keras 中lstm 和bidirectional lstm (Bilstm)的return_sequence和return_state之间的关系

import tensorflow as tf
tf.enable_eager_execution()



embedding = tf.Variable(tf.truncated_normal((2, 3, 4)))

lstm = tf.keras.layers.LSTM(units=5, return_sequences=False, return_state=False)
outputs = lstm(embedding)  # return_sequences=False, return_state=False
print(outputs)  # 只有每个样本的最后一个time step的输出， shape:(batch, hidden_units)

lstm = tf.keras.layers.LSTM(units=5, return_sequences=True, return_state=False)

outputs = lstm(embedding)
print(outputs)   # (batch, seq_len, hidden_units)


lstm = tf.keras.layers.LSTM(units=5, return_sequences=True, return_state=True)
outputs, hidden, state = lstm(embedding)
"""
outputs: (batch, seq_len, hidden_units)
hidden: (batch, hidden_units), 是每个样本对应最后一个time step的输出， 这个输出对应着各自样本的最后一个time step
state: (batch, hidden_units), 是每个样本对应的记忆状态， 代表着cell state。
"""
print(outputs)
print(hidden)
print(state)




bilstm = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(5, return_sequences=False, return_state=False), merge_mode="concat")
outputs = bilstm(embedding)  # shape: (batch, hidden_units * 2), 这个取决于merge_mode参数的设置
print(outputs)
"""
outputs: shape: (batch, hidden_units * 2), 这个取决于merge_mode参数的设置
"""




bilstm = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(5, return_sequences=True, return_state=False), merge_mode="concat")
outputs = bilstm(embedding)
print(outputs)
"""
outputs: shape=(batch,seq_len,hidden_units * 2), 最后一个维度值取决于merge_mode
"""
print("====")
bilstm = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(2, return_sequences=True, return_state=True), merge_mode="concat")
outputs = bilstm(embedding)
print(outputs)
"""
o, f_o, f_c, b_o, b_c = outputs
输出包括三部分：
1、layer output
2、 （h, c） for forward lstm
3、 （h, c） for backward lstm

"""