def __init__(self, visualize=False, token=None, max_obstacles=3):
logger.info("max_obstacles={}".format(max_obstacles))
if token is None:
self.remote_env = False
self.env = RunEnv(visualize=visualize, max_obstacles=max_obstacles)
else:
self.remote_env = True
self.local_env = RunEnv(visualize=False, max_obstacles=max_obstacles)
self.token = token
self.env = Client(GRADER_URL)
self.env_created = False
def __init__(self, game_name, display):
self.game_name = game_name
self.display = display
# self.env = gym.make(game_name)
self.env = RunEnv(self.display)
self.reset()
def main():
env = RunEnv(visualize=False)
population = [[NN(), 0] for _ in range(100)]
generation = 0
for _ in range(2000):
for i in range(len(population)):
print i
population[i][1] = run(population[i][0], env)
population = sorted(population, key=lambda x: x[1], reverse=True)
print np.mean([p[1] for p in population[:5]])
generation += 1
population = population[:50]
for _ in range(20):
population.append([random.choice(population[:50])[0].mutate(), 0])
for _ in range(20):
nn1 = random.choice(population[:20])[0]
nn2 = random.choice(population[:50])[0]
population.append([nn1.crossover(nn2), 0])
for _ in range(10):
population.append([NN(), 0])
with open('save.p', 'w') as f:
pickle.dump(population, f)
def create_env(args):
env = RunEnv(visualize=True, max_obstacles=args.max_obstacles)
if hasattr(args, "baseline_wrapper") or hasattr(args, "ddpg_wrapper"):
env = DdpgWrapper(env, args)
return env
def test():
task_fn = lambda: LTR()
task = task_fn()
state_dim = task.env.observation_space.shape[0]
action_dim = task.env.action_space.shape[0]
with open('data/ddpg-model-LearningToRun.bin', 'rb') as f:
model = pickle.load(f)
actor = DDPGActorNet(state_dim, action_dim)
actor.load_state_dict(model)
logger = Logger('./log')
env = RunEnv(visualize=False)
state = env.reset(difficulty=0)
print state
done = False
total_reward = 0.0
step = 0
while not done:
action = actor.predict(np.stack([state]), to_numpy=True).flatten()
state, reward, done, info = env.step(action)
total_reward += reward
step += 1
logger.histo_summary('input', actor.input, step)
logger.histo_summary('act1', actor.act1, step)
logger.histo_summary('act2', actor.act2, step)
logger.histo_summary('pre_act3', actor.pre_act3, step)
logger.histo_summary('act3', actor.act3, step)
for tag, value in actor.named_parameters():
tag = tag.replace('.', '/')
logger.histo_summary(tag, value.data.numpy(), step)
print total_reward
print step
def test(skip=4):
test_env = RunEnv(visualize=True, max_obstacles=0)
fast_env = FastEnv(test_env, skip) # 4 is skip factor
agent.training = False
agent.play(fast_env, noise_level=1e-11, episode_index=-1)
agent.training = True
del test_env
def main():
env = RunEnv(visualize=False)
s = socket.socket()
s.bind(("localhost", 8000))
s.listen(10) # max number of connections
while True:
sc, address = s.accept()
f = open("work.p", 'wb')
while (True):
l = sc.recv(1024)
while (l):
f.write(l)
l = sc.recv(1024)
f.close()
with open('work.p', 'r') as f:
nn = pickle.load(f)
reward = run(nn, env)
sc.send(str(reward))
sc.close()
s.close()
def test():
env = RunEnv(visualize=False)
observation_d = env.reset(project=False)
observation = process_obs_dict(observation_d)
total_reward = 0
steps = 0
while True:
#a = AGENT OUTPUT
a, q = agent.act(observation)
observation_d, reward, done, info = env.step(a, project=False)
observation = process_obs_dict(observation_d)
total_reward += reward
steps += 1
#print(observation)
print(steps, 'total reward:', total_reward)
if done:
break
print('finished testing!')
