def run_variant(experiment, variant):
lu.run_experiment(
experiment,
variant=variant,
# run_id=variant["run_id"],
mode=variant["mode"],
exp_prefix=variant["exp_prefix"],
exp_id=variant["exp_id"],
instance_type=variant["instance_type"],
use_gpu=variant["use_gpu"],
snapshot_mode=variant["snapshot_mode"],
snapshot_gap=variant["snapshot_gap"],
base_log_dir=variant["base_log_dir"],
num_exps_per_instance=variant.get("num_exps_per_instance", 1),
prepend_date_to_exp_prefix=False,
# spot_price=variant["spot_price"],
region=variant.get("region", "us-east-1"),
time_in_mins=variant.get("time_in_mins", 0),
ssh_host=variant.get("ssh_host", None),
)
# 5. Execute bash command: CUDA_VISIBLE_DEVICES=? python realworld_analyzer.py -da [cloth,poke,solid]
if __name__ == "__main__":
parser = ArgumentParser()
parser.add_argument('-da',
'--dataset',
type=str,
default=None,
required=True) # stack_o2p2_60k, pickplace_1env_1k
parser.add_argument('-m', '--mode', type=str, default='here_no_doodad')
args = parser.parse_args()
# models: list of dictionaries containing the information of the models to be loaded/run
# Images will be of form: T images wide, 1st row is true images, then next rows are dictated by models, where
# each model takes K+1 rows, where the 1st row is the combined reconstruction and the next K are the subimages
variant = dict(
models=dataset_to_models[args.dataset],
T=8, # Set this!
dataset=args.dataset)
# Relevant options: 'here_no_doodad', 'local_docker', 'ec2'
run_experiment(
create_images_from_dataset, # Set this! get_mse_from_dataset, create_images_from_dataset, analyze_mse, create_mse_graphs
exp_prefix='images-{}'.format(args.dataset),
mode=args.mode,
variant=variant,
use_gpu=True, # Turn on if you have a GPU
seed=None,
region='us-west-2')
parser.add_argument('-de', '--debug', type=int, default=1)
parser.add_argument('-m', '--mode', type=str, default='here_no_doodad')
args = parser.parse_args()
if args.variant == 'stack':
variant = stack_variant
elif args.variant == 'pickplace':
variant = pickplace_variant
elif args.variant == 'cloth':
variant = cloth_variant
else:
raise Exception("Exp variant not found")
variant['debug'] = args.debug
# Relevant options: 'here_no_doodad', 'local_docker', 'ec2'
run_experiment(
train_vae,
exp_prefix='{}'.format(args.variant),
mode=args.mode,
variant=variant,
use_gpu=True, # Turn on if you have a GPU
seed=None,
region='us-west-2' # only used if mode is ec2
)
op3_args=params_to_info[args.model_file]["op3_args"],
cem_args=dict(
cem_steps=5,
num_samples=1000,
time_horizon=1,
),
cost_args=dict(
core_type='final_recon', # "subimage", "final_recon", "latent"
# latent_or_subimage = 'subimage',
compare_func='mse',
post_process='raw',
aggregate='min',
),
filter_goal_image=False,
structure=s,
debug=args.debug,
model_file=args.model_file,
)
n_seeds = 1
exp_prefix = 'iodine-mpc-stage1-k7-v2-unfactorized-fixed'
run_experiment(
main,
exp_prefix=exp_prefix,
mode=args.mode,
variant=variant,
use_gpu=True, # Turn on if you have a GPU
region='us-west-2',
)
cem_args=dict(
cem_steps=5,
num_samples=200,
time_horizon=num_goal_objects,
action_type=
4, # Should be [None, 4], This controls if it is object oriented (when set to 4) or raw env action space (None)
),
num_actions_to_take=num_goal_objects * 2,
num_action_to_take_per_plan=num_goal_objects,
cost_args=dict(
core_type='final_recon', # "subimage", "final_recon", "latent"
compare_func='mse', # mse, psuedo_intersect
post_process='raw',
aggregate='sum',
),
goal_start_end_range=[0, 100],
debug=args.debug,
model_file=args.model_file,
number_goal_objects=num_goal_objects,
)
run_experiment(
main,
exp_prefix='iodine-mpc-stage3-n{}-{}-v2'.format(
variant["number_goal_objects"], args.model_file),
mode=args.mode,
variant=variant,
use_gpu=True, # Turn on if you have a GPU
region='us-west-2',
)