File size: 13,006 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 | """Regression tests for long-run training audit safeguards."""
from __future__ import annotations
import random
from unittest import mock
import numpy as np
import pytest
import torch
from torch import nn
from gmnet.data.build import _build_manifest, _validate_expected_samples
from gmnet.engine import state_dict_schema_sha256
from gmnet.models import SmoothClippedSelfGate
from gmnet.train import (
capture_rng_state,
checkpoint_path_for_progress,
collect_rng_states,
fail_if_nonfinite,
parameter_groups,
make_scheduler,
reduce_batchnorm_statistics,
restore_rng_state,
should_evaluate_epoch,
smooth_clip_diagnostics,
stable_config_fingerprint,
validate_resume_checkpoint,
validate_resume_data_manifest,
validate_resume_training_progress,
)
class _Parameters(nn.Module):
def __init__(self) -> None:
super().__init__()
self.weight = nn.Parameter(torch.ones(2, 2))
self.raw_clip = nn.Parameter(torch.ones(1, 2, 1, 1))
self.bias = nn.Parameter(torch.ones(2))
def test_raw_clip_pattern_disables_weight_decay() -> None:
model = _Parameters()
groups = parameter_groups(model, 0.03, ["raw_clip"])
assert [id(value) for value in groups[0]["params"]] == [id(model.weight)]
assert {id(value) for value in groups[1]["params"]} == {
id(model.raw_clip),
id(model.bias),
}
assert groups[1]["weight_decay"] == 0.0
def test_no_weight_decay_pattern_supports_glob() -> None:
model = nn.Sequential(_Parameters())
groups = parameter_groups(model, 0.03, ["*.raw_clip"])
assert id(model[0].raw_clip) in {id(value) for value in groups[1]["params"]}
def test_config_fingerprint_is_stable_and_sensitive() -> None:
first = {"train": {"epochs": 300}, "seed": 1}
reordered = {"seed": 1, "train": {"epochs": 300}}
changed = {"seed": 2, "train": {"epochs": 300}}
assert stable_config_fingerprint(first) == stable_config_fingerprint(reordered)
assert stable_config_fingerprint(first) != stable_config_fingerprint(changed)
def test_strict_resume_checks_all_experiment_identity_fields() -> None:
checkpoint = {
"config_fingerprint": "abc",
"run_name": "run",
"seed": 7,
"world_size": 8,
"epoch_complete": True,
"rng_state_by_rank": [{} for _ in range(8)],
}
validate_resume_checkpoint(
checkpoint, config_fingerprint="abc", run_name="run", seed=7, world_size=8
)
with pytest.raises(ValueError, match="config_fingerprint.*run_name.*seed.*world_size"):
validate_resume_checkpoint(
checkpoint,
config_fingerprint="different",
run_name="other",
seed=8,
world_size=4,
)
with pytest.raises(ValueError, match="missing config_fingerprint"):
validate_resume_checkpoint(
{}, config_fingerprint="abc", run_name="run", seed=7, world_size=8
)
def test_strict_resume_requires_a_complete_consistent_epoch() -> None:
checkpoint = {
"epoch": 2,
"global_step": 30,
"epoch_complete": True,
"steps_in_epoch": 10,
"expected_steps_per_epoch": 10,
"training_complete": False,
}
validate_resume_training_progress(
checkpoint, expected_steps_per_epoch=10, epochs=4
)
with pytest.raises(ValueError, match="steps_in_epoch.*global_step"):
validate_resume_training_progress(
{**checkpoint, "steps_in_epoch": 1, "global_step": 21},
expected_steps_per_epoch=10,
epochs=4,
)
def test_partial_epoch_uses_recovery_checkpoint_only(tmp_path) -> None:
assert checkpoint_path_for_progress(
tmp_path, epoch_complete=False
) == tmp_path / "checkpoint_recovery.pt"
assert checkpoint_path_for_progress(
tmp_path, epoch_complete=True
) == tmp_path / "checkpoint_last.pt"
def test_state_schema_hash_tracks_names_shapes_and_dtypes() -> None:
first = {"weight": torch.zeros(2, 3), "count": torch.zeros((), dtype=torch.int64)}
