这是代码，我想要具有相同符号（变量和函数..）的 pytorch 代码
这是使用 lstm 改进 DQN 算法的张量流实现
对我来说最重要的是翻译而不是细节。
import numpy as np
import tensorflow as tf
# import matplotlib.pyplot as plt
from collections import deque
class DeepQNetwork:
def __init__(self,
n_actions, # the number of actions
n_features,
n_lstm_features,
n_time,
learning_rate = 0.01,
reward_decay = 0.9,
e_greedy = 0.99,
replace_target_iter = 200, # each 200 steps, update target net
memory_size = 500, # maximum of memory
batch_size=32,
e_greedy_increment= 0.00025,
n_lstm_step = 10,
dueling = True,
double_q = True,
N_L1 = 20,
N_lstm = 20):
self.n_actions = n_actions
self.n_features = n_features
self.n_time = n_time
self.lr = learning_rate
self.gamma = reward_decay
self.epsilon_max = e_greedy
self.replace_target_iter = replace_target_iter
self.memory_size = memory_size
self.batch_size = batch_size # select self.batch_size number of time sequence for learning
self.epsilon_increment = e_greedy_increment
self.epsilon = 0 if e_greedy_increment is not None else self.epsilon_max
self.dueling = dueling
self.double_q = double_q
self.learn_step_counter = 0
self.N_L1 = N_L1
# lstm
self.N_lstm = N_lstm
self.n_lstm_step = n_lstm_step # step_size in lstm
self.n_lstm_state = n_lstm_features # [fog1, fog2, ...., fogn, M_n(t)]
# initialize zero memory np.hstack((s, [a, r], s_, lstm_s, lstm_s_))
self.memory = np.zeros((self.memory_size, self.n_features + 1 + 1
+ self.n_features + self.n_lstm_state + self.n_lstm_state))
# consist of [target_net, evaluate_net]
self._build_net()
# replace the parameters in target net
t_params = tf.get_collection('target_net_params') # obtain the parameters in target_net
e_params = tf.get_collection('eval_net_params') # obtain the parameters in eval_net
self.replace_target_op = [tf.assign(t, e) for t, e in
zip(t_params, e_params)] # update the parameters in target_net
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
self.reward_store = list()
self.action_store = list()
self.delay_store = list()
self.lstm_history = deque(maxlen=self.n_lstm_step)
for ii in range(self.n_lstm_step):
self.lstm_history.append(np.zeros([self.n_lstm_state]))
self.store_q_value = list()
def _build_net(self):
tf.reset_default_graph()
def build_layers(s,lstm_s,c_names, n_l1, n_lstm, w_initializer, b_initializer):
# lstm for load levels
with tf.variable_scope('l0'):
lstm_dnn = tf.contrib.rnn.BasicLSTMCell(n_lstm)
lstm_dnn.zero_state(self.batch_size, tf.float32)
lstm_output,lstm_state = tf.nn.dynamic_rnn(lstm_dnn, lstm_s, dtype=tf.float32)
lstm_output_reduced = tf.reshape(lstm_output[:, -1, :], shape=[-1, n_lstm])
# first layer
with tf.variable_scope('l1'):
w1 = tf.get_variable('w1',[n_lstm + self.n_features, n_l1], initializer=w_initializer,
collections=c_names)
b1 = tf.get_variable('b1',[1,n_l1],initializer=b_initializer, collections=c_names)
l1 = tf.nn.relu(tf.matmul(tf.concat([lstm_output_reduced, s],1), w1) + b1)
# second layer
with tf.variable_scope('l12'):
w12 = tf.get_variable('w12', [n_l1, n_l1], initializer=w_initializer,
collections=c_names)
b12 = tf.get_variable('b12', [1, n_l1], initializer=b_initializer, collections=c_names)
l12 = tf.nn.relu(tf.matmul(l1, w12) + b12)
# the second layer is different
if self.dueling:
# Dueling DQN
# a single output n_l1 -> 1
with tf.variable_scope('Value'):
w2 = tf.get_variable('w2',[n_l1,1],initializer=w_initializer,collections=c_names)
b2 = tf.get_variable('b2',[1,1],initializer=b_initializer,collections=c_names)
self.V = tf.matmul(l12,w2) + b2
# n_l1 -> n_actions
with tf.variable_scope('Advantage'):
