Fix efficientnet-lite0 fine-tuning regression from the timm backbone swap

[gomezzz]

Summary

Switching the efficientnet-lite0 backbone to timm (tf_efficientnet_lite0) silently dropped three model-level defaults the previous efficientnet_lite_pytorch package applied, substantially regressing FixMatch fine-tuning quality. This restores all three:

• Classifier head over-scaling (main cause). timm inits the head with TF-EfficientNet's 1/sqrt(fan_in+fan_out) scale (for the 1000-class ImageNet head) — ~num_classes× too large for the 2-class head (weight std ~0.4 vs PyTorch's ~0.016). The overconfident fresh head makes ~100% of unlabeled images clear the FixMatch p_cutoff (0.95) from the first step with random pseudo-labels, poisoning the backbone. Fix: reset the classifier to PyTorch's default init after create_model.
• cfg.bn_momentum never applied (BN ran at timm's 0.1 default, ~10x too fast for the small batch). Now applied to both models' BatchNorm layers.
• cfg.seed never applied (training was not reproducible). Now seeds all RNGs after config validation.

The timm and efficientnet_lite_pytorch weights are byte-identical and eval-forward is bit-identical, so no weight swap or dependency change is needed.

Validation (paper benchmark, miniImageNet AL, AUROC)

class	before (broken)	after fix
hourglass	0.81 / 0.84 / 0.88	0.96 / 0.94 / 0.94
piano	0.86 / 0.87 / 0.83	0.92 / 0.89 / 0.90

Test plan

• ruff check/ruff format clean on changed files
• New head-init regression test + existing unit tests pass (376 tests green)
• End-to-end benchmark retrain confirms recovery to ~0.94-0.96 AUROC

[github-actions]

Coverage Report

File	Stmts	Miss	Cover	Missing
__init__.py	8	0	100%
data_io
SessionIOHandler.py	326	41	87%	111, 128–129, 173–174, 184–185, 215–216, 282–283, 286, 326–328, 370, 388–390, 455, 530, 553, 567–568, 570–571, 573–575, 719–724, 738, 742–743, 747–749
checkpoint_io.py	148	25	83%	68, 74, 76, 78, 95–109, 120, 122, 161, 182, 228, 281
find_images_in_folder.py	21	0	100%
load_images.py	91	28	69%	51–53, 58–60, 81, 93, 110, 140–147, 150, 156–157, 189, 248, 256, 277, 285, 292–294
metadata_handler.py	80	8	90%	70–71, 117, 122–123, 170–172
save_config.py	33	3	91%	92–94
datasets
AnomalyDetectionDataset.py	248	39	84%	126–127, 356, 362–363, 365, 369, 371–372, 374, 395–396, 398, 404–405, 409, 422–423, 484–485, 491–496, 498–499, 503–506, 508–510, 512–514, 525
BasicDataset.py	52	4	92%	59, 61, 97, 103
Label.py	5	0	100%
SSL_Dataset.py	68	3	96%	136, 139, 209
__init__.py	0	0	100%
data_utils.py	56	2	96%	80, 199
datasets/augmentation
randaugment.py	92	11	88%	222, 225–226, 245, 328–330, 332–335
randaugment_multispectral.py	77	23	70%	77–79, 120–121, 134–135, 148–149, 166, 170, 174, 248, 253, 270–272, 274–277, 280–281
image_processing
transforms.py	57	3	95%	46–47, 76
models
FixMatch.py	214	29	86%	108, 201–202, 217, 222, 254, 281–282, 285, 288–290, 293, 301, 306–307, 366, 404, 430–434, 495, 497–498, 502, 519–520
pipeline
SessionTracker.py	122	3	98%	127, 219–220
session.py	541	85	84%	146–149, 399–402, 447, 460, 612, 669, 671–672, 680, 683, 694, 702, 732, 736, 740–741, 745, 750–751, 758, 762–763, 775, 785–786, 788, 792, 806–807, 809–811, 813, 816–817, 875–876, 878–881, 883–887, 893, 899, 904–908, 910, 918, 921–922, 938, 979–980, 986–987, 991–993, 995, 1020, 1046, 1080–1081, 1089, 1100, 1108, 1112, 1116, 1121, 1124, 1168, 1171
utils
accuracy.py	13	0	100%
consistency_loss.py	18	0	100%
create_model_string.py	4	0	100%
cross_entropy_loss.py	9	0	100%
cutana_stream_utils.py	70	9	87%	64–65, 88, 108, 118–122
get_cosine_schedule_with_warmup.py	11	1	91%	39
get_default_cfg.py	61	0	100%
get_net_builder.py	55	7	87%	81–82, 89, 92, 170–171, 175
get_optimizer.py	21	2	90%	54, 58
print_cfg.py	45	4	91%	89, 95, 117, 119
set_log_level.py	15	0	100%
set_seeds.py	13	2	85%	25–26
validate_config.py	131	11	92%	180, 199, 214, 228, 244, 248, 264, 290, 379–381
TOTAL	2705	343	87%

Tests	Skipped	Failures	Errors	Time
376	0 💤	0 ❌	0 🔥	1m 0s ⏱️