diff --git "a/applied/_index/APPLIED_RUNS_INDEX.json" "b/applied/_index/APPLIED_RUNS_INDEX.json" --- "a/applied/_index/APPLIED_RUNS_INDEX.json" +++ "b/applied/_index/APPLIED_RUNS_INDEX.json" @@ -1073,131 +1073,536 @@ } }, { - "key": "opt02_seenn_bayes/conf_temp_4", - "tag": "conf_temp_4", - "script": "opt02_seenn_bayes.py", - "hf_folder": "applied/opt02_seenn_bayes", - "phase": 7, + "key": "opt06_uticket_prune/prune_readout", + "tag": "prune_readout", + "script": "opt06_uticket_prune.py", + "hf_folder": "applied/opt06_uticket_prune", + "phase": 9, "status": "ok", - "duration_s": 616.4, + "duration_s": 582.6, "optimization_summary": { - "optimization": "seenn_early_exit_bayesian_fusion", - "research_section": "2. Temporal-dimension reduction", - "reference": "SEENN: Towards Temporal Spiking Early-Exit Neural Networks, arXiv:2304.01230 (NeurIPS 2023); BayesianSpikeFusion, Front. Neurosci. 2024 (PMC11330889)", - "script": "opt02_seenn_bayes.py", + "optimization": "uticket_workload_balanced_pruning", + "research_section": "3. Spike sparsity & MAC->AC restructuring", + "reference": "u-Ticket: Workload-Balanced Pruning for Sparse SNNs, arXiv:2302.06746", + "script": "opt06_uticket_prune.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.9278571428571428, - "test_target_acc": 0.9278571428571428, - "best_val_acc": 0.9264285714285714 + "test_acc": 0.4135714285714286, + "test_target_acc": 0.4135714285714286, + "best_val_acc": 0.9335714285714286 }, "efficiency": { "spikes_per_inference": null, - "syn_ops_per_inference": 3638325.674512987, - "dense_macs_per_inference": 123955200.0, - "event_fraction_of_dense": 0.029351940656890448, - "syn_ops_per_inference_dense_fanout": 65214943.461038955, - "weight_sparsity": 0.7324707846410685, - "effective_ticks": 4.575714285714286, + "syn_ops_per_inference": 104257.90097402596, + "dense_macs_per_inference": 122675200.0, + "event_fraction_of_dense": 0.0008498694191982239, + "syn_ops_per_inference_dense_fanout": 65062433.861201294, + "weight_sparsity": 0.9917057400354758, + "effective_ticks": 11.177142857142858, "num_ticks": 100, "membrane_bytes_per_inference": 451200, "weight_bytes": 4907008, "input_raster_bytes": null, - "train_ms_per_step": 187.23959922790527, - "infer_ms_per_sample": 0.6532642859310361, - "peak_vram_mb": 10914.677248, - "syn_ops_per_inference_early_exit": 1640777.8961038962 + "train_ms_per_step": 94.6925303999933, + "infer_ms_per_sample": 0.6165744149999597, + "peak_vram_mb": 2836.776448, + "weight_bytes_nonzero": 40703 }, "optimization_specific": { - "seenn_enabled": true, - "bayes_fusion_enabled": true, - "conf_threshold": 0.9, - "conf_temp": 4.0, - "anytime_fracs": "0.25,0.5,1.0", - "anytime_weights": "0.3,0.3,1.0", - "ic_loss_weight": 0.3, - "ic_layer": 0, - "mean_exit_tick": 4.575714285714286, - "mean_exit_ms": 228.78571428571428, - "ticks_saved_frac": 0.9542428571428572, - "exit_tick_histogram": { - "counts": [ - 1339, - 54, - 5, - 2, - 0, - 0, - 0, - 0, - 0, - 0 - ], - "edges": [ - 0.0, - 10.0, - 20.0, - 30.0, - 40.0, - 50.0, - 60.0, - 70.0, - 80.0, - 90.0, - 100.0 - ] + "pruning_enabled": true, + "balance_enabled": true, + "rounds": 5, + "final_sparsity": 0.9917049248747913, + "nnz_per_layer": { + "layer0": 640, + "layer1": 5120, + "layer2": 4096, + "readout": 320 }, - "acc_fc_only": 0.9278571428571428, - "acc_ic_only": 0.9364285714285714, - "acc_fused": 0.935, - "acc_early_exit": 0.25214285714285717, - "syn_ops_saved_frac": 0.5490294044874021, - "syn_per_inference_early_exit": 1640777.8961038962, - "anytime_acc": { - "0.25": 0.62, - "0.5": 0.8178571428571428, - "0.75": 0.8885714285714286, - "1.0": 0.9278571428571428 + "pe_stats_per_layer": { + "layer0": { + "mean_nnz": 2.0, + "max_nnz": 2.0, + "min_nnz": 2.0, + "total_nnz": 640, + "numel": 921600, + "sparsity": 0.9993055555555556, + "utilization": 1.0 + }, + "layer1": { + "mean_nnz": 10.0, + "max_nnz": 10.0, + "min_nnz": 10.0, + "total_nnz": 5120, + "numel": 163840, + "sparsity": 0.96875, + "utilization": 1.0 + }, + "layer2": { + "mean_nnz": 16.0, + "max_nnz": 16.0, + "min_nnz": 16.0, + "total_nnz": 4096, + "numel": 131072, + "sparsity": 0.96875, + "utilization": 1.0 + }, + "readout": { + "mean_nnz": 8.0, + "max_nnz": 8.0, + "min_nnz": 8.0, + "total_nnz": 320, + "numel": 10240, + "sparsity": 0.96875, + "utilization": 1.0 + } }, - "pareto": [ - { - "conf": 0.5, - "mean_ticks": 3.902142857142857, - "acc": 0.22357142857142856, - "syn_ops": 1613080.9172077922 - }, - { - "conf": 0.7, - "mean_ticks": 4.2371428571428575, - "acc": 0.23857142857142857, - "syn_ops": 1626802.577922078 - }, - { - "conf": 0.8, - "mean_ticks": 4.365, - "acc": 0.245, - "syn_ops": 1632438.9545454546 - }, - { - "conf": 0.9, - "mean_ticks": 4.575714285714286, - "acc": 0.25214285714285717, - "syn_ops": 1640777.924512987 - }, - { - "conf": 0.95, - "mean_ticks": 4.854285714285714, - "acc": 0.26071428571428573, - "syn_ops": 1651982.8961038962 - }, - { - "conf": 0.99, - "mean_ticks": 6.857857142857143, - "acc": 0.3357142857142857, - "syn_ops": 1728711.8571428573 + "pe_utilization_final": 1.0, + "pe_utilization_lth": 0.5667460317460318, + "pe_utilization_uticket": 1.0, + "latency_proxy_dense": 3968, + "latency_proxy_final": 36, + "latency_reduction_frac": 0.9909274193548387, + "rounds_history": [ + { + "round": 1, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 36.0, + "max_nnz": 47.0, + "min_nnz": 25.0, + "total_nnz": 11520, + "numel": 921600, + "sparsity": 0.9875, + "utilization": 0.7659574468085106 + }, + "uticket": { + "mean_nnz": 36.0, + "max_nnz": 36.0, + "min_nnz": 36.0, + "total_nnz": 11520, + "numel": 921600, + "sparsity": 0.9875, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 160.0, + "max_nnz": 180.0, + "min_nnz": 123.0, + "total_nnz": 81920, + "numel": 163840, + "sparsity": 0.5, + "utilization": 0.8888888888888888 + }, + "uticket": { + "mean_nnz": 160.0, + "max_nnz": 160.0, + "min_nnz": 160.0, + "total_nnz": 81920, + "numel": 163840, + "sparsity": 0.5, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 256.0, + "max_nnz": 288.0, + "min_nnz": 210.0, + "total_nnz": 65536, + "numel": 131072, + "sparsity": 0.5, + "utilization": 0.8888888888888888 + }, + "uticket": { + "mean_nnz": 256.0, + "max_nnz": 256.0, + "min_nnz": 256.0, + "total_nnz": 65536, + "numel": 131072, + "sparsity": 0.5, + "utilization": 1.0 + } + }, + "readout": { + "lth": { + "mean_nnz": 128.0, + "max_nnz": 143.0, + "min_nnz": 117.0, + "total_nnz": 5120, + "numel": 10240, + "sparsity": 0.5, + "utilization": 0.8951048951048951 + }, + "uticket": { + "mean_nnz": 128.0, + "max_nnz": 128.0, + "min_nnz": 128.0, + "total_nnz": 5120, + "numel": 10240, + "sparsity": 0.5, + "utilization": 1.0 + } + } + }, + "latency_proxy": 580, + "sparsity": 0.8662353923205343 + }, + { + "round": 2, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 18.0, + "max_nnz": 28.0, + "min_nnz": 10.0, + "total_nnz": 5760, + "numel": 921600, + "sparsity": 0.99375, + "utilization": 0.6428571428571429 + }, + "uticket": { + "mean_nnz": 18.0, + "max_nnz": 18.0, + "min_nnz": 18.0, + "total_nnz": 5760, + "numel": 921600, + "sparsity": 0.99375, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 80.0, + "max_nnz": 104.0, + "min_nnz": 33.0, + "total_nnz": 40960, + "numel": 163840, + "sparsity": 0.75, + "utilization": 0.7692307692307693 + }, + "uticket": { + "mean_nnz": 80.0, + "max_nnz": 80.0, + "min_nnz": 80.0, + "total_nnz": 40960, + "numel": 163840, + "sparsity": 0.75, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 128.0, + "max_nnz": 161.0, + "min_nnz": 93.0, + "total_nnz": 32768, + "numel": 131072, + "sparsity": 0.75, + "utilization": 0.7950310559006211 + }, + "uticket": { + "mean_nnz": 128.0, + "max_nnz": 128.0, + "min_nnz": 128.0, + "total_nnz": 32768, + "numel": 131072, + "sparsity": 0.75, + "utilization": 1.0 + } + }, + "readout": { + "lth": { + "mean_nnz": 64.0, + "max_nnz": 74.0, + "min_nnz": 56.0, + "total_nnz": 2560, + "numel": 10240, + "sparsity": 0.75, + "utilization": 0.8648648648648649 + }, + "uticket": { + "mean_nnz": 64.0, + "max_nnz": 64.0, + "min_nnz": 64.0, + "total_nnz": 2560, + "numel": 10240, + "sparsity": 0.75, + "utilization": 1.0 + } + } + }, + "latency_proxy": 290, + "sparsity": 0.9331176961602671 + }, + { + "round": 3, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 9.0, + "max_nnz": 18.0, + "min_nnz": 1.0, + "total_nnz": 2880, + "numel": 921600, + "sparsity": 0.996875, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 9.0, + "max_nnz": 9.0, + "min_nnz": 9.0, + "total_nnz": 2880, + "numel": 921600, + "sparsity": 0.996875, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 40.0, + "max_nnz": 64.0, + "min_nnz": 2.0, + "total_nnz": 20480, + "numel": 163840, + "sparsity": 0.875, + "utilization": 0.625 + }, + "uticket": { + "mean_nnz": 40.0, + "max_nnz": 40.0, + "min_nnz": 40.0, + "total_nnz": 20480, + "numel": 163840, + "sparsity": 0.875, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 64.0, + "max_nnz": 85.0, + "min_nnz": 38.0, + "total_nnz": 16384, + "numel": 131072, + "sparsity": 0.875, + "utilization": 0.7529411764705882 + }, + "uticket": { + "mean_nnz": 64.0, + "max_nnz": 64.0, + "min_nnz": 64.0, + "total_nnz": 16384, + "numel": 131072, + "sparsity": 0.875, + "utilization": 1.0 + } + }, + "readout": { + "lth": { + "mean_nnz": 32.0, + "max_nnz": 38.0, + "min_nnz": 26.0, + "total_nnz": 1280, + "numel": 10240, + "sparsity": 0.875, + "utilization": 0.8421052631578947 + }, + "uticket": { + "mean_nnz": 32.0, + "max_nnz": 32.0, + "min_nnz": 32.0, + "total_nnz": 1280, + "numel": 10240, + "sparsity": 0.875, + "utilization": 1.0 + } + } + }, + "latency_proxy": 145, + "sparsity": 0.9665588480801336 + }, + { + "round": 4, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 4.5, + "max_nnz": 9.0, + "min_nnz": 0.0, + "total_nnz": 1440, + "numel": 921600, + "sparsity": 0.9984375, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 4.0, + "max_nnz": 4.0, + "min_nnz": 4.0, + "total_nnz": 1280, + "numel": 921600, + "sparsity": 0.9986111111111111, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 20.0, + "max_nnz": 33.0, + "min_nnz": 0.0, + "total_nnz": 10240, + "numel": 163840, + "sparsity": 0.9375, + "utilization": 0.6060606060606061 + }, + "uticket": { + "mean_nnz": 20.0, + "max_nnz": 20.0, + "min_nnz": 