运行LSTM模型遇AttributeError：keras.layers无LSTMStateTuple属性

错误信息 #

AttributeError: module 'keras.api._v2.keras.layers' has no attribute 'LSTMStateTuple'

问题代码 #

# Create cell state and hidden state variables to maintain the state of the LSTM
c, h = [], []
initial_state = []

for li in range(n_layers):
    c.append(tf.Variable(tf.zeros([batch_size, num_nodes[li]]), trainable=False))
    h.append(tf.Variable(tf.zeros([batch_size, num_nodes[li]]), trainable=False))
    initial_state.append(tf.keras.layers.LSTMStateTuple(c[li], h[li]))

# Do several tensor transformations, because the function dynamic_rnn requires the output to be of
# a specific format. Read more at: https://www.tensorflow.org/api_docs/python/tf/nn/dynamic_rnn

all_inputs = tf.concat([tf.expand_dims(t, 0) for t in train_inputs], axis=0)

# Create LSTM layer
lstm_layer = tf.keras.layers.LSTM(num_nodes[-1], return_sequences=True, return_state=True, dropout=dropout)

# Pass inputs and initial state to the LSTM layer
all_lstm_outputs, final_state, _ = lstm_layer(all_inputs, initial_state=initial_state)

all_outputs = tf.keras.layers.Dense(1)(all_lstm_outputs)

split_outputs = tf.split(all_outputs, num_unrollings, axis=0)

解决方案 #

原因 #

LSTMStateTuple不属于tf.keras.layers模块，它是TensorFlow 1.x旧版RNN API的组件。在TensorFlow 2.x的原生Keras API中，LSTM层接受的初始状态可以直接用**(细胞状态c, 隐藏状态h)**的普通元组表示，无需依赖专门的LSTMStateTuple类。

修复代码 #

修改初始化状态的循环部分，用普通元组替代LSTMStateTuple：

# Create cell state and hidden state variables to maintain the state of the LSTM
c, h = [], []
initial_state = []

for li in range(n_layers):
    c.append(tf.Variable(tf.zeros([batch_size, num_nodes[li]]), trainable=False))
    h.append(tf.Variable(tf.zeros([batch_size, num_nodes[li]]), trainable=False))
    # 直接使用元组传递状态，替代LSTMStateTuple
    initial_state.append((c[li], h[li]))

备选方案（兼容旧代码） #

如果需要保留LSTMStateTuple的写法，可以导入正确的兼容模块：

from tensorflow.compat.v1.nn.rnn_cell import LSTMStateTuple

但更推荐使用普通元组的方式，适配TF2.x的原生Keras API。

内容的提问来源于stack exchange，提问作者Swapnil kumar Sahu