def Simulation(proxy_agent,index, return_dict, episodes, vis=False):
print('starting simulation')
env = RunEnv(visualize=vis)
observation = env.reset(difficulty=0)
rewards = np.zeros(episodes)
totalreward = 0
for episode in range(0, episodes):
action = env.action_space.sample()
observation, reward, done, info = env.step(action)
observation = np.array(observation)
Preprocess = Preprocessing(observation, delta=0.01)
prevState = Preprocess.GetState(observation)
for i in range(1,1000):
observation, reward, done, info = env.step(action)
observation = np.array(observation)
#means it didn't go the full simulation
if done and i < 1000:
reward = 0
state = Preprocess.GetState(observation)
s,a,r,sp = Preprocess.ConvertToTensor(prevState,action, reward, state)
totalreward += reward
if done:
env.reset(difficulty = 0, seed = None) #resets the environment if done is true
print("reseting environment" + str(episode))
rewards[episode] = totalreward
totalreward = 0
break
action = proxy_agent(Variable(s, volatile=True))
action = action.data.numpy()
prevState = state;
return_dict[index] = np.sum(rewards) / episodes
return np.sum(rewards) / episodes
def test(frameskip = 1, vis = False):
env = RunEnv(visualize=vis)
#env.change_model(model='2D', prosthetic=True, difficulty=0, seed=None)
observation_d = env.reset(project = False)
#observation = process_obs_dict(observation_d)
total_reward = 0
steps = 0
while True:
#a = AGENT OUTPUT
observation = process_obs_dict(observation_d)
a, q = agent.act(observation)
for _ in range(frameskip):
observation_d, reward, done, info = env.step(a, project = False)
#observation = process_obs_dict(observation_d)
total_reward += reward
steps += 1
#print(observation)
print(steps, 'total reward:', total_reward)
if done:
break
print('finished testing!')
def test1(self):
env = RunEnv(visualize=False)
observation = env.reset()
action = env.action_space.sample()
action[5] = np.NaN
self.assertRaises(ValueError, env.step, action)
def run(self):
self.env = RunEnv(visualize=False)
self.env.reset(difficulty = 2, seed = int(time.time()))
if self.monitor:
self.env.monitor.start('monitor/', force=True)
# tensorflow variables (same as in model.py)
self.observation_size = 55+7
self.action_size = np.prod(self.env.action_space.shape)
self.hidden_size = 128
weight_init = tf.random_uniform_initializer(-0.05, 0.05)
bias_init = tf.constant_initializer(0)
# tensorflow model of the policy
self.obs = tf.placeholder(tf.float32, [None, self.observation_size])
self.debug = tf.constant([2,2])
with tf.variable_scope("policy-a"):
h1 = fully_connected(self.obs, self.observation_size, self.hidden_size, weight_init, bias_init, "policy_h1")
h1 = tf.nn.relu(h1)
h2 = fully_connected(h1, self.hidden_size, self.hidden_size, weight_init, bias_init, "policy_h2")
h2 = tf.nn.relu(h2)
h3 = fully_connected(h2, self.hidden_size, self.action_size, weight_init, bias_init, "policy_h3_1")
h3 = tf.nn.tanh(h3,name="policy_h3")
action_dist_logstd_param = tf.Variable((.01*np.random.randn(1, self.action_size)).astype(np.float32), name="policy_logstd")
self.action_dist_mu = h3
self.action_dist_logstd = tf.tile(action_dist_logstd_param, tf.stack((tf.shape(self.action_dist_mu)[0], 1)))
config = tf.ConfigProto(
device_count = {'CPU': 0}
)
self.session = tf.Session()
self.session.run(tf.initialize_all_variables())
var_list = tf.trainable_variables()
self.set_policy = SetPolicyWeights(self.session, var_list)
while True:
# get a task, or wait until it gets one
next_task = self.task_q.get(block=True)
if next_task == 1:
# the task is an actor request to collect experience
path = self.rollout()
self.task_q.task_done()
self.result_q.put(path)
elif next_task == 2:
print "kill message"
if self.monitor:
self.env.monitor.close()
self.task_q.task_done()
break
else:
# the task is to set parameters of the actor policy
self.set_policy(next_task)
