diff options
| author | YurenHao0426 <Blackhao0426@gmail.com> | 2026-07-22 13:57:08 -0500 |
|---|---|---|
| committer | YurenHao0426 <Blackhao0426@gmail.com> | 2026-07-22 13:57:08 -0500 |
| commit | 80ead0dcdc410281894956402c08a2a26da63434 (patch) | |
| tree | 0037bffef4417c570123f0bfdfd3ca7543a97027 /results | |
| parent | 6206483780d75ecff61c396b7f1bf7e23e5b1ca5 (diff) | |
results: pass residual mirror short gate
Diffstat (limited to 'results')
| -rw-r--r-- | results/residual_mirror_short/rrm_lr0.03.json | 756 | ||||
| -rw-r--r-- | results/residual_mirror_short/rrm_lr0.1.json | 756 | ||||
| -rw-r--r-- | results/residual_mirror_short_gate.json | 77 |
3 files changed, 1589 insertions, 0 deletions
diff --git a/results/residual_mirror_short/rrm_lr0.03.json b/results/residual_mirror_short/rrm_lr0.03.json new file mode 100644 index 0000000..a11eafb --- /dev/null +++ b/results/residual_mirror_short/rrm_lr0.03.json @@ -0,0 +1,756 @@ +{ + "architecture": { + "adaptive_apical_parameters": 0, + "adaptive_feedback_parameters": 267904, + "base_width": 16, + "blocks_per_stage": 3, + "bn_eps": 1e-05, + "bn_momentum": 0.1, + "depth": 20, + "family": "CIFAR 6n+2 ResNet, option-A shortcuts", + "fixed_feedback_parameters": 0, + "fixed_traffic_coefficients": 0, + "forward_parameters": 269722, + "hidden_shapes": [ + [ + 16, + 32, + 32 + ], + [ + 16, + 32, + 32 + ], + [ + 16, + 32, + 32 + ], + [ + 16, + 32, + 32 + ], + [ + 16, + 32, + 32 + ], + [ + 16, + 32, + 32 + ], + [ + 16, + 32, + 32 + ], + [ + 32, + 16, + 16 + ], + [ + 32, + 16, + 16 + ], + [ + 32, + 16, + 16 + ], + [ + 32, + 16, + 16 + ], + [ + 32, + 16, + 16 + ], + [ + 32, + 16, + 16 + ], + [ + 64, + 8, + 8 + ], + [ + 64, + 8, + 8 + ], + [ + 64, + 8, + 8 + ], + [ + 64, + 8, + 8 + ], + [ + 64, + 8, + 8 + ], + [ + 64, + 8, + 8 + ] + ], + "normalization": "batchnorm", + "predictor_parameters": 0, + "residual_scale": 1.0, + "vectorizer_mode": null, + "vectorizer_parameters": 0 + }, + "args": { + "a_scale": 1.0, + "a_warmup_steps": 0, + "alignment_probe": 32, + "apical_calibration_mode": "unit_targets", + "apical_seed": null, + "augment_train": 1, + "batch_size": 128, + "bn_eps": 1e-05, + "bn_momentum": 0.1, + "data_dir": "/home/yurenh2/sdrn/data", + "depth": 20, + "device": "cuda", + "epochs": 20, + "eta_A": 0.01, + "eta_P": 0.01, + "eval_every": 0, + "eval_split": "validation", + "learn_P": 0, + "loader_seed": 0, + "lr": 0.03, + "lr_gamma": 0.1, + "lr_milestones": "100,150", + "lr_schedule": "cosine", + "max_steps": 0, + "mirror_batch_size": 1, + "mirror_eta": 0.1, + "mirror_every": 16, + "mirror_noise_std": 1.0, + "mirror_seed": 3000, + "mirror_warmup_steps": 20, + "mode": "rrm", + "momentum": 0.9, + "normalization": "batchnorm", + "nuisance_scale": 0.0, + "out": "results/residual_mirror_short/rrm_lr0.03.json", + "output_lr": 0.1, + "pert_directions": 1, + "pert_every": 4, + "pert_sigma": 0.01, + "perturb_seed": 1000, + "predictor_warmup_steps": 0, + "residual_scale": null, + "seed": 0, + "split_seed": 2027, + "train_limit": 10000, + "use_residual": 1, + "val_examples": 5000, + "vectorizer_mode": "spatial_template", + "warmup_epochs": 0, + "weight_decay": 0.0001, + "weight_scale": 1.0, + "width": 16 + }, + "calibration_metric_space": "local_parent_child_response_residual", + "counters": { + "apical_warmup_examples": 0, + "calibration_event_examples": 0, + "causal_scalar_observations": 0, + "logical_batch_loss_queries": 0, + "mirror_conv_examples": 119, + "mirror_events": 119, + "mirror_readout_examples": 121856, + "ordinary_examples": 200000, + "per_example_loss_terms": 0, + "perturbation_events": 0, + "perturbation_forward_examples": 0, + "predictor_warmup_examples": 0 + }, + "diagnostics": { + "early_third_mean": 0.9996043642361959, + "feedback_forward_cosine": [ + 0.9999768733978271, + 0.9999785423278809, + 0.9999861121177673, + 0.9999871850013733, + 0.9999870657920837, + 0.9999917149543762, + 0.9999925494194031, + 0.9999871850013733, + 0.9999890327453613, + 0.9999918937683105, + 0.999989926815033, + 0.9999934434890747, + 0.999988317489624, + 0.9999141693115234, + 0.9999170303344727, + 0.9999324679374695, + 0.9999240040779114, + 0.9999455809593201, + 0.9999633431434631 + ], + "feedback_forward_norm_ratio": [ + 0.9995005130767822, + 0.9996951818466187, + 0.9998293519020081, + 0.9998653531074524, + 0.9998900294303894, + 0.9999541640281677, + 1.0000402927398682, + 0.9999350905418396, + 0.9999354481697083, + 1.000056266784668, + 1.0000362396240234, + 1.0000970363616943, + 1.0001606941223145, + 1.0000756978988647, + 1.0002508163452148, + 1.0001662969589233, + 1.000160813331604, + 1.0001862049102783, + 0.9902048110961914 + ], + "innovation_negative_gradient_cosine": [ + 0.999528169631958, + 0.9995716214179993, + 0.9996142387390137, + 0.9996160268783569, + 0.9996427893638611, + 0.9996533393859863, + 0.9996694326400757, + 0.9996705055236816, + 