def main(config, gpu, docker, debug, exp_id, seed):
variant = default_config
if config:
with open(os.path.join(config)) as f:
exp_params = json.load(f)
variant = deep_update_dict(exp_params, variant)
variant['util_params']['gpu_id'] = gpu
# experiment(variant)
run_experiment_here(experiment, variant=variant, exp_id=exp_id, seed=seed)
def run_experiment_with_multiple_seeds(variant):
scores = []
for i in range(n_seeds):
variant['seed'] = str(int(variant['seed']) + i)
exp_prefix = variant['exp_prefix']
scores.append(
run_experiment_here(
get_ocm_score,
exp_prefix=exp_prefix,
variant=variant,
exp_id=i,
))
return np.mean(scores)
'SLURM_JOB_ID', None
)
if run_mode and run_mode == 'ec2':
try:
import urllib.request
instance_id = urllib.request.urlopen(
'http://169.254.169.254/latest/meta-data/instance-id'
).read().decode()
run_experiment_kwargs['variant']['EC2_instance_id'] = instance_id
except Exception as e:
print("Could not get instance ID. Error was...")
print(e)
if run_mode and (run_mode == 'ec2' or run_mode == 'gcp'):
# Do this in case base_log_dir was already set
run_experiment_kwargs['base_log_dir'] = output_dir
run_experiment_here(
method_call,
include_exp_prefix_sub_dir=False,
**run_experiment_kwargs
)
else:
# print("re kwargs")
# print(run_experiment_kwargs)
# print('run experiment from doodad / import mujoco')
# import mujoco_py
# print("import success")
run_experiment_here(
method_call,
log_dir=output_dir,
**run_experiment_kwargs
)
)
n_seeds = 1
mode = 'local'
exp_prefix = 'dev-{}'.format(
__file__.replace('/', '-').replace('_', '-').split('.')[0]
)
n_seeds = 3
mode = 'ec2'
exp_prefix = 'rlkit-skew-fit-pusher-reference-sample-from-true-prior-take2'
for exp_id, variant in enumerate(sweeper.iterate_hyperparameters()):
for _ in range(n_seeds):
run_experiment_here(
skewfit_full_experiment,
exp_prefix=exp_prefix,
mode=mode,
variant=variant,
use_gpu=True,
num_exps_per_instance=3,
gcp_kwargs=dict(
terminate=True,
zone='us-east1-c',
gpu_kwargs=dict(
gpu_model='nvidia-gtx-1080',
num_gpu=1,
)
)
)
vae_kwargs=dict(
decoder_distribution="gaussian_identity_variance",
input_channels=3,
architecture=imsize48_default_architecture,
),
algo_kwargs=dict(lr=1e-3),
save_period=1,
),
)
search_space = {}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space, default_parameters=variant
)
n_seeds = 1
mode = "local"
exp_prefix = "dev-{}".format(
__file__.replace("/", "-").replace("_", "-").split(".")[0]
)
for exp_id, variant in enumerate(sweeper.iterate_hyperparameters()):
for _ in range(n_seeds):
run_experiment_here(
skewfit_full_experiment,
exp_prefix=exp_prefix,
variant=variant,
use_gpu=False,
)
algorithm.train()
if __name__ == '__main__':
variant = dict(
algorithm="DQN", #
version="normal",
layer_size=256,
replay_buffer_size=int(1E6),
algorithm_kwargs=dict(
num_epochs=600,
num_eval_steps_per_epoch=1000,
num_trains_per_train_loop=200,
num_expl_steps_per_train_loop=1000,
min_num_steps_before_training=3000,
max_path_length=1000,
batch_size=1024,
),
trainer_kwargs=dict(
discount=0.99,
learning_rate=3E-4,
),
)
run_experiment_here(experiment,
variant=variant,
use_gpu=True,
exp_prefix='Test',
snapshot_mode='last',
snapshot_gap=5)
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space, default_parameters=variant,
)
n_seeds = 1
mode = 'local'
exp_prefix = 'dev-{}'.format(
__file__.replace('/', '-').replace('_', '-').split('.')[0]
)
# n_seeds = 3
# mode = 'gcp'
# exp_prefix = 'skew-fit-pickup-reference-post-refactor'
for exp_id, variant in enumerate(sweeper.iterate_hyperparameters()):
for _ in range(n_seeds):
run_experiment_here(
skewfit_full_experiment,
exp_prefix=exp_prefix,
mode=mode,
variant=variant,
use_gpu=True,
snapshot_gap=200,
snapshot_mode='gap_and_last',
num_exps_per_instance=3,
gcp_kwargs=dict(
zone='us-west1-b',
),
)
vae_kwargs=dict(
decoder_distribution='gaussian_identity_variance',
input_channels=3,
architecture=imsize48_default_architecture,
),
algo_kwargs=dict(lr=1e-3, ),
save_period=1,
),
)
search_space = {}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
n_seeds = 1
mode = 'local'
exp_prefix = 'dev-{}'.format(
__file__.replace('/', '-').replace('_', '-').split('.')[0])
for exp_id, variant in enumerate(sweeper.iterate_hyperparameters()):
for _ in range(n_seeds):
run_experiment_here(
skewfit_full_experiment,
exp_prefix=exp_prefix,
mode=mode,
variant=variant,
use_gpu=True,
)
def example(*_):
env = HalfCheetahEnv()
es = OUStrategy(env_spec=env.spec)
qf = FeedForwardCritic(
name_or_scope="critic",
env_spec=env.spec,
)
policy = FeedForwardPolicy(
name_or_scope="actor",
env_spec=env.spec,
)
algorithm = DDPG(
env,
es,
policy,
qf,
n_epochs=25,
epoch_length=1000,
batch_size=1024,
)
algorithm.train()
if __name__ == "__main__":
run_experiment_here(
example,
exp_prefix="3-5-profile-ddpg-half-cheetah",
seed=2,
)
max_path_length=max_path_length,
# qf_learning_rate=1e-1,
# policy_learning_rate=1e-1,
)
variant = dict(
H=H,
num_values=num_values,
exp_prefix=exp_prefix,
ddpg_params=ddpg_params,
)
for seed in range(n_seeds):
variant['seed'] = seed
variant['exp_id'] = exp_id
if USE_EC2:
run_experiment(
run_linear_ocm_exp,
exp_prefix=exp_prefix,
seed=seed,
mode="ec2",
variant=variant,
)
else:
run_experiment_here(
run_linear_ocm_exp,
exp_prefix=exp_prefix,
variant=variant,
exp_id=exp_id,
seed=seed,
)
instance_id = urllib.request.urlopen(
'http://169.254.169.254/latest/meta-data/instance-id'
).read().decode()
doodad_config.extra_launch_info[
'EC2_instance_id'] = instance_id
except Exception as e:
print("Could not get AWS instance ID. Error was...")
print(e)
if run_mode == 'gcp':
try:
import urllib.request
request = urllib.request.Request(
"http://metadata/computeMetadata/v1/instance/name", )
# See this URL for why we need this header:
# https://cloud.google.com/compute/docs/storing-retrieving-metadata
request.add_header("Metadata-Flavor", "Google")
instance_name = urllib.request.urlopen(request).read().decode()
doodad_config.extra_launch_info['GCP_instance_name'] = (
instance_name)
except Exception as e:
print("Could not get GCP instance name. Error was...")
print(e)
# Do this in case base_log_dir was already set
doodad_config = doodad_config._replace(base_log_dir=output_dir, )
run_experiment_here(
method_call,
doodad_config,
variant,
)
