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()
def opt_vpg(env,
baseline,
policy,
learning_rate=1e-5,
batch_size=4000,
**kwargs):
# no idea what batch size, learning rate, etc. should be
optimiser = FirstOrderOptimizer(
tf_optimizer_cls=tf.train.AdamOptimizer,
tf_optimizer_args=dict(learning_rate=learning_rate),
# batch_size actually gets passed to BatchPolopt (parent of VPG)
# instead of TF optimiser (makes sense, I guess)
batch_size=None,
max_epochs=1)
return VPG(env=env,
policy=policy,
baseline=baseline,
n_itr=int(1e9),
optimizer=optimiser,
batch_size=batch_size,
**kwargs)
policy = GaussianMLPPolicy( # random policy
name='policy',
env_spec=env.spec,
hidden_sizes=(100, 100),
)
if initial_params_file is not None:
policy = None
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,
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 get_algo(env, policy, es, qf, baseline, max_path_length,
batch_size, replay_pool_size, discount,
scale_reward, learning_rate, replacement_prob,
policy_updates_ratio,
step_size, gae_lambda,
sample_backups,
qprop_min_itr,
qf_updates_ratio,
qprop_use_qf_baseline,
qprop_eta_option,
algo_name,
qf_learning_rate,
n_itr,
**kwargs):
algo = None
min_pool_size = 1000
qf_batch_size = 64
qf_baseline = None
print('Creating algo=%s with n_itr=%d, max_path_length=%d...'%(
algo_name, n_itr, max_path_length))
if algo_name in ['ddpg']:
algo = DDPG(
env=env,
policy=policy,
es=es,
qf=qf,
batch_size=qf_batch_size,
max_path_length=max_path_length,
epoch_length=batch_size, # make comparable to batchopt methods
min_pool_size=min_pool_size,
replay_pool_size=replay_pool_size,
n_epochs=n_itr,
discount=discount,
scale_reward=scale_reward,
qf_learning_rate=qf_learning_rate,
policy_learning_rate=learning_rate,
replacement_prob=replacement_prob,
policy_updates_ratio=policy_updates_ratio,
# Uncomment both lines (this and the plot parameter below) to enable plotting
# plot=True,
)
elif algo_name in ['trpo', 'qprop']:
if qf is not None:
qf_baseline = QfunctionBaseline(env_spec=env.spec,
policy=policy, qf=qf)
algo = TRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=batch_size,
max_path_length=max_path_length,
n_itr=n_itr,
discount=discount,
step_size=step_size,
gae_lambda=gae_lambda,
# Uncomment both lines (this and the plot parameter below) to enable plotting
# plot=True,
sample_backups=sample_backups,
qf=qf,
qf_batch_size=qf_batch_size,
min_pool_size=min_pool_size,
scale_reward=scale_reward,
qprop_min_itr=qprop_min_itr,
qf_updates_ratio=qf_updates_ratio,
qprop_eta_option=qprop_eta_option,
replay_pool_size=replay_pool_size,
replacement_prob=replacement_prob,
qf_baseline=qf_baseline,
qf_learning_rate=qf_learning_rate,
qprop_use_qf_baseline=qprop_use_qf_baseline,
)
elif algo_name in ['vpg', 'qvpg']:
if qf is not None:
qf_baseline = QfunctionBaseline(env_spec=env.spec,
policy=policy, qf=qf)
algo = VPG(
env=env,
policy=policy,
baseline=baseline,
batch_size=batch_size,
max_path_length=max_path_length,
n_itr=n_itr,
discount=discount,
gae_lambda=gae_lambda,
optimizer_args=dict(
tf_optimizer_args=dict(
learning_rate=learning_rate,
)
),
qf=qf,
qf_batch_size=qf_batch_size,
min_pool_size=min_pool_size,
scale_reward=scale_reward,
qprop_min_itr=qprop_min_itr,
qf_updates_ratio=qf_updates_ratio,
qprop_eta_option=qprop_eta_option,
replay_pool_size=replay_pool_size,
qf_baseline=qf_baseline,
qf_learning_rate=qf_learning_rate,
qprop_use_qf_baseline=qprop_use_qf_baseline,
)
return algo
if initial_params_file is not None:
policy = None
make_video1 = True if goalnum in [0, 1, 2] else False
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=policy,
load_policy=initial_params_file,
baseline=baseline,
batch_size=2000,
max_path_length=100,
n_itr=n_itr,
#step_size=10.0,
sampler_cls=VectorizedSampler, # added by RK 6/19
sampler_args=dict(n_envs=1),
