def doTestDDPG(self):
np.random.seed(0)
env = gym.make("Pendulum-v0")
env.seed(0)
ddpg_g = tf.Graph()
with ddpg_g.as_default():
tf.set_random_seed(123)
agent = agents[DDPG_AGENT_CONFIG["type"]](
env.observation_space,
env.action_space,
DDPG_AGENT_CONFIG,
DDPG_MODEL_CONFIG,
distributed_spec={})
reward_window = WindowStat("reward", 25)
obs, actions, rewards, next_obs, dones = list(), list(), list(), list(
), list()
act_count = 0
for i in range(200):
ob = env.reset()
done = False
episode_reward = .0
while not done:
action, results = agent.act(
[ob], False, use_perturbed_action=False)
act_count += 1
next_ob, reward, done, info = env.step(action[0])
obs.append(ob)
actions.append(action[0])
rewards.append(0.1 * reward)
next_obs.append(next_ob)
dones.append(done)
if agent.ready_to_send:
agent.send_experience(
obs=obs,
actions=actions,
rewards=rewards,
dones=dones,
next_obs=next_obs)
if agent.ready_to_receive:
batch_data = agent.receive_experience()
res = agent.learn(batch_data)
if DDPG_AGENT_CONFIG.get("prioritized_replay", False):
agent.update_priorities(
indexes=batch_data["indexes"],
td_error=res["td_error"])
ob = next_ob
episode_reward += reward
if act_count % 1024 == 0:
print("timestep:", act_count, reward_window)
agent.add_episode(1)
reward_window.push(episode_reward)
return reward_window.stats()["reward_mean"]
def doTestPPO(self):
env = gym.make("CartPole-v0")
env.seed(0)
ppo_g = tf.Graph()
with ppo_g.as_default():
tf.set_random_seed(123)
agent = agents[PPO_AGENT_CONFIG["type"]](
env.observation_space,
env.action_space,
PPO_AGENT_CONFIG,
PPO_MODEL_CONFIG,
distributed_spec={})
reward_window = WindowStat("reward", 25)
obs, actions, rewards, next_obs, dones, value_preds, logits = list(
), list(), list(), list(), list(), list(), list()
act_count = 0
for i in range(300):
ob = env.reset()
done = False
episode_reward = .0
while not done:
action, results = agent.act([ob], False)
next_ob, reward, done, info = env.step(action[0])
act_count += 1
obs.append(ob)
actions.append(action[0])
rewards.append(0.1 * reward)
next_obs.append(next_ob)
dones.append(done)
logits.append(results["logits"][0])
value_preds.append(results["value_preds"][0])
if agent.ready_to_send:
agent.send_experience(
obs=obs,
actions=actions,
rewards=rewards,
dones=dones,
next_obs=next_obs,
value_preds=value_preds,
logits=logits)
if agent.ready_to_receive:
batch_data = agent.receive_experience()
res = agent.learn(batch_data)
ob = next_ob
episode_reward += reward
if act_count % 1024 == 0:
print("timestep:", act_count, reward_window)
reward_window.push(episode_reward)
return reward_window.stats()["reward_mean"]
def main():
env = gym.make("CartPole-v0")
agent_class = agents[AGENT_CONFIG["type"]]
agent = agent_class(env.observation_space,
env.action_space,
AGENT_CONFIG,
MODEL_CONFIG,
distributed_spec={},
export_dir="dump_dir",
checkpoint_dir="ckpt_dir")
reward_window = WindowStat("reward", 50)
length_window = WindowStat("length", 50)
loss_window = WindowStat("loss", 50)
obs, actions, rewards, next_obs, dones, value_preds = list(), list(), list(
), list(), list(), list()
act_count = 0
for i in range(600):
ob = env.reset()
done = False
episode_reward = .0
episode_len = 0
while not done:
action, results = agent.act([ob], False)
next_ob, reward, done, info = env.step(action[0])
act_count += 1
obs.append(ob)
actions.append(action[0])
rewards.append(0.1 * reward)
next_obs.append(next_ob)
dones.append(done)
value_preds.append(results["value_preds"][0])
if agent.ready_to_send:
agent.send_experience(obs=obs,
actions=actions,
rewards=rewards,
next_obs=next_obs,
dones=dones,
value_preds=value_preds)
if agent.ready_to_receive:
batch_data = agent.receive_experience()
res = agent.learn(batch_data)
loss_window.push(res['loss'])
ob = next_ob
episode_reward += reward
episode_len += 1
if act_count % 1000 == 0:
print("timestep:", act_count, reward_window, length_window)
reward_window.push(episode_reward)
length_window.push(episode_len)
agent.export_saved_model()
print("Done.")
def main():
with open(FLAGS.config, 'r') as ips:
config = json.load(ips)
print(config)
job_name = FLAGS.job_name
env = make_atari("PongNoFrameskip-v4")
env = wrap_deepmind(
env=env,
frame_stack=True,
clip_rewards=False,
episode_life=True,
wrap_frame=True,
frame_resize=42)
agent_class = agents[config["agent"]["type"]]
agent = agent_class(
env.observation_space,
env.action_space,
config["agent"],
config["model"],
distributed_spec={
"ps_hosts": FLAGS.ps_hosts,
"memory_hosts": FLAGS.memory_hosts,
"actor_hosts": FLAGS.actor_hosts,
"learner_hosts": FLAGS.learner_hosts,
"job_name": FLAGS.job_name,
"task_index": FLAGS.task_index
},
custom_model=MyVTmodel,
checkpoint_dir=None)
all_cost = time.time()
if job_name == "ps":
print("ps starts===>")
agent.join()
elif job_name == "memory":
start_tt = time.time()
log_count = 0
print("memory starts===>")
while not agent.should_stop():
agent.communicate()
if time.time() - start_tt > log_count:
log_count += 1
print(agent._receive_count, "actor2mem_q:",
agent._actor2mem_q.qsize(), "mem2learner_2:",
agent._mem2learner_q.qsize())
sys.stdout.flush()
elif job_name == "actor":
print("actor starts===>")
start_tt = time.time()
log_count = 0
act_log_count = 0
# create vectorized env
def make_env(rank):
def make_atari_env():
env = make_atari("PongNoFrameskip-v4")
env = wrap_deepmind(
env=env,
frame_stack=True,
clip_rewards=False,
episode_life=True,
wrap_frame=True,
frame_resize=42)
env.seed(rank)
return env
return make_atari_env
num_env = config["agent"].get("num_env", 1)
vec_env = VectorizedEnvironment(
make_env=make_env, num_env=num_env, seed=100 * FLAGS.task_index)
act_count = 0
reward_window = WindowStat("reward", 10)
length_window = WindowStat("length", 10)
obs, actions, rewards, dones, logits = list(), list(), list(), list(
), list()
agent.sync_vars()
while not agent.should_stop():
ob = vec_env.reset()
episode_reward = np.zeros(num_env, )
episode_len = np.zeros(num_env, )
while not agent.should_stop():
action, results = agent.act(ob, False)
act_count += 1
new_ob, reward, done, info = vec_env.step(action)
obs.append(ob)
actions.append(action)
rewards.append(reward)
dones.append(done)
logits.append(results["logits"])
if agent.ready_to_send:
agent.send_experience(
obs=obs,
actions=actions,
rewards=rewards,
dones=dones,
logits=logits,
vec_env=True,
num_env=num_env)
agent.sync_vars()
ob = new_ob
episode_reward += np.asarray(reward)
episode_len += 1
for i in range(num_env):
if done[i]:
reward_window.push(episode_reward[i])
length_window.push(episode_len[i])
episode_reward[i] = .0
episode_len[i] = 0
total_cost = time.time() - start_tt
if int(total_cost / 5) > log_count:
log_count += 1
print("act_count:", act_count, "actor2mem_q:",
agent._actor2mem_q.qsize(), "total:", total_cost)
print('total_cost:', total_cost, reward_window)
print(length_window)
sys.stdout.flush()
if int((act_count * num_env) / 10000) > act_log_count:
act_log_count += 1
print('timestep:', act_log_count * 10000, reward_window)
elif job_name == "learner":
print("learner starts===>")
start_tt = time.time()
train_count = 0
try:
while not agent.should_stop():
batch_data = agent.receive_experience()
if batch_data:
extra_data = agent.learn(batch_data)
train_count += 1
print("learning {}".format(extra_data), "receive_q:",
agent._receive_q.qsize())
print("train_count:", train_count, "total:",
time.time() - start_tt)
sys.stdout.flush()
except tf.errors.OutOfRangeError as e:
print("memory has stopped.")
