def __init__(self, save_path, observation_space, action_space, model, queue, config, logger):
self.config = config
self.queue = queue
self.logger = logger
self.save_path = save_path
self.batch_size = config.batch_size
with tf.device('/cpu:0'):
self.global_step = tf.train.create_global_step()
self.update_global_step = tf.assign_add(self.global_step, 1)
with tf.device('/gpu:0'):
queue_size_op = self.queue.size()
self.queue_size = U.function([], queue_size_op)
dequeue_op = self.queue.get()
self.act, self.train, self.update_target, self.debug = qdqn.build_train(
make_obs_ph=lambda name: U.Uint8Input(observation_space.shape, name=name),
q_func=model,
num_actions=action_space.n,
gamma=config.gamma,
optimizer=tf.train.AdamOptimizer(learning_rate=config.learning_rate, epsilon=1e-4),
train_inputs=dequeue_op,
scope="learner",
grad_norm_clipping=10,
reuse=False)
self.num_iters = 0
self.max_iteration_count = self.config.num_iterations
self.checkpoint_frequency = max(self.max_iteration_count / 1000, 10000)
self.log_frequency = 300
def main():
state = bk.create_default_state()
game = state.game_num
frame = 3
#TODO Change for your own file structure
directory = "gameImages/" + str(game).zfill(4)
filepaths = [
directory + "/" + str(i).zfill(5) + ".png"
for i in range(int(frame - 3), int(frame + 1))
]
observations = []
for i in filepaths:
observations.append(process_observation(cv2.imread(i)))
observations = np.reshape(np.asarray(observations), (84, 84, 4))
print(observations, observations.shape)
#TODO Load act function. See XAI_IS_test.py for guidance.
with U.make_session(4) as sess:
n_actions = 6
act = build_act(make_obs_ph=lambda name: U.Uint8Input(
observations.shape, name=name),
q_func=dueling_model,
num_actions=n_actions)
saver = tf.train.Saver()
saver.restore(sess, "model-atari-prior-duel-breakout-1/saved")
play(observations, act)
def __init__(self, is_chief, env, model, config, should_render=True):
self.config = config
self.is_chief = is_chief
self.env = env
self.should_render = should_render
self.act, self.train, self.update_target, self.debug = multi_deepq.build_train(
make_obs_ph=lambda name: U.Uint8Input(env.observation_space.shape, name=name),
q_func=model,
num_actions=env.action_space.n,
gamma=config.gamma,
optimizer=tf.train.AdamOptimizer(learning_rate=config.learning_rate),
reuse=(not is_chief),
)
self.max_iteraction_count = int(self.config.num_iterations)
# Create the replay buffer
self.replay_buffer = ReplayBuffer(config.replay_size)
if self.config.exploration_schedule == "constant":
self.exploration = ConstantSchedule(0.1)
elif self.config.exploration_schedule == "linear":
# Create the schedule for exploration starting from 1 (every action is random) down to
# 0.02 (98% of actions are selected according to values predicted by the model).
self.exploration = LinearSchedule(
schedule_timesteps=self.config.num_iterations / 4, initial_p=1.0, final_p=0.02)
elif self.config.exploration_schedule == "piecewise":
approximate_num_iters = self.config.num_iterations
self.exploration = PiecewiseSchedule([
(0, 1.0),
(approximate_num_iters / 50, 0.1),
(approximate_num_iters / 5, 0.01)
], outside_value=0.01)
else:
raise ValueError("Bad exploration schedule")
def main():
set_global_seeds(1)
args = parse_args()
with U.make_session(4): # noqa
_, env = make_env(args.env)
act = deepq.build_act(make_obs_ph=lambda name: U.Uint8Input(
env.observation_space.shape, name=name),
q_func=dueling_model if args.dueling else model,
num_actions=env.action_space.n)
U.load_state(os.path.join(args.model_dir, "saved"))
wang2015_eval(args.env, act, stochastic=args.stochastic)
def __init__(self, index, is_chief, env, model, queue, config, logger, episode_logger, should_render=False):
self.config = config
self.is_chief = is_chief
self.env = env
self.global_step = tf.train.get_global_step()
self.should_render = should_render
self.logger = logger
self.episode_logger = episode_logger
self.log_frequency = 10
with tf.device('/cpu:0'):
self.act, self.update_params, self.debug = qdqn.build_act(
make_obs_ph=lambda name: U.Uint8Input(self.env.observation_space.shape, name=name),
q_func=model,
num_actions=self.env.action_space.n,
scope="actor_{}".format(index),
learner_scope="learner",
reuse=False)
with tf.device('/cpu:0'):
obs_t_input = tf.placeholder(tf.uint8, self.env.observation_space.shape, name="obs_t")
act_t_ph = tf.placeholder(tf.int32, self.env.action_space.shape, name="action")
rew_t_ph = tf.placeholder(tf.float32, [], name="reward")
obs_tp1_input = tf.placeholder(tf.uint8, self.env.observation_space.shape, name="obs_tp1")
done_mask_ph = tf.placeholder(tf.float32, [], name="done")
global_step_ph = tf.placeholder(tf.int32, [], name="sample_global_step")
enqueue_op = queue.enqueue(
[obs_t_input, act_t_ph, rew_t_ph, obs_tp1_input, done_mask_ph, global_step_ph])
self.enqueue = U.function(
[obs_t_input, act_t_ph, rew_t_ph, obs_tp1_input, done_mask_ph, global_step_ph], enqueue_op)
self.max_iteration_count = self.config.num_iterations
if self.config.exploration_schedule == "constant":
self.exploration = ConstantSchedule(0.1)
elif self.config.exploration_schedule == "linear":
