File size: 18,230 Bytes
10979b5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 | """Unit tests for paper GmNet variants and gate ablations."""
from __future__ import annotations
import copy
import unittest
import torch
from gmnet.models import (
GATE_INTERVENTION_SEED_STRIDE,
GMNET_CONFIGS,
GmNet,
GmNetBlock,
SmoothClippedSelfGate,
build_gate,
gmnet_s1,
gmnet_s2,
gmnet_s3,
gmnet_s4,
)
class GateTests(unittest.TestCase):
def test_required_gate_semantics(self) -> None:
x = torch.tensor([[[[-2.0, -0.5, 0.5, 2.0, 8.0]]]])
torch.testing.assert_close(
build_gate("relu6_self", 1)(x), x * torch.clamp(x, 0.0, 6.0)
)
torch.testing.assert_close(
build_gate("relu6_only", 1)(x), torch.clamp(x, 0.0, 6.0)
)
torch.testing.assert_close(build_gate("relu_self", 1)(x), x * torch.relu(x))
torch.testing.assert_close(
build_gate("gelu_self", 1)(x), x * torch.nn.functional.gelu(x)
)
torch.testing.assert_close(build_gate("identity", 1)(x), x.square())
torch.testing.assert_close(build_gate("no_gate", 1)(x), x)
def test_smooth_clipped_gate_is_static_learnable_and_bounded(self) -> None:
gate = SmoothClippedSelfGate(3, init_clip=6.0, beta=10.0)
x = torch.linspace(-10.0, 12.0, 23).view(1, 1, 1, 23).expand(2, 3, 4, 23)
values = gate.gate_value(x)
self.assertEqual(gate.raw_clip.shape, (1, 3, 1, 1))
self.assertTrue(torch.all(gate.clip_value > 0))
self.assertGreaterEqual(float(values.detach().min()), -1e-6)
self.assertLessEqual(
float(values.detach().max()), float(gate.clip_value.detach().max()) + 1e-5
)
loss = gate(x).square().mean()
loss.backward()
self.assertIsNotNone(gate.raw_clip.grad)
self.assertTrue(torch.isfinite(gate.raw_clip.grad).all())
self.assertGreater(float(gate.raw_clip.grad.abs().sum()), 0.0)
def test_smooth_clipped_gate_supports_fixed_persistent_clip(self) -> None:
gate = SmoothClippedSelfGate(
3, init_clip=6.0, beta=10.0, per_channel=False, trainable=False
)
self.assertIsNone(gate.raw_clip)
self.assertEqual(sum(parameter.numel() for parameter in gate.parameters()), 0)
torch.testing.assert_close(gate.clip_value, torch.tensor([[[[6.0]]]]))
self.assertIn("fixed_clip", gate.state_dict())
restored = SmoothClippedSelfGate(
3, init_clip=6.0, beta=10.0, per_channel=False, trainable=False
)
restored.load_state_dict(gate.state_dict(), strict=True)
x = torch.randn(2, 3, 4, 4)
torch.testing.assert_close(restored(x), gate(x))
def test_smooth_checkpoint_topology_is_rebuilt_from_trainable_config(self) -> None:
common = {
"embed_dims": (8, 16, 24, 32),
"depths": (1, 1, 1, 1),
"mlp_ratios": (1, 1, 1, 1),
"stem_dim": 8,
"num_classes": 5,
"gate_type": "smooth_clipped_self",
"smooth_clip_per_channel": False,
}
for trainable, state_suffix in ((True, "raw_clip"), (False, "fixed_clip")):
with self.subTest(trainable=trainable):
config = {**common, "smooth_clip_trainable": trainable}
source = GmNet(**config)
checkpoint = {
"config": copy.deepcopy(config),
"model": copy.deepcopy(source.state_dict()),
}
restored = GmNet(**checkpoint["config"])
restored.load_state_dict(checkpoint["model"], strict=True)
gate_keys = [
key
for key in checkpoint["model"]
if key.endswith(("raw_clip", "fixed_clip"))
]
self.assertTrue(gate_keys)
self.assertTrue(all(key.endswith(state_suffix) for key in gate_keys))
def test_invalid_gate_is_rejected(self) -> None:
with self.assertRaisesRegex(ValueError, "unknown gate_type"):
build_gate("not_a_gate", 8)
def test_stop_gradient_preserves_forward_and_changes_gate_gradient(self) -> None:
baseline = build_gate("relu6_self", 3)