w2 = tf.get_variable('w2',[n_l1,self.n_actions],initializer=w_initializer,collections=c_names)
b2 = tf.get_variable('b2',[1,self.n_actions],initializer=b_initializer,collections=c_names)
self.A = tf.matmul(l12,w2) + b2
with tf.variable_scope('Q'):
out = self.V + (self.A - tf.reduce_mean(self.A,axis=1,keep_dims=True)) # Q = V(s) +A(s,a)
else:
with tf.variable_scope('Q'):
w2 = tf.get_variable('w2', [n_l1, self.n_actions], initializer=w_initializer, collections=c_names)
b2 = tf.get_variable('b2', [1, self.n_actions], initializer=b_initializer, collections=c_names)
out = tf.matmul(l1, w2) + b2
return out
# input for eval_net
self.s = tf.placeholder(tf.float32,[None,self.n_features], name = 's') # state (observation)
self.lstm_s = tf.placeholder(tf.float32,[None,self.n_lstm_step,self.n_lstm_state], name='lstm1_s')
self.q_target = tf.placeholder(tf.float32,[None,self.n_actions], name = 'Q_target') # q_target
# input for target_net
self.s_ = tf.placeholder(tf.float32, [None, self.n_features], name='s_')
self.lstm_s_ = tf.placeholder(tf.float32,[None,self.n_lstm_step,self.n_lstm_state], name='lstm1_s_')
# generate EVAL_NET, update parameters
with tf.variable_scope('eval_net'):
# c_names(collections_names), will be used when update target_net
# tf.random_normal_initializer(mean=0.0, stddev=1.0, seed=None, dtype=tf.float32), return a initializer
c_names, n_l1, n_lstm, w_initializer, b_initializer = \
['eval_net_params', tf.GraphKeys.GLOBAL_VARIABLES], self.N_L1, self.N_lstm,\
tf.random_normal_initializer(0., 0.3), tf.constant_initializer(0.1) # config of layers
# input (n_feature) -> l1 (n_l1) -> l2 (n_actions)
self.q_eval = build_layers(self.s, self.lstm_s, c_names, n_l1, n_lstm, w_initializer, b_initializer)
# generate TARGET_NET
with tf.variable_scope('target_net'):
c_names = ['target_net_params', tf.GraphKeys.GLOBAL_VARIABLES]
self.q_next = build_layers(self.s_, self.lstm_s_, c_names, n_l1, n_lstm, w_initializer, b_initializer)
# loss and train
with tf.variable_scope('loss'):
self.loss = tf.reduce_mean(tf.squared_difference(self.q_target,self.q_eval))
with tf.variable_scope('train'):
self._train_op = tf.train.RMSPropOptimizer(self.lr).minimize(self.loss)
def store_transition(self, s, lstm_s, a, r, s_, lstm_s_):
# RL.store_transition(observation,action,reward,observation_)
# hasattr(object, name), if object has name attribute
if not hasattr(self, 'memory_counter'):
self.memory_counter = 0
# store np.hstack((s, [a, r], s_, lstm_s, lstm_s_))
transition = np.hstack((s, [a, r], s_, lstm_s, lstm_s_)) # stack in horizontal direction
# if memory overflows, replace old memory with new one
index = self.memory_counter % self.memory_size
# print(transition)
self.memory[index, :] = transition
self.memory_counter += 1
def update_lstm(self, lstm_s):
self.lstm_history.append(lstm_s)
def choose_action(self, observation):
# the shape of the observation (1, size_of_observation)
# x1 = np.array([1, 2, 3, 4, 5]), x1_new = x1[np.newaxis, :], now, the shape of x1_new is (1, 5)
observation = observation[np.newaxis, :]
if np.random.uniform() < self.epsilon:
# lstm only contains history, there is no current observation
lstm_observation = np.array(self.lstm_history)
actions_value = self.sess.run(self.q_eval,
feed_dict={self.s: observation,
self.lstm_s: lstm_observation.reshape(1, self.n_lstm_step,
self.n_lstm_state),
})
self.store_q_value.append({'observation': observation, 'q_value': actions_value})
action = np.argmax(actions_value)
else:
action = np.random.randint(0, self.n_actions)
return action
def learn(self):
