def node_delete(project, label): e = edges(project) n = nodes(project) edata = load(e) ndata = load(n) node_del_record(ndata[0], ndata[1], edata[0], edata[1], label) save(n, ndata)
def node_delete(project, label):
e = edges(project)
n = nodes(project)
edata = load(e)
ndata = load(n)
node_del_record(ndata[0], ndata[1], edata[0], edata[1], label)
save(n, ndata)
def edge_update(project, record): e = edges(project) n = nodes(project) edata = load(e) ndata = load(n) edge_upd_record(ndata[0], ndata[1], edata[0], edata[1], record) save(e, ndata)
def edge_update(project, record):
e = edges(project)
n = nodes(project)
edata = load(e)
ndata = load(n)
edge_upd_record(ndata[0], ndata[1], edata[0], edata[1], record)
save(e, ndata)
def run(episodes=2500, #2500,
render=False,
experiment='ERA5-v0',
max_actions=1000,
knn=0.1,
save_dir=None):
env = gym.make(experiment)
print(env.observation_space)
print(env.action_space)
steps = env.spec.timestep_limit # pulls from the init file where it is registered
# agent = DDPGAgent(env)
agent = WolpertingerAgent(env, max_actions=max_actions, k_ratio=knn)
timer = Timer()
data = util.data.Data()
data.set_agent(agent.get_name(), int(agent.action_space.get_number_of_actions()),
agent.k_nearest_neighbors, 3)
data.set_experiment(experiment, agent.low.tolist(), agent.high.tolist(), episodes)
agent.add_data_fetch(data)
print(data.get_file_name())
full_epoch_timer = Timer()
reward_sum = 0
# EREZ ADD
num_avg = 40
recent_rew_list = []
for ep in range(episodes):
timer.reset()
observation = env.reset()
total_reward = 0
print('Episode ', ep, '/', episodes - 1, 'started...', end='')
for t in range(steps):
if render:
env.render()
action = agent.act(observation)
data.set_action(action.tolist())
data.set_state(observation.tolist())
prev_observation = observation
observation, reward, done, info = env.step(action[0] if len(action) == 1 else action)
data.set_reward(reward)
episode = {'obs': prev_observation,
'action': action,
'reward': reward,
'obs2': observation,
'done': done,
't': t}
agent.observe(episode)
total_reward += reward
if done or (t == steps - 1):
t += 1
reward_sum += total_reward
time_passed = timer.get_time()
# NOTE: shouldnt we be reporting average over recent?
# Added better print formating
#print('\tReward:{:04.4f} \tSteps:{} \tt:{} \t({}/step) \tCur avg={:04.4f}'.format(total_reward, t,
# time_passed, round(
# time_passed / t),
# reward_sum / (ep + 1)))
# EREZ ADDED
recent_rew_list.append(total_reward)
if ep < num_avg:
recent_avg = sum(recent_rew_list)/len(recent_rew_list)
else:
recent_avg = sum(recent_rew_list[-num_avg:])/num_avg
print('\tReward:{:05.3f} \tSteps:{} \tt:{} \t({}/step) \tCur avg={:04.4f}'.format(total_reward, t,
time_passed, round(
time_passed / t),
recent_avg))
#print('\tReward:{:04.4f} \tSteps:{} \tt:{} \t({}/step) \tCur avg={:04.4f}'.format(total_reward, t,
# time_passed, round(
# time_passed / t),
# round(reward_sum / (ep + 1))))
# TODO -- look into these and change how we write out directories
data.finish_and_store_episode()
break
# end of episodes
time = full_epoch_timer.get_time()
print('Run {} episodes in {} seconds and got {} average reward'.format(
episodes, time / 1000, reward_sum / episodes))
#data.save()
if save_dir is None:
data.save() # EREZ ADDED ARG
else:
data.save(path=save_dir) # EREZ ADDED ARG
agent.save(save_dir) # could add a seperate call to data,save from within this!