# super hacky method to make sure when we fill the queue with set parameter tasks,
# an actor doesn't finish updating before the other actors can accept their own tasks.
time.sleep(0.1)
self.task_q.task_done()
return
def build_model(shared_object):
shared_object['env'] = RunEnv(shared_object.get('visualize',False))
model_class_name = 'models.agents.' + shared_object.get('model_class',None)
log_info('importing class : {}'.format(model_class_name))
model_class = import_class(model_class_name)
log_info('{} successfuly imported'.format(model_class_name))
log_info('building model')
model = model_class(shared_object)
return model
def Simulation(proxy_agent, episodes, vis=False):
env = RunEnv(visualize=vis)
observation = env.reset(difficulty=0)
memory = random.randint(1000, 2000)
tau = random.uniform(0.01, .9)
epsilon = random.uniform(.15, .9)
target = proxy_agent.ProduceTargetActorCritic( memory, tau, epsilon )
batches = [ 16, 32, 64, 128]
batchsize = batches[random.randint(0,len(batches)-1)]
for episode in range(0, episodes):
action = env.action_space.sample()
observation, reward, done, info = env.step(action)
observation = np.array(observation)
Preprocess = Preprocessing(observation, delta=0.01)
prevState = Preprocess.GetState(observation)
if(vis):
target.OUprocess(0, 0.15, 0.0)
else:
target.OUprocess(random.random(), 0.15,0.0)
pelvis_y = 0
for i in range(1,1000):
observation, reward, done, info = env.step(action)
observation = np.array(observation)
#means it didn't go the full simulation
if i > 1:
reward += (observation[2] - pelvis_y)*0.01 #penalty for pelvis going down
reward = env.current_state[4] * 0.01
reward += 0.01 # small reward for still standing
reward += min(0, env.current_state[22] - env.current_state[1]) * 0.1 # penalty for head behind pelvis
reward -= sum([max(0.0, k - 0.1) for k in [env.current_state[7], env.current_state[10]]]) * 0.02 # penalty for straight legs
if done and i < 1000:
reward = 0
state = Preprocess.GetState(observation)
s,a,r,sp = Preprocess.ConvertToTensor(prevState,action, reward, state)
target.addToMemory(s,a,r,sp)
# env.render()
if done:
env.reset(difficulty = 0, seed = None) #resets the environment if done is true
if(target.primedToLearn()):
lock.acquire()
proxy_agent.PerformUpdate(batchsize, target)
target.UpdateTargetNetworks(agent.getCritic(), agent.getActor())
print("saving actor")
proxy_agent.saveActorCritic()
print("actor saved")
lock.release()
print("reseting environment" + str(episode))
break
action = target.selectAction(s)
action = action.numpy()
prevState = state;
def test_actions(self):
env = RunEnv(visualize=False)
env.reset()
v = env.action_space.sample()
v[0] = 1.5
v[1] = -0.5
observation, reward, done, info = env.step(v)
self.assertLessEqual(env.last_action[0],1.0)
self.assertGreaterEqual(env.last_action[1],0.0)
def __init__(self, visualize=False, difficulty=None):
super(LearnToRunEnv, self).__init__()
if difficulty == None:
self.difficulty = random.randint(0, 2)
else:
self.difficulty = difficulty
self.learntorun_env = RunEnv(visualize=visualize)
self.observation_space = self.learntorun_env.observation_space
self.action_space = self.learntorun_env.action_space
def standalone_headless_isolated(pq, cq, plock):
# locking to prevent mixed-up printing.
plock.acquire()
print('starting headless...', pq, cq)
try:
from osim.env import RunEnv
# RunEnv = runenv_with_alternative_obstacle_generation_scheme()
e = RunEnv(visualize=False, max_obstacles=0)
# bind_alternative_pelvis_judgement(e)
# use_alternative_episode_length(e)
except Exception as err:
print('error on start of standalone')
traceback.print_exc()
plock.release()
return
else:
plock.release()
def report(e):
# a way to report errors ( since you can't just throw them over a pipe )
# e should be a string
print('(standalone) got error!!!')