20.0, + "total_nnz": 10240, + "numel": 163840, + "sparsity": 0.9375, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 32.0, + "max_nnz": 45.0, + "min_nnz": 12.0, + "total_nnz": 8192, + "numel": 131072, + "sparsity": 0.9375, + "utilization": 0.7111111111111111 + }, + "uticket": { + "mean_nnz": 32.0, + "max_nnz": 32.0, + "min_nnz": 32.0, + "total_nnz": 8192, + "numel": 131072, + "sparsity": 0.9375, + "utilization": 1.0 + } + }, + "readout": { + "lth": { + "mean_nnz": 16.0, + "max_nnz": 28.0, + "min_nnz": 11.0, + "total_nnz": 640, + "numel": 10240, + "sparsity": 0.9375, + "utilization": 0.5714285714285714 + }, + "uticket": { + "mean_nnz": 16.0, + "max_nnz": 16.0, + "min_nnz": 16.0, + "total_nnz": 640, + "numel": 10240, + "sparsity": 0.9375, + "utilization": 1.0 + } + } + }, + "latency_proxy": 72, + "sparsity": 0.9834098497495827 + }, + { + "round": 5, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 2.0, + "max_nnz": 4.0, + "min_nnz": 0.0, + "total_nnz": 640, + "numel": 921600, + "sparsity": 0.9993055555555556, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 2.0, + "max_nnz": 2.0, + "min_nnz": 2.0, + "total_nnz": 640, + "numel": 921600, + "sparsity": 0.9993055555555556, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 10.0, + "max_nnz": 18.0, + "min_nnz": 0.0, + "total_nnz": 5120, + "numel": 163840, + "sparsity": 0.96875, + "utilization": 0.5555555555555556 + }, + "uticket": { + "mean_nnz": 10.0, + "max_nnz": 10.0, + "min_nnz": 10.0, + "total_nnz": 5120, + "numel": 163840, + "sparsity": 0.96875, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 16.0, + "max_nnz": 25.0, + "min_nnz": 0.0, + "total_nnz": 4096, + "numel": 131072, + "sparsity": 0.96875, + "utilization": 0.64 + }, + "uticket": { + "mean_nnz": 16.0, + "max_nnz": 16.0, + "min_nnz": 16.0, + "total_nnz": 4096, + "numel": 131072, + "sparsity": 0.96875, + "utilization": 1.0 + } + }, + "readout": { + "lth": { + "mean_nnz": 8.0, + "max_nnz": 14.0, + "min_nnz": 3.0, + "total_nnz": 320, + "numel": 10240, + "sparsity": 0.96875, + "utilization": 0.5714285714285714 + }, + "uticket": { + "mean_nnz": 8.0, + "max_nnz": 8.0, + "min_nnz": 8.0, + "total_nnz": 320, + "numel": 10240, + "sparsity": 0.96875, + "utilization": 1.0 + } + } + }, + "latency_proxy": 36, + "sparsity": 0.9917049248747913 } - ] + ], + "latency_proxy_lth": 61, + "latency_reduction_vs_lth_frac": 0.4098360655737705 }, "config": { "method": "harmonic_matched", @@ -1237,163 +1642,321 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "baseline": false, - "bench": true, - "bench_warmup": 10, - "bench_iters": 50, - "bench_train_steps": 20, - "anytime_fracs": "0.25,0.5,1.0", - "anytime_weights": "0.3,0.3,1.0", - "ic_layer": 0, - "ic_loss_weight": 0.3, - "bayes_fusion": true, - "fusion_prior_weight": 1.0, - "conf_threshold": 0.9, - "conf_temp": 4.0, - "conf_sweep": "0.5,0.7,0.8,0.9,0.95,0.99", - "hub_dir": "/kaggle/working/hub/opt02_seenn_bayes/conf_temp_4", + "hub_dir": "/kaggle/working/hub/opt06_uticket_prune/prune_readout", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, "no_resume": true, "selftest": false, + "baseline": false, + "bench": true, + "bench_warmup": 10, + "bench_iters": 50, + "bench_train_steps": 20, + "prune_rounds": 5, + "prune_rate": 0.5, + "rewind_epoch": 1, + "balance": true, + "balance_iters": 2, + "prune_readout": true, "sampling_rate": 250.0 } }, "zip": { - "name": "conf_temp_4.zip", + "name": "prune_readout.zip", "files": 81, - "bytes": 12640717, - "MB": 12.6, - "sha256": "a9f16ef1db8ee6c4c51113bece56694dc7dd17feb4b73d57c99a3dad1b66ca71" + "bytes": 8103773, + "MB": 8.1, + "sha256": "71a4d931df327896a7f5538c47ae56a8c6d2a773250c37ea42d325f453d2355a" } }, { - "key": "opt02_seenn_bayes/fusion_prior_2x", - "tag": "fusion_prior_2x", - "script": "opt02_seenn_bayes.py", - "hf_folder": "applied/opt02_seenn_bayes", - "phase": 7, + "key": "opt06_uticket_prune/prune_rounds_3", + "tag": "prune_rounds_3", + "script": "opt06_uticket_prune.py", + "hf_folder": "applied/opt06_uticket_prune", + "phase": 9, "status": "ok", - "duration_s": 609.0, + "duration_s": 320.7, "optimization_summary": { - "optimization": "seenn_early_exit_bayesian_fusion", - "research_section": "2. Temporal-dimension reduction", - "reference": "SEENN: Towards Temporal Spiking Early-Exit Neural Networks, arXiv:2304.01230 (NeurIPS 2023); BayesianSpikeFusion, Front. Neurosci. 2024 (PMC11330889)", - "script": "opt02_seenn_bayes.py", + "optimization": "uticket_workload_balanced_pruning", + "research_section": "3. Spike sparsity & MAC->AC restructuring", + "reference": "u-Ticket: Workload-Balanced Pruning for Sparse SNNs, arXiv:2302.06746", + "script": "opt06_uticket_prune.