stopped = build_gate("relu6_self", 3, gate_intervention="stop_gradient")
baseline_input = torch.tensor([0.5, 2.0, 4.0]).view(1, 3, 1, 1)
baseline_input.requires_grad_()
stopped_input = baseline_input.detach().clone().requires_grad_()
baseline_output = baseline(baseline_input)
stopped_output = stopped(stopped_input)
self.assertTrue(torch.equal(baseline_output, stopped_output))
baseline_output.sum().backward()
stopped_output.sum().backward()
self.assertTrue(torch.isfinite(stopped_input.grad).all())
self.assertFalse(torch.equal(baseline_input.grad, stopped_input.grad))
torch.testing.assert_close(
stopped_input.grad, torch.tensor([0.5, 2.0, 4.0]).view(1, 3, 1, 1)
)
def test_channel_derangement_is_deterministic_and_preserves_multiset(self) -> None:
first = build_gate(
"relu6_self",
8,
gate_intervention="channel_derangement",
gate_intervention_seed=41_041,
)
second = build_gate(
"relu6_self",
8,
gate_intervention="channel_derangement",
gate_intervention_seed=41_041,
)
permutation = first.gate_permutation
self.assertIsNotNone(permutation)
assert permutation is not None
self.assertTrue(torch.equal(permutation, second.gate_permutation))
self.assertTrue(
torch.equal(permutation.sort().values, torch.arange(permutation.numel()))
)
self.assertTrue(torch.all(permutation != torch.arange(permutation.numel())))
x = torch.arange(48, dtype=torch.float32).reshape(2, 8, 1, 3)
permuted = first.gate_input(x)
torch.testing.assert_close(permuted.sort(dim=1).values, x.sort(dim=1).values)
torch.testing.assert_close(first(x), x * torch.clamp(permuted, 0.0, 6.0))
def test_channel_derangement_does_not_consume_global_rng(self) -> None:
torch.manual_seed(73)
expected = torch.rand(7)
torch.manual_seed(73)
build_gate(
"relu6_self",
8,
gate_intervention="channel_derangement",
gate_intervention_seed=41_041,
)
actual = torch.rand(7)
self.assertTrue(torch.equal(actual, expected))
def test_batch_derangement_is_seeded_bijective_and_rng_free(self) -> None:
gate = build_gate(
"relu6_self",
3,
gate_intervention="batch_derangement",
gate_intervention_seed=41_041,
)
x = torch.arange(60, dtype=torch.float32).reshape(5, 3, 2, 2)
shift = 1 + 41_041 % 4
expected_input = x.roll(shifts=-shift, dims=0)
torch.manual_seed(919)
expected_rng = torch.rand(8)
torch.manual_seed(919)
actual_input = gate.gate_input(x)
actual_rng = torch.rand(8)
self.assertEqual(gate.batch_derangement_shift(5), shift)
self.assertTrue(torch.equal(actual_input, expected_input))
self.assertTrue(torch.equal(actual_rng, expected_rng))
self.assertTrue(
torch.equal(
actual_input.flatten(1).sort(dim=0).values,
x.flatten(1).sort(dim=0).values,
)
)
with self.assertRaisesRegex(ValueError, "at least two samples"):
gate.gate_input(x[:1])
def test_stop_gradient_channel_derangement_matches_forward_only(self) -> None:
channel = build_gate(
"relu6_self",
4,
gate_intervention="channel_derangement",
gate_intervention_seed=41_041,
)
combined = build_gate(
"relu6_self",
4,
gate_intervention="stop_gradient_channel_derangement",
gate_intervention_seed=41_041,
)
self.assertTrue(torch.equal(channel.gate_permutation, combined.gate_permutation))
channel_input = (
torch.tensor([0.5, 1.5, 2.5, 3.5])
.reshape(1, 4, 1, 1)
.requires_grad_()
)
combined_input = channel_input.detach().clone().requires_grad_()
channel_output = channel(channel_input)
combined_output = combined(combined_input)
self.assertTrue(torch.equal(channel_output, combined_output))
channel_output.sum().backward()
combined_output.sum().backward()
self.assertFalse(torch.equal(channel_input.grad, combined_input.grad))
