def run_experiment(args):
# def env_fn():
# import flexibility # register flexibility to gym env registry
# return gym.make(args.env_name)
eg = ExperimentGrid(name=args.exp_name)
eg.add('seed', [10*i for i in range(args.num_runs)] if args.seed is None else args.seed)
eg.add('epochs', args.epochs)
eg.add('steps_per_epoch', args.steps_per_epoch)
eg.add('save_freq', args.save_freq)
eg.add('max_ep_len', 200)
eg.add('ac_kwargs:activation', eval(args.act), '')
eg.add('custom_h', args.custom_h)
eg.add('do_checkpoint_eval', args.do_checkpoint_eval)
eg.add('eval_episodes', args.eval_episodes)
eg.add('train_v_iters', args.train_v_iters)
eg.add('eval_temp', args.eval_temp)
eg.add('train_starting_temp', args.train_starting_temp)
eg.add('gamma', args.gamma)
eg.add('env_version', args.env_version)
eg.add('env_name', args.env_name)
eg.add('env_subtract_full_flex', args.env_subtract_full_flex)
eg.add('meta_learning', args.meta_learning)
eg.add('finetune', args.finetune)
eg.add('finetune_model_path', args.finetune_model_path)
eg.add('lam', args.lam)
eg.add('early_stop_epochs', args.early_stop_epochs)
eg.add('save_all_eval', args.save_all_eval)
if args.episodes_per_epoch is not None:
eg.add('episodes_per_epoch', args.episodes_per_epoch)
if args.env_version >= 3:
# args.file_path = "/home/user/git/spinningup/spinup/FlexibilityEnv/input_m8n12_cv0.8.pkl"
prefix = os.getcwd().split('RL_flex_design')[0]
args.file_path = prefix + "RL_flex_design/spinup/FlexibilityEnv_input/{}".format(args.env_input)
m, n, mean_c, mean_d, sd_d, profit_mat, target_arcs, fixed_costs, flex_0 = load_FlexibilityEnv_input(args.file_path)
eg.add('env_input', args.file_path)
eg.add('env_n_sample', args.env_n_sample)
if args.target_arcs is None:
eg.add('target_arcs', target_arcs)
else: # target_arcs is explicitly specified by the scripts, which overrides the target_arc from the input file
eg.add('target_arcs', args.target_arcs)
if args.algo == "ppo":
eg.add('train_pi_iters', args.train_pi_iters)
eg.run(ppo)
elif args.algo == "vpg":
eg.run(vpg)
def get_custom_env_fn(env_name,
env_version=None,
target_arcs=None,
env_input=None,
env_n_sample=None,
subtract_full_flexibility_performance=False,
meta_leraning=False):
if env_version in (1, 2):
# parse FlexibilityEnv settings from env_name
n_plant, n_product, target_arcs, n_sample = _parse_attributes(env_name)
class CustomFlexibilityEnv(FlexibilityEnv):
def __init__(self):
super().__init__(n_plant=n_plant,
n_product=n_product,
target_arcs=target_arcs,
n_sample=n_sample,
env_version=env_version)
print(
'using custom env: {} | n_plant: {} | n_product: {} | target_arcs: {} | n_sample: {} | env_version: {}'
.format(env_name, n_plant, n_product, target_arcs,
n_sample, env_version))
else: # env_version in (3, 4, >40, 5):
# load FlexibilityEnv settings from env_input
n_plant, n_product, mean_c, mean_d, sd_d, profit_mat, _, fixed_costs, flex_0 = load_FlexibilityEnv_input(
env_input)
def to_numpy_array(obj):
if isinstance(obj, list):
obj_array = np.asarray(obj, dtype=np.float32)
return obj_array
else:
return obj
mean_c = to_numpy_array(mean_c)
mean_d = to_numpy_array(mean_d)
sd_d = to_numpy_array(sd_d)
class CustomFlexibilityEnv(FlexibilityEnv):
def __init__(self):
super().__init__(
n_plant=n_plant,
n_product=n_product,
target_arcs=target_arcs,
# for env_version=3, target_arcs is passed into the function call
n_sample=env_n_sample,
capacity_mean=mean_c,
env_version=env_version,
demand_mean=mean_d,
demand_std=sd_d,
profit_matrix=profit_mat,
fixed_costs=fixed_costs,
starting_structure=flex_0,
subtract_full_flexibility_performance=
subtract_full_flexibility_performance,
meta_learning=meta_leraning)
print(