0.9996938109397888, + 0.9996921420097351, + 0.9997087717056274, + 0.9997057318687439, + 0.9997020959854126, + 0.9995958209037781, + 0.9997575879096985, + 0.9997144937515259, + 0.9998520016670227, + 0.9998068809509277, + 0.9999513626098633 + ], + "normalization_state": "training_batch_stats_without_running_update", + "raw_negative_gradient_cosine": [ + 0.999528169631958, + 0.9995716214179993, + 0.9996142387390137, + 0.9996160268783569, + 0.9996427893638611, + 0.9996533393859863, + 0.9996694326400757, + 0.9996705055236816, + 0.9996938109397888, + 0.9996921420097351, + 0.9997087717056274, + 0.9997057318687439, + 0.9997020959854126, + 0.9995958209037781, + 0.9997575879096985, + 0.9997144937515259, + 0.9998520016670227, + 0.9998068809509277, + 0.9999513626098633 + ], + "teaching_negative_gradient_cosine": [ + 0.999528169631958, + 0.9995716214179993, + 0.9996142387390137, + 0.9996160268783569, + 0.9996427893638611, + 0.9996533393859863, + 0.9996694326400757, + 0.9996705055236816, + 0.9996938109397888, + 0.9996921420097351, + 0.9997087717056274, + 0.9997057318687439, + 0.9997020959854126, + 0.9995958209037781, + 0.9997575879096985, + 0.9997144937515259, + 0.9998520016670227, + 0.9998068809509277, + 0.9999513626098633 + ], + "wall_s": 0.09704756736755371 + }, + "epochs": [ + { + "epoch": 1, + "lr": 0.03, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.23311814341040318, + "mirror_response_residual_rms": 0.3156063292915284, + "mirror_update_rms": 0.07668940078842221, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.1, + "step": 79, + "train_examples": 10000, + "train_loss": 2.434635739135742 + }, + { + "epoch": 2, + "lr": 0.029815325108927065, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.20402604902340427, + "mirror_response_residual_rms": 0.27805371235373844, + "mirror_update_rms": 0.06168744020569984, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.0993844170297569, + "step": 158, + "train_examples": 10000, + "train_loss": 2.2998252067565916 + }, + { + "epoch": 3, + "lr": 0.029265847744427302, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.16542378611773906, + "mirror_response_residual_rms": 0.2263167759715689, + "mirror_update_rms": 0.048502232916747975, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.09755282581475769, + "step": 237, + "train_examples": 10000, + "train_loss": 2.2560854587554933 + }, + { + "epoch": 4, + "lr": 0.028365097862825515, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.13348651911792966, + "mirror_response_residual_rms": 0.18303278083837884, + "mirror_update_rms": 0.03808328121413277, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.09455032620941839, + "step": 316, + "train_examples": 10000, + "train_loss": 2.114266714668274 + }, + { + "epoch": 5, + "lr": 0.02713525491562421, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.11722695513112225, + "mirror_response_residual_rms": 0.16164586388549532, + "mirror_update_rms": 0.03078940263897857, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.09045084971874738, + "step": 395, + "train_examples": 10000, + "train_loss": 2.238841275024414 + }, + { + "epoch": 6, + "lr": 0.02560660171779821, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.10380505538209925, + "mirror_response_residual_rms": 0.14311962215491328, + "mirror_update_rms": 0.02516932064852458, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.08535533905932738, + "step": 474, + "train_examples": 10000, + "train_loss": 1.790627029800415 + }, + { + "epoch": 7, + "lr": 0.023816778784387097, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.09441984979763204, + "mirror_response_residual_rms": 0.1307853126690807, + "mirror_update_rms": 0.020591888086161463, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.07938926261462366, + "step": 553, + "train_examples": 10000, + "train_loss": 1.8233744716644287 + }, + { + "epoch": 8, + "lr": 0.0218098574960932, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.08766515886239068, + "mirror_response_residual_rms": 0.12221957734069559, + "mirror_update_rms": 0.017453160153299223, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.07269952498697733, + "step": 632, + "train_examples": 10000, + "train_loss": 1.7596716131210328 + }, + { + "epoch": 9, + "lr": 0.01963525491562421, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.08203623606337147, + "mirror_response_residual_rms": 0.11465820061754553, + "mirror_update_rms": 0.014978560638242636, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.06545084971874737, + "step": 711, + "train_examples": 10000, + "train_loss": 1.659573628807068 + }, + { + "epoch": 10, + "lr": 0.017346516975603462, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.073935835101102, + "mirror_response_residual_rms": 0.10348125796130887, + "mirror_update_rms": 0.012977098879007715, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.05782172325201155, + "step": 790, + "train_examples": 10000, + "train_loss": 1.477030182647705 + }, + { + "epoch": 11, + "lr": 0.015, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.06792754832642667, + "mirror_response_residual_rms": 0.09552965244148469, + "mirror_update_rms": 0.011105307961398627, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.05, + "step": 869, + "train_examples": 10000, + "train_loss": 1.470662978553772 + }, + { + "epoch": 12, + "lr": 0.012653483024396533, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.05899317007766277, + "mirror_response_residual_rms": 0.08345544686067433, + "mirror_update_rms": 0.009759097335255375, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.04217827674798845, + "step": 948, + "train_examples": 10000, + "train_loss": 1.3200396375656127 + }, + { + "epoch": 13, + "lr": 0.01036474508437579, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.05147646829178666, + "mirror_response_residual_rms": 0.07284069494475491, + "mirror_update_rms": 0.008473658038829504, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.03454915028125263, + "step": 1027, + "train_examples": 10000, + "train_loss": 1.2434486444473267 + }, + { + "epoch": 14, + "lr": 0.0081901425039068, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.04412920759833355, + "mirror_response_residual_rms": 0.0626119977827275, + "mirror_update_rms": 0.007263608936965866, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.027300475013022664, + "step": 1106, + "train_examples": 10000, + "train_loss": 1.1923627337455749 + }, + { + "epoch": 15, + "lr": 0.006183221215612904, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.03685247319401539, + "mirror_response_residual_rms": 0.05235124424310925, + "mirror_update_rms": 0.006354452829585038, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.02061073738537635, + "step": 1185, + "train_examples": 10000, + "train_loss": 1.116790009498596 + }, + { + "epoch": 16, + "lr": 0.004393398282201788, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.03072863093476618, + "mirror_response_residual_rms": 0.043720037119289186, + "mirror_update_rms": 0.0053212781897893045, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.014644660940672627, + "step": 1264, + "train_examples": 10000, + "train_loss": 1.092687600326538 + }, + { + "epoch": 17, + "lr": 0.00286474508437579, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.025222243959948938, + "mirror_response_residual_rms": 0.03586358349597533, + "mirror_update_rms": 0.004563395395546671, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.009549150281252633, + "step": 1343, + "train_examples": 10000, + "train_loss": 1.050999007129669 + }, + { + "epoch": 18, + "lr": 0.001634902137174483, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.019423717576147297, + "mirror_response_residual_rms": 0.0277487644354511, + "mirror_update_rms": 0.003734861372460549, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.005449673790581611, + "step": 1422, + "train_examples": 10000, + "train_loss": 1.0021559915542602 + }, + { + "epoch": 19, + "lr": 0.000734152255572697, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.013885959625617894, + "mirror_response_residual_rms": 0.019797627171206474, + "mirror_update_rms": 0.002950896392911573, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.0024471741852423235, + "step": 1501, + "train_examples": 10000, + "train_loss": 0.9745530992507935 + }, + { + "epoch": 20, + "lr": 0.0001846748910729351, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.009382662629116802, + "mirror_response_residual_rms": 0.013384784612837958, + "mirror_update_rms": 0.0022900423996206074, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.0006155829702431171, + "step": 1580, + "train_examples": 10000, + "train_loss": 0.9762974742889404 + } + ], + "evaluation_protocol": { + "test_evaluations": 0, + "test_used_for_selection": false, + "validation_evaluations": 1 + }, + "final": { + "accuracy": 0.6232, + "epoch": 20, + "evaluation_split": "validation", + "finite": true, + "loss": 1.066288330078125, + "step": 1580 + }, + "hardware": { + "cuda_device_name": "NVIDIA GeForce GTX 1080", + "cuda_visible_devices": "5", + "device": "cuda", + "device_total_memory_bytes": 8507949056, + "peak_memory_allocated_bytes": 1730831872, + "peak_memory_reserved_bytes": 2501902336, + "torch_version": "2.3.1+cu118" + }, + "mirror_warmup": { + "first": { + "conv_batch_size": 1, + "mirror_response_residual_fraction": 1.4124661316345333, + "mirror_response_residual_rms": 1.90927576279183, + "mirror_update_rms": 0.25520769706611623, + "readout_batch_size": 1024 + }, + "last": { + "conv_batch_size": 1, + "mirror_response_residual_fraction": 0.26271250078043745, + "mirror_response_residual_rms": 0.3557588547676734, + "mirror_update_rms": 0.0907495165488641, + "readout_batch_size": 1024 + }, + "mean": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.6719071790962137, + "mirror_response_residual_rms": 0.9100978838204868, + "mirror_update_rms": 0.15685383656211757, + "readout_batch_size": 1024.0 + }, + "steps": 20 + }, + "protocol_family": "oral_a_cifar_local_resnet_development", + "provenance": { + "git_commit": "6206483780d75ecff61c396b7f1bf7e23e5b1ca5", + "git_tracked_dirty": false + }, + "schema_version": 1, + "split": { + "cifar_source_files": [ + { + "bytes": 31035704, + "path": "/home/yurenh2/sdrn/data/cifar-10-batches-py/data_batch_1", + "sha256": "54636561a3ce25bd3e19253c6b0d8538147b0ae398331ac4a2d86c6d987368cd" + }, + { + "bytes": 31035320, + "path": "/home/yurenh2/sdrn/data/cifar-10-batches-py/data_batch_2", + "sha256": "766b2cef9fbc745cf056b3152224f7cf77163b330ea9a15f9392beb8b89bc5a8" + }, + { + "bytes": 31035999, + "path": "/home/yurenh2/sdrn/data/cifar-10-batches-py/data_batch_3", + "sha256": "0f00d98ebfb30b3ec0ad19f9756dc2630b89003e10525f5e148445e82aa6a1f9" + }, + { + "bytes": 31035696, + "path": "/home/yurenh2/sdrn/data/cifar-10-batches-py/data_batch_4", + "sha256": "3f7bb240661948b8f4d53e36ec720d8306f5668bd0071dcb4e6c947f78e9682b" + }, + { + "bytes": 31035623, + "path": "/home/yurenh2/sdrn/data/cifar-10-batches-py/data_batch_5", + "sha256": "d91802434d8376bbaeeadf58a737e3a1b12ac839077e931237e0dcd43adcb154" + }, + { + "bytes": 31035526, + "path": "/home/yurenh2/sdrn/data/cifar-10-batches-py/test_batch", + "sha256": "f53d8d457504f7cff4ea9e021afcf0e0ad8e24a91f3fc42091b8adef61157831" + } + ], + "dataset": "cifar10", + "evaluation_split": "validation", + "input_layout": "NCHW", + "input_shape": [ + 3, + 32, + 32 + ], + "loader_seed": 0, + "normalization_mean": [ + 0.49140000343322754, + 0.4821999967098236, + 0.4465000033378601 + ], + "normalization_std": [ + 0.24699999392032623, + 0.2434999942779541, + 0.26159998774528503 + ], + "split_from_training_only": true, + "split_seed": 2027, + "test_examples": 10000, + "train_examples": 10000, + "training_augmentation": "random_crop_32_padding_4_zero_then_horizontal_flip_p0.5", + "validation_class_counts": { + "0": 500, + "1": 500, + "2": 500, + "3": 500, + "4": 500, + "5": 500, + "6": 500, + "7": 500, + "8": 500, + "9": 500 + }, + "validation_examples": 5000, + "validation_index_sha256": "8328b206a97c420e49e54e3eca4abe3274c4756b084355784ea3fb8059e4515b" + }, + "timing": { + "apical_warmup_wall_s": 0.0, + "evaluation_wall_s": 0.36531639099121094, + "mirror_warmup_wall_s": 0.5950455665588379, + "predictor_warmup_wall_s": 0.0, + "timing_excludes_data_loading_hashing_and_model_construction": true, + "total_timed_wall_s": 90.97272872924805, + "train_wall_s": 90.00931930541992 + }, + "work": { + "apical_macs_per_example": 40108672, + "causal_scalar_observations": 0, + "components": { + "apical_projection_macs": 8021734400000, + "apical_regression_macs": 0, + "bp_reverse_macs_estimate": 0, + "local_weight_correlation_macs": 8110208000000, + "mirror_feedback_prediction_macs": 4850843648, + "mirror_local_correlation_macs": 4850843648, + "mirror_response_macs": 4850843648, + "ordinary_forward_macs": 8110208000000, + "perturbation_forward_macs": 0, + "warmup_clean_forward_macs": 0 + }, + "definition": "multiply-accumulates in conv/linear maps; one local weight correlation equals one forward-weight MAC count; BP reverse is estimated as one weight-gradient plus one activation-gradient convolution per forward convolution; elementwise nonlinearities and optimizer arithmetic excluded", + "forward_macs_per_example": 40551040, + "logical_batch_loss_queries": 0, + "mirror_probe_examples": 119, + "mirror_readout_probe_examples": 121856, + "per_example_cross_entropy_terms": 0, + "total_clean_forward_examples": 200000, + "total_forward_equivalent_examples": 200000, + "total_macs_estimate": 24256702930944 + } +} diff --git a/results/residual_mirror_short/rrm_lr0.1.json b/results/residual_mirror_short/rrm_lr0.1.json new file mode 100644 index 0000000..6f54c21 --- /dev/null +++ b/results/residual_mirror_short/rrm_lr0.1.json @@ -0,0 +1,756 @@ +{ + "architecture": { + "adaptive_apical_parameters": 0, + "adaptive_feedback_parameters": 267904, + "base_width": 16, + "blocks_per_stage": 3, + "bn_eps": 1e-05, + "bn_momentum": 0.1, + "depth": 20, + "family": "CIFAR 6n+2 ResNet, option-A shortcuts", + "fixed_feedback_parameters": 0, + "fixed_traffic_coefficients": 0, + "forward_parameters": 269722, + "hidden_shapes": [ + [ + 16, + 32, + 32 + ], + [ + 16, + 32, + 32 + ], + [ + 16, + 32, + 32 + ], + [ + 16, + 32, + 32 + ], + [ + 16, + 32, + 32 + ], + [ + 16, + 32, + 32 + ], + [ + 16, + 32, + 32 + ], + [ + 32, + 16, + 16 + ], + [ + 32, + 16, + 16 + ], + [ + 32, + 16, + 16 + ], + [ + 32, + 16, + 16 + ], + [ + 32, + 16, + 16 + ], + [ + 32, + 16, + 16 + ], + [ + 64, + 8, + 8 + ], + [ + 64, + 8, + 8 + ], + [ + 