objects
    reset_graph()

    # data placeholders
    x = tf.placeholder(tf.int32, [batch_size, num_steps], name='x')
    y = tf.placeholder(tf.int32, [batch_size, num_steps], name='y')

    # dropout placeholder
    keep_prob = tf.placeholder(tf.float32, name='keep_prob')

    # embeddings are more efficient than one hot encoding
    # create a lookup table to your inputs that can run in parallel
    embeddings = tf.get_variable('embedding_matrix', [num_classes, state_size])
    rnn_inputs = tf.nn.embedding_lookup(embeddings, x)

    # pick cell type
    if cell_type == 'GRU':
        cell = tf.nn.rnn_cell.GRUCell(state_size)
    elif cell_type == 'LSTM':
        cell = tf.nn.rnn_cell.LSTMCell(state_size, state_is_tuple=True)
    elif cell_type == 'LN_LSTM':
        cell = LayerNormalizedLSTMCell(state_size)
    else:
        cell = tf.nn.rnn_cell.BasicRNNCell(state_size)

    # add dropout    
    if build_with_dropout:
        cell = tf.nn.rnn_cell.DropoutWrapper(cell, input_keep_prob=keep_prob)

    # initialize state    
    init_state = tf.identity(cell.zero_state(batch_size, tf.int32),name="init_state")

    # dynamic_rnn
    rnn_outputs, states = tf.nn.dynamic_rnn(cell, rnn_inputs, initial_state=init_state)

    #with tf.control_dependencies([init_state.assign(states)]):
    #    rnn_outputs = tf.identity(rnn_outputs)

    # These are there simply to give the nodes a name. I.e. they take a node, add another node to it 
    # (which doesn't do much, just identity transform), but crucially, that node can have a name, which 
    # will be saved in the frozen graph can later be accessed in C++. (Note, tf.identity also merges 
    # multiple tensors together, but that's another detail).
    final_state = tf.identity(states,name="final_state")

    with tf.variable_scope('softmax'):
        W = tf.get_variable('W', [state_size, num_classes])
        b = tf.get_variable('b', [num_classes], initializer=tf.constant_initializer(0.0))

    # reshape to get last output
    rnn_outputs = tf.reshape(rnn_outputs, [-1, state_size])
    y_reshaped = tf.reshape(y, [-1])

    logits = tf.matmul(rnn_outputs, W) + b
    predictions = tf.nn.softmax(logits)

    # minimize the mean squared errors.
    total_loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=y_reshaped))

    # pick optimizer
    train_step = tf.train.AdamOptimizer(learning_rate).minimize(total_loss)

    return dict(
        x = x,
        y = y,
        keep_prob = keep_prob,
        init_state = init_state,
        final_state = final_state,
        total_loss = total_loss,
        train_step = train_step,
        preds = predictions,
        saver = tf.train.Saver()
    )

Answer 1

这样的参考代码， https://github.com/BrotherJing/RNN_tabletennis

        initial_state_c = tf.get_collection("initial_state_c")
        initial_state_h = tf.get_collection("initial_state_h")
        initial_state = []
        for i in range(len(initial_state_c)):
            initial_state.append((initial_state_c[i], initial_state_h[i]))

名称冻结RNN模型的初始状态

1 个答案: