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()
type=str,
default='discrete',
help="either 'discrete' or 'continuous'")
bullet_cartpole.add_opts(parser)
opts = parser.parse_args()
actions = map(int, opts.actions.split(","))
if opts.action_type == 'discrete':
discrete_actions = True
elif opts.action_type == 'continuous':
discrete_actions = False
else:
raise Exception("Unknown action type [%s]" % opts.action_type)
env = bullet_cartpole.BulletCartpole(opts=opts,
discrete_actions=discrete_actions)
for _ in xrange(opts.num_eval):
env.reset()
done = False
total_reward = 0
steps = 0
while not done:
if discrete_actions:
action = random.choice(actions)
else:
action = env.action_space.sample()
_state, reward, done, info = env.step(action)
steps += 1
total_reward += reward
if opts.max_episode_len is not None and steps > opts.max_episode_len:
type=str,
default='discrete',
help="either 'discrete' or 'continuous'")
opts = parser.parse_args()
actions = map(int, opts.actions.split(","))
if opts.action_type == 'discrete':
discrete_actions = True
elif opts.action_type == 'continuous':
discrete_actions = False
else:
raise Exception("Unknown action type [%s]" % opts.action_type)
env = bullet_cartpole.BulletCartpole(gui=opts.gui,
initial_force=opts.initial_force,
discrete_actions=discrete_actions,
event_log_file=opts.event_log)
for _ in xrange(opts.num_eval):
env.reset()
done = False
steps = 0
while not done:
if discrete_actions:
action = random.choice(actions)
else:
action = env.action_space.sample()
_state, _reward, done, info = env.step(action)
steps += 1
if opts.delay > 0:
time.sleep(opts.delay)
type=float,
default=50.0,
help="magnitude of action push")
parser.add_argument('--num-train', type=int, default=100)
parser.add_argument('--num-eval', type=int, default=0)
parser.add_argument('--load-file', type=str, default=None)
parser.add_argument('--save-file', type=str, default=None)
parser.add_argument('--delay', type=float, default=0.0)
opts = parser.parse_args()
print "OPTS", opts
ENV_NAME = 'BulletCartpole'
# Get the environment and extract the number of actions.
env = bullet_cartpole.BulletCartpole(gui=opts.gui,
action_force=opts.action_force,
initial_force=opts.initial_force,
delay=opts.delay)
nb_actions = env.action_space.n
# Next, we build a very simple model.
model = Sequential()
model.add(Flatten(input_shape=(1, ) + env.observation_space.shape))
model.add(Dense(32))
model.add(Activation('tanh'))
#model.add(Dense(16))
#model.add(Activation('relu'))
#model.add(Dense(16))
#model.add(Activation('relu'))
model.add(Dense(nb_actions))
model.add(Activation('linear'))
print(model.summary())