64, + 8, + 8 + ], + [ + 64, + 8, + 8 + ], + [ + 64, + 8, + 8 + ], + [ + 64, + 8, + 8 + ] + ], + "normalization": "batchnorm", + "predictor_parameters": 0, + "residual_scale": 1.0, + "vectorizer_mode": null, + "vectorizer_parameters": 0 + }, + "args": { + "a_scale": 1.0, + "a_warmup_steps": 0, + "alignment_probe": 32, + "apical_calibration_mode": "unit_targets", + "apical_seed": null, + "augment_train": 1, + "batch_size": 128, + "bn_eps": 1e-05, + "bn_momentum": 0.1, + "data_dir": "/home/yurenh2/sdrn/data", + "depth": 20, + "device": "cuda", + "epochs": 20, + "eta_A": 0.01, + "eta_P": 0.01, + "eval_every": 0, + "eval_split": "validation", + "learn_P": 0, + "loader_seed": 0, + "lr": 0.1, + "lr_gamma": 0.1, + "lr_milestones": "100,150", + "lr_schedule": "cosine", + "max_steps": 0, + "mirror_batch_size": 1, + "mirror_eta": 0.1, + "mirror_every": 16, + "mirror_noise_std": 1.0, + "mirror_seed": 3000, + "mirror_warmup_steps": 20, + "mode": "rrm", + "momentum": 0.9, + "normalization": "batchnorm", + "nuisance_scale": 0.0, + "out": "results/residual_mirror_short/rrm_lr0.1.json", + "output_lr": 0.1, + "pert_directions": 1, + "pert_every": 4, + "pert_sigma": 0.01, + "perturb_seed": 1000, + "predictor_warmup_steps": 0, + "residual_scale": null, + "seed": 0, + "split_seed": 2027, + "train_limit": 10000, + "use_residual": 1, + "val_examples": 5000, + "vectorizer_mode": "spatial_template", + "warmup_epochs": 0, + "weight_decay": 0.0001, + "weight_scale": 1.0, + "width": 16 + }, + "calibration_metric_space": "local_parent_child_response_residual", + "counters": { + "apical_warmup_examples": 0, + "calibration_event_examples": 0, + "causal_scalar_observations": 0, + "logical_batch_loss_queries": 0, + "mirror_conv_examples": 119, + "mirror_events": 119, + "mirror_readout_examples": 121856, + "ordinary_examples": 200000, + "per_example_loss_terms": 0, + "perturbation_events": 0, + "perturbation_forward_examples": 0, + "predictor_warmup_examples": 0 + }, + "diagnostics": { + "early_third_mean": 0.9991278549035391, + "feedback_forward_cosine": [ + 0.9999154806137085, + 0.9999306797981262, + 0.9999545812606812, + 0.9999644160270691, + 0.9999690055847168, + 0.9999793767929077, + 0.9999752044677734, + 0.9999589920043945, + 0.9999452233314514, + 0.9999603033065796, + 0.9999622702598572, + 0.9999719858169556, + 0.9999478459358215, + 0.9997169375419617, + 0.9997090101242065, + 0.9997721910476685, + 0.9997538328170776, + 0.9998348951339722, + 0.9999200105667114 + ], + "feedback_forward_norm_ratio": [ + 0.9980469942092896, + 0.9989184737205505, + 0.9994211196899414, + 0.9998089671134949, + 0.9999980926513672, + 1.0001499652862549, + 1.0000603199005127, + 1.0000534057617188, + 0.9997336864471436, + 1.0000892877578735, + 1.000082015991211, + 1.000295639038086, + 1.0003461837768555, + 1.0006914138793945, + 1.0006706714630127, + 1.0003938674926758, + 1.0005216598510742, + 1.0005003213882446, + 0.9909704923629761 + ], + "innovation_negative_gradient_cosine": [ + 0.9988824725151062, + 0.9990805387496948, + 0.9991225004196167, + 0.9991109371185303, + 0.9992891550064087, + 0.9992815256118774, + 0.9993298649787903, + 0.9993308782577515, + 0.999434232711792, + 0.9993816614151001, + 0.9995160102844238, + 0.9995108842849731, + 0.9995355010032654, + 0.9993031024932861, + 0.99957275390625, + 0.9993000030517578, + 0.9997174739837646, + 0.9995701313018799, + 0.9999008178710938 + ], + "normalization_state": "training_batch_stats_without_running_update", + "raw_negative_gradient_cosine": [ + 0.9988824725151062, + 0.9990805387496948, + 0.9991225004196167, + 0.9991109371185303, + 0.9992891550064087, + 0.9992815256118774, + 0.9993298649787903, + 0.9993308782577515, + 0.999434232711792, + 0.9993816614151001, + 0.9995160102844238, + 0.9995108842849731, + 0.9995355010032654, + 0.9993031024932861, + 0.99957275390625, + 0.9993000030517578, + 0.9997174739837646, + 0.9995701313018799, + 0.9999008178710938 + ], + "teaching_negative_gradient_cosine": [ + 0.9988824725151062, + 0.9990805387496948, + 0.9991225004196167, + 0.9991109371185303, + 0.9992891550064087, + 0.9992815256118774, + 0.9993298649787903, + 0.9993308782577515, + 0.999434232711792, + 0.9993816614151001, + 0.9995160102844238, + 0.9995108842849731, + 0.9995355010032654, + 0.9993031024932861, + 0.99957275390625, + 0.9993000030517578, + 0.9997174739837646, + 0.9995701313018799, + 0.9999008178710938 + ], + "wall_s": 0.11099910736083984 + }, + "epochs": [ + { + "epoch": 1, + "lr": 0.1, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.2601846737040344, + "mirror_response_residual_rms": 0.3535099094948956, + "mirror_update_rms": 0.0794724507423343, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.1, + "step": 79, + "train_examples": 10000, + "train_loss": 2.6401633117675782 + }, + { + "epoch": 2, + "lr": 0.0993844170297569, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.2320270030739035, + "mirror_response_residual_rms": 