same = {"weight": torch.ones(2, 3), "count": torch.ones((), dtype=torch.int64)}
changed = {"weight": torch.ones(3, 2), "count": torch.ones((), dtype=torch.int64)}
assert state_dict_schema_sha256(first) == state_dict_schema_sha256(same)
assert state_dict_schema_sha256(first) != state_dict_schema_sha256(changed)
def test_scheduler_supports_epoch_steps_and_cooldown() -> None:
parameter = nn.Parameter(torch.ones(()))
optimizer = torch.optim.SGD([parameter], lr=1.0)
scheduler = make_scheduler(
optimizer,
epochs=6,
steps_per_epoch=10,
warmup_epochs=1,
warmup_lr=0.1,
base_lr=1.0,
min_lr=0.01,
step_unit="epoch",
cooldown_epochs=2,
)
observed = [optimizer.param_groups[0]["lr"]]
for _ in range(6):
optimizer.step()
scheduler.step()
observed.append(optimizer.param_groups[0]["lr"])
assert observed[0] == pytest.approx(0.1)
assert observed[1] == pytest.approx(1.0)
assert observed[4:] == pytest.approx([0.01, 0.01, 0.01])
def test_scheduler_rejects_invalid_protocol() -> None:
optimizer = torch.optim.SGD([nn.Parameter(torch.ones(()))], lr=1.0)
with pytest.raises(ValueError, match="step_unit"):
make_scheduler(
optimizer,
epochs=2,
steps_per_epoch=1,
warmup_epochs=0,
warmup_lr=0.0,
base_lr=1.0,
min_lr=0.0,
step_unit="batch",
)
def test_strict_resume_checks_data_manifest() -> None:
manifest = {"manifest_sha256": "same"}
validate_resume_data_manifest({"data_manifest": manifest}, manifest)
with pytest.raises(ValueError, match="data manifest changed"):
validate_resume_data_manifest(
{"data_manifest": {"manifest_sha256": "old"}}, manifest
)
def test_rank_local_rng_state_round_trip() -> None:
generator = torch.Generator().manual_seed(41)
random.seed(11)
np.random.seed(21)
torch.manual_seed(31)
with mock.patch("gmnet.train.torch.cuda.is_available", return_value=False):
state = capture_rng_state(generator)
expected = (
random.random(),
float(np.random.random()),
float(torch.rand(())),
float(torch.rand((), generator=generator)),
)
random.random()
np.random.random()
torch.rand(())
torch.rand((), generator=generator)
restore_rng_state(
{"rng_state_by_rank": [state]},
rank=0,
world_size=1,
generator=generator,
)
actual = (
random.random(),
float(np.random.random()),
float(torch.rand(())),
float(torch.rand((), generator=generator)),
)
assert actual == pytest.approx(expected)
def test_rng_collection_keeps_one_state_per_rank() -> None:
local = {"rank": 0}
def gather(output: list[object], value: object) -> None:
assert value is local
output[:] = [value, {"rank": 1}]
with (
mock.patch("gmnet.train.capture_rng_state", return_value=local),
mock.patch("gmnet.train.dist.all_gather_object", side_effect=gather),
):
states = collect_rng_states(
torch.Generator(), distributed=True, world_size=2
)
assert states == [{"rank": 0}, {"rank": 1}]
def test_rng_restore_rejects_wrong_world_size_or_schema() -> None:
with pytest.raises(ValueError, match="exactly 2 entries"):
restore_rng_state(
{"rng_state_by_rank": [{}]},
rank=0,
world_size=2,
generator=torch.Generator(),
)
with pytest.raises(ValueError, match="rank-local RNG state schema"):
restore_rng_state(
{"rng_state_by_rank": [{}]},
rank=0,
world_size=1,
generator=torch.Generator(),
)
def test_eval_interval_always_evaluates_final_or_stopped_epoch() -> None:
assert not should_evaluate_epoch(
epoch=0, epochs=12, eval_interval=5, stop_after_epoch=False
)
assert should_evaluate_epoch(
epoch=4, epochs=12, eval_interval=5, stop_after_epoch=False
)
assert should_evaluate_epoch(
epoch=11, epochs=12, eval_interval=5, stop_after_epoch=False
)
assert should_evaluate_epoch(
epoch=2, epochs=12, eval_interval=5, stop_after_epoch=True
)
with pytest.raises(ValueError, match="positive"):
should_evaluate_epoch(
