def run_task(variant):
from rllab.baselines.linear_feature_baseline import LinearFeatureBaseline
from sandbox.rocky.tf.algos.vpg import VPG
from sandbox.rocky.tf.policies.gaussian_mlp_policy import GaussianMLPPolicy
from rllab.envs.box2d.cartpole_env import CartpoleEnv
from sandbox.rocky.tf.envs.base import TfEnv
env_name = variant['Environment']
if env_name == 'Cartpole':
env = TfEnv(CartpoleEnv())
policy = GaussianMLPPolicy(name="policy",
env_spec=env.spec,
hidden_sizes=(100, 100))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algorithm = VPG(
env=env,
policy=policy,
baseline=baseline,
n_itr=100,
start_itr=0,
batch_size=1000,
max_path_length=1000,
discount=0.99,
)
algorithm.train()
def experiment(variant, saveDir):
initial_params_file = variant['initial_params_file']
goalIndex = variant['goalIndex']
init_step_size = variant['init_step_size']
baseEnv = SawyerPickPlace_finnMAMLEnv()
env = TfEnv(NormalizedBoxEnv(baseEnv))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=None,
load_policy=initial_params_file,
baseline=baseline,
batch_size=10000, # 2x
max_path_length=150,
n_itr=10,
reset_arg=goalIndex,
optimizer_args={
'init_learning_rate': init_step_size,
'tf_optimizer_args': {
'learning_rate': 0.5 * init_step_size
},
'tf_optimizer_cls': tf.train.GradientDescentOptimizer
})
import os
saveDir = variant['saveDir']
if os.path.isdir(saveDir) == False:
os.mkdir(saveDir)
logger.set_snapshot_dir(saveDir)
logger.add_tabular_output(saveDir + 'progress.csv')
algo.train()
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=policy,
load_policy=initial_params_file,
baseline=baseline,
batch_size=4000, # 2x
max_path_length=100,
n_itr=n_itr,
optimizer_args={'init_learning_rate': step_sizes[step_i], 'tf_optimizer_args': {'learning_rate': 0.5*step_sizes[step_i]}, 'tf_optimizer_cls': tf.train.GradientDescentOptimizer}
)
run_experiment_lite(
algo.train(),
# Number of parallel workers for sampling
n_parallel=4,
# Only keep the snapshot parameters for the last iteration
snapshot_mode="last",
# Specifies the seed for the experiment. If this is not provided, a random seed
# will be used
seed=4,
exp_prefix='trpopoint2d_test',
exp_name='test',
#plot=True,
)
import pdb; pdb.set_trace()
# get return from the experiment
with open('data/local/trpopoint2d-test/test/progress.csv', 'r') as f:
reader = csv.reader(f, delimiter=',')
from sandbox.rocky.tf.algos.vpg import VPG
from rllab.baselines.linear_feature_baseline import LinearFeatureBaseline
from rllab.envs.box2d.cartpole_env import CartpoleEnv
from rllab.envs.normalized_env import normalize
from sandbox.rocky.tf.policies.gaussian_mlp_policy import GaussianMLPPolicy
from sandbox.rocky.tf.envs.base import TfEnv
from rllab.misc.instrument import stub, run_experiment_lite
env = TfEnv(normalize(CartpoleEnv()))
policy = GaussianMLPPolicy(
name="policy",
env_spec=env.spec,
# The neural network policy should have two hidden layers, each with 32 hidden units.
hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(env=env,
policy=policy,
baseline=baseline,
batch_size=10000,
max_path_length=100,
n_itr=4,
discount=0.99,
optimizer_args=dict(tf_optimizer_args=dict(learning_rate=0.01, )))
algo.train()
def experiment(variant):
seed = variant['seed']
tf.set_random_seed(seed)
np.random.seed(seed)
random.seed(seed)
initial_params_file = variant['initial_params_file']
goalIndex = variant['goalIndex']
init_step_size = variant['init_step_size']
regionSize = variant['regionSize']
mode = variant['mode']
if 'docker' in mode:
taskFilePrefix = '/root/code'
else:
taskFilePrefix = '/home/russellm'
if variant['valRegionSize'] != None:
valRegionSize = variant['valRegionSize']
tasksFile = taskFilePrefix + '/multiworld/multiworld/envs/goals/pickPlace_' + valRegionSize + '_val.pkl'
else:
tasksFile = taskFilePrefix + '/multiworld/multiworld/envs/goals/pickPlace_' + regionSize + '.pkl'
tasks = pickle.load(open(tasksFile, 'rb'))
envType = variant['envType']
if envType == 'Push':
baseEnv = SawyerPushEnv(tasks=tasks)
else:
assert (envType) == 'PickPlace'
baseEnv = SawyerPickPlaceEnv(tasks=tasks)
env = FinnMamlEnv(
FlatGoalEnv(baseEnv,
obs_keys=['state_observation', 'state_desired_goal']))
env = TfEnv(NormalizedBoxEnv(env))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=None,
load_policy=initial_params_file,
baseline=baseline,
batch_size=7500, # 2x
max_path_length=150,
n_itr=10,
reset_arg=goalIndex,
optimizer_args={
'init_learning_rate': init_step_size,
'tf_optimizer_args': {
'learning_rate': 0.1 * init_step_size
},
'tf_optimizer_cls': tf.train.GradientDescentOptimizer
})
import os
saveDir = variant['saveDir']
currPath = ''
for _dir in saveDir.split('/'):
currPath += _dir + '/'
if os.path.isdir(currPath) == False:
os.mkdir(currPath)
logger.set_snapshot_dir(saveDir)
logger.add_tabular_output(saveDir + 'progress.csv')
algo.train()
from sandbox.rocky.tf.policies.gaussian_mlp_policy import GaussianMLPPolicy
from sandbox.rocky.tf.envs.base import TfEnv
from rllab.misc.instrument import stub, run_experiment_lite
env = TfEnv(normalize(CartpoleEnv()))
policy = GaussianMLPPolicy(
name="policy",
env_spec=env.spec,
# The neural network policy should have two hidden layers, each with 32 hidden units.
hidden_sizes=(32, 32)
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=policy,
baseline=baseline,
batch_size=10000,
max_path_length=100,
n_itr=40,
discount=0.99,
optimizer_args=dict(
tf_optimizer_args=dict(
learning_rate=0.01,
)
)
)
algo.train()
