def demo():
T = 256
n_env = 128
n_stabled = 8
worlds = MockWorlds.initial(n_env)
agent = MockAgent(0)
network = agent.network
league = League(agent, MockAgent, worlds.n_envs, n_stabled=n_stabled, stable_interval=8, verbose=False)
fielded = np.zeros((T, T+1))
stabled = np.zeros((T, T+1))
for t in range(T):
skills = agent(worlds)
new_worlds, transitions = worlds.step(skills)
league.update(agent, worlds.seats, transitions)
worlds = new_worlds
network.skill[()] = 10. if t == 6 else 0.
for n, s in zip(league.splitter.names, league.splitter.slices):
fielded[t, n] = s.stop - s.start
for n in league.stable.names:
stabled[t, n] = 1.
parts = dotdict.dotdict(
agent=agent,
league=league,
worlds=worlds)
trace = dotdict.dotdict(
fielded=fielded,
stabled=stabled)
return parts, trace
def state(self, e):
"""Returns a :class:`~rebar.dotdict.dotdict` tree representing the state of environment ``e``.
A typical state looks like this::
arrdict:
n_envs 1
n_agents 4
res 512
fov 60
agent_radius 0.10606601717798211
fps 10
scenery arrdict:
model Tensor((8, 2, 2), torch.float32)
lines Tensor((307, 2, 2), torch.float32)
lights Tensor((21, 3), torch.float32)
textures <megastepcuda.Ragged2D object at 0x7fba34112eb0>
baked <megastepcuda.Ragged1D object at 0x7fba34112670>
agents arrdict:
angles Tensor((4,), torch.float32)
positions Tensor((4, 2), torch.float32)
progress Tensor((4,), torch.float32)
This state tree is usually passed onto a :ref:`plotting` function."""
options = ('n_envs', 'n_agents', 'res', 'fov', 'agent_radius', 'fps')
options = {k: getattr(self, k) for k in options}
return arrdict.clone(dotdict.dotdict(
**options,
scenery=self.scenery.state(e),
agents=self.agents.state(e),
progress=self.progress[e]))
def geometry_data(regenerate=False):
# Why .npz.gz? Because applying gzip manually manages x10 better compression than
# np.savez_compressed. They use the same compression alg, so I assume the difference
# is in the default compression setting - which isn't accessible in np.savez_compressec.
p = Path('.cache/cubicasa-geometry.npz.gz')
if not p.exists() or regenerate:
p.parent.mkdir(exist_ok=True, parents=True)
if regenerate:
log.info('Regenerating geometry cache from SVG cache.')
with parallel.parallel(safe_geometry) as pool:
gs = pool.wait({
str(row.id): pool(row.id, row.svg)
for _, row in svg_data().iterrows()
})
gs = flatten({k: v for k, v in gs.items() if v is not None})
bs = BytesIO()
np.savez(bs, **gs)
p.write_bytes(gzip.compress(bs.getvalue()))
else:
#TODO: Shift this to Github
url = 'https://www.dropbox.com/s/3ohut8lvmr8lkwg/cubicasa-geometry.npz.gz?raw=1'
p.write_bytes(download(url))
# np.load is kinda slow.
raw = gzip.decompress(p.read_bytes())
with ZipFile(BytesIO(raw)) as zf:
flat = dotdict.dotdict(
{n[:-4]: fastload(zf.read(n))
for n in zf.namelist()})
return unflatten(flat)
def __init__(self, n_envs, *args, **kwargs):
geometries = cubicasa.sample(n_envs)
scenery = scene.scenery(geometries, 1)
self.core = core.Core(scenery, *args, res=4 * 64, fov=130, **kwargs)
self._rgb = modules.RGB(self.core, n_agents=1, subsample=4)
self._depth = modules.Depth(self.core, n_agents=1, subsample=4)
self._mover = modules.MomentumMovement(self.core)
self._imu = modules.IMU(self.core)
self._respawner = modules.RandomSpawns(geometries, self.core)
self.action_space = self._mover.space
self.obs_space = dotdict.dotdict(rgb=self._rgb.space,
d=self._depth.space,
imu=self._imu.space)
self._tex_to_env = self.core.scenery.lines.inverse[
self.core.scenery.textures.inverse.long()].long()
self._seen = torch.full_like(self._tex_to_env, False)
self._potential = self.core.env_full(0.)