reset_arg=goal,
optimizer=None,
optimizer_args={
'init_learning_rate': step_size,
'tf_optimizer_args': {
'learning_rate': 0.5 * step_size
},
'tf_optimizer_cls': tf.train.GradientDescentOptimizer
},
# extra_input="onehot_exploration", # added by RK 6/19
# extra_input_dim=5, # added by RK 6/19
make_video=make_video1)
exp_name = 'mamlil' + desc + str(run_id) + "_n_itr" + str(
n_itr) + "_goal" + str(goalnum)
run_experiment_lite(
for goal in goals:
goal = list(goal)
env = TfEnv(normalize(RandomBanditEnv(k=args.k, n=args.n, goal=goal)))
baseline = LinearFeatureBaseline(env_spec=env.spec)
if not args.path:
raise Exception("Please enter a valid path for the parameter file")
algo = VPG(env=env,
policy=None,
load_policy=args.path,
baseline=baseline,
batch_size=args.n * args.num_goals,
max_path_length=args.n,
n_itr=n_itr,
optimizer_args={
'init_learning_rate': args.step_size,
'tf_optimizer_args': {
'learning_rate': 0.5 * args.step_size
},
'tf_optimizer_cls': tf.train.GradientDescentOptimizer
})
run_experiment_lite(
algo.train(),
# Number of parallel workers for sampling
n_parallel=4,
snapshot_mode="last",
seed=4,
exp_prefix='trpobandit_test',
exp_name='test',
def get_algo(env, policy, es, qf, baseline, max_path_length, batch_size,
replay_pool_size, discount, scale_reward, learning_rate,
replacement_prob, policy_updates_ratio, step_size, gae_lambda,
sample_backups, kl_sample_backups, qprop_eta_option, qprop_unbias,
qprop_nu, algo_name, n_itr, recurrent, updates_ratio,
policy_use_target, policy_batch_size, policy_sample_last,
ac_delta, ac_sample_backups, save_freq, restore_auto,
qf_learning_rate, qf_use_target, qf_mc_ratio, qf_batch_size,
qf_residual_phi, **kwargs):
algo = None
algo_class = None
min_pool_size = 1000
qf_baseline = None
extra_kwargs = dict()
print('Creating algo=%s with n_itr=%d, max_path_length=%d...' %
(algo_name, n_itr, max_path_length))
if algo_name in [
'ddpg',
'dspg',
'dspgoff',
'dqn',
'dsqn',
'trpg',
'trpgoff',
]:
if algo_name in [
'trpg',
]:
extra_kwargs['policy_update_method'] = 'cg'
algo = DDPG(
env=env,
policy=policy,
policy_use_target=policy_use_target,
es=es,
qf=qf,
qf_use_target=qf_use_target,
policy_batch_size=policy_batch_size,
qf_batch_size=qf_batch_size,
qf_mc_ratio=qf_mc_ratio,
qf_residual_phi=qf_residual_phi,
max_path_length=max_path_length,
epoch_length=batch_size, # make comparable to batchopt methods
min_pool_size=min_pool_size,
replay_pool_size=replay_pool_size,
n_epochs=n_itr,
discount=discount,
scale_reward=scale_reward,
qf_learning_rate=qf_learning_rate,
policy_learning_rate=learning_rate,
policy_step_size=step_size,
policy_sample_last=policy_sample_last,
replacement_prob=replacement_prob,
policy_updates_ratio=policy_updates_ratio,
updates_ratio=updates_ratio,
save_freq=save_freq,
restore_auto=restore_auto,
**extra_kwargs,
)
algo_class = 'DDPG'
elif algo_name in [
'trpo',
'nuqprop',
'nuqfqprop',
'actrpo',
'acqftrpo',
'qprop',
'mqprop',
'qfqprop',
'nafqprop',
]:
if recurrent:
extra_kwargs['optimizer'] = \
ConjugateGradientOptimizer(hvp_approach=FiniteDifferenceHvp(base_eps=1e-5))
if algo_name in [
'actrpo',
'acqftrpo',
]:
extra_kwargs['ac_delta'] = ac_delta
extra_kwargs['qprop'] = False # disable qprop
if ac_delta == 0: qf = None
if algo_name in [
'mqprop',
]:
extra_kwargs['mqprop'] = True
if algo_name in [
'nuqprop',
'nuqfqprop',
]:
extra_kwargs['qprop_nu'] = qprop_nu
if qf is not None:
qf_baseline = QfunctionBaseline(env_spec=env.spec,
policy=policy,
qf=qf)
algo = TRPO(env=env,
policy=policy,
baseline=baseline,
batch_size=batch_size,
max_path_length=max_path_length,
n_itr=n_itr,
discount=discount,
step_size=step_size,
gae_lambda=gae_lambda,
sample_backups=sample_backups,
kl_sample_backups=kl_sample_backups,
qf=qf,
qf_use_target=qf_use_target,
qf_batch_size=qf_batch_size,
qf_mc_ratio=qf_mc_ratio,
