Tensorflow 1.0：从恢复的 RNN 中检索隐藏状态

Question

我想恢复 RNN 并获取隐藏状态。

我做了类似的事情来保存 RNN：

loc="path/to/save/rnn"
with tf.variable_scope("lstm") as scope:
    outputs, state = tf.nn.dynamic_rnn(..)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
save_path = saver.save(sess,loc)

现在我想取回state。

graph = tf.Graph()
sess = tf.Session(graph=graph)
with graph.as_default():
      saver = tf.train.import_meta_graph(loc + '.meta', clear_devices=True)
      saver.restore(sess, loc)
      state= ...

Answer 1

您可以将 state 张量添加到图 collection, which is basically a key value store to track tensors, using tf.add_to_collection and retrieve it later using tf.get_collection 中。例如：

loc="path/to/save/rnn"
with tf.variable_scope("lstm") as scope:
    outputs, state = tf.nn.dynamic_rnn(..)
    tf.add_to_collection('state', state)


graph = tf.Graph()
with graph.as_default():
      saver = tf.train.import_meta_graph(loc + '.meta', clear_devices=True)
      state = tf.get_collection('state')[0]  # Note: tf.get_collection returns a list.

Tensorflow 1.0：从恢复的 RNN 中检索隐藏状态

Tensorflow 1.0 : Retreive hidden state from restored RNN

restore

tensorflow

recurrent-neural-network