else:
raise ValueError("Invalid job_name.")
all_cost = time.time() - all_cost
print("done. all_cost:", all_cost)
def main():
env = gym.make("CartPole-v0")
env.seed(0)
agent_class = agents[AGENT_CONFIG["type"]]
agent = agent_class(env.observation_space,
env.action_space,
AGENT_CONFIG,
MODEL_CONFIG,
distributed_spec={},
checkpoint_dir="ckpt_dir",
custom_model=MyESmodel)
reward_window = WindowStat("reward", 50)
length_window = WindowStat("length", 50)
init_perturbation_scale = 0.1
seeds, rewards, perturbation_scales = list(), list(), list()
is_positive_direction = list()
# how many episodes needed for one trial
episode_per_perturbation = 1
returns = list()
for i in range(4000):
ob = env.reset()
done = False
episode_reward = .0
episode_len = 0
if i % episode_per_perturbation == 0:
# perturb parameters every `episode_per_seed` episodes
is_positive = True if len(
is_positive_direction
) == 0 else is_positive_direction[-1] != True
# each seed twice
seed = np.random.randint(1000000) if is_positive else seeds[-1]
perturbation_scale = max(
init_perturbation_scale * (1 - i / 2000.0), 0.02)
feed = agent.model.perturbation_feed
fetch = [agent.model.reset_perturbation_op]
agent.executor.run(fetches=fetch,
feed_dict={
feed['perturbation_seeds']: [seed],
feed['perturbation_scales']:
[perturbation_scale],
feed['positive_perturbation']:
is_positive
})
if is_positive:
seeds.append(seed)
perturbation_scales.append(perturbation_scale)
is_positive_direction.append(is_positive)
while not done:
action, result = agent.act([ob], True, use_perturbed_action=True)
next_ob, reward, done, info = env.step(action[0])
ob = next_ob
episode_reward += reward
episode_len += 1
rewards.append(episode_reward)
reward_window.push(episode_reward)
length_window.push(episode_len)
if len(rewards) == episode_per_perturbation:
returns.append(np.mean(rewards))
rewards = []
if len(returns) == 2 * agent.config.get('sample_batch_size', 100):
print(reward_window)
assert len(seeds) == (len(returns) / 2)
assert len(perturbation_scales) == (len(returns) / 2)
agent.learn(
batch_data=dict(perturbation_seeds=seeds,
perturbation_scales=perturbation_scales,
returns=np.reshape(returns, [-1, 2])))
seeds = []
perturbation_scales = []
returns = []
is_positive_direction = []
# evaluation 20 episodes
test_rewards = list()
for j in range(20):
done = False
ob = env.reset()
episode_reward = 0
episode_len = 0
while not done:
action, result = agent.act([ob],
True,
use_perturbed_action=False)
next_ob, reward, done, info = env.step(action[0])
ob = next_ob
episode_reward += reward
episode_len += 1
test_rewards.append(episode_reward)
print("[evaluation] average reward of 20 episodes:",
np.mean(test_rewards))
print('train at ', i)
agent.export_saved_model(export_dir="dump_dir")
print("Done.")
def main(_):
with open(FLAGS.config, 'r') as ips:
config = json.load(ips)
print(config)
env = gym.make("CartPole-v0")
env.seed(0)
agent_class = agents[config["agent"]["type"]]
agent = agent_class(env.observation_space,
env.action_space,
config["agent"],
config["model"],
distributed_spec={
"ps_hosts": FLAGS.ps_hosts,
"memory_hosts": FLAGS.memory_hosts,
"actor_hosts": FLAGS.actor_hosts,
"learner_hosts": FLAGS.learner_hosts,
"job_name": FLAGS.job_name,
"task_index": FLAGS.task_index
},
custom_model=MyDQN)
if FLAGS.job_name == "ps":
print("ps starts===>")
agent.join()
elif FLAGS.job_name == "memory":
print("memory starts===>")
while not agent.should_stop():
agent.communicate()
sys.stdout.flush()
elif FLAGS.job_name == "actor":
print("actor starts===>")
reward_window = WindowStat("reward", 50)
length_window = WindowStat("length", 50)
obs, actions, rewards, new_obs, dones = list(), list(), list(), list(
), list()
agent.sync_vars()
while not agent.should_stop():
ob = env.reset()
done = False
episode_reward = .0
episode_len = 0
while not done and not agent.should_stop():
action, results = agent.act([ob], False)
new_ob, reward, done, info = env.step(action[0])
obs.append(ob)
actions.append(action[0])
rewards.append(reward)
new_obs.append(new_ob)
dones.append(done)
if agent.ready_to_send:
agent.send_experience(obs=obs,
actions=actions,
rewards=rewards,
next_obs=new_obs,
dones=dones)
agent.sync_vars()
ob = new_ob
episode_reward += reward
episode_len += 1
reward_window.push(episode_reward)
length_window.push(episode_len)
print(reward_window)
print(length_window)
sys.stdout.flush()
elif FLAGS.job_name == "learner":
print("learner starts===>")
while not agent.should_stop():
batch_data = agent.receive_experience()
if batch_data:
extra_data = agent.learn(batch_data)
print("learning {}".format(extra_data))
sys.stdout.flush()
else:
raise ValueError("Invalid job_name.")
print("done.")
def main():
env = make_atari("PongNoFrameskip-v4")
env = wrap_deepmind(env=env,
frame_stack=True,
clip_rewards=False,
episode_life=True,
wrap_frame=True,
frame_resize=42)
agent_class = agents[AGENT_CONFIG["type"]]
agent = agent_class(env.observation_space,
env.action_space,
AGENT_CONFIG,
MODEL_CONFIG,
distributed_spec={},
custom_model=MyDQNModel)
reward_window = WindowStat("reward", 10)
length_window = WindowStat("length", 10)
loss_window = WindowStat("loss", 10)
obs, actions, rewards, next_obs, dones = list(), list(), list(), list(
), list()
for i in range(2000):
ob = env.reset()
ob = np.asarray(ob)
done = False
episode_reward = .0
episode_len = 0
while not done:
action, results = agent.act([ob],
deterministic=False,
use_perturbed_action=False)
next_ob, reward, done, info = env.step(action[0])
next_ob = np.asarray(next_ob)
obs.append(ob)
actions.append(action[0])
rewards.append(reward)
next_obs.append(next_ob)
dones.append(done)
if agent.ready_to_send:
agent.send_experience(obs=obs,
actions=actions,
rewards=rewards,
next_obs=next_obs,
dones=dones)
if agent.ready_to_receive:
batch_data = agent.receive_experience()
res = agent.learn(batch_data)
loss_window.push(res['loss'])
if AGENT_CONFIG.get("prioritized_replay", False):
agent.update_priorities(indexes=batch_data["indexes"],
td_error=res["td_error"])
ob = next_ob
episode_reward += reward
episode_len += 1
agent.add_episode(1)
reward_window.push(episode_reward)
length_window.push(episode_len)
if i % 10 == 0:
print('episode at', i)
print(reward_window)
print(length_window)
print(loss_window)
print("Done.")
def doTestCkpt(self):
trial_timestamp = time.strftime("%Y%m%d-%H%M%S")
np.random.seed(0)
env = gym.make("CartPole-v0")
env.seed(0)
dqn_g = tf.Graph()
with dqn_g.as_default():
tf.set_random_seed(123)
agent = agents[DQN_AGENT_CONFIG["type"]](
env.observation_space,
env.action_space,
DQN_AGENT_CONFIG,
DQN_MODEL_CONFIG,
checkpoint_dir="ckpt_dir_{}".format(trial_timestamp),
distributed_spec={})
reward_window = WindowStat("reward", 50)
obs, actions, rewards, next_obs, dones = list(), list(), list(), list(
), list()
act_count = 0
for i in range(500):
ob = env.reset()
done = False
episode_reward = .0
while not done:
action, results = agent.act([ob],
deterministic=False,
use_perturbed_action=False)
next_ob, reward, done, info = env.step(action[0])
act_count += 1
obs.append(ob)
actions.append(action[0])
rewards.append(reward)
next_obs.append(next_ob)
dones.append(done)
if agent.ready_to_send:
agent.send_experience(obs=obs,
actions=actions,
rewards=rewards,
next_obs=next_obs,
dones=dones)
if agent.ready_to_receive:
batch_data = agent.receive_experience()
res = agent.learn(batch_data)
if DQN_AGENT_CONFIG.get("prioritized_replay", False):
agent.update_priorities(indexes=batch_data["indexes"],
td_error=res["td_error"])
ob = next_ob
episode_reward += reward
if act_count % 1024 == 0:
print("timestep:", act_count, reward_window)
agent.add_episode(1)
reward_window.push(episode_reward)
prev_perf = reward_window.stats()["reward_mean"]
print("Performance before saving is {}".format(prev_perf))
new_dqn_g = tf.Graph()
with new_dqn_g.as_default():
agent = agents[DQN_AGENT_CONFIG["type"]](
env.observation_space,
env.action_space,
DQN_AGENT_CONFIG,
DQN_MODEL_CONFIG,
checkpoint_dir="ckpt_dir_{}".format(trial_timestamp),
distributed_spec={})
reward_window = WindowStat("reward", 10)
ob = env.reset()
for i in range(10):
ob = env.reset()
done = False
episode_reward = .0
while not done:
action, results = agent.act([ob],
deterministic=True,
use_perturbed_action=False)
next_ob, reward, done, info = env.step(action[0])
act_count += 1
ob = next_ob
episode_reward += reward
agent.add_episode(1)
reward_window.push(episode_reward)
cur_perf = reward_window.stats()["reward_mean"]
print("Performance after restore is {}".format(cur_perf))
return prev_perf - cur_perf
def main():
# create offline_env
env = offline_env(FLAGS.train_data_file, batch_size=128, n_step=MODEL_CONFIG.get("n_step", 1))
eval_env = offline_env(FLAGS.eval_data_file, batch_size=128, n_step=MODEL_CONFIG.get("n_step", 1))
agent_class = agents[AGENT_CONFIG["type"]]
agent = agent_class(
env.observation_space,
env.action_space,
AGENT_CONFIG,
MODEL_CONFIG,
distributed_spec={},
export_dir="bcq_tmp",
checkpoint_dir="bcq_tmp",
custom_model=MyBCQModel)
clone_loss_window = WindowStat("clone_loss", 50)
clone_reg_loss_window = WindowStat("clone_reg_loss", 50)
loss_window = WindowStat("loss", 50)
total_cost = time.time()
clone_learn_count = 0
# first, train a generative model by behavior clone
for i in range(1000):
table_data = env.reset()
# store raw data in replay buffer
agent.send_experience(
obs=table_data["obs"],
actions=table_data["actions"],
rewards=table_data["rewards"],
dones=table_data["dones"],
next_obs=table_data["next_obs"])
# sample from replay buffer
# the size of sampled data is equal to `AGENT_CONFIG["batch_size"]`
batch_data = agent.receive_experience()
clone_loss, clone_reg_loss = agent.behavior_learn(batch_data=batch_data)
clone_learn_count += 1
clone_loss_window.push(clone_loss)
clone_reg_loss_window.push(clone_reg_loss)
if i % 50 == 0:
print(clone_loss_window)
print(clone_reg_loss_window)
# eval generative model
all_clone_act, gd_act = [], []
for i in range(100):
table_data = env.reset()
clone_act = agent.behavior_act(table_data["obs"])
all_clone_act.extend(np.argsort(-1.0 * clone_act, axis=1).tolist())
gd_act.extend(table_data["actions"])
acc1 = np.sum(np.array(all_clone_act)[:, 0] == np.array(gd_act))*1.0/len(gd_act)
print("acc @top1:", acc1)
# second, train bcq
agent.reset_global_step()
epochs_to_end = 10
max_globel_steps_to_end = 10000
learn_count = 0
env.num_epoch = 0