'using env: {} | n_plant: {} | n_product: {} | target_arcs: {} | n_sample: {} | env_version: {} '
'| subtract_full_flex: {} | meta_leraning: {}'.format(
env_name, n_plant, n_product, target_arcs,
env_n_sample, env_version,
subtract_full_flexibility_performance, meta_leraning))
return CustomFlexibilityEnv
def generate_scripts_for_multiple_target_arcs(experiment,
env_input,
env_version_list,
epoch_episodes,
num_tars_per_script,
num_batches,
num_runs,
gamma=None,
lam=None,
variance_reduction=False,
env_n_sample=50,
custom_h=None,
cpu=2,
tar_list=None,
early_stop=None,
epoch=800,
save_freq=10,
save_all_eval=None,
included_tars=None):
m, n, mean_c, mean_d, sd_d, profit_mat, target_arcs, fixed_costs, flex_0 = load_FlexibilityEnv_input(
_get_full_path(env_input))
print("number of existing arcs {}".format(flex_0.sum()))
if included_tars is not None:
# filter out tars that is not in included_tars
target_arcs = [tar for tar in target_arcs if tar in included_tars]
if tar_list is None or (tar_list is not None and len(tar_list) == 0):
tar_list = get_tars_list(num_tars_per_script, target_arcs)
print("target arcs to be run: {}".format(tar_list))
for batch in range(num_batches):
starting_seed = 100 * batch
for env_version in env_version_list:
# create entrypoint script
# !/bin/bash
path = 'run_{}_ENV{}_batch{}_entrypoint.sh'.format(
experiment, env_version, batch)
python_string = 'for((i=0;i < {};i++)); do bash run_{}_ENV{}_batch{}_'.format(len(tar_list),
experiment,
env_version,
batch) \
+ '$' + '{' + 'i' + '}' + '.sh & done'
with open(path, 'w') as f:
f.write('#!/bin/bash\n\n')
f.write(python_string)
make_executable(path)
print(python_string)
# create scripts to be called in parallel
for idx, target_arcs in enumerate(tar_list):
assert len(target_arcs) >= 1
if len(target_arcs) == 1:
# add 'tar' to exp_name explicitely
exp_name = 'F{}_CH{}_ENV{}_tar{}'.format(
experiment,
'1024-128' if custom_h is None else custom_h,
env_version, target_arcs[0])
else:
exp_name = 'F{}_CH{}_ENV{}'.format(
experiment,
'1024-128' if custom_h is None else custom_h,
env_version)
python_string = "python -m spinup.run_flexibility \\\n \
--algo ppo \\\n \
--env_name F{}-v{} \\\n \
--exp_name {} \\\n \
--cpu {} \\\n \
--epochs {} \\\n \
--custom_h 1024-128 \\\n \
--env_version {} \\\n \
--env_input {} \\\n \
--target_arcs {} \\\n \
--seed {} \\\n \
--save_freq {} \\\n \
--steps_per_epoch {} \\\n \
--do_checkpoint_eval \\\n".format(
experiment, env_version, exp_name, cpu, epoch, env_version,
env_input, _get_target_arcs_string(target_arcs),
_get_seed_str(starting_seed, num_runs), save_freq,
int(
np.ceil(
(int(statistics.mean(target_arcs)) - flex_0.sum())
* epoch_episodes)))
if variance_reduction:
python_string += ' --env_subtract_full_flex \\\n'
if env_n_sample != 50:
python_string += ' --env_n_sample {} \\\n'.format(
env_n_sample)
if custom_h is not None:
python_string += ' --custom_h {} \\\n'.format(
custom_h)
if early_stop is not None:
python_string += ' --early_stop {} \\\n'.format(
early_stop)
if save_all_eval is not None:
python_string += ' --save_all_eval \\\n'
if gamma is not None:
python_string += ' --gamma {} \\\n'.format(
gamma)
if lam is None:
python_string += ' ;'
else:
python_string += ' --lam {};'.format(
lam)
path = 'run_{}_ENV{}_batch{}_{}.sh'.format(
experiment, env_version, batch, idx)
with open(path, 'w') as f:
f.write('#!/bin/bash\n\n')
f.write(python_string)
make_executable(path)
print(python_string)
else:
exp = experiment
return exp
from spinup.FlexibilityEnv_input.FlexibilityEnv_input_files import INPUTS
if __name__ == "__main__":
experiment = '10x10a-lspe'
envs = ['ENV5']
print()
input_path = get_input_path(INPUTS[get_input_key(experiment)])