torch.testing.assert_close(
combined_input.grad,
torch.clamp(
combined.gate_input(combined_input.detach()), min=0.0, max=6.0
),
)
def test_intervention_rejects_non_self_gate_and_invalid_mode(self) -> None:
for gate_type in ("relu6_only", "no_gate"):
with self.subTest(gate_type=gate_type):
with self.assertRaisesRegex(ValueError, "multiplicative self-gate"):
build_gate(gate_type, 8, gate_intervention="channel_derangement")
with self.assertRaisesRegex(ValueError, "unknown gate_intervention"):
build_gate("relu6_self", 8, gate_intervention="not_an_intervention")
class GmNetTests(unittest.TestCase):
FACTORIES = {
"s1": gmnet_s1,
"s2": gmnet_s2,
"s3": gmnet_s3,
"s4": gmnet_s4,
}
EXPECTED_PARAMETERS = {
"s1": 3_661_600,
"s2": 6_206_296,
"s3": 7_791_544,
"s4": 17_061_460,
}
def test_paper_variant_topology_parameters_and_forward(self) -> None:
for name, factory in self.FACTORIES.items():
with self.subTest(variant=name):
model = factory().eval()
config = GMNET_CONFIGS[name]
self.assertEqual(model.embed_dims, config.embed_dims)
self.assertEqual(model.depths, config.depths)
self.assertEqual(model.mlp_ratios, config.mlp_ratios)
self.assertEqual(
tuple(len(stage) - 1 for stage in model.stages), config.depths
)
parameter_count = sum(
parameter.numel() for parameter in model.parameters()
)
self.assertEqual(parameter_count, self.EXPECTED_PARAMETERS[name])
model.reset_classifier(17)
with torch.inference_mode():
output = model(torch.randn(1, 3, 64, 64))
self.assertEqual(output.shape, (1, 17))
self.assertTrue(torch.isfinite(output).all())
def test_all_gate_types_run_in_a_model(self) -> None:
gate_types = (
"relu6_self",
"relu6_only",
"relu_self",
"gelu_self",
"identity",
"no_gate",
"smooth_clipped_self",
)
for gate_type in gate_types:
with self.subTest(gate_type=gate_type):
model = GmNet(
embed_dims=(8, 16, 24, 32),
depths=(1, 1, 1, 1),
mlp_ratios=(1, 1, 1, 1),
stem_dim=8,
num_classes=5,
gate_type=gate_type,
).eval()
with torch.inference_mode():
output = model(torch.randn(1, 3, 64, 64))
self.assertEqual(output.shape, (1, 5))
def test_experiment_registry_gate_aliases(self) -> None:
aliases = {
"relu6": "relu6_self",
"activation_only": "relu6_only",
"relu6_activation_only": "relu6_only",
"relu": "relu_self",
"gelu": "gelu_self",
"smooth_clipped_static": "smooth_clipped_self",
}
x = torch.randn(2, 3, 4, 4)
for alias, canonical in aliases.items():
with self.subTest(alias=alias):
torch.testing.assert_close(
build_gate(alias, 3)(x), build_gate(canonical, 3)(x)
)
def test_stage_specific_gates_and_feature_head(self) -> None:
model = GmNet(
embed_dims=(8, 16, 24, 32),
depths=(1, 1, 1, 1),
mlp_ratios=(1, 1, 1, 1),
stem_dim=8,
num_classes=0,
gate_type=("no_gate", "relu_self", "gelu_self", "smooth_clipped_self"),
).eval()
self.assertEqual(
model.gate_types,
("no_gate", "relu_self", "gelu_self", "smooth_clipped_self"),
)
with torch.inference_mode():
output = model(torch.randn(2, 3, 64, 64))
self.assertEqual(output.shape, (2, 32))
model.reset_classifier(7)
with torch.inference_mode():
output = model(torch.randn(2, 3, 64, 64))
self.assertEqual(output.shape, (2, 7))
@staticmethod
def _tiny_model(**kwargs: object) -> GmNet:
return GmNet(
embed_dims=(8, 16, 24, 32),
depths=(1, 1, 1, 1),
mlp_ratios=(2, 2, 2, 2),
stem_dim=8,
num_classes=5,
**kwargs,
)
def test_default_and_explicit_baseline_are_bitwise_compatible(self) -> None:
torch.manual_seed(991)
default = self._tiny_model().eval()
torch.manual_seed(991)
explicit = self._tiny_model(
gate_intervention="baseline", gate_intervention_seed=41_041