while env.num_epoch < epochs_to_end and learn_count < max_globel_steps_to_end:
table_data = env.reset()
# store raw data in replay buffer
agent.send_experience(
obs=table_data["obs"],
actions=table_data["actions"],
rewards=table_data["rewards"],
dones=table_data["dones"],
next_obs=table_data["next_obs"])
# sample from replay buffer
# the size of sampled data is equal to `AGENT_CONFIG["batch_size"]`
batch_data = agent.receive_experience()
# update the model
res = agent.learn(batch_data)
# record the loss
loss_window.push(res["loss"])
learn_count += 1
if AGENT_CONFIG.get("prioritized_replay", False):
# update priorities
agent.update_priorities(
indexes=batch_data["indexes"],
td_error=res["td_error"])
if learn_count % 50 == 0:
print("learn_count:", learn_count)
print(loss_window)
# offline evaluation
batch_weights = []
batch_rewards = []
eval_num = 50
for _ in range(eval_num):
batch_data = eval_env.reset()
importance_ratio = agent.importance_ratio(batch_data)
batch_weights.append(importance_ratio)
batch_rewards.append(batch_data["rewards"])
ips, ips_sw, wips, wips_sw, wips_sw_mean = ips_eval(
batch_weights=batch_weights, batch_rewards=batch_rewards, gamma=MODEL_CONFIG.get("gamma", 0.95))
agent.add_extra_summary({agent.model.ips_score_op:ips,
agent.model.ips_score_stepwise_op:ips_sw,
agent.model.wnorm_ips_score_op:wips,
agent.model.wnorm_ips_score_stepwise_op:wips_sw,
agent.model.wnorm_ips_score_stepwise_mean_op:wips_sw_mean})
print("[IPS Policy Evaluation @learn_count={}] ips={}, ips_stepwise={}, wnorm_ips={}, wnorm_ips_stepwise={}, wnorm_ips_stepwise_mean={}".format(
learn_count, ips, ips_sw, wips, wips_sw, wips_sw_mean))
if learn_count % 2000 == 0:
# export saved model at any time
# AssertionError will occur if the export_dir already exists.
agent.export_saved_model("bcq_export_dir{}".format(learn_count))
if learn_count % 200 == 0:
# test with simulator
gym_env = gym.make("CartPole-v0")
for ix in range(10):
ob = gym_env.reset()
done = False
episode_reward = .0
while not done:
action, results = agent.act(
[ob], deterministic=False, use_perturbed_action=False)
next_ob, reward, done, info = gym_env.step(action[0])
episode_reward += reward
ob = next_ob
print("train@", learn_count, "test@", ix, "reward:", episode_reward)
print("Done.", "num_epoch:", env.num_epoch, "learn_count:", learn_count,
"total_cost:", time.time() - total_cost)
class PrioritizedReplayBuffer(ReplayBuffer):
def __init__(self, size, alpha):
"""Create Prioritized Replay buffer.
Parameters
----------
size: int
Max number of transitions to store in the buffer. When the buffer
overflows the old memories are dropped.
alpha: float
how much prioritization is used
(0 - no prioritization, 1 - full prioritization)
See Also
--------
ReplayBuffer.__init__
"""
super(PrioritizedReplayBuffer, self).__init__(size)
assert alpha > 0
self._alpha = alpha
it_capacity = 1
while it_capacity < size:
it_capacity *= 2
self._it_sum = SumSegmentTree(it_capacity)
self._it_min = MinSegmentTree(it_capacity)
self._max_priority = 1.0
self._prio_change_stats = WindowStat("reprio", 1000)
self._debug_cost = 0
def add(self, obs, actions, rewards, dones, next_obs, weights, **kwargs):
"""See ReplayBuffer.store_effect"""
super(PrioritizedReplayBuffer, self).add(
obs=obs,
actions=actions,
rewards=rewards,
dones=dones,
next_obs=next_obs,
**{})
if weights is None:
weights = self._max_priority
constant_weight = weights**self._alpha
for idx in self._cover_indices:
self._it_sum[idx] = constant_weight
self._it_min[idx] = constant_weight
else:
weights = np.power(weights, self._alpha)
for n, idx in enumerate(self._cover_indices):
self._it_sum[idx] = weights[n]
self._it_min[idx] = weights[n]
def _sample_proportional(self, batch_size):
res = []
sum_value = self._it_sum.sum(0, len(self))
mass = np.random.random(size=batch_size) * sum_value
for i in range(batch_size):
# TODO(szymon): should we ensure no repeats?
idx = self._it_sum.find_prefixsum_idx(mass[i])
res.append(idx)
return res
def sample(self, batch_size, beta):
"""Sample a batch of experiences.
compared to ReplayBuffer.sample
it also returns importance weights and idxes
of sampled experiences.
Parameters
----------
batch_size: int
How many transitions to sample.
beta: float
To what degree to use importance weights
(0 - no corrections, 1 - full correction)
Returns
-------
obs_batch: np.array
batch of observations
act_batch: np.array
batch of actions executed given obs_batch
rew_batch: np.array
rewards received as results of executing act_batch
next_obs_batch: np.array
next set of observations seen after executing act_batch
done_mask: np.array
done_mask[i] = 1 if executing act_batch[i] resulted in
the end of an episode and 0 otherwise.
weights: np.array
Array of shape (batch_size,) and dtype np.float32
denoting importance weight of each sampled transition
idxes: np.array
Array of shape (batch_size,) and dtype np.int32
idexes in buffer of sampled experiences
"""
assert beta > 0
self._num_sampled += batch_size
start = time.time()
idxes = self._sample_proportional(batch_size)
self._debug_cost += time.time() - start
sum_value = self._it_sum.sum()
weights = []
p_min = self._it_min.min() / sum_value
max_weight = (p_min * len(self))**(-beta)
for idx in idxes:
p_sample = self._it_sum[idx] / sum_value
weight = (p_sample * len(self))**(-beta)
weights.append(weight / max_weight)
weights = np.asarray(weights)
encoded_sample = self._encode_sample(idxes)
encoded_sample["weights"] = weights
encoded_sample["indexes"] = idxes
return encoded_sample
def update_priorities(self, indexes, priorities):
"""Update priorities of sampled transitions.
sets priority of transition at index idxes[i] in buffer
to priorities[i].