cq.put(('error', e))
def floatify(np):
return [float(np[i]) for i in range(len(np))]
try:
while True:
msg = pq.get()
# messages should be tuples,
# msg[0] should be string
# isinstance is dangerous, commented out
# if not isinstance(msg,tuple):
# raise Exception('pipe message received by headless is not a tuple')
if msg[0] == 'reset':
o = e.reset(difficulty=0)
cq.put(floatify(o))
elif msg[0] == 'step':
o, r, d, i = e.step(msg[1])
o = floatify(o) # floatify the observation
cq.put((o, r, d, i))
else:
cq.close()
pq.close()
del e
break
except Exception as e:
traceback.print_exc()
report(str(e))
return # end process
def standalone_headless_isolated(conn, plock):
# locking to prevent mixed-up printing.
plock.acquire()
print('starting headless...', conn)
try:
import traceback
from osim.env import RunEnv
e = RunEnv(visualize=False)
except Exception as e:
print('error on start of standalone')
traceback.print_exc()
plock.release()
return
else:
plock.release()
def report(e):
# a way to report errors ( since you can't just throw them over a pipe )
# e should be a string
print('(standalone) got error!!!')
conn.send(('error', e))
def floatify(np):
return [float(np[i]) for i in range(len(np))]
try:
while True:
msg = conn.recv()
# messages should be tuples,
# msg[0] should be string
# isinstance is dangerous, commented out
# if not isinstance(msg,tuple):
# raise Exception('pipe message received by headless is not a tuple')
if msg[0] == 'reset':
o = e.reset(difficulty=2)
conn.send(floatify(o))
elif msg[0] == 'step':
ordi = e.step(msg[1])
ordi[0] = floatify(ordi[0])
conn.send(ordi)
else:
conn.close()
del e
break
except Exception as e:
traceback.print_exc()
report(str(e))
return # end process
def test(skip=1):
# e = p.env
te = RunEnv(visualize=True, max_obstacles=10)
from multi import fastenv
fenv = fastenv(te, skip) # 4 is skip factor
agent.render = True
try:
agent.play(fenv, realtime=True, max_steps=-1, noise_level=1e-11)
except:
pass
finally:
del te
def __init__(self,
game='l2r',
visualize=False,
max_obstacles=10,
skip_count=1):
self.env = RunEnv(visualize=visualize, max_obstacles=max_obstacles)
self.step_count = 0
self.old_observation = None
self.skip_count = 1 # skip_count # 4
self.last_x = 0
self.current_x = 0
self.observation_space_shape = (76, )
self.action_space = self.env.action_space
self.difficulty = 2
def __init__(self,
visualize=True,
test=False,
step_size=0.01,
processor=None,
timestep_limit=1000):
self.visualize = visualize
self._osim_env = RunEnv(visualize=visualize)
self._osim_env.stepsize = step_size
self._osim_env.spec.timestep_limit = timestep_limit
self._osim_env.horizon = timestep_limit
# self._osim_env.integration_accuracy = 1e-1
if test:
self._osim_env.timestep_limit = 1000
self.processor = processor
print "stepsize: " + str(self._osim_env.stepsize)
def test(skip=1):
# e = p.env
te = RunEnv(visualize=False)
from multi import fastenv
fenv = fastenv(te, skip) # 4 is skip factor
agent.render = True
agent.training = False
try:
#print('playing')
#agent.play(fenv,realtime=True,max_steps=-1,noise_level=1e-11)
playifavailable(0)
except:
pass
finally:
del te
def test_reset(self):
env = RunEnv(visualize=False)
observation = env.reset(difficulty=2, seed=123)
env1 = env.env_desc
observation = env.reset(difficulty=2, seed=3)
observation = env.reset(difficulty=2, seed=3)
observation = env.reset(difficulty=2, seed=3)
observation = env.reset(difficulty=2, seed=3)
observation = env.reset(difficulty=2, seed=123)
env2 = env.env_desc
s = map(lambda x: x[0] - x[1], list(zip(env1["obstacles"][1],env2["obstacles"][1])))
self.assertAlmostEqual(sum([k**2 for k in s]), 0.0)
action = env.action_space.sample()
action[5] = np.NaN
self.assertRaises(ValueError, env.step, action)
def env(chrom):
from osim.env import L2RunEnv as RunEnv
e = RunEnv(visualize=False)
e.reset()
T = 2
total_reward = 0
for t in range(500):
obs, reward, done, _ = e.step(