qf_residual_phi=qf_residual_phi,
min_pool_size=min_pool_size,
scale_reward=scale_reward,
qf_updates_ratio=updates_ratio,
qprop_eta_option=qprop_eta_option,
qprop_unbias=qprop_unbias,
replay_pool_size=replay_pool_size,
replacement_prob=replacement_prob,
qf_baseline=qf_baseline,
qf_learning_rate=qf_learning_rate,
ac_sample_backups=ac_sample_backups,
policy_sample_last=policy_sample_last,
save_freq=save_freq,
restore_auto=restore_auto,
**extra_kwargs)
algo_class = 'TRPO'
elif algo_name in [
'vpg',
'qvpg',
]:
if qf is not None:
qf_baseline = QfunctionBaseline(env_spec=env.spec,
policy=policy,
qf=qf)
algo = VPG(
env=env,
policy=policy,
baseline=baseline,
batch_size=batch_size,
max_path_length=max_path_length,
n_itr=n_itr,
discount=discount,
gae_lambda=gae_lambda,
optimizer_args=dict(
tf_optimizer_args=dict(learning_rate=learning_rate, )),
qf=qf,
qf_use_target=qf_use_target,
qf_batch_size=qf_batch_size,
qf_mc_ratio=qf_mc_ratio,
qf_residual_phi=qf_residual_phi,
min_pool_size=min_pool_size,
scale_reward=scale_reward,
qf_updates_ratio=updates_ratio,
qprop_eta_option=qprop_eta_option,
qprop_unbias=qprop_unbias,
replay_pool_size=replay_pool_size,
qf_baseline=qf_baseline,
qf_learning_rate=qf_learning_rate,
save_freq=save_freq,
restore_auto=restore_auto,
)
algo_class = 'VPG'
print('[get_algo] Instantiating %s.' % algo_class)
return algo
for goal in range(10, 20):
stub(globals())
env = TfEnv(SawyerPickEnv(goal_idx=goal))
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=100,
n_itr=10,
optimizer_args={
'init_learning_rate': 0.1,
'tf_optimizer_args': {
'learning_rate': 0.01
},
'tf_optimizer_cls': tf.train.GradientDescentOptimizer
},
reset_arg=goal,
)
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="all",
from sandbox.rocky.tf.envs.base import TfEnv
stub(globals())
#env = TfEnv(normalize(PointEnv()))
env = TfEnv(normalize(PointEnvRandGoal()))
policy = GaussianMLPPolicy(
name="policy",
env_spec=env.spec,
)
#baseline = LinearFeatureBaseline(env_spec=env.spec)
baseline = ZeroBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=policy,
baseline=baseline,
#batch_size=20,
max_path_length=5,
n_itr=100,
#plot=True,
)
run_experiment_lite(
algo.train(),
n_parallel=1,
snapshot_mode="last",
seed=1,
exp_prefix='deleteme',
exp_name='deleteme',
#plot=True,
)
from examples.point_env_randgoal import PointEnvRandGoal
from rllab.envs.normalized_env import normalize
from rllab.misc.instrument import stub, run_experiment_lite
#from rllab.policies.gaussian_mlp_policy import GaussianMLPPolicy
#from sandbox.rocky.tf.policies.gaussian_mlp_policy import GaussianMLPPolicy
from sandbox.rocky.tf.policies.minimal_gauss_mlp_policy import GaussianMLPPolicy
from sandbox.rocky.tf.envs.base import TfEnv
stub(globals())
env = TfEnv(normalize(PointEnv()))
#env = TfEnv(normalize(PointEnvRandGoal()))
policy = GaussianMLPPolicy(
name="policy",
env_spec=env.spec,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=policy,
baseline=baseline,
#plot=True,
)
run_experiment_lite(
algo.train(),
n_parallel=1,
snapshot_mode="last",
seed=1,
#plot=True,
)
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()
batch_size = 200
else:
env = TfEnv(normalize(SwimmerEnv()))
batch_size = 20
policy = GaussianMLPPolicy(
name="policy",
env_spec=env.spec,
hidden_sizes=(100,100),
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
#baseline = ZeroBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=policy,
baseline=baseline,
batch_size=500*batch_size,
max_path_length=500,
n_itr=500,
#plot=True,
optimizer_args={'tf_optimizer_args':{'learning_rate': 1e-3}},
)
run_experiment_lite(
algo.train(),
n_parallel=1, # try increasing this to make it faster??? (Maybe need to modify code for this)
snapshot_mode="last",
seed=1,
exp_prefix='vpgswimmer',
#exp_name='basic',
exp_name='randomenv',
#plot=True,
)