File size: 18,028 Bytes
aecbd27 | 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 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 | """Engine tests on tiny networks: exact dynamics, Brian2 step order, and equality with a
naive implementation written independently of the kernel."""
import math
import numpy as np
import pytest
from flysim.engine import (
CpuEngine,
InputSchedule,
ModelParams,
SynapseMatrix,
poisson_schedule,
trial_rng,
)
from flysim.engine.inputs import merge
P = ModelParams()
# First draw of trial 0 of one condition, pinned so a change to the seeding scheme
# cannot pass silently.
PINNED_DRAW = 1489115183
def _matrix(edges: list[tuple[int, int, int]], n: int) -> SynapseMatrix:
pre, post, w = zip(*edges, strict=True) if edges else ((), (), ())
return SynapseMatrix.from_edges(np.array(pre), np.array(post), np.array(w), n)
def _inputs(at: dict[int, list[int]]) -> InputSchedule:
"""{neuron: [steps]} -> schedule."""
pairs = sorted((s, n) for n, steps in at.items() for s in steps)
return InputSchedule(
targets=np.array(sorted(at), dtype=np.int32),
steps=np.array([s for s, _ in pairs], dtype=np.int64),
neurons=np.array([n for _, n in pairs], dtype=np.int32),
)
KERNELS = ("plain", "skip_rest", "vec")
def _run(matrix, schedule, steps, dtype=np.float32, params=P, skip_rest=False, kernel=None):
eng = CpuEngine(dtype, skip_rest=skip_rest, kernel=kernel)
eng.load(matrix, params)
eng.schedule_inputs(schedule)
eng.advance_to(steps)
return eng
def _spike_steps(eng, neuron):
s, n = eng.spikes_since(0)
return s[n == neuron].tolist()
def test_params_match_the_preregistration():
assert P.rfc_steps == 22
assert P.delay_steps == 18
assert P.w_input == pytest.approx(68.75)
def test_input_spike_fires_next_step_and_reset_discards_same_step_input():
# Input at every step 0..9: an input fires the neuron in the next step, and the
# input arriving in the spiking step is wiped by the reset (Brian2 order).
eng = _run(_matrix([], 1), _inputs({0: list(range(10))}), 12)
assert _spike_steps(eng, 0) == [1, 3, 5, 7, 9]
def test_delay_and_exact_propagator():
# 0 spikes at step 1, so its synapse onto 1 lands at step 1 + 18 = 19.
w = 100
eng = _run(_matrix([(0, 1, w)], 2), _inputs({0: [0]}), 19, dtype=np.float64)
assert eng.g[1] == 0 and eng.v[1] == P.v_0
eng.advance_to(20)
g0 = P.w_syn * w
assert eng.g[1] == pytest.approx(g0)
# From the end of step 19, neuron 1 follows the closed-form solution.
for k in range(1, 200):
eng.advance_to(20 + k)
t = k * P.dt
u = g0 * P.tau / (P.tau - P.t_mbr) * (math.exp(-t / P.tau) - math.exp(-t / P.t_mbr))
assert eng.v[1] - P.v_0 == pytest.approx(u, rel=1e-9, abs=1e-12)
assert eng.g[1] == pytest.approx(g0 * math.exp(-t / P.tau), rel=1e-9)
assert _spike_steps(eng, 1) == []
def test_refractory_period_is_22_steps_and_drops_input():
# 0 spikes at every odd step; each spike gives 1 a huge kick 18 steps later, at
# odd steps from 19. 1 spikes at 20 and is refractory for steps 21-41; the kicks
# in that window are dropped (Brian2's conditional write), so it is at rest at
# step 42 and only the kick at 43 fires it, at 44. Period: 24 steps.
eng = _run(_matrix([(0, 1, 20000)], 2), _inputs({0: list(range(0, 400, 2))}), 120)
assert _spike_steps(eng, 1) == [20, 44, 68, 92, 116]
def test_input_to_refractory_neuron_is_dropped_not_deferred():