Parameters
----------
indexes: [int]
List of idxes of sampled transitions
priorities: [float]
List of updated priorities corresponding to
transitions at the sampled idxes denoted by
variable `idxes`.
"""
assert len(indexes) == len(priorities)
pvs = np.power(priorities, self._alpha).astype(np.float64)
for idx, priority, pv in zip(indexes, priorities, pvs):
assert priority > 0
assert 0 <= idx < len(self)
delta = pv - self._it_sum[idx]
self._prio_change_stats.push(delta)
self._it_sum[idx] = pv
self._it_min[idx] = pv
self._max_priority = max(self._max_priority, np.max(priorities))
def stats(self, debug=False):
parent = ReplayBuffer.stats(self, debug)
if debug:
parent.update(self._prio_change_stats.stats())
return parent
class ReplayBuffer(object):
"""Basic replay buffer.
Support O(1) `add` and O(1) `sample` operations (w.r.t. each transition).
The buffer is implemented as a fixed-length list where the index of insertion is reset to zero,
once the list length is reached.
"""
def __init__(self, size):
"""Create the replay buffer.
Parameters
----------
size: int
Max number of transitions to store in the buffer. When the buffer
overflows the old memories are dropped.
"""
self._maxsize = size
self._next_idx = 0
self._hit_count = np.zeros(size)
self._eviction_started = False
self._num_added = 0
self._num_sampled = 0
self._evicted_hit_stats = WindowStat("evicted_hit", 1000)
self._est_size_bytes = 0
self._extra_fields = None
self._first_add = True
def __len__(self):
return min(self._num_added, self._maxsize)
def add(self,
obs,
actions,
rewards,
dones,
next_obs=None,
weights=None,
**kwargs):
batch_size = np.shape(rewards)[0]
assert batch_size < self._maxsize, "size of data added in buffer is too big at once"
truncated_size = min(batch_size, self._maxsize - self._next_idx)
extra_size = max(0, batch_size - (self._maxsize - self._next_idx))
if self._extra_fields is None:
self._extra_fields = list(kwargs.keys())
if self._first_add:
self._obs = np.zeros(
shape=((self._maxsize, ) + np.shape(obs)[1:]), dtype=obs.dtype)
self._actions = np.zeros(
shape=((self._maxsize, ) + np.shape(actions)[1:]),
dtype=actions.dtype)
self._rewards = np.zeros(shape=(self._maxsize, ), dtype=np.float32)
if next_obs is not None:
self._next_obs = np.zeros(
shape=((self._maxsize, ) + np.shape(next_obs)[1:]),
dtype=next_obs.dtype)
if weights is not None:
self._weights = np.zeros(
shape=((self._maxsize, )), dtype=np.float32)
self._dones = np.zeros(shape=(self._maxsize, ), dtype=np.float32)
self._extras = {
name: np.zeros(
shape=((self._maxsize, ) + np.shape(kwargs[name])[1:]),
dtype=kwargs[name].dtype)
for name in self._extra_fields
}
self._first_add = False
self._num_added += batch_size
#if self._num_added <= self._maxsize:
#self._est_size_bytes += sum(sys.getsizeof(d) for d in data)
self._obs[self._next_idx:self._next_idx +
truncated_size] = obs[:truncated_size]
self._actions[self._next_idx:self._next_idx +
truncated_size] = actions[:truncated_size]
self._rewards[self._next_idx:self._next_idx +
truncated_size] = rewards[:truncated_size]
self._dones[self._next_idx:self._next_idx +
truncated_size] = dones[:truncated_size]
if next_obs is not None:
self._next_obs[self._next_idx:self._next_idx +
truncated_size] = next_obs[:truncated_size]
if weights is not None:
self._weights[self._next_idx:self._next_idx +
truncated_size] = weights[:truncated_size]
for name in self._extras.keys():
self._extras[name][self._next_idx:self._next_idx +
truncated_size] = kwargs[name][:truncated_size]
if extra_size > 0:
self._obs[:extra_size] = obs[truncated_size:]
self._actions[:extra_size] = actions[truncated_size:]
self._rewards[:extra_size] = rewards[truncated_size:]
self._dones[:extra_size] = dones[truncated_size:]
if next_obs is not None:
self._next_obs[:extra_size] = next_obs[truncated_size:]
if weights is not None:
self._weights[:extra_size] = weights[truncated_size:]
for name in self._extras.keys():
self._extras[name][:extra_size] = kwargs[name][truncated_size:]
if self._next_idx + batch_size >= self._maxsize:
self._eviction_started = True
self._cover_indices = [
self._next_idx + i for i in range(truncated_size)
]
if extra_size > 0:
self._cover_indices += [i for i in range(extra_size)]
self._next_idx = (self._next_idx + batch_size) % self._maxsize
if self._eviction_started:
for i in self._cover_indices:
self._evicted_hit_stats.push(self._hit_count[i])
self._hit_count[i] = 0
def _encode_sample(self, idxes):
idxes = np.asarray(idxes)
obs = np.take(self._obs, indices=idxes, axis=0)
actions = np.take(self._actions, indices=idxes, axis=0)
rewards = np.take(self._rewards, indices=idxes, axis=0)
next_obs = np.take(self._next_obs, indices=idxes, axis=0)
dones = np.take(self._dones, indices=idxes, axis=0)
batch_data = dict(
obs=obs,
actions=actions,
rewards=rewards,
dones=dones,
next_obs=next_obs)
return batch_data
def sample(self, batch_size):
"""Sample a batch of experiences.
Parameters
----------
batch_size: int
How many transitions to sample.
Returns
-------
obs_batch: np.array
batch of observations
act_batch: np.array
batch of actions executed given obs_batch
rew_batch: np.array
rewards received as results of executing act_batch
next_obs_batch: np.array
next set of observations seen after executing act_batch
done_mask: np.array
done_mask[i] = 1 if executing act_batch[i] resulted in
the end of an episode and 0 otherwise.
