def main():
config = tf.ConfigProto()
# config.gpu_options.allow_growth = True
# config.log_device_placement = True
with tf.Session(config=config) as sess:
agent = DeepDeterministicPolicyGradientAgent(env=env)
# setup saver util and either load latest ckpt or init variables
saver_util = None
if opts.ckpt_dir is not None:
saver_util = util.SaverUtil(sess, opts.ckpt_dir, opts.ckpt_freq)
else:
sess.run(tf.global_variables_initializer())
for v in tf.global_variables():
print(v.name, util.shape_and_product_of(v), file=sys.stderr)
# print >>sys.stderr, v.name, util.shape_and_product_of(v)
# now that we've either init'd from scratch, or loaded up a checkpoint,
# we can do any required post init work.
agent.post_var_init_setup()
#opts.num_eval = 100
# run either eval or training
if opts.num_eval > 0:
agent.run_eval(opts.num_eval, opts.eval_action_noise)
else:
agent.run_training(opts.max_num_actions, opts.max_run_time,
opts.batch_size, opts.batches_per_step,
saver_util)
if saver_util is not None:
saver_util.force_save()
env.reset() # just to flush logging, clumsy :/
def main():
env = bullet_cartpole.BulletCartpole(gui=opts.gui, action_force=opts.action_force,
max_episode_len=opts.max_episode_len,
initial_force=opts.initial_force, delay=opts.delay,
discrete_actions=False, event_log_file=opts.event_log)
with tf.Session() as sess: #config=tf.ConfigProto(log_device_placement=True)) as sess:
agent = DeepDeterministicPolicyGradientAgent(env=env, agent_opts=opts)
# setup saver util and either load latest ckpt, or init if none...
saver_util = None
ckpt_dir = None
if opts.run_id is not None:
ckpt_dir = "ckpts/%s" % opts.run_id
elif opts.ckpt_dir is not None:
ckpt_dir = opts.ckpt_dir
if ckpt_dir is not None:
saver_util = util.SaverUtil(sess, ckpt_dir, opts.ckpt_freq)
else:
sess.run(tf.initialize_all_variables())
# now that we've either init'd from scratch, or loaded up a checkpoint,
# we can hook together target networks
agent.hook_up_target_networks(opts.target_update_rate)
# run either eval or training
if opts.num_eval > 0:
agent.run_eval(opts.num_eval)
else:
agent.run_training(opts.max_num_actions, opts.batch_size, saver_util, opts.run_id)
if saver_util is not None:
saver_util.force_save()
def main():
env = bullet_cartpole.BulletCartpole(gui=opts.gui,
action_force=opts.action_force,
max_episode_len=opts.max_episode_len,
initial_force=opts.initial_force,
delay=opts.delay,
discrete_actions=True)
with tf.Session() as sess:
agent = LikelihoodRatioPolicyGradientAgent(
env=env,
gui=opts.gui,
hidden_dim=opts.num_hidden,
optimiser=tf.train.AdamOptimizer())
# setup saver util; will load latest ckpt, or init if none...
saver_util = None
ckpt_dir = None
if opts.run_id is not None:
ckpt_dir = "ckpts/%s" % opts.run_id
elif opts.ckpt_dir is not None:
ckpt_dir = opts.ckpt_dir
if ckpt_dir is not None:
saver_util = util.SaverUtil(sess, ckpt_dir, opts.ckpt_freq)
else:
sess.run(tf.initialize_all_variables())
# run either eval or training
if opts.num_eval > 0:
agent.run_eval(opts.num_eval)
else:
agent.run_training(opts.num_train_batches, opts.rollouts_per_batch,
saver_util)
if saver_util is not None:
saver_util.force_save()
def __init__(self, opts):
self.opts = opts
config = tf.ConfigProto()
#config.gpu_options.allow_growth = True
#config.log_device_placement = True
config.gpu_options.per_process_gpu_memory_fraction = 0.5 #opts.gpu_mem_fraction
self.sess = tf.Session(config=config)
render_shape = (opts.height, opts.width, 3)
self.replay_memory = replay_memory.ReplayMemory(
opts=opts, state_shape=render_shape, action_dim=2, load_factor=1.2)
if opts.event_log_in:
self.replay_memory.reset_from_event_log(opts.event_log_in,
opts.event_log_in_num)
# s1 and s2 placeholders
batched_state_shape = [None] + list(render_shape)
s1 = tf.placeholder(shape=batched_state_shape, dtype=tf.float32)
s2 = tf.placeholder(shape=batched_state_shape, dtype=tf.float32)
# initialise base models for value & naf networks. value subportion of net is
# explicitly created seperate because it has a target network note: in the case of
# --share-input-state-representation the input state network of the value_net will
# be reused by the naf.l_value and naf.output_actions net
self.value_net = models.ValueNetwork("value", s1, opts)
self.target_value_net = models.ValueNetwork("target_value", s2, opts)
self.network = models.NafNetwork("naf",
s1,
s2,
self.value_net,
self.target_value_net,
action_dim=2,
opts=opts)
with self.sess.as_default():
# setup saver util and either load latest ckpt or init variables
self.saver_util = None
if opts.ckpt_dir is not None:
self.saver_util = util.SaverUtil(self.sess, opts.ckpt_dir,
opts.ckpt_freq)
else:
self.sess.run(tf.initialize_all_variables())
for v in tf.all_variables():
print >> sys.stderr, v.name, util.shape_and_product_of(v)
# setup target network
self.target_value_net.set_as_target_network_for(
self.value_net, 0.01)