# 1 spikes at 20 (kick from 0's spike at 1). A second, smaller kick lands at 25,
# inside the refractory window: Brian2 drops it, so 1 stays exactly at rest.
m = _matrix([(0, 1, 20000), (2, 1, 20000)], 3)
eng = _run(m, _inputs({0: [0], 2: [5]}), 60, dtype=np.float64)
assert _spike_steps(eng, 1) == [20]
assert eng.v[1] == P.v_0 and eng.g[1] == 0
def test_inhibition_and_stimulated_neurons_have_no_refractory_period():
m = _matrix([(0, 2, 30000), (1, 2, -30000)], 3)
both = _run(m, _inputs({0: [0], 1: [0]}), 100)
only_exc = _run(m, _inputs({0: [0]}), 100)
assert _spike_steps(only_exc, 2) == [20]
assert _spike_steps(both, 2) == []
# 0 Hz targets still lose their refractory period (Brian2's poi() sets rfc = 0).
eng = _run(m, InputSchedule(np.array([0, 1], np.int32), np.zeros(0, np.int64),
np.zeros(0, np.int32)), 1)
assert eng.rfc.tolist() == [0, 0, P.rfc_steps]
def _naive(matrix: SynapseMatrix, schedule: InputSchedule, steps: int, dtype) -> list[tuple]:
"""Straightforward per-step implementation of the same schedule, no shortcuts."""
f = np.dtype(dtype).type
n = matrix.n
a, b, c = (f(x) for x in P.coefficients())
v = np.full(n, P.v_0, dtype)
g = np.zeros(n, dtype)
last = np.full(n, -(10**9))
rfc = np.full(n, P.rfc_steps)
rfc[schedule.targets] = 0
history: list[np.ndarray] = []
out = []
for s in range(steps):
active = (s - last) >= rfc
v_new = f(P.v_0) + (v - f(P.v_0)) * a + g * b
v = np.where(active, v_new, v).astype(dtype)
g = np.where(active, g * c, g).astype(dtype)
spk = np.flatnonzero(active & (v > f(P.v_th)))
last[spk] = s
history.append(spk)
# Brian2's not_refractory after the thresholder: writes to v and g are
# conditional on it.
writable = active.copy()
writable[spk] = False
if s >= P.delay_steps:
for j in history[s - P.delay_steps]:
lo, hi = matrix.indptr[j], matrix.indptr[j + 1]
for post, w in zip(matrix.indices[lo:hi], matrix.weights[lo:hi], strict=True):
if writable[post]:
g[post] += f(P.w_syn) * f(w)
for k in np.flatnonzero(schedule.steps == s):
if writable[schedule.neurons[k]]:
v[schedule.neurons[k]] += f(P.w_input)
v[spk] = f(P.v_rst)
g[spk] = 0
out += [(s, int(i)) for i in spk]
return out
@pytest.mark.parametrize("kernel", KERNELS)
@pytest.mark.parametrize("dtype", [np.float32, np.float64])
def test_kernel_matches_naive_implementation(dtype, kernel):
rng = np.random.default_rng(1)
n, n_edges = 300, 6000
pre = rng.integers(0, n, n_edges)
post = rng.integers(0, n, n_edges)
w = rng.integers(1, 40, n_edges) * np.where(rng.random(n_edges) < 0.25, -1, 1)
m = SynapseMatrix.from_edges(pre, post, w, n)
sched = poisson_schedule(rng, np.arange(20), 150.0, 0, 1500, P.dt)
eng = _run(m, sched, 1500, dtype=dtype, kernel=kernel)
s, i = eng.spikes_since(0)
ours = list(zip(s.tolist(), i.tolist(), strict=True))
assert len(ours) > 1000 # the network is actually active
assert ours == _naive(m, sched, 1500, dtype)