# Create the schedule for exploration starting from 1 (every action is random) down to
# 0.02 (98% of actions are selected according to values predicted by the model).
self.exploration = LinearSchedule(
schedule_timesteps=self.config.num_iterations / 4, initial_p=1.0, final_p=0.02)
elif self.config.exploration_schedule == "piecewise":
approximate_num_iters = self.config.num_iterations
self.exploration = PiecewiseSchedule([
(0, 1.0),
(approximate_num_iters / 50, 0.1),
(approximate_num_iters / 5, 0.01)
], outside_value=0.01)
else:
raise ValueError("Bad exploration schedule")
def main():
with U.make_session(4) as sess:
env = make_env(args.env_name)
n_actions = env.action_space.n
print(env.observation_space.shape)
if args.env_name == "Breakout":
n_actions = 6
act = build_act(make_obs_ph=lambda name: U.Uint8Input(
env.observation_space.shape, name=name),
q_func=dueling_model,
num_actions=n_actions)
saver = tf.train.Saver()
saver.restore(sess, args.model_path)
play(args.env_name, env, act)
def main():
set_global_seeds(1)
args = parse_args()
with U.make_session(4) as sess: # noqa
_, env = make_env(args.env)
model_parent_path = distdeepq.parent_path(args.model_dir)
old_args = json.load(open(model_parent_path + '/args.json'))
act = distdeepq.build_act(make_obs_ph=lambda name: U.Uint8Input(
env.observation_space.shape, name=name),
p_dist_func=distdeepq.models.atari_model(),
num_actions=env.action_space.n,
dist_params={
'Vmin': old_args['vmin'],
'Vmax': old_args['vmax'],
'nb_atoms': old_args['nb_atoms']
})
U.load_state(os.path.join(args.model_dir, "saved"))
wang2015_eval(args.env, act, stochastic=args.stochastic)
inRtd = inr + args.gamma * inRtd
z = np.array(z).reshape((args.latent_dim))
# q, _ = ec_buffer[a].peek(z, exRtd, inRtd, True)
qs, inrs = zip(*[
ec_buffer[i].peek(z, exRtd, inRtd, False)
for i in range(env.action_space.n)
])
q = qs[a]
inr = np.max(inrs)
if q is None: # new action
ec_buffer[a].add(z, exRtd, inRtd)
inr = ec_buffer[a].rmax
# Create training graph and replay buffer
act, train, update_target = deepq.build_train_dueling_true(
make_obs_ph=lambda name: U.Uint8Input(env.observation_space.shape,
name=name),
model_func=rp_model if args.rp else contrastive_model,
q_func=model,
imitate=args.imitate,
num_actions=env.action_space.n,
optimizer=tf.train.AdamOptimizer(learning_rate=args.lr,
epsilon=1e-4),
gamma=args.gamma,
grad_norm_clipping=10,
)
approximate_num_iters = args.num_steps
exploration = PiecewiseSchedule(
[
(0, 1),
video_path,
enabled=video_path is not None)
obs = env.reset()
while True:
env.unwrapped.render()
video_recorder.capture_frame()
action = act(np.array(obs)[None], stochastic=stochastic)[0]
obs, rew, done, info = env.step(action)
if done:
obs = env.reset()
if len(info["rewards"]) > num_episodes:
if len(info["rewards"]) == 1 and video_recorder.enabled:
# save video of first episode
print("Saved video.")