0.3178695360002074, + "mirror_update_rms": 0.06566347399745359, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.0993844170297569, + "step": 158, + "train_examples": 10000, + "train_loss": 2.428320087814331 + }, + { + "epoch": 3, + "lr": 0.09755282581475769, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.1902612302737704, + "mirror_response_residual_rms": 0.261214385527155, + "mirror_update_rms": 0.052451028617311414, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.09755282581475769, + "step": 237, + "train_examples": 10000, + "train_loss": 2.5726836837768556 + }, + { + "epoch": 4, + "lr": 0.09455032620941839, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.17018607732039362, + "mirror_response_residual_rms": 0.23495350486028804, + "mirror_update_rms": 0.042288586572611774, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.09455032620941839, + "step": 316, + "train_examples": 10000, + "train_loss": 2.528295587730408 + }, + { + "epoch": 5, + "lr": 0.09045084971874738, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.15231621565559716, + "mirror_response_residual_rms": 0.21180007147453325, + "mirror_update_rms": 0.03544802639828195, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.09045084971874738, + "step": 395, + "train_examples": 10000, + "train_loss": 2.218377275466919 + }, + { + "epoch": 6, + "lr": 0.08535533905932738, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.14108329109632886, + "mirror_response_residual_rms": 0.19726323418619918, + "mirror_update_rms": 0.02993521390942853, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.08535533905932738, + "step": 474, + "train_examples": 10000, + "train_loss": 2.0937036754608154 + }, + { + "epoch": 7, + "lr": 0.07938926261462366, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.13834582525482075, + "mirror_response_residual_rms": 0.19534085520608319, + "mirror_update_rms": 0.025985602855510263, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.07938926261462366, + "step": 553, + "train_examples": 10000, + "train_loss": 2.218831299972534 + }, + { + "epoch": 8, + "lr": 0.07269952498697733, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.13345085752752778, + "mirror_response_residual_rms": 0.1909307264239541, + "mirror_update_rms": 0.0240043080261321, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.07269952498697733, + "step": 632, + "train_examples": 10000, + "train_loss": 2.274434162902832 + }, + { + "epoch": 9, + "lr": 0.06545084971874737, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.12035600897902161, + "mirror_response_residual_rms": 0.17304416237519363, + "mirror_update_rms": 0.02157844918317908, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.06545084971874737, + "step": 711, + "train_examples": 10000, + "train_loss": 1.8082092502593994 + }, + { + "epoch": 10, + "lr": 0.05782172325201155, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.1096965221826743, + "mirror_response_residual_rms": 0.15870117165633163, + "mirror_update_rms": 0.01960604709124913, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.05782172325201155, + "step": 790, + "train_examples": 10000, + "train_loss": 1.6496932834625244 + }, + { + "epoch": 11, + "lr": 0.05, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.10060992128674659, + "mirror_response_residual_rms": 0.14640697910423445, + "mirror_update_rms": 0.017701212220667454, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.05, + "step": 869, + "train_examples": 10000, + "train_loss": 1.5478480306625366 + }, + { + "epoch": 12, + "lr": 0.04217827674798845, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.09085091944110543, + "mirror_response_residual_rms": 0.13333263756453764, + "mirror_update_rms": 0.01605992521153775, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.04217827674798845, + "step": 948, + "train_examples": 10000, + "train_loss": 1.3823965042114257 + }, + { + "epoch": 13, + "lr": 0.03454915028125263, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.07902996300142369, + "mirror_response_residual_rms": 0.11615116335525824, + "mirror_update_rms": 0.014287858257776787, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.03454915028125263, + "step": 1027, + "train_examples": 10000, + "train_loss": 1.34172273979187 + }, + { + "epoch": 14, + "lr": 0.027300475013022664, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.06905828831434452, + "mirror_response_residual_rms": 0.10180783214913795, + "mirror_update_rms": 0.012587746597357776, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.027300475013022664, + "step": 1106, + "train_examples": 10000, + "train_loss": 1.2290157025337218 + }, + { + "epoch": 15, + "lr": 0.02061073738537635, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.05757729123066844, + "mirror_response_residual_rms": 