@pytest.mark.parametrize("kernel", KERNELS)
def test_chunked_advance_and_log_growth_change_nothing(kernel):
rng = np.random.default_rng(2)
n = 200
m = SynapseMatrix.from_edges(rng.integers(0, n, 4000), rng.integers(0, n, 4000),
rng.integers(1, 30, 4000), n)
sched = poisson_schedule(rng, np.arange(15), 200.0, 0, 3000, P.dt)
whole = _run(m, sched, 3000) # plain kernel, in one go
chunked = CpuEngine(kernel=kernel)
chunked.load(m, P)
chunked._log = np.empty(n + 1, np.int32) # force repeated growth
chunked.schedule_inputs(sched)
seen_s, seen_n = [], []
for stop in range(100, 3001, 100): # 10 ms chunks, as the live runs step
before = chunked.step
chunked.advance_to(stop)
s, i = chunked.spikes_since(before)
seen_s.append(s)
seen_n.append(i)
ws, wn = whole.spikes_since(0)
np.testing.assert_array_equal(np.concatenate(seen_s), ws)
np.testing.assert_array_equal(np.concatenate(seen_n), wn)
np.testing.assert_array_equal(chunked.snapshot().v, whole.snapshot().v)
def test_reset_restores_rest_and_inputs_cannot_go_back_in_time():
eng = _run(_matrix([(0, 1, 5)], 2), _inputs({0: [0, 5]}), 50)
with pytest.raises(ValueError):
eng.schedule_inputs(_inputs({0: [10]}))
eng.reset()
assert eng.step == 0 and np.all(eng.v == P.v_0) and np.all(eng.g == 0)
assert eng.spikes_since(0)[0].size == 0
assert eng.rfc.tolist() == [P.rfc_steps] * 2
def test_seeding_is_stable_and_keyed():
a = trial_rng("E1", "E1/sugar/100Hz", 3).random(4)
b = trial_rng("E1", "E1/sugar/100Hz", 3).random(4)
np.testing.assert_array_equal(a, b)
assert not np.array_equal(a, trial_rng("E1", "E1/sugar/100Hz", 4).random(4))
assert not np.array_equal(a, trial_rng("E1", "E1/sugar/50Hz", 3).random(4))
assert trial_rng("E1", "E1/sugar/100Hz", 0).integers(0, 2**31) == PINNED_DRAW
def test_named_seed_namespaces_keep_the_recorded_seeds():
"""The public names of the chain tests and the live worker map to the namespace
strings the recorded runs were seeded with (draws pinned before the renaming)."""
assert trial_rng("chain-tests", "LC9/100Hz", 0).integers(0, 2**31) == 1436011335
assert trial_rng("chain-tests", "control-55/03", 7).integers(0, 2**31) == 707105761
assert trial_rng("live", "LC4/35Hz", 12).integers(0, 2**31) == 249271961
def test_poisson_rate_and_schedule_merge():
rng = trial_rng("test", "rate", 0)
s = poisson_schedule(rng, np.arange(50), 100.0, 0, 10000, P.dt)
rate = len(s.steps) / 50 / 1.0
assert 90 < rate < 110
s2 = poisson_schedule(rng, np.array([60, 61]), 0.0, 0, 10000, P.dt)
assert len(s2.steps) == 0
merged = merge(s, s2)
assert set(merged.targets.tolist()) == set(range(50)) | {60, 61}
assert np.all(np.diff(merged.steps) >= 0)
def _sparse_net(seed: int, n: int = 3000, n_edges: int = 30000) -> SynapseMatrix:
"""Mostly quiet network: activity spreads from a few stimulated neurons, as in the
connectome, so neurons leave rest, return to it and leave it again."""
