def __init__(self, A_DIM, **kws):
#=====================================
# SETUP
#=====================================
self.sess = tf.Session()
self.A_DIM = A_DIM
self.actor = a3c.ActorNetwork(self.sess,
state_dim=[S_INFO, S_LEN],
action_dim=A_DIM,
learning_rate=ACTOR_LR_RATE)
self.critic = a3c.CriticNetwork(self.sess,
state_dim=[S_INFO, S_LEN],
action_dim=A_DIM,
learning_rate=CRITIC_LR_RATE)
self.sess.run(tf.initialize_all_variables())
self.saver = tf.train.Saver() # save neural net parameters
# restore neural net parameters
nn_model = NN_MODEL
if nn_model is not None: # nn_model is the path to file
self.saver.restore(self.sess, nn_model)
print("Model restored.")
def __init__(self, actionset = [], infoDept=S_LEN, infoDim=S_INFO, log_path=None, summary_dir=None, nn_model=None):
assert summary_dir
myprint("Central init Params:", actionset, infoDept, log_path, summary_dir, nn_model)
self.summary_dir = os.path.join(summary_dir, "rnnQuality")
self.nn_model = nn_model
self.a_dim = len(actionset)
self._vActionset = actionset
self._vInfoDim = infoDim
self._vInfoDept = infoDept
if not os.path.exists(self.summary_dir):
os.makedirs(self.summary_dir)
self.sess = tf.Session()
# log_file = open(os.path.join(log_path, "PensiveLearner", "wb"))
self.actor = a3c.ActorNetwork(self.sess,
state_dim=[self._vInfoDim, self._vInfoDept], action_dim=self.a_dim,
learning_rate=ACTOR_LR_RATE)
self.critic = a3c.CriticNetwork(self.sess,
state_dim=[self._vInfoDim, self._vInfoDept], action_dim=self.a_dim,
learning_rate=CRITIC_LR_RATE)
self.summary_ops, self.summary_vars = a3c.build_summaries()
self.sess.run(tf.global_variables_initializer())
self.writer = tf.summary.FileWriter(self.summary_dir, self.sess.graph) # training monitor
self.saver = tf.train.Saver() # save neural net parameters
self.epoch = 0
# restore neural net parameters
if self.nn_model is None:
nn_model, epoch = guessSavedSession(self.summary_dir)
if nn_model:
self.nn_model = nn_model
self.epoch = epoch
# nn_model = NN_MODEL
if self.nn_model is not None: # nn_model is the path to file
self.saver.restore(self.sess, self.nn_model)
myprint("Model restored.")
self.actor_gradient_batch = []
self.critic_gradient_batch = []
def __init__(self,
actionset=[],
infoDept=S_LEN,
infoDim=S_INFO,
log_path=None,
summary_dir=None,
nn_model=None,
ipcQueue=None,
ipcId=None,
readOnly=False):
assert summary_dir
assert (not ipcQueue and not ipcId) or (ipcQueue and ipcId)
myprint("Pensieproc init Params:", actionset, infoDept, log_path,
summary_dir, nn_model)
self.ipcQueue = ipcQueue
self.pid = os.getpid()
self.ipcId = ipcId
self.summary_dir = os.path.join(summary_dir, "rnnBuffer")
self.nn_model = None if not nn_model else os.path.join(
self.summary_dir, nn_model)
self.a_dim = len(actionset)
self._vActionset = actionset
self._vInfoDim = infoDim
self._vInfoDept = infoDept
self._vReadOnly = readOnly
if not os.path.exists(self.summary_dir):
os.makedirs(self.summary_dir)
self.sess = tf.Session()
# log_file = open(os.path.join(log_path, "PensiveLearner", "wb"))
self.actor = a3c.ActorNetwork(
self.sess,
state_dim=[self._vInfoDim, self._vInfoDept],
action_dim=self.a_dim,
learning_rate=ACTOR_LR_RATE)
self.critic = a3c.CriticNetwork(
self.sess,
state_dim=[self._vInfoDim, self._vInfoDept],
action_dim=self.a_dim,
learning_rate=CRITIC_LR_RATE)
self.summary_ops, self.summary_vars = a3c.build_summaries()
self.sess.run(tf.global_variables_initializer())
self.writer = tf.summary.FileWriter(
self.summary_dir, self.sess.graph) # training monitor
self.saver = tf.train.Saver() # save neural net parameters
# restore neural net parameters
self.epoch = 0
if self.nn_model is None and not self.ipcQueue:
nn_model, epoch = guessSavedSession(self.summary_dir)
if nn_model:
self.nn_model = nn_model
self.epoch = epoch
# nn_model = NN_MODEL
if self.nn_model is not None and not self.ipcQueue: # nn_model is the path to file
self.saver.restore(self.sess, self.nn_model)
myprint("Model restored with `" + self.nn_model + "'")
if self.ipcQueue:
self.ipcQueue[0].put({
"id": self.ipcId,
"pid": self.pid,
"cmd": IPC_CMD_PARAM
})
myprint("=" * 50)
myprint(self.ipcId, ": waiting for ipc")
myprint("=" * 50)
res = None
while True:
res = self.ipcQueue[1].get()
pid = res["pid"]
res = res["res"]
if pid == self.pid:
break
actor_net_params, critic_net_params = res
self.actor.set_network_params(actor_net_params)
self.critic.set_network_params(critic_net_params)
myprint("=" * 50)
myprint(self.ipcId, ": ipcOver")
myprint("=" * 50)
self.s_batch = []
self.a_batch = []
self.r_batch = []
self.entropy_record = []
self.actor_gradient_batch = []
self.critic_gradient_batch = []
self.keyedSBatch = {}
self.keyedActionProb = {}
self.keyedAction = {}
self.keyedInputParam = {}
def __init__(self,
actionset=[],
infoDept=S_LEN,
infoDim=S_INFO,
log_path=None,
summary_dir=None,
nn_model=None,
readOnly=False,
master=False,
ipcQueues=None):
assert summary_dir
assert readOnly
assert master == (ipcQueues is not None)
self.master = master
cprint.cyan("!!CREATING OBJECT!!")