0.08502710898907911, + "mirror_update_rms": 0.01099315299825323, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.02061073738537635, + "step": 1185, + "train_examples": 10000, + "train_loss": 1.1287344594955444 + }, + { + "epoch": 16, + "lr": 0.014644660940672627, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.04826044807946187, + "mirror_response_residual_rms": 0.07131268925507378, + "mirror_update_rms": 0.009481739617492971, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.014644660940672627, + "step": 1264, + "train_examples": 10000, + "train_loss": 1.0884003732681276 + }, + { + "epoch": 17, + "lr": 0.009549150281252633, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.03974752585157935, + "mirror_response_residual_rms": 0.05869449700782102, + "mirror_update_rms": 0.008195443513961818, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.009549150281252633, + "step": 1343, + "train_examples": 10000, + "train_loss": 1.0395743354797364 + }, + { + "epoch": 18, + "lr": 0.005449673790581611, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.0302168111419576, + "mirror_response_residual_rms": 0.044779826338416255, + "mirror_update_rms": 0.006658613012222196, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.005449673790581611, + "step": 1422, + "train_examples": 10000, + "train_loss": 0.9841103965759277 + }, + { + "epoch": 19, + "lr": 0.0024471741852423235, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.02212630175876517, + "mirror_response_residual_rms": 0.03275805770984623, + "mirror_update_rms": 0.00537444741444945, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.0024471741852423235, + "step": 1501, + "train_examples": 10000, + "train_loss": 0.9521100430488586 + }, + { + "epoch": 20, + "lr": 0.0006155829702431171, + "mirror": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.015118729649761752, + "mirror_response_residual_rms": 0.022381774527825898, + "mirror_update_rms": 0.004188984685734049, + "readout_batch_size": 1024.0 + }, + "output_lr": 0.0006155829702431171, + "step": 1580, + "train_examples": 10000, + "train_loss": 0.938697664642334 + } + ], + "evaluation_protocol": { + "test_evaluations": 0, + "test_used_for_selection": false, + "validation_evaluations": 1 + }, + "final": { + "accuracy": 0.6508, + "epoch": 20, + "evaluation_split": "validation", + "finite": true, + "loss": 1.01295, + "step": 1580 + }, + "hardware": { + "cuda_device_name": "NVIDIA GeForce GTX 1080", + "cuda_visible_devices": "7", + "device": "cuda", + "device_total_memory_bytes": 8507949056, + "peak_memory_allocated_bytes": 1730831872, + "peak_memory_reserved_bytes": 2501902336, + "torch_version": "2.3.1+cu118" + }, + "mirror_warmup": { + "first": { + "conv_batch_size": 1, + "mirror_response_residual_fraction": 1.4124661316345333, + "mirror_response_residual_rms": 1.90927576279183, + "mirror_update_rms": 0.25520769706611623, + "readout_batch_size": 1024 + }, + "last": { + "conv_batch_size": 1, + "mirror_response_residual_fraction": 0.26271250078043745, + "mirror_response_residual_rms": 0.3557588547676734, + "mirror_update_rms": 0.0907495165488641, + "readout_batch_size": 1024 + }, + "mean": { + "conv_batch_size": 1.0, + "mirror_response_residual_fraction": 0.6719071790962137, + "mirror_response_residual_rms": 0.9100978838204868, + "mirror_update_rms": 0.15685383656211757, + "readout_batch_size": 1024.0 + }, + "steps": 20 + }, + "protocol_family": "oral_a_cifar_local_resnet_development", + "provenance": { + "git_commit": "6206483780d75ecff61c396b7f1bf7e23e5b1ca5", + "git_tracked_dirty": false + }, + "schema_version": 1, + "split": { + "cifar_source_files": [ + { + "bytes": 31035704, + "path": "/home/yurenh2/sdrn/data/cifar-10-batches-py/data_batch_1", + "sha256": "54636561a3ce25bd3e19253c6b0d8538147b0ae398331ac4a2d86c6d987368cd" + }, + { + "bytes": 31035320, + "path": "/home/yurenh2/sdrn/data/cifar-10-batches-py/data_batch_2", + "sha256": "766b2cef9fbc745cf056b3152224f7cf77163b330ea9a15f9392beb8b89bc5a8" + }, + { + "bytes": 31035999, + "path": "/home/yurenh2/sdrn/data/cifar-10-batches-py/data_batch_3", + "sha256": "0f00d98ebfb30b3ec0ad19f9756dc2630b89003e10525f5e148445e82aa6a1f9" + }, + { + "bytes": 31035696, + "path": "/home/yurenh2/sdrn/data/cifar-10-batches-py/data_batch_4", + "sha256": "3f7bb240661948b8f4d53e36ec720d8306f5668bd0071dcb4e6c947f78e9682b" + }, + { + "bytes": 31035623, + "path": "/home/yurenh2/sdrn/data/cifar-10-batches-py/data_batch_5", + "sha256": "d91802434d8376bbaeeadf58a737e3a1b12ac839077e931237e0dcd43adcb154" + }, + { + "bytes": 31035526, + "path": "/home/yurenh2/sdrn/data/cifar-10-batches-py/test_batch", + "sha256": "f53d8d457504f7cff4ea9e021afcf0e0ad8e24a91f3fc42091b8adef61157831" + } + ], + "dataset": "cifar10", + "evaluation_split": "validation", + "input_layout": "NCHW", + "input_shape": [ + 3, + 32, + 32 + ], + "loader_seed": 