rng = np.random.default_rng(seed)
pre, post = rng.integers(0, n, n_edges), rng.integers(0, n, n_edges)
w = rng.integers(1, 60, n_edges) * np.where(rng.random(n_edges) < 0.3, -1, 1)
return SynapseMatrix.from_edges(pre, post, w, n)
@pytest.mark.parametrize("seed", [3, 4, 5])
def test_skip_rest_is_bit_identical_to_the_plain_kernel(seed):
m = _sparse_net(seed)
rng = np.random.default_rng(seed)
sched = poisson_schedule(rng, rng.choice(m.n, 40, replace=False), 120.0, 0, 3000, P.dt)
plain = _run(m, sched, 6000)
fast = CpuEngine(kernel="skip_rest")
fast.dense_div = 1 # stay sparse: this test checks the list path
fast.load(m, P)
fast.schedule_inputs(sched)
for stop in (1, 17, 18, 100, 2999, 3000, 3001, 4500, 6000): # odd chunk edges
fast.advance_to(stop)
ps, pn = plain.spikes_since(0)
fs, fn = fast.spikes_since(0)
assert len(ps) > 1000
np.testing.assert_array_equal(fs, ps)
np.testing.assert_array_equal(fn, pn)
# the whole state, bit for bit (v and g compared as raw bits, NaN-proof)
a, b = plain.snapshot(), fast.snapshot()
np.testing.assert_array_equal(a.v.view(np.int32), b.v.view(np.int32))
np.testing.assert_array_equal(a.g.view(np.int32), b.g.view(np.int32))
np.testing.assert_array_equal(a.last_spike, b.last_spike)
# list invariant (sparse mode): every neuron off rest is listed, the list is sorted
assert not fast._dense
off = np.flatnonzero((fast.v != np.float32(P.v_0)) | (fast.g != 0))
listed = np.union1d(fast._act[fast._cur, : fast._n_act], fast._new[: fast._n_new])
assert np.isin(off, listed).all()
cur = fast._act[fast._cur, : fast._n_act]
assert np.all(np.diff(cur) > 0)
assert 0 < fast._n_act < m.n # it actually skips neurons
def test_skip_rest_survives_reset():
m = _sparse_net(6)
rng = np.random.default_rng(6)
sched = poisson_schedule(rng, np.arange(30), 150.0, 0, 1000, P.dt)
eng = CpuEngine(kernel="skip_rest")
eng.load(m, P)
eng.schedule_inputs(sched)
eng.advance_to(1500)
first = eng.spikes_since(0)
eng.reset()
assert eng._n_act == 0 and eng._n_new == 0 and not eng._flag.any()
eng.schedule_inputs(sched)
eng.advance_to(1500)
for x, y in zip(first, eng.spikes_since(0), strict=True):
np.testing.assert_array_equal(x, y)
@pytest.mark.parametrize("dense_div", [1, 2, 50, 10**6])
def test_skip_rest_dense_switch_is_bit_identical(dense_div):
m = _sparse_net(7)
rng = np.random.default_rng(7)
sched = poisson_schedule(rng, rng.choice(m.n, 60, replace=False), 150.0, 0, 3000, P.dt)
plain = _run(m, sched, 6000)
fast = CpuEngine(kernel="skip_rest")
fast.dense_div = dense_div
fast.load(m, P)
fast.schedule_inputs(sched)
for stop in (250, 2000, 6000):
fast.advance_to(stop)
for x, y in zip(plain.spikes_since(0), fast.spikes_since(0), strict=True):
np.testing.assert_array_equal(x, y)
np.testing.assert_array_equal(plain.v.view(np.int32), fast.v.view(np.int32))
np.testing.assert_array_equal(plain.g.view(np.int32), fast.g.view(np.int32))
if dense_div in (1, 10**6): # 1 never switches; 10**6 switches at the first input
assert fast._dense == (dense_div == 10**6)
def _bits_equal(a: CpuEngine, b: CpuEngine) -> None:
"""Spikes and the whole state, bit for bit (v and g as raw bits, NaN-proof)."""
for x, y in zip(a.spikes_since(0), b.spikes_since(0), strict=True):
np.testing.assert_array_equal(x, y)
bits = np.int32 if a.dtype == np.float32 else np.int64
np.testing.assert_array_equal(a.v.view(bits), b.v.view(bits))
np.testing.assert_array_equal(a.g.view(bits), b.g.view(bits))
np.testing.assert_array_equal(a.last, b.last)
def _until_matches(eng: CpuEngine) -> None:
"""vec's refractory state is last + max(rfc, 1) for every neuron that has spiked."""
