编辑：我已经确定了以下行作为错误的来源，
unstacked_observation_outputs = tf.nn.relu(tf.nn.xw_plus_b(unstacked_observation_outputs[
tf.reduce_max(length(observation_history)).eval(session = self.sess) - 1],
weights['second_observation'], biases['second_observation']))
特别是，在tf.reduce_max(length(observation_history)).eval(session = self.sess) - 1中应用.eval()是问题。
我有两个非常相似的类。一个叫Actor((和另一个评论家((。当我实例化演员对象如下：
class Actor(object):
def __init__(self, sess, observation_dimension, action_dimension, learning_rate, tau):
self.sess = sess
self.observation_dimension = observation_dimension
self.action_dimension = action_dimension
self.learning_rate = learning_rate
self.tau = tau
self.lstm_cell = rnn.BasicLSTMCell(lstm_hidden_units)
with tf.variable_scope('lstm') as scope:
self.observation_history, self.action_history = self.depthless_lstm()
self.network_parameters = tf.trainable_variables()
scope.reuse_variables()
self.target_observation_history, self.target_action_history = self.depthless_lstm()
self.target_network_parameters = tf.trainable_variables()[len(self.network_parameters):]
def depthless_lstm(self):
state = tf.placeholder(tf.float32, [None, 3600, self.observation_dimension])
weights = {
'0': tf.get_variable('weights_0', [lstm_hidden_units, hidden_units_01], initializer = tf.random_normal_initializer()),
'1': tf.get_variable('weights_1', [hidden_units_01, 1], initializer = tf.random_normal_initializer())
}
biases = {
'0': tf.get_variable('biases_0', [hidden_units_01], initializer = tf.random_normal_initializer()),
'1': tf.get_variable('biases_1', [1], initializer = tf.random_normal_initializer())
}
lstm_outputs, learned_state = tf.nn.dynamic_rnn(self.lstm_cell, state, dtype = tf.float32)
unstacked_lstm_outputs = tf.unstack(lstm_outputs, axis = 1)
list_of_hidden_01_outputs = [tf.nn.relu(tf.nn.xw_plus_b(matrix, weights['0'], biases['0']))
for matrix in unstacked_lstm_outputs]
actions = [tf.nn.softmax(tf.nn.xw_plus_b(matrix, weights['1'], biases['1']))
for matrix in list_of_hidden_01_outputs]
actions = [actions[n] for n in range(len(actions)) if (n % frequency) == 1]
return state, actions
actor = Actor(sess, observation_dimension, action_dimension, learning_rate, tau)
我没有收到任何错误。但是，当我实例化评论家对象时：
class Critic(object):
def __init__(self, sess, observation_dimension, action_dimension, learning_rate, tau,
model, number_actor_variables):
self.sess = sess
self.observation_dimension = observation_dimension
self.action_dimension = action_dimension
self.learning_rate = learning_rate
self.tau = tau
self.lstm_observation_cell = rnn.BasicLSTMCell(lstm_observation_units)
with tf.variable_scope('next_lstm') as scope:
self.observation_history, self.action, self.critique = self.depthless_lstm_critic()
self.network_parameters = tf.trainable_variables()[number_actor_variables:]
scope.reuse_variables()
self.target_observation_history, self.target_action, self.target_critique = self.depthless_lstm_critic()
self.target_network_parameters = tf.trainable_variables()[(len(self.network_parameters)
+ number_actor_variables):]
def depthless_lstm_critic(self):
observation_history = tf.placeholder(tf.float32, [None, 3600, self.observation_dimension])
action = tf.placeholder(tf.float32, [None, self.action_dimension])
weights = {
'first_observation': tf.get_variable('first_observation', [self.observation_dimension, observation_hidden_units01],
initializer = tf.random_normal_initializer()),
'first_action': tf.get_variable('first_action', [self.action_dimension, action_hidden_units01],
initializer = tf.random_normal_initializer()),
'second_observation': tf.get_variable('second_observation', [lstm_observation_units, observation_hidden_units02],
initializer = tf.random_normal_initializer()),
