def main(args):
experiment_title = "demo_experiment"
affinity_code = quick_affinity_code(n_parallel=8)
# NOTE: you can also use encode_affinity to specifying how to distribute each
# experiment in your computing nodes.
# set up variants
variant_levels = list()
values = [
[
"one",
],
[
"two",
],
]
dir_names = ["{}".format(*v) for v in values]
keys = [("optionA", "choiceB")
] # each entry in the list is the string path to your config
variant_levels.append(VariantLevel(keys, values, dir_names))
values = [
["good", int(1e-3)],
["better", int(1e3)],
]
dir_names = ["{}".format(*v) for v in values]
keys = [("optionA", "choiceB"), ("optionB", )]
variant_levels.append(VariantLevel(keys, values, dir_names))
# get all variants and their own log directory
variants, log_dirs = make_variants(*variant_levels)
for i, variant in enumerate(variants):
variants[i] = update_config(default_config, variant)
run_experiments(
script="demo.py",
affinity_code=affinity_code,
experiment_title=experiment_title,
runs_per_setting=1, # how many times to run repeated experiments
variants=variants,
log_dirs=log_dirs,
debug_mode=args.
debug, # if greater than 0, the launcher will run one variant in this process)
)
# env_ids = ["Hopper-v2"] # , "Swimmer-v3"]
# env_ids = ["Ant-v3", "HalfCheetah-v3"]
env_ids = ["Walker2d-v3", "Hopper-v3"]
values = list(zip(env_ids))
dir_names = ["env_{}".format(*v) for v in values]
keys = [("env", "id")]
variant_levels_1M.append(VariantLevel(keys, values, dir_names))
env_ids = ["Ant-v3", "HalfCheetah-v3"]
values = list(zip(env_ids))
dir_names = ["env_{}".format(*v) for v in values]
keys = [("env", "id")]
variant_levels_3M.append(VariantLevel(keys, values, dir_names))
variants_3M, log_dirs_3M = make_variants(*variant_levels_3M)
variants_1M, log_dirs_1M = make_variants(*variant_levels_1M)
variants = variants_3M + variants_1M
log_dirs = log_dirs_3M + log_dirs_1M
run_experiments(
script=script,
affinity_code=affinity_code,
experiment_title=experiment_title,
runs_per_setting=runs_per_setting,
variants=variants,
log_dirs=log_dirs,
common_args=(default_config_key, ),
)
n_gpu=8,
contexts_per_gpu=2,
hyperthread_offset=24,
n_socket=2,
# cpu_per_run=2,
)
runs_per_setting = 5
variant_levels = list()
env_ids = ["Hopper-v3", "HalfCheetah-v3", "Walker2d-v3", "Ant-v3"]
values = list(zip(env_ids))
dir_names = ["env_{}".format(*v) for v in values]
keys = [("env", "id")]
variant_levels.append(VariantLevel(keys, values, dir_names))
variants, log_dirs = make_variants(*variant_levels)
# default_config_key = "ppo_1M_serial"
# script = "rlpyt/experiments/scripts/mujoco/pg/train/mujoco_ppo_serial.py"
# experiment_title = "ppo_mujoco"
# run_experiments(
# script=script,
# affinity_code=affinity_code,
# experiment_title=experiment_title,
# runs_per_setting=runs_per_setting,
# variants=variants,
# log_dirs=log_dirs,
# common_args=(default_config_key,),
# )
def main(args):
# Either manually set the resources for the experiment:
affinity_code = encode_affinity(
n_cpu_core=16,
n_gpu=1,
contexts_per_gpu=16,
# hyperthread_offset=8, # if auto-detect doesn't work, number of CPU cores
# n_socket=1, # if auto-detect doesn't work, can force (or force to 1)
cpu_per_run=1,
)
# Or try an automatic one, but results may vary:
# affinity_code = quick_affinity_code(n_parallel=None, use_gpu=True)
default_config = make_default_config()
# start building variants
variant_levels = list()
variant_choice = 3
############ experiments for eGreedy ############################
if variant_choice == 0:
values = [
[
"eGreedy",
0.1,
],
[
"eGreedy",
0.5,
],
[
"eGreedy",
0.9,
],
]
dir_names = ["eGreedy-e{}".format(v[1]) for v in values]
keys = [
("solution", ),
("agent_kwargs", "epsilon"),
] # each entry in the list is the string path to your config
variant_levels.append(VariantLevel(keys, values, dir_names))
############ experiments for UCB ################################
elif variant_choice == 1:
values = [
[
"ucb",
1,
],
[
"ucb",
5,
],
[
"ucb",
10,
],
]
dir_names = ["{}-c{}".format(*v) for v in values]
keys = [
("solution", ),
("agent_kwargs", "c"),
] # each entry in the list is the string path to your config
variant_levels.append(VariantLevel(keys, values, dir_names))
############ experiments for Thompson sampling ##################
elif variant_choice == 2:
values = [
[
"thompson",
[[1, 1], [1, 1], [1, 1]],
],
[
"thompson",
[[601, 401], [401, 601], [2, 3]],
],
]
dir_names = ["{}-prior{}".format(v[0], v[1][0][0]) for v in values]
keys = [
("solution", ),
("agent_kwargs", "prior"),
] # each entry in the list is the string path to your config
variant_levels.append(VariantLevel(keys, values, dir_names))
########## experiments for graident bandit ######################
elif variant_choice == 3:
values = [
[
"gradientBandit",
],
]
dir_names = ["{}".format(*v) for v in values]
keys = [
("solution", ),
]
variant_levels.append(VariantLevel(keys, values, dir_names))
values = [
[
0.2,
],
[
1.0,
],
[
2.0,
],
[
5.0,
],
]
dir_names = ["beta{}".format(*v) for v in values]
keys = [
("agent_kwargs", "beta"),
] # each entry in the list is the string path to your config
variant_levels.append(VariantLevel(keys, values, dir_names))
values = [
[
0.0,
],
[
0.8,
],
[
5.0,
],
[
20.0,
],
]
dir_names = ["b{}".format(*v) for v in values]
keys = [
("agent_kwargs", "b"),
] # each entry in the list is the string path to your config
variant_levels.append(VariantLevel(keys, values, dir_names))
######### Done setting hyper-parameters #########################
else:
raise ValueError("Wrong experiment choice {}".format(variant_choice))
# get all variants and their own log directory
variants, log_dirs = make_variants(*variant_levels)
for i, variant in enumerate(variants):
variants[i] = update_config(default_config, variant)
run_experiments(
script="girl/experiments/bandit/bandit.py",
affinity_code=affinity_code,
experiment_title="Bandit",
runs_per_setting=200,
variants=variants,
log_dirs=log_dirs, # the directory under "${experiment title}"
debug_mode=args.
debug, # if greater than 0, the launcher will run one variant in this process)
)