File size: 3,327 Bytes
5a42b66 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 | """The batched sampler must reproduce propagating each sample alone.
All J BALLAST samples share one fused scan over time (a ~20x win, since the
per-sample scans are latency-bound). Key derivation is arranged to match the
per-sample loop exactly, so this is a refactor, not a change of method -- and
that is checked here rather than assumed.
Agreement is to ~1e-14 rather than bit-exact: the batched step contracts
L_space @ Z @ L_Q as one einsum, so XLA associates the products differently from
the per-sample `L_space @ Z @ L_Q.T`. Trajectories are chaotic (a drifter's path
is an Euler integration through a sampled field), so this tolerance is asserted
on the *positions*, where it stays at 1e-14 over the horizon tested; the
validity masks must match exactly.
"""
import jax
import jax.numpy as jnp
import numpy as np
jax.config.update("jax_enable_x64", True)
from ballast.experiment import SYNTH_PARAMS as P
from ballast.gp import posterior_ext_state, sample_ext_state
from ballast.policies import sample_and_project
from ballast.spde import make_ops, propagate
from ballast.trajectory import Grid, advect
DT, OBS_EVERY, J = 0.01, 5, 3
def _setup():
grid = Grid(jnp.linspace(-2, 2, 7), jnp.linspace(-2, 2, 7))
ops = make_ops(grid.R, P, DT)
k = jax.random.PRNGKey(0)
S = jax.random.uniform(k, (9, 2), minval=-2, maxval=2)
t = jax.random.uniform(jax.random.fold_in(k, 1), (9,), minval=0, maxval=1.0)
y = jax.random.normal(jax.random.fold_in(k, 2), (9, 2))
mean, chol = posterior_ext_state(S, t, y, grid.R, 1.0, P, 0.1)
exist = jnp.array([[0.3, -0.2], [-1.0, 0.7]])
return grid, ops, mean, chol, exist
def test_batched_matches_per_sample_loop():
grid, ops, mean, chol, exist = _setup()
n_steps, t_m = 60, 1.0
key = jax.random.PRNGKey(7)
cpos, cval, epos, eval_, t_traj, raw = sample_and_project(
key, ops, mean, chol, exist, t_m, grid=grid, n_steps=n_steps,
obs_every=OBS_EVERY, n_samples=J, dt=DT,
)
# reference: the original one-sample-at-a-time route
keys = jax.random.split(key, J)
for j in range(J):
ka, kb = jax.random.split(keys[j])
X0 = sample_ext_state(ka, mean, chol, grid.n)
fields = propagate(X0, kb, ops, n_steps)
s0 = jnp.concatenate([grid.R, exist], axis=0)
pos_r, val_r, tidx_r = advect(
grid, fields, s0, jnp.ones(s0.shape[0], dtype=bool), DT, OBS_EVERY
)
np.testing.assert_allclose(raw[:, j], pos_r, rtol=1e-12, atol=1e-12)
np.testing.assert_array_equal(
np.concatenate([cval[:, j], eval_[:, j]], axis=1), val_r
)
np.testing.assert_allclose(t_traj, t_m + tidx_r * DT, rtol=1e-12)
snapped = grid.snap(pos_r)
np.testing.assert_allclose(cpos[:, j], snapped[:, : grid.n], rtol=0, atol=0) # snapped -> exact
np.testing.assert_allclose(epos[:, j], snapped[:, grid.n :], rtol=0, atol=0) # snapped -> exact
def test_grid_is_hashable_by_value():
"""Static jit args must hash by value, or every campaign recompiles."""
a = Grid(jnp.linspace(-2, 2, 5), jnp.linspace(-2, 2, 5))
b = Grid(jnp.linspace(-2, 2, 5), jnp.linspace(-2, 2, 5))
c = Grid(jnp.linspace(-2, 2, 6), jnp.linspace(-2, 2, 5))
assert hash(a) == hash(b) and a == b
assert hash(a) != hash(c) and a != c
|