'second_action': tf.get_variable('second_action', [action_hidden_units01, action_hidden_units02],
initializer = tf.random_normal_initializer()),
'combined_observation': tf.get_variable('combined_observation', [observation_hidden_units02, combined_units],
initializer = tf.random_normal_initializer()),
'combined_action': tf.get_variable('combined_action', [action_hidden_units02, combined_units],
initializer = tf.random_normal_initializer()),
'final': tf.get_variable('final', [combined_units, 1],
initializer = tf.random_normal_initializer())
}
biases = {
'first_observation': tf.get_variable('_first_observation', [observation_hidden_units01],
initializer = tf.random_normal_initializer()),
'first_action': tf.get_variable('_first_action', [action_hidden_units01],
initializer = tf.random_normal_initializer()),
'second_observation': tf.get_variable('_second_observation', [observation_hidden_units02],
initializer = tf.random_normal_initializer()),
'second_action': tf.get_variable('_second_action', [action_hidden_units01],
initializer = tf.random_normal_initializer()),
'combined': tf.get_variable('_combined', [combined_units],
initializer = tf.random_normal_initializer()),
'final': tf.get_variable('_final', [1],
initializer = tf.random_normal_initializer())
}
unstacked_observation_history = tf.unstack(observation_history, axis = 1)
unstacked_observation_history = [tf.nn.relu(tf.nn.xw_plus_b(vector,
weights['first_observation'], biases['first_observation']))
for vector in unstacked_observation_history]
learned_action = tf.nn.relu(tf.nn.xw_plus_b(action, weights['first_action'], biases['first_action']))
stacked_observation_history = tf.stack(unstacked_observation_history, axis = 1)
lstm_observation_outputs, observation_states = tf.nn.dynamic_rnn(self.lstm_observation_cell,
stacked_observation_history, dtype = tf.float32,
sequence_length = length(stacked_observation_history))
unstacked_observation_outputs = tf.unstack(lstm_observation_outputs, axis = 1)
unstacked_observation_outputs = tf.nn.relu(tf.nn.xw_plus_b(unstacked_observation_outputs[
tf.reduce_max(length(observation_history)).eval(session = self.sess) - 1],
weights['second_observation'], biases['second_observation']))
learned_action = tf.nn.relu(tf.nn.xw_plus_b(learned_action, weights['second_action'],
biases['second_action']))
combined_results = tf.nn.relu(tf.matmul(weights['combined_observation'],
unstacked_observation_outputs) + tf.nn.xw_plus_b(learned_action, weights['combined_action'],
biases['combined']))
critique = tf.nn.xw_plus_b(combined_results, weights['final'], biases['final'])
return observation_history, action, critique
critic = Critic(sess, observation_dimension, action_dimension, learning_rate, tau,
model, number_actor_variables)
我收到以下错误消息：
InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'next_lstm/Placeholder' with dtype float
[[Node: next_lstm/Placeholder = Placeholder[dtype=DT_FLOAT, shape=[], _device="/job:localhost/replica:0/task:0/cpu:0"]()]]
和追溯
Caused by op 'next_lstm/Placeholder', defined at:
File "test_critic.py", line 89, in <module>
model, number_actor_variables)
File "test_critic.py", line 15, in __init__
self.observation_history, self.action, self.critique =
self.depthless_lstm_critic()
File "test_critic.py", line 23, in depthless_lstm_critic
observation_history = tf.placeholder(tf.float32, [None, 3600,
self.observation_dimension])
追溯表明问题是占位符observation_history。但是，代码的那一部分似乎与我为Actor((类写的内容相同，该类别没有产生错误消息。
observation_history是一个占位符，这意味着在构建图形并等待一些实际值时，它是图中的一个空节点。因此，您需要将一些价值传递给此占位符。
Tensor.eval({self.placeholderA : someValueA, self.placeholderB : someValueB ...}, self.sess)
在您的情况下，您需要将一些实际值传递给observation_history。