0, + "normalization_mean": [ + 0.49140000343322754, + 0.4821999967098236, + 0.4465000033378601 + ], + "normalization_std": [ + 0.24699999392032623, + 0.2434999942779541, + 0.26159998774528503 + ], + "split_from_training_only": true, + "split_seed": 2027, + "test_examples": 10000, + "train_examples": 10000, + "training_augmentation": "random_crop_32_padding_4_zero_then_horizontal_flip_p0.5", + "validation_class_counts": { + "0": 500, + "1": 500, + "2": 500, + "3": 500, + "4": 500, + "5": 500, + "6": 500, + "7": 500, + "8": 500, + "9": 500 + }, + "validation_examples": 5000, + "validation_index_sha256": "8328b206a97c420e49e54e3eca4abe3274c4756b084355784ea3fb8059e4515b" + }, + "timing": { + "apical_warmup_wall_s": 0.0, + "evaluation_wall_s": 0.36290621757507324, + "mirror_warmup_wall_s": 0.630126953125, + "predictor_warmup_wall_s": 0.0, + "timing_excludes_data_loading_hashing_and_model_construction": true, + "total_timed_wall_s": 93.04161047935486, + "train_wall_s": 92.04554319381714 + }, + "work": { + "apical_macs_per_example": 40108672, + "causal_scalar_observations": 0, + "components": { + "apical_projection_macs": 8021734400000, + "apical_regression_macs": 0, + "bp_reverse_macs_estimate": 0, + "local_weight_correlation_macs": 8110208000000, + "mirror_feedback_prediction_macs": 4850843648, + "mirror_local_correlation_macs": 4850843648, + "mirror_response_macs": 4850843648, + "ordinary_forward_macs": 8110208000000, + "perturbation_forward_macs": 0, + "warmup_clean_forward_macs": 0 + }, + "definition": "multiply-accumulates in conv/linear maps; one local weight correlation equals one forward-weight MAC count; BP reverse is estimated as one weight-gradient plus one activation-gradient convolution per forward convolution; elementwise nonlinearities and optimizer arithmetic excluded", + "forward_macs_per_example": 40551040, + "logical_batch_loss_queries": 0, + "mirror_probe_examples": 119, + "mirror_readout_probe_examples": 121856, + "per_example_cross_entropy_terms": 0, + "total_clean_forward_examples": 200000, + "total_forward_equivalent_examples": 200000, + "total_macs_estimate": 24256702930944 + } +} diff --git a/results/residual_mirror_short_gate.json b/results/residual_mirror_short_gate.json new file mode 100644 index 0000000..1c84e29 --- /dev/null +++ b/results/residual_mirror_short_gate.json @@ -0,0 +1,77 @@ +{ + "checks": { + "accuracy_at_least_0.65": true, + "both_records_finite": true, + "candidate_selected": true, + "early_alignment_at_least_0.50": true, + "macs_at_most_1.15x_bp": true, + "within_10_points_of_bp": true, + "zero_task_loss_queries": true + }, + "confirmation_test_seeds_touched": false, + "matched_bp": { + "accuracy": 0.7494, + "total_macs": 24330624000000 + }, + "protocol": "residual_response_mirror_short_v1", + "review_score_after": 5, + "review_score_before": 5, + "rows": [ + { + "accuracy": 0.6232, + "all_layer_alignment": 0.9996919380991083, + "early_third_alignment": 0.9996043642361959, + "finite": true, + "logical_batch_loss_queries": 0, + "loss": 1.066288330078125, + "lr": 0.03, + "max_feedback_forward_norm_ratio": 1.0002508163452148, + "mean_feedback_forward_cosine": 0.9999703388465079, + "min_feedback_forward_norm_ratio": 0.9902048110961914, + "mirror_events": 119, + "path": "results/residual_mirror_short/rrm_lr0.03.json", + "peak_memory_allocated_bytes": 1730831872, + "source_commit": "6206483780d75ecff61c396b7f1bf7e23e5b1ca5", + "total_macs": 24256702930944, + "wall_s": 90.97272872924805 + }, + { + "accuracy": 0.6508, + "all_layer_alignment": 0.9993773918402823, + "early_third_alignment": 0.9991278549035391, + "finite": true, + "logical_batch_loss_queries": 0, + "loss": 1.01295, + "lr": 0.1, + "max_feedback_forward_norm_ratio": 1.0006914138793945, + "mean_feedback_forward_cosine": 0.9999022232858759, + "min_feedback_forward_norm_ratio": 0.9909704923629761, + "mirror_events": 119, + "path": "results/residual_mirror_short/rrm_lr0.1.json", + "peak_memory_allocated_bytes": 1730831872, + "source_commit": "6206483780d75ecff61c396b7f1bf7e23e5b1ca5", + "total_macs": 24256702930944, + "wall_s": 93.04161047935486 + } + ], + "score_change_rule": "inherited short baseline cannot raise score", + "selected_rrm": { + "accuracy": 0.6508, + "all_layer_alignment": 0.9993773918402823, + "early_third_alignment": 0.9991278549035391, + "finite": true, + "logical_batch_loss_queries": 0, + "loss": 1.01295, + "lr": 0.1, + "max_feedback_forward_norm_ratio": 1.0006914138793945, + "mean_feedback_forward_cosine": 0.9999022232858759, + "min_feedback_forward_norm_ratio": 0.9909704923629761, + "mirror_events": 119, + "path": "results/residual_mirror_short/rrm_lr0.1.json", + "peak_memory_allocated_bytes": 1730831872, + "source_commit": "6206483780d75ecff61c396b7f1bf7e23e5b1ca5", + "total_macs": 24256702930944, + "wall_s": 93.04161047935486 + }, + "status": "passed" +} |