if not master:
cprint.blue("!!SLAVE!!")
self.recv, self.send = [mp.Queue(), mp.Queue()]
self.proc = mp.Process(
target=PensiveLearner,
args=(actionset, infoDept, infoDim, log_path, summary_dir,
nn_model, readOnly, True, [self.send, self.recv]))
self.proc.start()
while True:
cmd = self.recv.get()
if cmd == "ready":
break
return
cprint.red("!!MASTER!!")
self.recv, self.send = ipcQueues
# myprint("Pensieproc init Params:", actionset, infoDept, log_path, summary_dir, nn_model)
self.pid = os.getpid()
self.summary_dir = summary_dir
self.nn_model = None if not nn_model else os.path.join(
self.summary_dir, nn_model)
self.a_dim = len(actionset)
self._vActionset = actionset
self._vInfoDim = infoDim
self._vInfoDept = infoDept
self._vReadOnly = readOnly
if not os.path.exists(self.summary_dir):
os.makedirs(self.summary_dir)
self.sess = tf.Session()
self.actor = a3c.ActorNetwork(
self.sess,
state_dim=[self._vInfoDim, self._vInfoDept],
action_dim=self.a_dim,
learning_rate=ACTOR_LR_RATE)
self.critic = a3c.CriticNetwork(
self.sess,
state_dim=[self._vInfoDim, self._vInfoDept],
action_dim=self.a_dim,
learning_rate=CRITIC_LR_RATE)
self.summary_ops, self.summary_vars = a3c.build_summaries()
self.sess.run(tf.global_variables_initializer())
self.writer = tf.summary.FileWriter(
self.summary_dir, self.sess.graph) # training monitor
self.saver = tf.train.Saver() # save neural net parameters
# restore neural net parameters
self.epoch = 0
if self.nn_model is None and self.master:
nn_model, epoch = guessSavedSession(self.summary_dir)
if nn_model:
self.nn_model = nn_model
self.epoch = epoch
if self.nn_model is not None: # nn_model is the path to file
self.saver.restore(self.sess, self.nn_model)
cprint.red("Model restored with `" + self.nn_model + "'")
try:
self.runCmd()
except Exception as ex:
track = tb.format_exc()
cprint.red(track)
def __init__(self, videoInfo, agent, log_file_path=LOG_FILE, *kw, **kws):
self.video = None
#=====================================
# SETUP
#=====================================
log_file = None if not log_file_path else open(log_file_path, 'wb')
sess = tf.Session()
A_DIM = len(videoInfo.bitratesKbps)
actor = a3c.ActorNetwork(sess,
state_dim=[S_INFO, S_LEN],
action_dim=A_DIM,
learning_rate=ACTOR_LR_RATE)
critic = a3c.CriticNetwork(sess,
state_dim=[S_INFO, S_LEN],
action_dim=A_DIM,
learning_rate=CRITIC_LR_RATE)
sess.run(tf.initialize_all_variables())
saver = tf.train.Saver() # save neural net parameters
# restore neural net parameters
nn_model = NN_MODEL
if nn_model is not None: # nn_model is the path to file
saver.restore(sess, nn_model)
print("Model restored.")
init_action = np.zeros(A_DIM)
init_action[DEFAULT_QUALITY] = 1
s_batch = [np.zeros((S_INFO, S_LEN))]
a_batch = [init_action]
r_batch = []
train_counter = 0
# need this storage, because observation only contains total rebuffering time
# we compute the difference to get
input_dict = {
'sess': sess,
'log_file': log_file,
'actor': actor,
'critic': critic,
'saver': saver,
'train_counter': train_counter,
's_batch': s_batch,
'a_batch': a_batch,
'r_batch': r_batch,
}
#=====================================
# INIT
#=====================================
self.input_dict = input_dict
self.sess = input_dict['sess']
self.log_file = input_dict['log_file']
self.actor = input_dict['actor']
self.critic = input_dict['critic']
self.saver = input_dict['saver']
self.s_batch = input_dict['s_batch']
self.a_batch = input_dict['a_batch']
self.r_batch = input_dict['r_batch']