"""
idxes = np.random.randint(
0, min(self._num_added, self._maxsize) - 1, size=(batch_size, ))
self._num_sampled += batch_size
return self._encode_sample(idxes)
def stats(self, debug=False):
data = {
"added_count": self._num_added,
"sampled_count": self._num_sampled,
"est_size_bytes": self._est_size_bytes,
"num_entries": len(self),
}
if debug:
data.update(self._evicted_hit_stats.stats())
return data
def main():
gym_env = gym.make("CartPole-v0")
atari_env = make_atari("PongNoFrameskip-v4")
atari_env = wrap_deepmind(env=atari_env,
frame_stack=True,
clip_rewards=False,
episode_life=True,
wrap_frame=True,
frame_resize=42)
# replace the following env according to your saved_model
# env = atari_env
env = gym_env
with tf.Session() as sess:
path = 'dump_dir'
MetaGraphDef = tf.saved_model.loader.load(
sess, tags=[sm.tag_constants.SERVING], export_dir=path)
# get SignatureDef protobuf
SignatureDef_d = MetaGraphDef.signature_def
SignatureDef = SignatureDef_d["predict_results"]
# get inputs/outputs TensorInfo protobuf
ph_inputs = {}
for name, ts_info in SignatureDef.inputs.items():
ph_inputs[name] = sm.utils.get_tensor_from_tensor_info(
ts_info, sess.graph)
outputs = {}
for name, ts_info in SignatureDef.outputs.items():
outputs[name] = sm.utils.get_tensor_from_tensor_info(
ts_info, sess.graph)
for name, ph in ph_inputs.items():
print(name, ph)
for name, ts in outputs.items():
print(name, ts)
len_window = WindowStat("length", 50)
reward_window = WindowStat("reward", 50)
for i in range(100):
ob = env.reset()
env.render()
time.sleep(0.2)
done = False
episode_len = 0
episode_reward = .0
while not done:
action = sess.run(outputs["output_actions"],
feed_dict={
ph_inputs["obs_ph"]: [np.asarray(ob)],
ph_inputs["deterministic_ph"]: True
})
next_ob, reward, done, info = env.step(action[0])
env.render()
time.sleep(0.1)
episode_reward += reward
episode_len += 1
ob = next_ob
len_window.push(episode_len)
reward_window.push(episode_reward)
print(reward_window)
print(len_window)
def main():
with open(FLAGS.config, 'r') as ips:
config = json.load(ips)
print(config)
env = gym.make("CartPole-v0")
env.seed(0)
agent_class = agents[config["agent"]["type"]]
agent = agent_class(env.observation_space,
env.action_space,
config["agent"],
config["model"],
distributed_spec={
"ps_hosts": FLAGS.ps_hosts,
"memory_hosts": FLAGS.memory_hosts,
"actor_hosts": FLAGS.actor_hosts,
"learner_hosts": FLAGS.learner_hosts,
"job_name": FLAGS.job_name,
"task_index": FLAGS.task_index
},
custom_model=MyESmodel,
checkpoint_dir=None)
all_cost = time.time()
if FLAGS.job_name == "ps":
print("ps starts===>")
agent.join()
elif FLAGS.job_name == "memory":
print("memory starts===>")
while not agent.should_stop():
agent.communicate()
print("communicating")
time.sleep(0.1)
elif FLAGS.job_name == "actor":
print("actor starts===>")
reward_window = WindowStat("reward", 50)
length_window = WindowStat("length", 50)
perturbation_scale = 0.1
run_episode_per_perturbation = config["agent"].get(
"run_episode_per_perturbation", 1)
seeds, rewards, perturbation_scales = list(), list(), list()
is_positive = False
returns = list()
agent.sync_vars()
episode_count = 0
try:
while not agent.should_stop():
# do perturbation
is_positive = False if is_positive else True
# each seed will be used twice
seed = np.random.randint(1000000) if is_positive else seeds[-1]
perturbation_scale = max(
perturbation_scale * (1 - episode_count / 2000.0), 0.02)
feed = agent.behavior_model.perturbation_feed
fetch = [agent.behavior_model.reset_perturbation_op]
agent.executor.run(fetches=fetch,
feed_dict={
feed['perturbation_seeds']: [seed],
feed['perturbation_scales']:
[perturbation_scale],
feed['positive_perturbation']:
is_positive
})
if is_positive:
seeds.append(seed)
perturbation_scales.append(perturbation_scale)
rewards, episode_lens = rollout(
agent,
env,
episode_num=run_episode_per_perturbation,
use_perturbed_action=True)
episode_count += run_episode_per_perturbation
# calculate the average reward from a specific perturbation with one direction
if len(returns) == 0:
returns.append([np.mean(rewards)])
elif len(returns[-1]) < 2:
returns[-1].append(np.mean(rewards))
else:
returns.append([np.mean(rewards)])
if len(returns) == agent.config.get(
'sample_batch_size', 100) and len(returns[-1]) == 2:
# send out the results for the latest `sample_batch_size` * 2 trials
print(reward_window)
assert len(seeds) == len(returns)
assert len(perturbation_scales) == len(returns)
agent.send_experience(
**dict(perturbation_seeds=seeds,
perturbation_scales=perturbation_scales,
returns=returns))
# reset the direction
is_positive = False
# synchronize the weights from parameter server to local behavior_model
agent.sync_vars()
# do evaluation for 20 episode
evaluation_num = 20
evl_returns, _ = rollout(agent,
env,
episode_num=evaluation_num,
use_perturbed_action=False)
print(
"evaluation at episode:", episode_count,
",avg episode reward of {} evaluation:".format(
evaluation_num), np.mean(evl_returns))
reward_window.push(rewards)
length_window.push(episode_lens)
if episode_count % 50 == 0:
print(reward_window)
print(length_window)
sys.stdout.flush()
except tf.errors.OutOfRangeError as e:
print("memory has stopped.")
elif FLAGS.job_name == "learner":
print("learner starts===>")
train_count = 0
try:
while not agent.should_stop():
batch_data = agent.receive_experience()
if batch_data:
extra_data = agent.learn(batch_data)
train_count += 1
print("learning {}".format(extra_data))
sys.stdout.flush()
except tf.errors.OutOfRangeError as e:
print("memory has stopped.")
else:
raise ValueError("Invalid job_name.")
all_cost = time.time() - all_cost
print("done. all_cost:", all_cost)
def main():
env = MaxComponentEnv(num_arms=6)
agent_class = agents[AGENT_CONFIG["type"]]
agent = agent_class(env.observation_space,
env.action_space,
AGENT_CONFIG,
MODEL_CONFIG,
distributed_spec={},
export_dir="hook_dump_dir")
reward_window = WindowStat("reward", 50)
length_window = WindowStat("length", 50)
loss_window = WindowStat("loss", 50)
obs, actions, rewards, next_obs, dones = list(), list(), list(), list(
), list()
act_count = 0
for i in range(100):
ob = env.reset()
done = False
episode_reward = .0
episode_len = 0
while not done:
action, results = agent.act([ob],
deterministic=False,
use_perturbed_action=False)
next_ob, reward, done, info = env.step(action[0])
act_count += 1
obs.append(ob)
actions.append(action[0])
rewards.append(reward)
next_obs.append(next_ob)
dones.append(done)
if agent.ready_to_send:
agent.send_experience(obs=obs,
actions=actions,
rewards=rewards,
next_obs=next_obs,
dones=dones)
if agent.ready_to_receive:
batch_data = agent.receive_experience()
res = agent.learn(batch_data)
loss_window.push(res['loss'])
if AGENT_CONFIG.get("prioritized_replay", False):
agent.update_priorities(indexes=batch_data["indexes"],
td_error=res["td_error"])
ob = next_ob
episode_reward += reward
episode_len += 1
if act_count % 5 == 0:
print("timestep:", act_count, reward_window, length_window)
agent.add_episode(1)
reward_window.push(episode_reward)
length_window.push(episode_len)
agent.export_saved_model()
print("Done.")