# Code added below
return agent
def run(experiment,
episodes,
max_actions,
knn,
action_space_config=['off', 'square', 1000, 10],
result_dir=PROJECT_DIR,
render=False,
load_agent=True,
agent_to_load=None,
save_agent=True,
save_data=True,
training_flag=True,
id=0,
comment="run",
close_session=True,
silent=False,
tempsave=True,
save_action_space=False):
env = gym.make(experiment)
print(env.observation_space)
print(env.action_space)
steps = env.spec.timestep_limit
agent = WolpertingerAgent(env,
result_dir,
max_actions=max_actions,
k_ratio=knn,
training_flag=training_flag,
action_space_config=action_space_config,
save_action_space=save_action_space)
if load_agent:
if agent_to_load is not None:
agent.load_agent(agent_name=agent_to_load[0],
comment=agent_to_load[1])
else:
agent.load_agent(comment=comment)
timer = Timer()
if save_data:
data = util.data.Data(agent.get_dir(),
comment=comment,
tempsave=tempsave)
data.set_agent(agent.get_name(), int(agent.action_space.get_size()),
agent.k_nearest_neighbors, agent.get_version())
data.set_experiment(experiment, agent.low.tolist(),
agent.high.tolist(), episodes)
data.set_id(id)
agent.add_data_fetch(data)
print(data.get_file_name())
# if render:
# monitor = Monitor(400, env.observation_space.shape[0], env.action_space.shape[0], 50,
# [agent.low.tolist()[0], agent.high.tolist()[0]])
full_epoch_timer = Timer()
reward_sum = 0
ou = OUNoise(1, mu=0, theta=0.5, sigma=.1)
# temp_buffer = [0] * 150
for ep in range(episodes):
timer.reset()
observation = env.reset()
total_reward = 0
if not silent:
print('Episode ', ep, '/', episodes - 1, end='. ')
for t in range(steps):
if render:
env.render()
action = agent.act(observation)
if save_data:
data.set_action(action.tolist())
data.set_state(observation.tolist())
prev_observation = observation
# some environments need the action as scalar valua, and other as array
# for scalar: action[0] if len(action) == 1 else action
observation, reward, done, info = env.step(action.flatten())
# if render:
# monitor.add_data(observation, action, reward)
# monitor.repaint()
if save_data:
data.set_reward(reward)
# if render:
# monitor.add_data(observation, action, reward)
episode = {
'obs': prev_observation,
'action': action,
'reward': reward,
'obs2': observation,
'done': done,
't': t
}
agent.observe(episode)
total_reward += reward
# print(episode['obs'], episode['action'], episode['obs2'], episode['reward'])
if done or (t == steps - 1):
# if render:
# monitor.end_of_episode()
t += 1
reward_sum += total_reward
time_passed = timer.get_time()
if not silent:
print(
'Reward:{} Steps:{} t:{} ({}/step) Curr avg={}, {} actions({})'
.format(total_reward, t, time_passed,
round(time_passed / t),
round(reward_sum / (ep + 1)),
agent.get_action_space_size(),
agent.get_action_space_size() / max_actions))
if save_data:
data.finish_and_store_episode()
break
# end of episodes
time = full_epoch_timer.get_time()
print('Run {} episodes in {} seconds and got {} average reward'.format(
episodes, time / 1000, reward_sum / episodes))
if save_data:
data.save()
if save_agent:
agent.save_agent(force=True, comment=comment)
if close_session:
agent.close_session()
print("END")
return agent
def edge_delete(project, source, target): e = edges(project) data = load(e) edge_del_record(data[0], data[1], source, target) save(e, data)
def node_create(project, record): n = nodes(project) data = load(n) node_add_record(data[0], data[1], record) save(n, data)
def create(which, f, attr): data = load(f) create_it(which, data[0], attr) save(f, data)
def create(which, f, attr):
data = load(f)
create_it(which, data[0], attr)
save(f, data)
def run(episodes=2500,
render=False,
experiment='InvertedPendulum-v1',
max_actions=1000,
knn=0.1):
env = gym.make(experiment)
print(env.observation_space)
print(env.action_space)
steps = env.spec.timestep_limit
# agent = DDPGAgent(env)
agent = WolpertingerAgent(env, max_actions=max_actions, k_ratio=knn)
timer = Timer()
data = util.data.Data()
data.set_agent(agent.get_name(),
int(agent.action_space.get_number_of_actions()),
agent.k_nearest_neighbors, 3)
data.set_experiment(experiment, agent.low.tolist(), agent.high.tolist(),
episodes)
agent.add_data_fetch(data)
print(data.get_file_name())
full_epoch_timer = Timer()
reward_sum = 0
for ep in range(episodes):
timer.reset()
observation = env.reset()
total_reward = 0
print('Episode ', ep, '/', episodes - 1, 'started...', end='')
for t in range(steps):
if render:
env.render()
action = agent.act(observation)
data.set_action(action.tolist())
data.set_state(observation.tolist())
prev_observation = observation
observation, reward, done, info = env.step(
action[0] if len(action) == 1 else action)
data.set_reward(reward)
episode = {
'obs': prev_observation,
'action': action,
'reward': reward,
'obs2': observation,
'done': done,
't': t
}
agent.observe(episode)
total_reward += reward
if done or (t == steps - 1):
t += 1
reward_sum += total_reward
time_passed = timer.get_time()
print('Reward:{} Steps:{} t:{} ({}/step) Cur avg={}'.format(
total_reward, t, time_passed, round(time_passed / t),
round(reward_sum / (ep + 1))))
data.finish_and_store_episode()
break
# end of episodes
time = full_epoch_timer.get_time()
print('Run {} episodes in {} seconds and got {} average reward'.format(
episodes, time / 1000, reward_sum / episodes))
data.save()
def edge_delete(project, source, target):
e = edges(project)
data = load(e)
edge_del_record(data[0], data[1], source, target)
save(e, data)
def node_create(project, record):
n = nodes(project)
data = load(n)
node_add_record(data[0], data[1], record)
save(n, data)
