def load_history_policy_model(self, model_file):
AI_net.create_policy_network(3, "/cpu:0", "/cpu:0")
history_policy = PolicyRLNetwork(self.planes,
None,
filters=self.filters,
board_size=self.board_size,
model_dir=self.model_dir,
device="cpu",
distributed_train=False)
history_policy.session = tf.Session(config=tf.ConfigProto(
log_device_placement=False, allow_soft_placement=True))
history_policy.session.run(tf.initialize_all_variables())
history_policy.restore_model(model_file=model_file)
return history_policy
def __init__(self, planes, args, phase=1, filters=192, board_size=15, model_dir="./value_net_models",
model_file=None,
device="gpu", gpu=1, optimizer="sgd", learn_rate=1e-6, distributed_train=False):
self.board_size = board_size
self.phase = phase
self.planes = planes
# init network
if distributed_train:
ps_device = "/job:ps/task:0/cpu:0"
worker_device = "/job:worker/task:%d/gpu:%d" % (args.task_index, args.gpu_id)
else:
ps_device = "/cpu:0"
if device == "cpu":
worker_device = "/cpu:0"
else:
worker_device = "/gpu:%d" % gpu
self.tf_var = dict()
self.tf_var["in"], self.tf_var["out"] = AI_net.create_value_network(
planes, ps_device, worker_device, filters=filters, board_size=self.board_size, name_prefix="value_net")
# super init
AI_net.SuperNetwork.__init__(self, model_dir=model_dir)
history_step = int(self.param_unserierlize(init_params={"global_step": 0})["global_step"])
with tf.device(ps_device):
self.global_step = tf.Variable(history_step)
# loss function
with tf.device(worker_device):
self.loss_function(optimizer, learn_rate, args.values_net_batch_size)
def __init__(self, planes, corpus, policy_args, filters=192, board_size=15, model_dir="./policy_dl_models",
device="gpu", gpu=1, optimizer="sgd", learn_rate=1e-6, distributed_train=False):
self.board_size = board_size
# init network
if distributed_train:
ps_device = "/job:ps/task:0/cpu:0"
worker_device = "/job:worker/task:%d/gpu:%d" % (policy_args.task_index, policy_args.gpu_id)
else:
# ps_device = "/cpu:0"
if device == "cpu":
worker_device = "/cpu:0"
else:
worker_device = "/gpu:%d" % gpu
ps_device = worker_device
self.tf_var = dict({"in": tf.placeholder(tf.float32, [None, board_size, board_size, planes])})
with tf.device(worker_device):
with tf.name_scope('tower_%d' % 0) as scope:
self.tf_var["out"] = AI_net.create_policy_network(self.tf_var["in"],
planes, filters=filters, board_size=self.board_size,
layers=5)
# super init
AI_net.SuperNetwork.__init__(self, model_dir=model_dir)
history_step = int(self.param_unserierlize(init_params={"global_step": 0})["global_step"])
with tf.device(ps_device):
self.global_step = tf.Variable(history_step)
# loss function
with tf.device(worker_device):
# num_batchs = corpus.num_batchs_per_epochs(policy_args.policy_dl_batch_size)
# decay_steps = policy_args.policy_dl_epochs_per_decay * num_batchs
# learn_rate = tf.train.exponential_decay(learn_rate, self.global_step, decay_steps,
# policy_args.policy_dl_decay_rate, staircase=True)
self.tf_var["lr"] = tf.Variable(learn_rate)
self.loss_function(optimizer, learn_rate)
def __init__(self,
planes,
corpus,
policy_args,
filters=192,
board_size=15,
model_dir="./policy_dl_models",
device="gpu",
gpu=1,
optimizer="sgd",
learn_rate=1e-6,
distributed_train=False):
self.board_size = board_size
# init network
if distributed_train:
ps_device = "/job:ps/task:0/cpu:0"
worker_device = "/job:worker/task:%d/gpu:%d" % (