video_recorder.close()
video_recorder.enabled = False
print(info["rewards"][-1])
num_episodes = len(info["rewards"])
if __name__ == '__main__':
with U.make_session(4) as sess:
args = parse_args()
env = make_env(args.env)
act = deepq.build_act(make_obs_ph=lambda name: U.Uint8Input(
env.observation_space.shape, name=name),
q_func=dueling_model if args.dueling else model,
num_actions=env.action_space.n)
U.load_state(os.path.join(args.model_dir, "saved"))
play(env, act, args.stochastic, args.video)
def learn(env,
q_func,
lr=5e-4,
max_timesteps=100000,
buffer_size=50000,
exploration_fraction=0.1,
exploration_final_eps=0.02,
train_freq=1,
batch_size=32,
print_freq=100,
checkpoint_freq=10000,
checkpoint_path=None,
learning_starts=1000,
gamma=1.0,
target_network_update_freq=500,
prioritized_replay=False,
prioritized_replay_alpha=0.6,
prioritized_replay_beta0=0.4,
prioritized_replay_beta_iters=None,
prioritized_replay_eps=1e-6,
param_noise=False,
callback=None,
bootstrap=False,
noisy=False,
greedy=False):
"""Train a deepq model.
Parameters
-------
env: gym.Env
environment to train on
q_func: (tf.Variable, int, str, bool) -> tf.Variable
the model that takes the following inputs:
observation_in: object
the output of observation placeholder
num_actions: int
number of actions
scope: str
reuse: bool
should be passed to outer variable scope
and returns a tensor of shape (batch_size, num_actions) with values of every action.
lr: float
learning rate for adam optimizer
max_timesteps: int
number of env steps to optimizer for
buffer_size: int
size of the replay buffer
exploration_fraction: float
fraction of entire training period over which the exploration rate is annealed
exploration_final_eps: float
final value of random action probability
train_freq: int
update the model every `train_freq` steps.
set to None to disable printing
batch_size: int
size of a batched sampled from replay buffer for training
print_freq: int
how often to print out training progress
set to None to disable printing
checkpoint_freq: int
how often to save the model. This is so that the best version is restored
at the end of the training. If you do not wish to restore the best version at
the end of the training set this variable to None.
learning_starts: int
how many steps of the model to collect transitions for before learning starts
gamma: float
discount factor
target_network_update_freq: int
update the target network every `target_network_update_freq` steps.
prioritized_replay: True
if True prioritized replay buffer will be used.
prioritized_replay_alpha: float
alpha parameter for prioritized replay buffer
prioritized_replay_beta0: float
initial value of beta for prioritized replay buffer
prioritized_replay_beta_iters: int
number of iterations over which beta will be annealed from initial value
to 1.0. If set to None equals to max_timesteps.
prioritized_replay_eps: float
epsilon to add to the TD errors when updating priorities.
callback: (locals, globals) -> None
function called at every steps with state of the algorithm.
If callback returns true training stops.
Returns
-------
act: ActWrapper
Wrapper over act function. Adds ability to save it and load it.
See header of baselines/deepq/categorical.py for details on the act function.
"""
# Create all the functions necessary to train the model
sess = tf.Session()
sess.__enter__()
# capture the shape outside the closure so that the env object is not serialized
# by cloudpickle when serializing make_obs_ph
act, train, update_target, debug = deepq.build_train(
make_obs_ph=lambda name: U.Uint8Input(env.observation_space.shape,
name=name),
q_func=q_func,
num_actions=env.action_space.n,
optimizer=tf.train.AdamOptimizer(learning_rate=lr),
gamma=gamma,
grad_norm_clipping=10,
noisy=noisy,
bootstrap=bootstrap)
logger.configure('models', ['json', 'stdout'])
# Create the replay buffer
if prioritized_replay:
replay_buffer = PrioritizedReplayBuffer(buffer_size,
alpha=prioritized_replay_alpha)
if prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = max_timesteps
beta_schedule = LinearSchedule(prioritized_replay_beta_iters,
initial_p=prioritized_replay_beta0,
final_p=1.0)
else:
replay_buffer = ReplayBuffer(buffer_size)
beta_schedule = None
# Create the schedule for exploration starting from 1.
exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction *
max_timesteps),
initial_p=1.0,
final_p=exploration_final_eps)
# Initialize the parameters and copy them to the target network.
U.initialize()
update_target()
episode_rewards = [0.0]
saved_mean_reward = None
obs = env.reset()
reset = True
head = np.random.randint(10) #Initial head initialisation
with tempfile.TemporaryDirectory() as td:
td = checkpoint_path or td
model_file = os.path.join(td, "model")
model_saved = False
if tf.train.latest_checkpoint(td) is not None:
load_state(model_file)
logger.log('Loaded model from {}'.format(model_file))
model_saved = True
for t in range(max_timesteps):
if callback is not None:
if callback(locals(), globals()):
break
# Take action and update exploration to the newest value
kwargs = {}
if not param_noise:
update_eps = exploration.value(t)
update_param_noise_threshold = 0.
else:
update_eps = 0.
