def test_variant_generator():
vg = instrument.VariantGenerator()
vg.add("key1", [1, 2, 3])
vg.add("key2", [True, False])
vg.add("key3", lambda key2: [1] if key2 else [1, 2])
it.assertEqual(len(vg.variants()), 9)
class VG(instrument.VariantGenerator):
@instrument.variant
def key1(self):
return [1, 2, 3]
@instrument.variant
def key2(self):
yield True
yield False
@instrument.variant
def key3(self, key2):
if key2:
yield 1
else:
yield 1
yield 2
it.assertEqual(len(VG().variants()), 9)
from rllab.misc import instrument
import sys
instrument.stub(globals())
env = normalize(GymEnv("Swimmer-v1", record_video=False))
policy = GaussianMLPPolicy(
env_spec=env.spec,
# The neural network policy should have two hidden layers, each with 42 hidden units.
hidden_sizes=(42, 42))
# baseline = LinearFeatureBaseline(env_spec=env.spec)
baseline = ZeroBaseline(env_spec=env.spec)
vg = instrument.VariantGenerator()
vg.add("seed", range(1))
variants = vg.variants()
for variant in variants:
algo = ERWR(
env=env,
policy=policy,
baseline=baseline,
batch_size=50000,
max_path_length=500,
n_itr=500,
discount=0.99,
plot=True,
)