def doTestES(self):
np.random.seed(0)
env = gym.make("CartPole-v0")
env.seed(0)
es_g = tf.Graph()
with es_g.as_default():
tf.set_random_seed(123)
agent = agents[ES_AGENT_CONFIG["type"]](
env.observation_space,
env.action_space,
ES_AGENT_CONFIG,
ES_MODEL_CONFIG,
distributed_spec={},
custom_model=MyESmodel)
reward_window = WindowStat("reward", 25)
perturbation_scale = 0.1
seeds, rewards, perturbation_scales = list(), list(), list()
is_positive_direction = list()
episode_per_perturbation = 1
returns = list()
for i in range(5000):
ob = env.reset()
done = False
episode_reward = .0
if i % episode_per_perturbation == 0:
# perturb parameters every `episode_per_seed` episodes
is_positive = True if len(
is_positive_direction
) == 0 else is_positive_direction[-1] != True
# each seed twice
seed = np.random.randint(1000000) if is_positive else seeds[-1]
perturbation_scale = max(perturbation_scale * (1 - i / 2000.0),
0.02)
feed = agent.model.perturbation_feed
fetch = [agent.model.reset_perturbation_op]
agent.executor.run(
fetches=fetch,
feed_dict={
feed['perturbation_seeds']: [seed],
feed['perturbation_scales']: [perturbation_scale],
feed['positive_perturbation']: is_positive
})
if is_positive:
seeds.append(seed)
perturbation_scales.append(perturbation_scale)
is_positive_direction.append(is_positive)
while not done:
action, result = agent.act(
[ob], True, use_perturbed_action=True)
next_ob, reward, done, info = env.step(action[0])
ob = next_ob
episode_reward += reward
rewards.append(episode_reward)
reward_window.push(episode_reward)
if len(rewards) == episode_per_perturbation:
returns.append(np.mean(rewards))
rewards = []
if len(returns) == 2 * agent.config.get(
'sample_batch_size', 100):
print(reward_window)
assert len(seeds) == (len(returns) / 2)
assert len(perturbation_scales) == (len(returns) / 2)
agent.learn(
batch_data=dict(
perturbation_seeds=seeds,
perturbation_scales=perturbation_scales,
returns=np.reshape(returns, [-1, 2])))
seeds = []
perturbation_scales = []
returns = []
is_positive_direction = []
# evaluation 20 episodes
test_rewards = list()
for j in range(10):
done = False
ob = env.reset()
episode_reward = 0
while not done:
action, result = agent.act(
[ob], True, use_perturbed_action=False)
next_ob, reward, done, info = env.step(action[0])
ob = next_ob
episode_reward += reward
test_rewards.append(episode_reward)
print("[evaluation] average reward of 20 episodes:",
np.mean(test_rewards))
print('train at ', i)
return np.mean(test_rewards)
def main():
env = gym.make("Pendulum-v0")
env.seed(0)
agent_class = agents[AGENT_CONFIG["type"]]
agent = agent_class(env.observation_space,
env.action_space,
AGENT_CONFIG,
MODEL_CONFIG,
distributed_spec={},
checkpoint_dir="ckpt_dir",
export_dir="dump_dir",
custom_model=MyDDPG)
reward_window = WindowStat("reward", 50)
length_window = WindowStat("length", 50)
loss_window = WindowStat("loss", 50)
actor_loss = WindowStat("actor_loss", 50)
obs, actions, rewards, next_obs, dones = list(), list(), list(), list(
), list()
act_count = 0
train_count = 0
total_cost = time.time()
for i in range(500):
ob = env.reset()
done = False
episode_reward = .0
episode_len = 0
while not done:
action, results = agent.act([ob],
False,
use_perturbed_action=False)
act_count += 1
next_ob, reward, done, info = env.step(action[0])
obs.append(ob)
actions.append(action[0])
rewards.append(0.1 * reward)
next_obs.append(next_ob)
dones.append(done)
if agent.ready_to_send:
agent.send_experience(obs=obs,
actions=actions,
rewards=rewards,
dones=dones,
next_obs=next_obs)
if agent.ready_to_receive:
batch_data = agent.receive_experience()
res = agent.learn(batch_data)
loss_window.push(res["critic_loss"])
actor_loss.push(res["actor_loss"])
train_count += 1
if AGENT_CONFIG.get("prioritized_replay", False):
agent.update_priorities(indexes=batch_data["indexes"],
td_error=res["td_error"])
ob = next_ob
episode_reward += reward
episode_len += 1
agent.add_episode(1)
reward_window.push(episode_reward)
length_window.push(episode_len)
if act_count % 200 == 0:
print("timestep:", act_count, reward_window, loss_window,
actor_loss)
agent.export_saved_model()
print("Done.", "act_count:", act_count, "train_count:", train_count,
"total_cost:",
time.time() - total_cost)
def main():
with open(FLAGS.config, 'r') as ips:
config = json.load(ips)
print(config)
env = gym.make("CartPole-v0")
env.seed(0)
agent_class = agents[config["agent"]["type"]]
agent = agent_class(env.observation_space,
env.action_space,
agent_config=config["agent"],
model_config=config["model"],
distributed_spec={
"ps_hosts": FLAGS.ps_hosts,
"worker_hosts": FLAGS.worker_hosts,
"job_name": FLAGS.job_name,
"task_index": FLAGS.task_index
},
custom_model=MyPGModel,
checkpoint_dir="")
all_cost = time.time()
if FLAGS.job_name == "ps":
print("ps starts===>")
agent.join()
elif FLAGS.job_name == "worker":
print("actor starts===>")
act_count = 0
train_count = 0
reward_window = WindowStat("reward", 50)
length_window = WindowStat("length", 50)
obs, actions, rewards, dones, value_preds = list(), list(), list(
), list(), list()
while not agent.should_stop():
ob = env.reset()
done = False
episode_reward = 0.0
episode_len = 0.0
while not done and not agent.should_stop():
action, results = agent.act([ob], False)
act_count += 1
new_ob, reward, done, info = env.step(action[0])
obs.append(ob)
actions.append(action[0])
rewards.append(0.1 * reward)
dones.append(done)
value_preds.append(results["value_preds"][0])
if agent.ready_to_send:
agent.send_experience(obs=obs,
actions=actions,
rewards=rewards,
dones=dones,
value_preds=value_preds)
batch_data = agent.receive_experience()
if batch_data:
extra_data = agent.learn(batch_data)
train_count += 1
print("learning {}".format(extra_data))
ob = new_ob
episode_reward += reward
episode_len += 1
print("act_count:", act_count)
reward_window.push(episode_reward)
length_window.push(episode_len)
print(reward_window)
print(length_window)
sys.stdout.flush()
if FLAGS.task_index == 0:
agent.export_saved_model(export_dir="a3c_export_dir")
print("export savedmodel finish")
else:
raise ValueError("Invalid job_name.")
