def all_models(cls, env):
choices = cls.param_choices(env)
if choices:
for prm in dict_product(choices):
yield cls(**prm)
else:
yield cls()
def generate_configs(output_dir,
ngen=int(1.0e+5),
pixDepth=1,
pixWidth=1.5,
npix=5000,
spectra_path=None,
jobs=1):
hist_dir = get_dir('../data/background_spectra')
hists = [
osp.join(hist_dir, item) for item in os.listdir(hist_dir)
if item.endswith('.root')
]
particle_names = [osp.basename(hist)[:-len('.root')] for hist in hists]
fluxes = [get_total_flux(hist) for hist in hists]
print("Total flux: %.3e" % np.sum(fluxes))
priors = [flux / np.sum(fluxes) for flux in fluxes]
print("Priors:\n %s" % '\n '.join([
"%s: %.3e" % (particle, prior)
for particle, prior in zip(particle_names, priors)
]))
assert len(hists), 'there is no data for cosmic background spectra!'
if spectra_path is None:
spectra_path = hist_dir
runtime_hists = [
osp.join(spectra_path, osp.basename(hist)) for hist in hists
]
configs = dict(
beamEnergy=-1,
particle_meta=[
dict(output=osp.join(
output_dir,
'%s_%04d' % (osp.basename(hist)[:-len('.root')], job)),
particle=particle_name,
energyHisto=hist,
job=job)
for particle_name, hist in zip(particle_names, runtime_hists)
for job in range(jobs)
],
ngen=ngen,
pixDepth=pixDepth,
pixWidth=pixWidth,
npix=npix,
)
return dict_product(configs)
def conditions(self):
conds = dict_product({
'creation_date': str(datetime.now()),
'clickDelay': 0,
'moveDelay': 500,
'encourage_planning': [False, True],
# 'timeLimit': 240,
# 'energy': 200,
# 'moveEnergy': 2,
# 'clickEnergy': 1,
'depth': 3,
'breadth': 2,
'inspectCost': 1,
'bonus_rate': 0.001
})
for c in conds:
if c['encourage_planning']:
c['mu'] = -4
c['sigma'] = 16
else:
c['mu'] = 8
c['sigma'] = 3
yield c
def trials(self, params):
for _ in range(params['n_trial']):
yield from dict_product(params)
BASE_CONFIG = json.load(f)
PARAMS = {
"game": ["Humanoid-v2"],
"mode": ["ppo"],
"clip_eps": [1e32],
"out_dir": ["results/ppo_noclip_humanoid/agents"],
"norm_rewards": ["returns"],
"initialization": ["xavier"],
"anneal_lr": [False],
"value_clipping": [True],
"entropy_coeff": [0.005],
"ppo_lr_adam": [2e-5] * 40,
"clip_grad_norm": [0.5],
"val_lr": [5e-5],
"lambda": [0.85],
"cpu": [True],
"advanced_logging": [True]
}
all_configs = [{**BASE_CONFIG, **p} for p in dict_product(PARAMS)]
if os.path.isdir("agent_configs/") or os.path.isdir("agents/"):
raise ValueError(
"Please delete the 'agent_configs/' and 'agents/' directories")
os.makedirs("agent_configs/")
os.makedirs("agents/")
for i, config in enumerate(all_configs):
with open(f"agent_configs/{i}.json", "w") as f:
json.dump(config, f)
# print(f'exact solution computed in {t.elapsed} seconds')
# agents['optimal'] = FunctionPolicy(pi)
for name, pol in agents.items():
np.random.seed(45)
df = pd.DataFrame(Agent(env, pol).run_many(5000, pbar=False))
df['name'] = name
df['n_city'] = n_city
df['max_sims'] = max_sims
df = df.set_index(['n_city', 'max_sims', 'name'])
df.to_csv(f'data/weather/sims/{n_city}_{max_sims}_{name}.pkl')
print(f'wrote {fn}')
from utils import dict_product
params = list(
dict_product({
'n_city': [10, 20, 30],
'max_sims': [
1,
2,
3,
4,
]
}))
print(f'running {len(params)} simulations')
jobs = [delayed(simulate)(**prm) for prm in params]
Parallel(min(len(jobs), 45))(jobs)