epoch=0, epochs=1, eval_interval=0, stop_after_epoch=False
)
def test_nonfinite_guard_rejects_nonfinite_gradient() -> None:
model = nn.Linear(2, 1)
loss = model(torch.ones(1, 2)).sum()
loss.backward()
fail_if_nonfinite(loss, model, epoch=1, step=2)
model.weight.grad[0, 0] = torch.inf
with pytest.raises(FloatingPointError, match="weight"):
fail_if_nonfinite(loss, model, epoch=1, step=2)
def test_bn_reduction_averages_running_statistics() -> None:
model = nn.BatchNorm1d(2)
model.running_mean.copy_(torch.tensor([2.0, 4.0]))
model.running_var.copy_(torch.tensor([6.0, 8.0]))
def double(tensor: torch.Tensor, **_: object) -> None:
tensor.mul_(2)
with mock.patch("gmnet.train.dist.all_reduce", side_effect=double) as reduce:
reduce_batchnorm_statistics(model, world_size=2)
assert reduce.call_count == 2
assert torch.equal(model.running_mean, torch.tensor([2.0, 4.0]))
assert torch.equal(model.running_var, torch.tensor([6.0, 8.0]))
def test_smooth_clip_diagnostics_include_stage_statistics() -> None:
model = nn.Module()
model.stages = nn.ModuleList(
[
nn.Sequential(SmoothClippedSelfGate(2, init_clip=2.0)),
nn.Sequential(SmoothClippedSelfGate(3, init_clip=3.0)),
]
)
result = smooth_clip_diagnostics(model)
assert result is not None
assert result["module_count"] == 2
assert result["overall"]["count"] == 5
assert set(result["by_stage"]) == {"stage1", "stage2"}
assert result["by_stage"]["stage1"]["count"] == 2
def test_expected_sample_counts_are_strict() -> None:
_validate_expected_samples(
{"expected_samples": {"train": 10, "val": 4}},
train_samples=10,
val_samples=4,
)
with pytest.raises(ValueError, match="train dataset has 10"):
_validate_expected_samples(
{"expected_train_samples": 11}, train_samples=10, val_samples=4
)
def test_imagefolder_manifest_is_root_independent(tmp_path) -> None:
from PIL import Image
from torchvision import datasets
manifests = []
for root_name in ("first", "second"):
root = tmp_path / root_name
for split in ("train", "val"):
directory = root / split / "class0"
directory.mkdir(parents=True)
Image.new("RGB", (2, 2), color="white").save(directory / "a.png")
train = datasets.ImageFolder(root / "train")
val = datasets.ImageFolder(root / "val")
manifests.append(
_build_manifest(
dataset_name="imagefolder",
train_split="train",
val_split="val",
train_dataset=train,
val_dataset=val,
train_root=root / "train",
val_root=root / "val",
class_to_idx=train.class_to_idx,
)
)
assert manifests[0] == manifests[1]
assert manifests[0]["schema_version"] == 2
assert manifests[0]["sampled_content_samples"] == {"train": 1, "val": 1}
def test_imagefolder_manifest_detects_sampled_file_content_change(tmp_path) -> None:
from PIL import Image
from torchvision import datasets
root = tmp_path / "data"
for split in ("train", "val"):
directory = root / split / "class0"
directory.mkdir(parents=True)
Image.new("RGB", (2, 2), color="white").save(directory / "a.png")
def build() -> dict[str, object]:
train = datasets.ImageFolder(root / "train")
val = datasets.ImageFolder(root / "val")
return _build_manifest(
dataset_name="imagefolder",
train_split="train",
val_split="val",
train_dataset=train,
val_dataset=val,
train_root=root / "train",
val_root=root / "val",
class_to_idx=train.class_to_idx,
)
before = build()
Image.new("RGB", (2, 2), color="black").save(
root / "val" / "class0" / "a.png"
)
after = build()
assert before["sample_index_sha256"] == after["sample_index_sha256"]
assert before["sampled_content_sha256"]["train"] == after[
"sampled_content_sha256"
]["train"]
assert before["sampled_content_sha256"]["val"] != after[
"sampled_content_sha256"
]["val"]
assert before["manifest_sha256"] != after["manifest_sha256"]
|