def main(dir, interval):
with logger.session(dir):
saver = SnapshotSaver(dir, interval=interval)
state = saver.get_state()
alg_state = state['alg_state']
env = alg_state['env_maker'].make()
alg = state['alg']
alg(env=env, snapshot_saver=saver, **alg_state)
def main(dir, interval):
with logger.session(dir):
saver = SnapshotSaver(dir, interval=interval)
state = saver.get_state()
alg_state = state['alg_state']
env = alg_state['env_maker'].make()
alg = state['alg']
alg(env=env, snapshot_saver=saver, **alg_state)
def main(dir):
env = None
while True:
saver = SnapshotSaver(dir)
state = saver.get_state()
if state is None:
time.sleep(1)
continue
alg_state = state['alg_state']
if env is None:
env = alg_state['env_maker'].make()
policy = alg_state['policy']
ob = env.reset()
done = False
while not done:
action, _ = policy.get_action(ob)
ob, _, done, _ = env.step(action)
env.render()
def main(dir):
env = None
while True:
saver = SnapshotSaver(dir)
state = saver.get_state()
if state is None:
time.sleep(1)
continue
alg_state = state['alg_state']
if env is None:
env = alg_state['env_maker'].make()
policy = alg_state['policy']
ob = env.reset()
done = False
while not done:
action, _ = policy.get_action(ob)
ob, _, done, _ = env.step(action)
env.render()
def main(dir):
env = None
while True:
saver = SnapshotSaver(dir)
state = saver.get_state()
if state is None:
time.sleep(1)
continue
alg_state = state['alg_state']
if env is None:
# save videos of all episodes to monitor_dir
env = alg_state['env_maker'].make(
video_callable=lambda episode_id: True)
policy = alg_state['policy']
ob = env.reset()
done = False
while not done:
action, _ = policy.get_action(ob)
ob, _, done, _ = env.step(action)
env.render()
def run(v):
np.random.seed(v['seed'])
env_maker = EnvMaker('Pendulum-v0')
env = env_maker.make()
policy = GaussianMLPPolicy(
observation_space=env.observation_space,
action_space=env.action_space,
env_spec=env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=chainer.functions.tanh,
)
if v['baseline'] == 'mlp':
baseline = MLPBaseline(
observation_space=env.observation_space,
action_space=env.action_space,
env_spec=env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=chainer.functions.tanh,
)
elif v['baseline'] == 'time_dependent':
baseline = TimeDependentBaseline(
observation_space=env.observation_space,
action_space=env.action_space,
env_spec=env.spec,
)
elif v['baseline'] == 'linear_feature':
baseline = LinearFeatureBaseline(
observation_space=env.observation_space,
action_space=env.action_space,
env_spec=env.spec,
)
else:
raise ValueError
trpo(
env=env,
env_maker=env_maker,
n_envs=16,
policy=policy,
baseline=baseline,
batch_size=10000,
n_iters=100,
snapshot_saver=SnapshotSaver(logger.get_dir()),
)
def run(v):
np.random.seed(v['seed'])
env_maker = EnvMaker('CartPole-v0')
env = env_maker.make()
policy = CategoricalMLPPolicy(
observation_space=env.observation_space,
action_space=env.action_space,
env_spec=env.spec
)
baseline = MLPBaseline(
observation_space=env.observation_space,
action_space=env.action_space,
env_spec=env.spec
)
trpo(
env=env,
env_maker=env_maker,
n_envs=16,
policy=policy,
baseline=baseline,
batch_size=2000,
n_iters=100,
snapshot_saver=SnapshotSaver(logger.get_dir())
)
#!/usr/bin/env python
from algs import a2c
from env_makers import EnvMaker
from models import CategoricalCNNPolicy
from utils import SnapshotSaver
import numpy as np
import os
import logger
log_dir = "data/local/a2c-pong"
np.random.seed(42)
# Clean up existing logs
os.system("rm -rf {}".format(log_dir))
with logger.session(log_dir):
env_maker = EnvMaker('PongNoFrameskip-v4')
env = env_maker.make()
policy = CategoricalCNNPolicy(env.observation_space, env.action_space,
env.spec)
vf = policy.create_vf()
a2c(
env=env,
env_maker=env_maker,
n_envs=16,
policy=policy,
vf=vf,
snapshot_saver=SnapshotSaver(log_dir, interval=10),
)
log_dir = "data/local/trpo-cartpole"
np.random.seed(42)
# Clean up existing logs
os.system("rm -rf {}".format(log_dir))
with logger.session(log_dir):
env_maker = EnvMaker('CartPole-v0')
env = env_maker.make()
policy = CategoricalMLPPolicy(
observation_space=env.observation_space,
action_space=env.action_space,
env_spec=env.spec
)
baseline = MLPBaseline(
observation_space=env.observation_space,
action_space=env.action_space,
env_spec=env.spec
)
trpo(
env=env,
env_maker=env_maker,
n_envs=16,
policy=policy,
baseline=baseline,
batch_size=2000,
n_iters=100,
snapshot_saver=SnapshotSaver(log_dir)
)