controller.input(chrom.allele, T, t * 0.01))
total_reward += reward
if done:
break
# print("HEADLESS: The reward is {}".format(total_reward))
# enables to calculate accumulated fitness
if total_reward < 0: total_reward = 0
del e
return total_reward
def test(args):
print('start testing')
ddpg = DDPG()
ddpg.load_model(args.model, load_memory=False)
env = RunEnv(visualize=args.visualize, max_obstacles=args.max_obs)
np.random.seed(args.seed)
for i in range(1):
step = 0
state = env.reset(difficulty=2)
fg = FeatureGenerator()
state = fg.gen(state)
#obs = fg.traj[0]
#print(obs.left_knee_r, obs.right_knee_r)
ep_reward = 0
ep_memories = []
while True:
action = ddpg.select_action(list(state))
next_state, reward, done, info = env.step(action.tolist())
next_state = fg.gen(next_state)
#obs = fg.traj[0]
#print(obs.left_knee_r, obs.right_knee_r)
print('step: {0:03d}'.format(step), end=', action: ')
for act in action:
print('{0:.3f}'.format(act), end=', ')
print()
state = next_state
ep_reward += reward
step += 1
print('reward:', ep_reward)
if done:
break
print('\nEpisode: {} Reward: {}, n_steps: {}'.format(
i, ep_reward, step))
def create(self, env_id, seed=None):
try:
if (env_id == 'osim'):
from osim.env import RunEnv
env = RunEnv(visualize=True)
else:
env = gym.make(env_id)
print('making environment')
if seed:
env.seed(seed)
except gym.error.Error:
raise InvalidUsage(
"Attempted to look up malformed environment ID '{}'".format(
env_id))
instance_id = str(uuid.uuid4().hex)[:self.id_len]
self.envs[instance_id] = env
self.envs_id[instance_id] = env_id
return instance_id
def submit(self):
remote_base = 'http://grader.crowdai.org:1729'
env = RunEnv(visualize=self.visualize)
client = Client(remote_base)
# Create environment
observation = client.env_create(self.submit_token)
# Run a single step
#
# The grader runs 3 simulations of at most 1000 steps each. We stop after the last one
while True:
[observation, reward, done,
info] = client.env_step(self.agent.forward(observation))
if done:
observation = client.env_reset()
if not observation:
break
client.submit()
def standalone(conn,visualize=True):
from osim.env import RunEnv
re = RunEnv(visualize=visualize)
e = fastenv(re,4)
while True:
msg = conn.recv()
# messages should be tuples,
# msg[0] should be string
if msg[0] == 'reset':
obs = e.reset()
conn.send(obs)
elif msg[0] == 'step':
four = e.step(msg[1])
conn.send(four)
else:
conn.close()
del e
return
def standalone_headless_isolated(conn,
visualize,
n_obstacles,
run_logs_dir,
additional_info,
higher_pelvis=0.65):
try:
e = RunEnv(visualize=visualize, max_obstacles=n_obstacles)
if higher_pelvis != 0.65:
bind_alternative_pelvis_judgement(e, higher_pelvis)
e = MyRunEnvLogger(e,
log_dir=run_logs_dir,
additional_info=additional_info)
while True:
msg = conn.recv()
# messages should be tuples,
# msg[0] should be string
if msg[0] == 'reset':
o = e.reset(difficulty=msg[1], seed=msg[2])
conn.send(o)
elif msg[0] == 'step':
ordi = e.step(msg[1])
conn.send(ordi)
elif msg[0] == 'close':
e.close()
conn.send(None)
import psutil
current_process = psutil.Process()
children = current_process.children(recursive=True)
for child in children:
child.terminate()
return
except Exception as e:
import traceback
print(traceback.format_exc())
conn.send(e)
def __init__(self, shared_object):
self.env = shared_object.get("env", None)
if self.env:
self.env = RunEnv(shared_object.get('visualize', False))
self.nb_actions = self.env.action_space.shape[0]
## memory parameters
self.memoryLimit = shared_object.get('memoryLimit', 100000)
self.window_length = shared_object.get('window_length', 1)
## random process parameters
self.random_process_theta = shared_object.get('random_process_theta',
.15)
self.random_process_mu = shared_object.get('random_process_mu', 0.)
self.random_process_sigma = shared_object.get('random_process_sigma',
.2)
## building the networks
super(example, self).__init__(shared_object)