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.9271428571428572, - "test_target_acc": 0.9271428571428572, - "best_val_acc": 0.9307142857142857 + "test_acc": 0.9057142857142857, + "test_target_acc": 0.9057142857142857, + "best_val_acc": 0.9407142857142857 }, "efficiency": { "spikes_per_inference": null, - "syn_ops_per_inference": 3657479.1964285714, - "dense_macs_per_inference": 123955200.0, - "event_fraction_of_dense": 0.02950646036978337, - "syn_ops_per_inference_dense_fanout": 65234096.98295455, - "weight_sparsity": 0.7324707846410685, - "effective_ticks": 7.784285714285715, + "syn_ops_per_inference": 550376.6811079546, + "dense_macs_per_inference": 122675200.0, + "event_fraction_of_dense": 0.004486454320905567, + "syn_ops_per_inference_dense_fanout": 65058668.56331169, + "weight_sparsity": 0.9592550083472454, + "effective_ticks": 13.450714285714286, "num_ticks": 100, "membrane_bytes_per_inference": 451200, "weight_bytes": 4907008, "input_raster_bytes": null, - "train_ms_per_step": 185.8240008354187, - "infer_ms_per_sample": 0.6510175053198733, - "peak_vram_mb": 10914.386944, - "syn_ops_per_inference_early_exit": 1785947.325487013 + "train_ms_per_step": 94.40872569999783, + "infer_ms_per_sample": 0.6202037340624855, + "peak_vram_mb": 2837.544448, + "weight_bytes_nonzero": 160314 }, "optimization_specific": { - "seenn_enabled": true, - "bayes_fusion_enabled": true, - "conf_threshold": 0.9, - "conf_temp": 2.0, - "anytime_fracs": "0.25,0.5,1.0", - "anytime_weights": "0.3,0.3,1.0", - "ic_loss_weight": 0.3, - "ic_layer": 0, - "mean_exit_tick": 7.784285714285715, - "mean_exit_ms": 389.2142857142857, - "ticks_saved_frac": 0.9221571428571429, - "exit_tick_histogram": { - "counts": [ - 1048, - 306, - 34, - 11, - 1, - 0, - 0, - 0, - 0, - 0 - ], - "edges": [ - 0.0, - 10.0, - 20.0, - 30.0, - 40.0, - 50.0, - 60.0, - 70.0, - 80.0, - 90.0, - 100.0 - ] + "pruning_enabled": true, + "balance_enabled": true, + "rounds": 3, + "final_sparsity": 0.9673295454545454, + "nnz_per_layer": { + "layer0": 2880, + "layer1": 20480, + "layer2": 16384 }, - "acc_fc_only": 0.9271428571428572, - "acc_ic_only": 0.9357142857142857, - "acc_fused": 0.9357142857142857, - "acc_early_exit": 0.37714285714285717, - "syn_ops_saved_frac": 0.5116999360567951, - "syn_per_inference_early_exit": 1785947.325487013, - "anytime_acc": { - "0.25": 0.6264285714285714, - "0.5": 0.8178571428571428, - "0.75": 0.8921428571428571, - "1.0": 0.9271428571428572 + "pe_stats_per_layer": { + "layer0": { + "mean_nnz": 9.0, + "max_nnz": 9.0, + "min_nnz": 9.0, + "total_nnz": 2880, + "numel": 921600, + "sparsity": 0.996875, + "utilization": 1.0 + }, + "layer1": { + "mean_nnz": 40.0, + "max_nnz": 40.0, + "min_nnz": 40.0, + "total_nnz": 20480, + "numel": 163840, + "sparsity": 0.875, + "utilization": 1.0 + }, + "layer2": { + "mean_nnz": 64.0, + "max_nnz": 64.0, + "min_nnz": 64.0, + "total_nnz": 16384, + "numel": 131072, + "sparsity": 0.875, + "utilization": 1.0 + } }, - "pareto": [ - { - "conf": 0.5, - "mean_ticks": 4.289285714285715, - "acc": 0.24571428571428572, - "syn_ops": 1647281.4212662338 - }, - { - "conf": 0.7, - "mean_ticks": 4.978571428571429, - "acc": 0.275, - "syn_ops": 1678001.2727272727 - }, - { - "conf": 0.8, - "mean_ticks": 5.852857142857143, - "acc": 0.3028571428571429, - "syn_ops": 1714795.320616883 - }, - { - "conf": 0.9, - "mean_ticks": 7.785, - "acc": 0.37714285714285717, - "syn_ops": 1785960.5300324673 - }, - { - "conf": 0.95, - "mean_ticks": 10.002142857142857, - "acc": 0.45071428571428573, - "syn_ops": 1850843.49512987 - }, - { - "conf": 0.99, - "mean_ticks": 14.499285714285714, - "acc": 0.5742857142857143, - "syn_ops": 1959679.1558441557 + "pe_utilization_final": 1.0, + "pe_utilization_lth": 0.623062015503876, + "pe_utilization_uticket": 1.0, + "latency_proxy_dense": 3712, + "latency_proxy_final": 113, + "latency_reduction_frac": 0.9695581896551724, + "rounds_history": [ + { + "round": 1, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 36.0, + "max_nnz": 49.0, + "min_nnz": 25.0, + "total_nnz": 11520, + "numel": 921600, + "sparsity": 0.9875, + "utilization": 0.7346938775510204 + }, + "uticket": { + "mean_nnz": 36.0, + "max_nnz": 36.0, + "min_nnz": 36.0, + "total_nnz": 11520, + "numel": 921600, + "sparsity": 0.9875, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 160.0, + "max_nnz": 190.0, + "min_nnz": 93.0, + "total_nnz": 81920, + "numel": 163840, + "sparsity": 0.5, + "utilization": 0.8421052631578947 + }, + "uticket": { + "mean_nnz": 160.0, + "max_nnz": 160.0, + "min_nnz": 160.0, + "total_nnz": 81920, + "numel": 163840, + "sparsity": 0.5, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 256.0, + "max_nnz": 297.0, + "min_nnz": 193.0, + "total_nnz": 65536, + "numel": 131072, + "sparsity": 0.5, + "utilization": 0.8619528619528619 + }, + "uticket": { + "mean_nnz": 256.0, + "max_nnz": 256.0, + "min_nnz": 256.0, + "total_nnz": 65536, + "numel": 131072, + "sparsity": 0.5, + "utilization": 1.0 + } + } + }, + "latency_proxy": 452, + "sparsity": 0.8693181818181819 + }, + { + "round": 2, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 18.0, + "max_nnz": 29.0, + "min_nnz": 8.0, + "total_nnz": 5760, + "numel": 921600, + "sparsity": 0.99375, + "utilization": 0.6206896551724138 + }, + "uticket": { + "mean_nnz": 18.0, + "max_nnz": 18.0, + "min_nnz": 18.0, + "total_nnz": 5760, + "numel": 921600, + "sparsity": 0.99375, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 80.0, + "max_nnz": 115.0, + "min_nnz": 28.0, + "total_nnz": 40960, + "numel": 163840, + "sparsity": 0.75, + "utilization": 0.6956521739130435 + }, + "uticket": { + "mean_nnz": 80.0, + "max_nnz": 80.0, + "min_nnz": 80.0, + "total_nnz": 40960, + "numel": 163840, + "sparsity": 0.75, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 128.0, + "max_nnz": 157.0, + "min_nnz": 99.0, + "total_nnz": 32768, + "numel": 131072, + "sparsity": 0.75, + "utilization": 0.8152866242038217 + }, + "uticket": { + "mean_nnz": 128.0, + "max_nnz": 128.0, + "min_nnz": 128.0, + "total_nnz": 32768, + "numel": 131072, + "sparsity": 0.75, + "utilization": 1.0 + } + } + }, + "latency_proxy": 226, + "sparsity": 0.9346590909090909 + }, + { + "round": 3, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 9.0, + "max_nnz": 18.0, + "min_nnz": 1.0, + "total_nnz": 2880, + "numel": 921600, + "sparsity": 0.996875, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 9.0, + "max_nnz": 9.0, + "min_nnz": 9.0, + "total_nnz": 2880, + "numel": 921600, + "sparsity": 0.996875, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 40.0, + "max_nnz": 64.0, + "min_nnz": 6.0, + "total_nnz": 20480, + "numel": 163840, + "sparsity": 0.875, + "utilization": 0.625 + }, + "uticket": { + "mean_nnz": 40.0, + "max_nnz": 40.0, + "min_nnz": 40.0, + "total_nnz": 20480, + "numel": 163840, + "sparsity": 0.875, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 64.0, + "max_nnz": 86.0, + "min_nnz": 45.0, + "total_nnz": 16384, + "numel": 131072, + "sparsity": 0.875, + "utilization": 0.7441860465116279 + }, + "uticket": { + "mean_nnz": 64.0, + "max_nnz": 64.0, + "min_nnz": 64.0, + "total_nnz": 16384, + "numel": 131072, + "sparsity": 0.875, + "utilization": 1.0 + } + } + }, + "latency_proxy": 113, + "sparsity": 0.9673295454545454 } - ] + ], + "latency_proxy_lth": 168, + "latency_reduction_vs_lth_frac": 0.32738095238095233 }, "config": { "method": "harmonic_matched", @@ -1433,163 +1996,656 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "baseline": false, - "bench": true, - "bench_warmup": 10, - "bench_iters": 50, - "bench_train_steps": 20, - "anytime_fracs": "0.25,0.5,1.0", - "anytime_weights": "0.3,0.3,1.0", - "ic_layer": 0, - "ic_loss_weight": 0.3, - "bayes_fusion": true, - "fusion_prior_weight": 2.0, - "conf_threshold": 0.9, - "conf_temp": 2.0, - "conf_sweep": "0.5,0.7,0.8,0.9,0.95,0.99", - "hub_dir": "/kaggle/working/hub/opt02_seenn_bayes/fusion_prior_2x", + "hub_dir": "/kaggle/working/hub/opt06_uticket_prune/prune_rounds_3", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, "no_resume": true, "selftest": false, + "baseline": false, + "bench": true, + "bench_warmup": 10, + "bench_iters": 50, + "bench_train_steps": 20, + "prune_rounds": 3, + "prune_rate": 0.5, + "rewind_epoch": 1, + "balance": true, + "balance_iters": 2, + "prune_readout": false, "sampling_rate": 250.0 } }, "zip": { - "name": "fusion_prior_2x.zip", + "name": "prune_rounds_3.zip", "files": 81, - "bytes": 12637636, - "MB": 12.6, - "sha256": "fc6050fb762f5c1e2b45dd261c259362087f7598c1df1884b7b7b97c019e13f4" + "bytes": 8378826, + "MB": 8.4, + "sha256": "4f95cb87dd3ae4936d31d4c1e94317cb6a26fbeb7da655363811df3a91cd6bf8" } }, { - "key": "opt02_seenn_bayes/ic_layer_1", - "tag": "ic_layer_1", - "script": "opt02_seenn_bayes.py", - "hf_folder": "applied/opt02_seenn_bayes", - "phase": 7, + "key": "opt06_uticket_prune/prune_rounds_8", + "tag": "prune_rounds_8", + "script": "opt06_uticket_prune.py", + "hf_folder": "applied/opt06_uticket_prune", + "phase": 9, "status": "ok", - "duration_s": 609.7, + "duration_s": 322.7, "optimization_summary": { - "optimization": "seenn_early_exit_bayesian_fusion", - "research_section": "2. Temporal-dimension reduction", - "reference": "SEENN: Towards Temporal Spiking Early-Exit Neural Networks, arXiv:2304.01230 (NeurIPS 2023); BayesianSpikeFusion, Front. Neurosci. 2024 (PMC11330889)", - "script": "opt02_seenn_bayes.py", + "optimization": "uticket_workload_balanced_pruning", + "research_section": "3. Spike sparsity & MAC->AC restructuring", + "reference": "u-Ticket: Workload-Balanced Pruning for Sparse SNNs, arXiv:2302.06746", + "script": "opt06_uticket_prune.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.9278571428571428, - "test_target_acc": 0.9278571428571428, - "best_val_acc": 0.9321428571428572 + "test_acc": 0.18428571428571427, + "test_target_acc": 0.18428571428571427, + "best_val_acc": 0.9142857142857143 }, "efficiency": { "spikes_per_inference": null, - "syn_ops_per_inference": 3613774.1201298703, - "dense_macs_per_inference": 124723200.0, - "event_fraction_of_dense": 0.02897435377002731, - "syn_ops_per_inference_dense_fanout": 65190391.90665584, - "weight_sparsity": 0.7324707846410685, - "effective_ticks": 8.52, + "syn_ops_per_inference": 129797.51948051948, + "dense_macs_per_inference": 122675200.0, + "event_fraction_of_dense": 0.0010580583482278365, + "syn_ops_per_inference_dense_fanout": 65036167.55113637, + "weight_sparsity": 0.9907585233201168, + "effective_ticks": 16.09857142857143, "num_ticks": 100, "membrane_bytes_per_inference": 451200, "weight_bytes": 4907008, "input_raster_bytes": null, - "train_ms_per_step": 185.69024801254272, - "infer_ms_per_sample": 0.6491338937126812, - "peak_vram_mb": 10916.063744, - "syn_ops_per_inference_early_exit": 