def act(self, ac: Any) -> None:
self._firsts[0] = False
state = self._q.popleft()
rews = []
def add_reward(subspace, substate, subval):
if isinstance(subspace.eltype, types.Discrete):
r = 1 if (substate == subval).all() else 0
elif isinstance(subspace.eltype, types.Real):
diff = subval - substate
diff = diff[:]
r = -0.5 * np.dot(diff, diff)
else:
raise Exception(
f"unrecognized action space eltype {subspace.eltype}")
rews.append(r)
types.multimap(add_reward, self.ac_space, state, ac)
rew = sum(rews) / len(rews)
if self._step < self._delay_steps:
# don't give any reward for guessing un-observed states
rew = 0
self._rews[0] = rew
self._q.append(
types_np.sample(self.ac_space, bshape=(self.num, ), rng=self._rng))
self._step += 1
if self._step >= self._episode_len:
self._reset()
def _reset(self) -> None:
self._q.clear()
for _ in range(self._delay_steps + 1):
self._q.append(
types_np.sample(self.ac_space,
bshape=(self.num, ),
rng=self._rng))
self._step = 0
self._firsts[0] = True
def main():
import procgen
env = procgen.ProcgenGym3Env(num=1, env_name="coinrun", render_human=True)
env = ViewerWrapper(env=env, info_key="rgb")
start = time.time()
for i in range(10000):
env.act(types_np.sample(env.ac_space, bshape=(env.num, )))
print("step", i, i / (time.time() - start))
def test_works(make_env):
"""
Make sure the environment works at all and that we can instantiate multiple copies
"""
envs = []
for _ in range(3):
env = make_env()
envs.append(env)
for _ in range(10):
ac = types_np.sample(env.ac_space, bshape=(env.num,))
env.act(ac)
def test_fast_env():
num_env = 2
num_steps = 10000
episode_len = 100
start = time.time()
env = TimingEnv(num=num_env, episode_len=episode_len)
episode_count = 0
expected_episode_count = num_env * num_steps / episode_len
for i in range(num_steps):
env.act(types_np.sample(env.ac_space, bshape=(env.num, )))
_rew, _obs, first = env.observe()
for f in first:
if f:
episode_count += 1
if i == num_steps - 2:
assert episode_count == expected_episode_count - num_env
elapsed = time.time() - start
assert elapsed / num_steps < 1e-3
assert episode_count == expected_episode_count
def gym3_rollout(e):
for _ in range(10):
rew, ob, done = e.observe()
print(multimap(lambda x: x.shape, ob), rew.shape, done.shape)
e.act(sample(e.ac_space, (e.num, )))
"""
Example random agent script using the gym3 API to demonstrate that procgen works
"""
from gym3 import types_np
from procgen import ProcgenGym3Env
env = ProcgenGym3Env(num=1, env_name="coinrun")
step = 0
while True:
env.act(types_np.sample(env.ac_space, bshape=(env.num, )))
rew, obs, first = env.observe()
print(f"step {step} reward {rew} first {first}")
if step > 0 and first:
break
step += 1