).eval()
self.assertEqual(default.state_dict().keys(), explicit.state_dict().keys())
for key, value in default.state_dict().items():
self.assertTrue(torch.equal(value, explicit.state_dict()[key]), key)
x = torch.randn(2, 3, 64, 64)
with torch.inference_mode():
self.assertTrue(torch.equal(default(x), explicit(x)))
def test_interventions_preserve_parameters_and_checkpoint_schema(self) -> None:
models = {
mode: self._tiny_model(
gate_intervention=mode, gate_intervention_seed=41_041
)
for mode in (
"baseline",
"stop_gradient",
"channel_derangement",
"batch_derangement",
"stop_gradient_channel_derangement",
)
}
baseline = models["baseline"]
expected_keys = tuple(baseline.state_dict())
expected_parameters = sum(p.numel() for p in baseline.parameters())
for mode, model in models.items():
with self.subTest(mode=mode):
self.assertEqual(tuple(model.state_dict()), expected_keys)
self.assertEqual(
sum(p.numel() for p in model.parameters()), expected_parameters
)
model.load_state_dict(baseline.state_dict(), strict=True)
self.assertFalse(
any("gate_permutation" in key for key in model.state_dict())
)
def test_channel_derangement_uses_per_block_seed_and_auditable_metadata(
self,
) -> None:
model = gmnet_s3(
gate_intervention="channel_derangement",
gate_intervention_seed=41_041,
)
metadata = model.gate_intervention_metadata()
blocks = [
module for module in model.modules() if isinstance(module, GmNetBlock)
]
self.assertEqual(len(metadata), 17)
self.assertEqual(len(blocks), 17)
for index, (block, item) in enumerate(zip(blocks, metadata, strict=True)):
with self.subTest(block=index):
expected_seed = 41_041 + index * GATE_INTERVENTION_SEED_STRIDE
self.assertEqual(block.gate_intervention_seed, expected_seed)
self.assertEqual(item["seed"], expected_seed)
self.assertEqual(item["global_block_index"], index)
self.assertEqual(item["mode"], "channel_derangement")
self.assertEqual(item["fixed_points"], 0)
self.assertIs(item["is_bijection"], True)
self.assertEqual(len(item["permutation_sha256"]), 64)
self.assertEqual(
item["permutation_hash_encoding"],
"little_endian_int64_c_order",
)
self.assertIn(
"gate_permutation", block.gate._non_persistent_buffers_set
)
def test_combined_intervention_reuses_every_channel_permutation(self) -> None:
channel = gmnet_s3(
gate_intervention="channel_derangement",
gate_intervention_seed=41_041,
)
combined = gmnet_s3(
gate_intervention="stop_gradient_channel_derangement",
gate_intervention_seed=41_041,
)
channel_rows = channel.gate_intervention_metadata()
combined_rows = combined.gate_intervention_metadata()
self.assertEqual(len(channel_rows), 17)
self.assertEqual(len(combined_rows), 17)
for channel_row, combined_row in zip(
channel_rows, combined_rows, strict=True
):
self.assertEqual(
channel_row["permutation_sha256"],
combined_row["permutation_sha256"],
)
self.assertIs(combined_row["stops_gate_gradient"], True)
self.assertIs(combined_row["channel_derangement"], True)
self.assertIs(combined_row["batch_derangement"], False)
def test_stop_gradient_model_has_finite_parameter_gradients(self) -> None:
model = self._tiny_model(gate_intervention="stop_gradient")
loss = model(torch.randn(2, 3, 64, 64)).square().mean()
loss.backward()
gradients = [
parameter.grad
for parameter in model.parameters()
if parameter.requires_grad
]
self.assertTrue(all(gradient is not None for gradient in gradients))
self.assertTrue(all(torch.isfinite(gradient).all() for gradient in gradients))
if __name__ == "__main__":
unittest.main()
|