1842881.0933441557 + "train_ms_per_step": 94.60634120000577, + "infer_ms_per_sample": 0.6177368868750222, + "peak_vram_mb": 2839.143424, + "weight_bytes_nonzero": 4453 }, "optimization_specific": { - "seenn_enabled": true, - "bayes_fusion_enabled": true, - "conf_threshold": 0.9, - "conf_temp": 2.0, - "anytime_fracs": "0.25,0.5,1.0", - "anytime_weights": "0.3,0.3,1.0", - "ic_loss_weight": 0.3, - "ic_layer": 1, - "mean_exit_tick": 8.52, - "mean_exit_ms": 426.0, - "ticks_saved_frac": 0.9148000000000001, - "exit_tick_histogram": { - "counts": [ - 970, - 375, - 48, - 7, - 0, - 0, - 0, - 0, - 0, - 0 - ], - "edges": [ - 0.0, - 10.0, - 20.0, - 30.0, - 40.0, - 50.0, - 60.0, - 70.0, - 80.0, - 90.0, - 100.0 - ] + "pruning_enabled": true, + "balance_enabled": true, + "rounds": 8, + "final_sparsity": 0.9990924873737373, + "nnz_per_layer": { + "layer0": 80, + "layer1": 512, + "layer2": 512 }, - "acc_fc_only": 0.9278571428571428, - "acc_ic_only": 0.935, - "acc_fused": 0.935, - "acc_early_exit": 0.39785714285714285, - "syn_ops_saved_frac": 0.4900397667140505, - "syn_per_inference_early_exit": 1842881.0933441557, - "anytime_acc": { - "0.25": 0.6442857142857142, - "0.5": 0.8335714285714285, - "0.75": 0.8935714285714286, - "1.0": 0.9278571428571428 + "pe_stats_per_layer": { + "layer0": { + "mean_nnz": 0.25, + "max_nnz": 1.0, + "min_nnz": 0.0, + "total_nnz": 80, + "numel": 921600, + "sparsity": 0.9999131944444445, + "utilization": 0.25 + }, + "layer1": { + "mean_nnz": 1.0, + "max_nnz": 1.0, + "min_nnz": 1.0, + "total_nnz": 512, + "numel": 163840, + "sparsity": 0.996875, + "utilization": 1.0 + }, + "layer2": { + "mean_nnz": 2.0, + "max_nnz": 2.0, + "min_nnz": 2.0, + "total_nnz": 512, + "numel": 131072, + "sparsity": 0.99609375, + "utilization": 1.0 + } }, - "pareto": [ - { - "conf": 0.5, - "mean_ticks": 3.99, - "acc": 0.2307142857142857, - "syn_ops": 1674774.195616883 - }, - { - "conf": 0.7, - "mean_ticks": 5.156428571428571, - "acc": 0.2742857142857143, - "syn_ops": 1725540.0787337662 - }, - { - "conf": 0.8, - "mean_ticks": 6.271428571428571, - "acc": 0.32, - "syn_ops": 1770739.0535714284 - }, - { - "conf": 0.9, - "mean_ticks": 8.519285714285715, - "acc": 0.39785714285714285, - "syn_ops": 1842866.4797077922 - }, - { - "conf": 0.95, - "mean_ticks": 10.811428571428571, - "acc": 0.4764285714285714, - "syn_ops": 1903597.862012987 - }, - { - "conf": 0.99, - "mean_ticks": 14.98857142857143, - "acc": 0.5942857142857143, - "syn_ops": 1996385.5430194808 + "pe_utilization_final": 0.75, + "pe_utilization_lth": 0.4166666666666667, + "pe_utilization_uticket": 0.75, + "latency_proxy_dense": 3712, + "latency_proxy_final": 4, + "latency_reduction_frac": 0.9989224137931034, + "rounds_history": [ + { + "round": 1, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 36.0, + "max_nnz": 49.0, + "min_nnz": 24.0, + "total_nnz": 11520, + "numel": 921600, + "sparsity": 0.9875, + "utilization": 0.7346938775510204 + }, + "uticket": { + "mean_nnz": 36.0, + "max_nnz": 36.0, + "min_nnz": 36.0, + "total_nnz": 11520, + "numel": 921600, + "sparsity": 0.9875, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 160.0, + "max_nnz": 194.0, + "min_nnz": 134.0, + "total_nnz": 81920, + "numel": 163840, + "sparsity": 0.5, + "utilization": 0.8247422680412371 + }, + "uticket": { + "mean_nnz": 160.0, + "max_nnz": 160.0, + "min_nnz": 160.0, + "total_nnz": 81920, + "numel": 163840, + "sparsity": 0.5, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 256.0, + "max_nnz": 289.0, + "min_nnz": 203.0, + "total_nnz": 65536, + "numel": 131072, + "sparsity": 0.5, + "utilization": 0.8858131487889274 + }, + "uticket": { + "mean_nnz": 256.0, + "max_nnz": 256.0, + "min_nnz": 256.0, + "total_nnz": 65536, + "numel": 131072, + "sparsity": 0.5, + "utilization": 1.0 + } + } + }, + "latency_proxy": 452, + "sparsity": 0.8693181818181819 + }, + { + "round": 2, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 18.0, + "max_nnz": 25.0, + "min_nnz": 9.0, + "total_nnz": 5760, + "numel": 921600, + "sparsity": 0.99375, + "utilization": 0.72 + }, + "uticket": { + "mean_nnz": 18.0, + "max_nnz": 18.0, + "min_nnz": 18.0, + "total_nnz": 5760, + "numel": 921600, + "sparsity": 0.99375, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 80.0, + "max_nnz": 102.0, + "min_nnz": 41.0, + "total_nnz": 40960, + "numel": 163840, + "sparsity": 0.75, + "utilization": 0.7843137254901961 + }, + "uticket": { + "mean_nnz": 80.0, + "max_nnz": 80.0, + "min_nnz": 80.0, + "total_nnz": 40960, + "numel": 163840, + "sparsity": 0.75, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 128.0, + "max_nnz": 158.0, + "min_nnz": 80.0, + "total_nnz": 32768, + "numel": 131072, + "sparsity": 0.75, + "utilization": 0.810126582278481 + }, + "uticket": { + "mean_nnz": 128.0, + "max_nnz": 128.0, + "min_nnz": 128.0, + "total_nnz": 32768, + "numel": 131072, + "sparsity": 0.75, + "utilization": 1.0 + } + } + }, + "latency_proxy": 226, + "sparsity": 0.9346590909090909 + }, + { + "round": 3, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 9.0, + "max_nnz": 15.0, + "min_nnz": 4.0, + "total_nnz": 2880, + "numel": 921600, + "sparsity": 