def nested_sample(param_dist, n_samples=1):
""" Generate all samples from a distribution.
Distribution must be specified as a dictionary mapping
from names to either a list of possible values or a distribution
(i.e. has a method `rvs`).
"""
assert isinstance(param_dist, dict)
config = Config(param_dist)
flat = config.flatten()
other = {}
samples = []
sampled_keys = []
for k in sorted(flat.keys()):
v = flat[k]
try:
samples.append(list(np.random.choice(list(v), size=n_samples)))
except (TypeError, ValueError):
if hasattr(v, 'rvs'):
samples.append(v.rvs(n_samples))
sampled_keys.append(k)
else:
other[k] = v
else:
sampled_keys.append(k)
samples = sorted(zip(*samples))
configs = []
for sample in samples:
new = Config(deepcopy(other.copy()))
for k, s in zip(sampled_keys, sample):
new[k] = s
configs.append(type(param_dist)(new))
return configs
def generate_all(param_dist):
""" Generate all samples from a parameter distribution.
Distribution must be specified as a dictionary mapping
from names to lists of possible values.
"""
assert isinstance(param_dist, dict)
config = Config(param_dist)
flat = config.flatten()
other = {}
sampled_keys = []
lists = []
for k in sorted(flat.keys()):
v = flat[k]
try:
lists.append(list(v))
except (TypeError, ValueError):
if hasattr(v, 'rvs'):
raise Exception(
"Attempting to generate all samples, but element {} "
"with key {} is a continuous distribution.".format(v, k))
other[k] = v
else:
sampled_keys.append(k)
param_sets = sorted(product(*lists))
configs = []
for pset in param_sets:
new = Config(deepcopy(other.copy()))
for k, p in zip(sampled_keys, pset):
new[k] = p
configs.append(type(param_dist)(new))
return configs