policy_args.task_index, policy_args.gpu_id)
else:
# ps_device = "/cpu:0"
if device == "cpu":
worker_device = "/cpu:0"
else:
worker_device = "/gpu:%d" % gpu
ps_device = worker_device
self.tf_var = dict({
"in":
tf.placeholder(tf.float32, [None, board_size, board_size, planes])
})
with tf.device(worker_device):
with tf.name_scope('tower_%d' % 0) as scope:
self.tf_var["out"] = AI_net.create_policy_network(
self.tf_var["in"],
planes,
filters=filters,
board_size=self.board_size,
layers=5)
# super init
AI_net.SuperNetwork.__init__(self, model_dir=model_dir)
history_step = int(
self.param_unserierlize(
init_params={"global_step": 0})["global_step"])
with tf.device(ps_device):
self.global_step = tf.Variable(history_step)
# loss function
with tf.device(worker_device):
# num_batchs = corpus.num_batchs_per_epochs(policy_args.policy_dl_batch_size)
# decay_steps = policy_args.policy_dl_epochs_per_decay * num_batchs
# learn_rate = tf.train.exponential_decay(learn_rate, self.global_step, decay_steps,
# policy_args.policy_dl_decay_rate, staircase=True)
self.tf_var["lr"] = tf.Variable(learn_rate)
self.loss_function(optimizer, learn_rate)
def __init__(self,
planes,
policy_args,
phase=1,
filters=192,
board_size=15,
model_dir="./policy_rl_models",
device="gpu",
gpu=1,
optimizer="sgd",
learn_rate=1e-6,
distributed_train=False):
self.board_size = board_size
self.phase = phase
self.planes = planes
self.filters = filters
# init network
if distributed_train:
ps_device = "/job:ps/task:0/cpu:0"
worker_device = "/job:worker/task:%d/gpu:%d" % (
policy_args.task_index, policy_args.gpu_id)
else:
ps_device = "/cpu:0"
if device == "cpu":
worker_device = "/cpu:0"
else:
worker_device = "/gpu:%d" % gpu
self.tf_var = dict({
"in":
tf.placeholder(tf.float32, [None, board_size, board_size, planes])
})
with tf.device(worker_device):
with tf.name_scope('tower_%d' % 0) as scope:
self.tf_var["out"] = AI_net.create_policy_network(
self.tf_var["in"],
planes,
filters=filters,
board_size=self.board_size,
layers=5)
# super init
AI_net.SuperNetwork.__init__(self, model_dir=model_dir)
history_step = int(
self.param_unserierlize(
init_params={"global_step": 0})["global_step"])
with tf.device(ps_device):
self.global_step = tf.Variable(history_step)
# loss function
with tf.device(worker_device):
self.loss_function(optimizer, learn_rate)
def __init__(self,
planes,
args,
phase=1,
filters=192,
board_size=15,
model_dir="./value_net_models",
model_file=None,
device="gpu",
gpu=1,
optimizer="sgd",
learn_rate=1e-6,
distributed_train=False):
self.board_size = board_size
self.phase = phase
self.planes = planes
# init network
if distributed_train:
ps_device = "/job:ps/task:0/cpu:0"
worker_device = "/job:worker/task:%d/gpu:%d" % (args.task_index,
args.gpu_id)
else:
ps_device = "/cpu:0"
if device == "cpu":
worker_device = "/cpu:0"
else:
worker_device = "/gpu:%d" % gpu
self.tf_var = dict()
self.tf_var["in"], self.tf_var["out"] = AI_net.create_value_network(
planes,
ps_device,
worker_device,
filters=filters,
board_size=self.board_size,
name_prefix="value_net")
# super init
AI_net.SuperNetwork.__init__(self, model_dir=model_dir)
history_step = int(
self.param_unserierlize(
init_params={"global_step": 0})["global_step"])
with tf.device(ps_device):
self.global_step = tf.Variable(history_step)
# loss function
with tf.device(worker_device):
self.loss_function(optimizer, learn_rate,
args.values_net_batch_size)