0.996875, + "utilization": 0.6 + }, + "uticket": { + "mean_nnz": 9.0, + "max_nnz": 9.0, + "min_nnz": 9.0, + "total_nnz": 2880, + "numel": 921600, + "sparsity": 0.996875, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 40.0, + "max_nnz": 57.0, + "min_nnz": 11.0, + "total_nnz": 20480, + "numel": 163840, + "sparsity": 0.875, + "utilization": 0.7017543859649122 + }, + "uticket": { + "mean_nnz": 40.0, + "max_nnz": 40.0, + "min_nnz": 40.0, + "total_nnz": 20480, + "numel": 163840, + "sparsity": 0.875, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 64.0, + "max_nnz": 84.0, + "min_nnz": 32.0, + "total_nnz": 16384, + "numel": 131072, + "sparsity": 0.875, + "utilization": 0.7619047619047619 + }, + "uticket": { + "mean_nnz": 64.0, + "max_nnz": 64.0, + "min_nnz": 64.0, + "total_nnz": 16384, + "numel": 131072, + "sparsity": 0.875, + "utilization": 1.0 + } + } + }, + "latency_proxy": 113, + "sparsity": 0.9673295454545454 + }, + { + "round": 4, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 4.5, + "max_nnz": 9.0, + "min_nnz": 0.0, + "total_nnz": 1440, + "numel": 921600, + "sparsity": 0.9984375, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 4.0, + "max_nnz": 4.0, + "min_nnz": 4.0, + "total_nnz": 1280, + "numel": 921600, + "sparsity": 0.9986111111111111, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 20.0, + "max_nnz": 34.0, + "min_nnz": 1.0, + "total_nnz": 10240, + "numel": 163840, + "sparsity": 0.9375, + "utilization": 0.5882352941176471 + }, + "uticket": { + "mean_nnz": 20.0, + "max_nnz": 20.0, + "min_nnz": 20.0, + "total_nnz": 10240, + "numel": 163840, + "sparsity": 0.9375, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 32.0, + "max_nnz": 45.0, + "min_nnz": 18.0, + "total_nnz": 8192, + "numel": 131072, + "sparsity": 0.9375, + "utilization": 0.7111111111111111 + }, + "uticket": { + "mean_nnz": 32.0, + "max_nnz": 32.0, + "min_nnz": 32.0, + "total_nnz": 8192, + "numel": 131072, + "sparsity": 0.9375, + "utilization": 1.0 + } + } + }, + "latency_proxy": 56, + "sparsity": 0.9837962962962963 + }, + { + "round": 5, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 2.0, + "max_nnz": 4.0, + "min_nnz": 0.0, + "total_nnz": 640, + "numel": 921600, + "sparsity": 0.9993055555555556, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 2.0, + "max_nnz": 2.0, + "min_nnz": 2.0, + "total_nnz": 640, + "numel": 921600, + "sparsity": 0.9993055555555556, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 10.0, + "max_nnz": 17.0, + "min_nnz": 0.0, + "total_nnz": 5120, + "numel": 163840, + "sparsity": 0.96875, + "utilization": 0.5882352941176471 + }, + "uticket": { + "mean_nnz": 10.0, + "max_nnz": 10.0, + "min_nnz": 10.0, + "total_nnz": 5120, + "numel": 163840, + "sparsity": 0.96875, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 16.0, + "max_nnz": 26.0, + "min_nnz": 9.0, + "total_nnz": 4096, + "numel": 131072, + "sparsity": 0.96875, + "utilization": 0.6153846153846154 + }, + "uticket": { + "mean_nnz": 16.0, + "max_nnz": 16.0, + "min_nnz": 16.0, + "total_nnz": 4096, + "numel": 131072, + "sparsity": 0.96875, + "utilization": 1.0 + } + } + }, + "latency_proxy": 28, + "sparsity": 0.9918981481481481 + }, + { + "round": 6, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 1.0, + "max_nnz": 2.0, + "min_nnz": 0.0, + "total_nnz": 320, + "numel": 921600, + "sparsity": 0.9996527777777777, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 1.0, + "max_nnz": 1.0, + "min_nnz": 1.0, + "total_nnz": 320, + "numel": 921600, + "sparsity": 0.9996527777777777, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 5.0, + "max_nnz": 10.0, + "min_nnz": 0.0, + "total_nnz": 2560, + "numel": 163840, + "sparsity": 0.984375, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 5.0, + "max_nnz": 5.0, + "min_nnz": 5.0, + "total_nnz": 2560, + "numel": 163840, + "sparsity": 0.984375, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 8.0, + "max_nnz": 14.0, + "min_nnz": 3.0, + "total_nnz": 2048, + "numel": 131072, + "sparsity": 0.984375, + "utilization": 0.5714285714285714 + }, + "uticket": { + "mean_nnz": 8.0, + "max_nnz": 8.0, + "min_nnz": 8.0, + "total_nnz": 2048, + "numel": 131072, + "sparsity": 0.984375, + "utilization": 1.0 + } + } + }, + "latency_proxy": 14, + "sparsity": 0.9959490740740741 + }, + { + "round": 7, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 0.5, + "max_nnz": 1.0, + "min_nnz": 0.0, + "total_nnz": 160, + "numel": 921600, + "sparsity": 0.9998263888888889, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 0.5, + "max_nnz": 1.0, + "min_nnz": 0.0, + "total_nnz": 160, + "numel": 921600, + "sparsity": 0.9998263888888889, + "utilization": 0.5 + } + }, + "layer1": { + "lth": { + "mean_nnz": 2.5, + "max_nnz": 5.0, + "min_nnz": 0.0, + "total_nnz": 1280, + "numel": 163840, + "sparsity": 0.9921875, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 2.0, + "max_nnz": 2.0, + "min_nnz": 2.0, + "total_nnz": 1024, + "numel": 163840, + "sparsity": 0.99375, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 4.0, + "max_nnz": 7.0, + "min_nnz": 1.0, + "total_nnz": 1024, + "numel": 131072, + "sparsity": 0.9921875, + "utilization": 