# check if replace target_net parameters
if self.learn_step_counter % self.replace_target_iter == 0:
# run the self.replace_target_op in __int__
self.sess.run(self.replace_target_op)
print('\ntarget_params_replaced\n')
# randomly pick [batch_size] memory from memory np.hstack((s, [a, r], s_, lstm_s, lstm_s_))
if self.memory_counter > self.memory_size:
sample_index = np.random.choice(self.memory_size - self.n_lstm_step, size=self.batch_size)
else:
sample_index = np.random.choice(self.memory_counter - self.n_lstm_step, size=self.batch_size)\
# transition = np.hstack(s, [a, r], s_, lstm_s, lstm_s_)
batch_memory = self.memory[sample_index, :self.n_features+1+1+self.n_features]
lstm_batch_memory = np.zeros([self.batch_size, self.n_lstm_step, self.n_lstm_state * 2])
for ii in range(len(sample_index)):
for jj in range(self.n_lstm_step):
lstm_batch_memory[ii,jj,:] = self.memory[sample_index[ii]+jj,
self.n_features+1+1+self.n_features:]
# obtain q_next (from target_net) (to q_target) and q_eval (from eval_net)
# minimize（target_q - q_eval）^2
# q_target = reward + gamma * q_next
# in the size of bacth_memory
# q_next, given the next state from batch, what will be the q_next from q_next
# q_eval4next, given the next state from batch, what will be the q_eval4next from q_eval
q_next, q_eval4next = self.sess.run(
[self.q_next, self.q_eval], # output
feed_dict={
# [s, a, r, s_]
# input for target_q (last)
self.s_: batch_memory[:, -self.n_features:], self.lstm_s_: lstm_batch_memory[:,:,self.n_lstm_state:],
# input for eval_q (last)
self.s: batch_memory[:, -self.n_features:], self.lstm_s: lstm_batch_memory[:,:,self.n_lstm_state:],
}
)
# q_eval, given the current state from batch, what will be the q_eval from q_eval
q_eval = self.sess.run(self.q_eval, {self.s: batch_memory[:, :self.n_features],
self.lstm_s: lstm_batch_memory[:,:,:self.n_lstm_state]})
q_target = q_eval.copy()
batch_index = np.arange(self.batch_size, dtype=np.int32)
eval_act_index = batch_memory[:, self.n_features].astype(int) # action with a single value (int action)
reward = batch_memory[:, self.n_features + 1] # reward with a single value
# update the q_target at the particular batch at the correponding action
if self.double_q:
max_act4next = np.argmax(q_eval4next, axis=1)
selected_q_next = q_next[batch_index, max_act4next]
else:
selected_q_next = np.max(q_next, axis=1)
q_target[batch_index, eval_act_index] = reward + self.gamma * selected_q_next
# both self.s and self.q_target belong to eval_q
# input self.s and self.q_target, output self._train_op, self.loss (to minimize the gap)
# self.sess.run: given input (feed), output the required element
_, self.cost = self.sess.run([self._train_op, self.loss],
feed_dict={self.s: batch_memory[:, :self.n_features],
self.lstm_s: lstm_batch_memory[:, :, :self.n_lstm_state],
self.q_target: q_target})
# gradually increase epsilon
self.epsilon = self.epsilon + self.epsilon_increment if self.epsilon < self.epsilon_max else self.epsilon_max
self.learn_step_counter += 1
def do_store_reward(self, episode, time, reward):
while episode >= len(self.reward_store):
self.reward_store.append(np.zeros([self.n_time]))
self.reward_store[episode][time] = reward
def do_store_action(self,episode,time, action):
while episode >= len(self.action_store):
self.action_store.append(- np.ones([self.n_time]))
self.action_store[episode][time] = action
def do_store_delay(self, episode, time, delay):
while episode >= len(self.delay_store):
self.delay_store.append(np.zeros([self.n_time]))
self.delay_store[episode][time] = delay
pytorch写的代码