self._lengths = torch.zeros(self.core.n_envs,
device=self.core.device,
dtype=torch.int)
self.device = self.core.device
def record(self, transitions, live, start, end):
#TODO: Figure out how to get scatter_add_ to work on vector-valued vals
wins = (transitions.rewards == 1).int()
scatter_add_(self.stats.wins[:, :, 0], live, wins[:, 0])
scatter_add_(self.stats.wins[:, :, 1], live, wins[:, 1])
scatter_add_(self.stats.moves, live, 1)
scatter_add_(self.stats.times, live,
(end - start) / transitions.terminal.size(0))
done = self.stats.wins.sum(-1) == self.tracker.n_envs_per
stats = self.stats[done].cpu()
results = []
for idx in range(stats.indices.size(0)):
item = stats[idx]
names = tuple(self.tracker.names[i] for i in item.indices)
results.append(
dotdict.dotdict(names=names,
wins=tuple(map(float, item.wins)),
moves=float(item.moves),
games=float(sum(item.wins)),
times=float(item.times),
boardsize=self.worlds.boardsize))
self.stats.wins[done] = -1
return results
def initial_stats(n_jobs):
return dotdict.dotdict(finished=0,
total=n_jobs,
moves=0,
games=0,
matchups=0,
start=time.time(),
end=time.time() + 1)
def to(self, state, action=0, reward=0., weight=1.):
action = int(action)
self._builder._trans.append(
dotdict.dotdict(prev=self._name,
action=action,
next=state,
reward=reward,
weight=weight))
return self
def gather(wins, moves, matchup_idxs, names, boardsize):
names = np.array(names)
n_envs, n_seats = matchup_idxs.shape
results = []
for p in matchup_patterns(n_seats):
ws = wins[(matchup_idxs == p).all(-1)].sum(0)
ms = moves[(matchup_idxs == p).all(-1)].sum(0)
results.append(
dotdict.dotdict(names=tuple(names[p]),
wins=tuple(map(float, ws)),
moves=float(ms[0]),
games=float(ws.sum()),
boardsize=boardsize))
return results
def build(self):
states = ({x.state
for x in self._obs} | {x.prev
for x in self._trans}
| {x.next
for x in self._trans})
actions = {x.action for x in self._trans}
assert max(actions) == len(actions) - 1, 'Action set isn\'t contiguous'
indices = {s: i for i, s in enumerate(states)}
names = np.array(list(states))
n_states = len(states)
n_actions = len(actions)
(d_obs, ) = {len(x.obs) for x in self._obs}
obs = torch.full((n_states, d_obs), np.nan)
start = torch.full((n_states, ), 0.)
for x in self._obs:
obs[indices[x.state]] = torch.as_tensor(x.obs)
start[indices[x.state]] = x.start
trans = torch.full((n_states, n_actions, n_states), 0.)
reward = torch.full((n_states, n_actions), 0.)
for x in self._trans:
trans[indices[x.prev], x.action, indices[x.next]] = x.weight
reward[indices[x.prev], x.action] = x.reward
terminal = (trans.sum(-1).max(-1).values == 0)
assert start.sum() > 0, 'No start state declared'
return dotdict.dotdict(obs=obs,
trans=trans,
reward=reward,
terminal=terminal,
start=start,
indices=indices,
names=names,
n_states=n_states,
n_actions=n_actions,
d_obs=d_obs)
def __init__(self, n_envs, n_agents, *args, **kwargs):
geometries = cubicasa.sample(max(n_envs // 4, 1))
scenery = scene.scenery(geometries, n_agents)
self.core = core.Core(scenery, *args, res=4 * 128, fov=70, **kwargs)
self._rgb = modules.RGB(self.core, n_agents=1, subsample=4)
self._depth = modules.Depth(self.core, n_agents=1, subsample=4)
self._imu = modules.IMU(self.core, n_agents=1)
self._movement = modules.MomentumMovement(self.core, n_agents=1)
self._spawner = modules.RandomSpawns(geometries, self.core)
self.action_space = self._movement.space
self.obs_space = dotdict.dotdict(rgb=self._rgb.space,
d=self._depth.space,
imu=self._imu.space,
health=spaces.MultiVector(1, 1))
self._bounds = arrdict.torchify(
np.stack([g.masks.shape * g.res
for g in geometries])).to(self.core.device)
self._health = self.core.agent_full(np.nan)
self._damage = self.core.agent_full(np.nan)
self.n_envs = self.core.n_envs * self.core.n_agents
self.device = self.core.device
def auxinfo(f, K=1001, S=50):
μ = np.linspace(*μ_lims, K)
σ2 = np.logspace(*σ2_lims, K, base=10)
zs, ws = np.polynomial.hermite_e.hermegauss(S)
ds = (μ[:, None, None] + zs[None, None, :] * σ2[None, :, None]**.5)
scale = 1 / (2 * np.pi)**.5
fs = scale * (f(ds) * ws).sum(-1)
f = sp.interpolate.RectBivariateSpline(μ, σ2, fs, kx=1, ky=1)
dμs = (fs[2:, :] - fs[:-2, :]) / (μ[2:] - μ[:-2])[:, None]
dμ = sp.interpolate.RectBivariateSpline(μ[1:-1], σ2, dμs, kx=2, ky=2)
dσ2s = (fs[:, 2:] - fs[:, :-2]) / (σ2[2:] - σ2[:-2])[None, :]
dσ2 = sp.interpolate.RectBivariateSpline(μ, σ2[1:-1], dσ2s, kx=2, ky=2)
return dotdict.dotdict(μ=μ,
σ2=σ2,
f=f,
dμ=dμ,
dσ2=dσ2,
fs=fs,
dμs=dμs,
dσs=dσ2s)
def state(self, e=0):
return dotdict.dotdict(
core=self.core.state(e),
rgb=self.rgb.state(e))
def pandas(soln, names):
return dotdict.dotdict(μ=pd.Series(soln.μ, names),
Σ=pd.DataFrame(soln.Σ, names, names))
def state(self, name, obs, start=0.):
if isinstance(obs, (int, float, bool)):
obs = (obs, )
self._obs.append(dotdict.dotdict(state=name, obs=obs, start=start))
return State(name, self)