spiked = eng.last > -(2**40)
np.testing.assert_array_equal(eng._until[spiked],
eng.last[spiked] + np.maximum(eng.rfc[spiked], 1))
assert np.all(eng._until[~spiked] < -(2**29))
def test_kernel_choice():
assert CpuEngine().kernel == "vec" and CpuEngine(skip_rest=True).kernel == "vec"
assert CpuEngine(kernel="skip_rest").kernel == "skip_rest"
assert CpuEngine(skip_rest=False).kernel == "plain"
assert CpuEngine(kernel="vec").kernel == "vec"
with pytest.raises(ValueError):
CpuEngine(kernel="fast")
@pytest.mark.parametrize("dtype", [np.float32, np.float64])
@pytest.mark.parametrize("dense_div", [1, 2, 50, 10**6])
@pytest.mark.parametrize("seed", [3, 7])
def test_vec_is_bit_identical_to_the_plain_kernel(seed, dense_div, dtype):
# dense_div 1 never leaves the list path; 10**6 sweeps densely from the first input
m = _sparse_net(seed)
rng = np.random.default_rng(seed)
sched = poisson_schedule(rng, rng.choice(m.n, 60, replace=False), 150.0, 0, 3000, P.dt)
plain = _run(m, sched, 6000, dtype=dtype)
fast = CpuEngine(dtype, kernel="vec")
fast.dense_div = dense_div
fast.load(m, P)
fast._log = np.empty(m.n + 1, np.int32) # force repeated log growth
fast.schedule_inputs(sched)
for stop in (1, 17, 18, 100, 511, 2999, 3000, 3001, 4500, 6000): # odd chunk edges
fast.advance_to(stop)
assert len(plain.spikes_since(0)[0]) > 1000
_bits_equal(plain, fast)
_until_matches(fast)
if dense_div in (1, 10**6):
assert fast._dense == (dense_div == 10**6)
def test_vec_dense_sweep_with_blocks_and_ragged_tail():
# more neurons than one block, not a multiple of it, dense from the start; a long
# quiet tail lets g decay into the subnormal range, where it gets stuck: m * c
# rounds back to m for subnormal mantissas m <= 25, so g never reaches 0
from flysim.engine.kernel import VEC_BLOCK
n = 3 * VEC_BLOCK + 77
m = _sparse_net(8, n=n, n_edges=12 * n)
rng = np.random.default_rng(8)
sched = poisson_schedule(rng, rng.choice(n, 40, replace=False), 150.0, 0, 1500, P.dt)
plain = CpuEngine(kernel="plain")
fast = CpuEngine(kernel="vec")
fast.dense_div = 10**6
subnormal = 0
for eng in (plain, fast):
eng.load(m, P)
eng.schedule_inputs(sched)
for stop in range(500, 12001, 500):
plain.advance_to(stop)
fast.advance_to(stop)
_bits_equal(plain, fast)
gb = plain.g.view(np.int32) & 0x7FFFFFFF
subnormal += int(np.sum((gb > 0) & (gb < 0x00800000)))
assert fast._dense and subnormal > 0
gb = plain.g.view(np.int32) & 0x7FFFFFFF
assert np.all(gb < 0x00800000) and np.any(gb > 0) # every g is 0 or subnormal
@pytest.mark.parametrize("kernel", ["skip_rest", "vec"])
def test_rfc_change_after_a_spike_is_honoured(kernel):
# 1 spikes at 20 (kick from 0's spike at 1), so it is refractory for 22 steps. At
# step 22 it becomes a stimulus target, so rfc drops to 0 and the inputs at 25 and
# later count at once, as in the plain kernel (rfc is read when it is used).
m = _matrix([(0, 1, 20000)], 2)
for k in ("plain", kernel):
eng = CpuEngine(np.float64, kernel=k)
eng.load(m, P)
eng.schedule_inputs(_inputs({0: [0]}))
eng.advance_to(22)
eng.schedule_inputs(_inputs({1: [25, 26, 30]}))
eng.advance_to(60)
if k == "plain":
plain = eng
assert _spike_steps(plain, 1) == [20, 26, 31]
_bits_equal(plain, eng)
if kernel == "vec":
_until_matches(eng)
def test_vec_survives_reset():
m = _sparse_net(6)
rng = np.random.default_rng(6)
sched = poisson_schedule(rng, np.arange(30), 150.0, 0, 1000, P.dt)
eng = CpuEngine(kernel="vec")
eng.dense_div = 10**6
eng.load(m, P)
eng.schedule_inputs(sched)
eng.advance_to(1500)
first = eng.spikes_since(0)
assert eng._dense
eng.reset()
assert not eng._dense and eng._n_act == 0 and not eng._flag.any()
assert np.all(eng._until < -(2**29)) and eng.rfc.tolist() == [P.rfc_steps] * m.n
eng.schedule_inputs(sched)
eng.advance_to(1500)
for x, y in zip(first, eng.spikes_since(0), strict=True):
np.testing.assert_array_equal(x, y)
|