# Compute the threshold such that the KL divergence between perturbed and non-perturbed
# policy is comparable to eps-greedy exploration with eps = exploration.value(t).
# See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017
# for detailed explanation.
update_param_noise_threshold = -np.log(1. - exploration.value(
t) + exploration.value(t) / float(env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
if bootstrap:
action = act(np.array(obs)[None],
head=head,
update_eps=update_eps)[0]
elif noisy:
action = act(np.array(obs)[None], stochastic=False)[0]
elif greedy:
action = act(np.array(obs)[None], stochastic=False)[0]
else:
action = act(np.array(obs)[None], update_eps=update_eps)[0]
env_action = action
reset = False
new_obs, rew, done, _ = env.step(env_action)
# Store transition in the replay buffer.
replay_buffer.add(obs, action, rew, new_obs, float(done))
obs = new_obs
episode_rewards[-1] += rew
if done:
ep_rew = episode_rewards[-1]
obs = env.reset()
episode_rewards.append(0.0)
reset = True
if t > learning_starts and t % train_freq == 0:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
if prioritized_replay:
experience = replay_buffer.sample(
batch_size, beta=beta_schedule.value(t))
(obses_t, actions, rewards, obses_tp1, dones, weights,
batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(
batch_size)
weights, batch_idxes = np.ones_like(rewards), None
if bootstrap:
td_errors = train(obses_t, actions, rewards, obses_tp1,
dones, weights, lr)
else:
td_errors = train(obses_t, actions, rewards, obses_tp1,
dones, weights)
if prioritized_replay:
new_priorities = np.abs(td_errors) + prioritized_replay_eps
replay_buffer.update_priorities(batch_idxes,
new_priorities)
if t > learning_starts and t % target_network_update_freq == 0:
# Update target network periodically.
update_target()
mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1)
num_episodes = len(episode_rewards)
if done and print_freq is not None and len(
episode_rewards) % print_freq == 0:
logger.record_tabular("steps", t)
logger.record_tabular("episodes", num_episodes)
logger.record_tabular("reward", ep_rew)
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.record_tabular("% time spent exploring",
int(100 * exploration.value(t)))
if bootstrap:
logger.record_tabular("head for episode", (head + 1))
logger.dump_tabular()
head = np.random.randint(10)
if (checkpoint_freq is not None and t > learning_starts
and num_episodes > 100 and t % checkpoint_freq == 0):
if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward:
if print_freq is not None:
logger.log(
"Saving model due to mean reward increase: {} -> {}"
.format(saved_mean_reward, mean_100ep_reward))
save_state(model_file)
model_saved = True
saved_mean_reward = mean_100ep_reward
if model_saved:
if print_freq is not None:
logger.log("Restored model with mean reward: {}".format(
saved_mean_reward))
load_state(model_file)
return act
account_key=account_key,
container_name=container_name,
maybe_create=True)
if savedir is None:
# Careful! This will not get cleaned up. Docker spoils the developers.
savedir = tempfile.TemporaryDirectory().name
else:
container = None
env = multidim_mdp(args.mdp_arity, args.mdp_dimension, args.mdp_state_size)
with U.make_session(120) as sess:
# Create training graph and replay buffer
if args.bootstrap:
act, train, update_target, debug = deepq.build_train(
make_obs_ph=lambda name: U.Uint8Input(
(args.mdp_arity**args.mdp_dimension, ), name=name),
q_func=simple_bootstrap_model,
bootstrap=args.bootstrap,
num_actions=2 * args.mdp_dimension,
optimizer=tf.train.AdamOptimizer(learning_rate=args.lr,
epsilon=1e-4),
gamma=0.99,
grad_norm_clipping=10,
double_q=args.double_q,
heads=args.heads,
swarm=args.swarm,
device=args.device,
voting=args.voting)
else:
act, train, update_target, debug = deepq.build_train(
make_obs_ph=lambda name: U.Uint8Input(