exclude = ['abcdef']
m, n, mean_c, mean_d, sd_d, profit_mat, target_arcs, fixed_costs, flex_0 = load_FlexibilityEnv_input(input_path)
perf_dicts = []
files_dicts = []
for env in envs:
print("==== processing files for {}".format(env))
files = collect_best_structures(experiment, env, exclude)
dict_tar_perf = {}
dict_tar_file = {}
files = sorted(files)
for file in files:
tar = file.split('T')[1].split('_SP')[0]
with open(file, 'rb') as f:
def generate_scripts_for_one_target_arcs(experiment,
env_input,
env_version_list,
epoch_episodes,
target_arcs,
num_runs,
starting_seed,
gamma=None,
lam=None,
variance_reduction=False,
env_n_sample=50,
early_stop=None,
cpu=8,
epoch=800,
save_freq=10,
save_all_eval=None,
custom_h=None,
meta_learning=False,
finetune=False,
finetune_path=None,
finetune_meta_trained_epoch=None):
m, n, mean_c, mean_d, sd_d, profit_mat, _, fixed_costs, flex_0 = load_FlexibilityEnv_input(
_get_full_path(env_input))
print("number of existing arcs {}".format(flex_0.sum()))
target_arcs_in_names = 0 if meta_learning else target_arcs
for env_version in env_version_list:
# create entrypoint script
# !/bin/bash
path = 'run_{}_ENV{}_tar{}_entrypoint.sh'.format(
experiment, env_version, target_arcs_in_names)
python_string = 'for((i=0;i < {};i++)); do bash run_{}_ENV{}_tar{}_'.format(num_runs,
experiment,
env_version,
target_arcs_in_names) \
+ '$' + '{' + 'i' + '}' + '.sh & done'
with open(path, 'w') as f:
f.write('#!/bin/bash\n\n')
f.write(python_string)
make_executable(path)
print(python_string)
# create scripts to be called in parallel
for idx in range(num_runs):
target_arcs = target_arcs if meta_learning is False else (
13 if idx % 2 == 1 else 22)
python_string = "python -m spinup.run_flexibility \\\n \
--algo ppo \\\n \
--env_name F{}-v{} \\\n \
--exp_name F{}_CH{}_ENV{}_tar{} \\\n \
--cpu {} \\\n \
--epochs {} \\\n \
--custom_h 1024-128 \\\n \
--env_version {} \\\n \
--env_input {} \\\n \
--target_arcs {} \\\n \
--seed {} \\\n \
--save_freq {} \\\n \
--steps_per_epoch {} \\\n \
--do_checkpoint_eval \\\n".format(
experiment, env_version, experiment,
'1024-128' if custom_h is None else custom_h, env_version,
target_arcs_in_names, cpu, epoch, env_version, env_input,
target_arcs if meta_learning is False else
(13 if idx % 2 == 1 else 22), starting_seed + 10 * idx,
save_freq,
int(np.ceil(
(target_arcs - flex_0.sum()) *
epoch_episodes)) if meta_learning is False else 20 *
epoch_episodes)
if variance_reduction:
python_string += ' --env_subtract_full_flex \\\n'
if env_n_sample != 50:
python_string += ' --env_n_sample {} \\\n'.format(
env_n_sample)
if early_stop is not None:
python_string += ' --early_stop {} \\\n'.format(
early_stop)
if save_all_eval is not None:
python_string += ' --save_all_eval \\\n'
if gamma is not None:
python_string += ' --gamma {} \\\n'.format(
gamma)
if custom_h is not None:
python_string += ' --custom_h {} \\\n'.format(
custom_h)
if meta_learning:
python_string += ' --meta_learning \\\n'
if finetune:
python_string += ' --finetune \\\n'
if finetune_meta_trained_epoch is None:
python_string += ' --finetune_model_path {} \\\n'.format(
'{}_s{}'.format(finetune_path,
starting_seed + 10 * idx))
else:
python_string += ' --finetune_model_path {} \\\n'.format(
'{}_s{}/simple_save{}'.format(
finetune_path, starting_seed + 10 * idx,
finetune_meta_trained_epoch))
if lam is None:
python_string += ' ;'
else:
python_string += ' --lam {};'.format(
lam)
path = 'run_{}_ENV{}_tar{}_{}.sh'.format(experiment, env_version,
target_arcs_in_names, idx)
with open(path, 'w') as f:
f.write('#!/bin/bash\n\n')
f.write(python_string)
make_executable(path)
print(python_string)