all_cost = time.time() - all_cost
print("done. all_cost:", all_cost)
def main():
with open(FLAGS.config, 'r') as ips:
config = json.load(ips)
print(config)
env = gym.make("CartPole-v0")
env.seed(0)
agent_class = agents[config["agent"]["type"]]
agent = agent_class(env.observation_space,
env.action_space,
agent_config=config["agent"],
model_config=config["model"],
distributed_spec={
"ps_hosts": FLAGS.ps_hosts,
"memory_hosts": FLAGS.memory_hosts,
"actor_hosts": FLAGS.actor_hosts,
"learner_hosts": FLAGS.learner_hosts,
"job_name": FLAGS.job_name,
"task_index": FLAGS.task_index
},
custom_model=MyPPOModel,
checkpoint_dir=None)
all_cost = time.time()
if FLAGS.job_name == "ps":
print("ps starts===>")
agent.join()
elif FLAGS.job_name == "memory":
print("memory starts===>")
while not agent.should_stop():
agent.communicate()
print(agent._receive_count, "actor2mem_q:",
agent._actor2mem_q.qsize(), "mem2learner_q:",
agent._mem2learner_q.qsize())
sys.stdout.flush()
time.sleep(0.1)
elif FLAGS.job_name == "actor":
print("actor starts===>")
act_count = 0
reward_window = WindowStat("reward", 50)
length_window = WindowStat("length", 50)
obs, actions, rewards, dones, value_preds, logits = list(), list(
), list(), list(), list(), list()
agent.sync_vars()
while not agent.should_stop():
ob = env.reset()
done = False
episode_reward = .0
episode_len = 0
print("begin an episode.")
sys.stdout.flush()
while not done and not agent.should_stop():
action, results = agent.act([ob], False)
act_count += 1
new_ob, reward, done, info = env.step(action[0])
obs.append(ob)
actions.append(action[0])
rewards.append(0.1 * reward)
dones.append(done)
logits.append(results["logits"][0])
value_preds.append(results["value_preds"][0])
if agent.ready_to_send:
print("to send exp.")
sys.stdout.flush()
agent.send_experience(obs=obs,
actions=actions,
rewards=rewards,
dones=dones,
logits=logits,
value_preds=value_preds)
agent.sync_vars()
print("sent")
sys.stdout.flush()
ob = new_ob
episode_reward += reward
episode_len += 1
print("act_count:", act_count)
reward_window.push(episode_reward)
length_window.push(episode_len)
print(reward_window)
print(length_window)
sys.stdout.flush()
elif FLAGS.job_name == "learner":
print("learner starts===>")
train_count = 0
while not agent.should_stop():
batch_data = agent.receive_experience()
if batch_data:
extra_data = agent.learn(batch_data)
train_count += 1
print("learning {}".format(extra_data))
sys.stdout.flush()
else:
raise ValueError("Invalid job_name.")
all_cost = time.time() - all_cost
print("done. all_cost:", all_cost)
def doTestSavedModel(self):
trial_timestamp = time.strftime("%Y%m%d-%H%M%S")
model_dir = "model_dir_{}".format(trial_timestamp)
os.system("mkdir {}".format(model_dir))
np.random.seed(0)
env = gym.make("CartPole-v0")
env.seed(0)
dqn_g = tf.Graph()
with dqn_g.as_default():
tf.set_random_seed(123)
agent = agents[DQN_AGENT_CONFIG["type"]](env.observation_space,
env.action_space,
DQN_AGENT_CONFIG,
DQN_MODEL_CONFIG,
export_dir=model_dir,
distributed_spec={})
reward_window = WindowStat("reward", 50)
obs, actions, rewards, next_obs, dones = list(), list(), list(), list(
), list()
act_count = 0
for i in range(500):
ob = env.reset()
done = False
episode_reward = .0
while not done:
action, results = agent.act([ob],
deterministic=False,
use_perturbed_action=False)
next_ob, reward, done, info = env.step(action[0])
act_count += 1
obs.append(ob)
actions.append(action[0])
rewards.append(reward)
next_obs.append(next_ob)
dones.append(done)
if agent.ready_to_send:
agent.send_experience(obs=obs,
actions=actions,
rewards=rewards,
next_obs=next_obs,
dones=dones)
if agent.ready_to_receive:
batch_data = agent.receive_experience()
res = agent.learn(batch_data)
if DQN_AGENT_CONFIG.get("prioritized_replay", False):
agent.update_priorities(indexes=batch_data["indexes"],
td_error=res["td_error"])
ob = next_ob
episode_reward += reward
if act_count % 1024 == 0:
print("timestep:", act_count, reward_window)
agent.add_episode(1)
reward_window.push(episode_reward)
prev_perf = reward_window.stats()["reward_mean"]
print("Performance before saving is {}".format(prev_perf))
with tf.Session() as sess:
path = model_dir
MetaGraphDef = tf.saved_model.loader.load(
sess, tags=[sm.tag_constants.SERVING], export_dir=path)
# get SignatureDef protobuf
SignatureDef_d = MetaGraphDef.signature_def
SignatureDef = SignatureDef_d["predict_results"]
# get inputs/outputs TensorInfo protobuf
ph_inputs = {}
for name, ts_info in SignatureDef.inputs.items():
ph_inputs[name] = sm.utils.get_tensor_from_tensor_info(
ts_info, sess.graph)
outputs = {}
for name, ts_info in SignatureDef.outputs.items():
outputs[name] = sm.utils.get_tensor_from_tensor_info(
ts_info, sess.graph)
for name, ph in ph_inputs.items():
print(name, ph)
for name, ts in outputs.items():
print(name, ts)
reward_window = WindowStat("reward", 10)
for i in range(10):
ob = env.reset()
done = False
episode_reward = .0
while not done:
action = sess.run(outputs["output_actions"],
feed_dict={
ph_inputs["obs_ph"]:
[np.asarray(ob)],
ph_inputs["deterministic_ph"]: True
})
next_ob, reward, done, info = env.step(action[0])
episode_reward += reward
ob = next_ob
reward_window.push(episode_reward)
cur_perf = reward_window.stats()["reward_mean"]
print("Performance after restore is {}".format(cur_perf))
return prev_perf - cur_perf