0.5714285714285714 + }, + "uticket": { + "mean_nnz": 4.0, + "max_nnz": 4.0, + "min_nnz": 4.0, + "total_nnz": 1024, + "numel": 131072, + "sparsity": 0.9921875, + "utilization": 1.0 + } + } + }, + "latency_proxy": 7, + "sparsity": 0.9981849747474747 + }, + { + "round": 8, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 0.25, + "max_nnz": 1.0, + "min_nnz": 0.0, + "total_nnz": 80, + "numel": 921600, + "sparsity": 0.9999131944444445, + "utilization": 0.25 + }, + "uticket": { + "mean_nnz": 0.25, + "max_nnz": 1.0, + "min_nnz": 0.0, + "total_nnz": 80, + "numel": 921600, + "sparsity": 0.9999131944444445, + "utilization": 0.25 + } + }, + "layer1": { + "lth": { + "mean_nnz": 1.0, + "max_nnz": 2.0, + "min_nnz": 0.0, + "total_nnz": 512, + "numel": 163840, + "sparsity": 0.996875, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 1.0, + "max_nnz": 1.0, + "min_nnz": 1.0, + "total_nnz": 512, + "numel": 163840, + "sparsity": 0.996875, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 2.0, + "max_nnz": 4.0, + "min_nnz": 0.0, + "total_nnz": 512, + "numel": 131072, + "sparsity": 0.99609375, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 2.0, + "max_nnz": 2.0, + "min_nnz": 2.0, + "total_nnz": 512, + "numel": 131072, + "sparsity": 0.99609375, + "utilization": 1.0 + } + } + }, + "latency_proxy": 4, + "sparsity": 0.9990924873737373 } - ] + ], + "latency_proxy_lth": 7, + "latency_reduction_vs_lth_frac": 0.4285714285714286 }, "config": { "method": "harmonic_matched", @@ -1629,163 +2685,455 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "baseline": false, - "bench": true, - "bench_warmup": 10, - "bench_iters": 50, - "bench_train_steps": 20, - "anytime_fracs": "0.25,0.5,1.0", - "anytime_weights": "0.3,0.3,1.0", - "ic_layer": 1, - "ic_loss_weight": 0.3, - "bayes_fusion": true, - "fusion_prior_weight": 1.0, - "conf_threshold": 0.9, - "conf_temp": 2.0, - "conf_sweep": "0.5,0.7,0.8,0.9,0.95,0.99", - "hub_dir": "/kaggle/working/hub/opt02_seenn_bayes/ic_layer_1", + "hub_dir": "/kaggle/working/hub/opt06_uticket_prune/prune_rounds_8", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, "no_resume": true, "selftest": false, - "sampling_rate": 250.0 - } - }, + "baseline": false, + "bench": true, + "bench_warmup": 10, + "bench_iters": 50, + "bench_train_steps": 20, + "prune_rounds": 8, + "prune_rate": 0.5, + "rewind_epoch": 1, + "balance": true, + "balance_iters": 2, + "prune_readout": false, + "sampling_rate": 250.0 + } + }, "zip": { - "name": "ic_layer_1.zip", + "name": "prune_rounds_8.zip", "files": 81, - "bytes": 12859971, - "MB": 12.9, - "sha256": "98a8ce45dac9959cbbf7cd753b3711e1819bd3c4c7d767ae7eb71ba456bc39eb" + "bytes": 7983037, + "MB": 8.0, + "sha256": "477e6120957d9a5697e3cc386ecc6dce0c2c61b98a383a8e903d9e83590962cc" } }, { - "key": "opt02_seenn_bayes/ic_loss_0.1", - "tag": "ic_loss_0.1", - "script": "opt02_seenn_bayes.py", - "hf_folder": "applied/opt02_seenn_bayes", - "phase": 7, + "key": "opt06_uticket_prune/rewind_epoch_0", + "tag": "rewind_epoch_0", + "script": "opt06_uticket_prune.py", + "hf_folder": "applied/opt06_uticket_prune", + "phase": 9, "status": "ok", - "duration_s": 606.7, + "duration_s": 321.6, "optimization_summary": { - "optimization": "seenn_early_exit_bayesian_fusion", - "research_section": "2. Temporal-dimension reduction", - "reference": "SEENN: Towards Temporal Spiking Early-Exit Neural Networks, arXiv:2304.01230 (NeurIPS 2023); BayesianSpikeFusion, Front. Neurosci. 2024 (PMC11330889)", - "script": "opt02_seenn_bayes.py", + "optimization": "uticket_workload_balanced_pruning", + "research_section": "3. Spike sparsity & MAC->AC restructuring", + "reference": "u-Ticket: Workload-Balanced Pruning for Sparse SNNs, arXiv:2302.06746", + "script": "opt06_uticket_prune.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.9271428571428572, - "test_target_acc": 0.9271428571428572, - "best_val_acc": 0.9242857142857143 + "test_acc": 0.7814285714285715, + "test_target_acc": 0.7814285714285715, + "best_val_acc": 0.9264285714285714 }, "efficiency": { "spikes_per_inference": null, - "syn_ops_per_inference": 3436030.0146103892, - "dense_macs_per_inference": 123955200.0, - "event_fraction_of_dense": 0.027719934416711757, - "syn_ops_per_inference_dense_fanout": 65012647.80113637, - "weight_sparsity": 0.7324707846410685, - "effective_ticks": 7.478571428571429, + "syn_ops_per_inference": 216478.7705965909, + "dense_macs_per_inference": 122675200.0, + "event_fraction_of_dense": 0.0017646498281363382, + "syn_ops_per_inference_dense_fanout": 65202982.11444805, + "weight_sparsity": 0.983618530884808, + "effective_ticks": 12.964285714285714, "num_ticks": 100, "membrane_bytes_per_inference": 451200, "weight_bytes": 4907008, "input_raster_bytes": null, - "train_ms_per_step": 184.88277196884155, - "infer_ms_per_sample": 0.6457533518872668, - "peak_vram_mb": 10912.55296, - "syn_ops_per_inference_early_exit": 1629213.7150974027 + "train_ms_per_step": 94.85441450001417, + "infer_ms_per_sample": 0.6180852540624926, + "peak_vram_mb": 2837.810688, + "weight_bytes_nonzero": 39755 }, "optimization_specific": { - "seenn_enabled": true, - "bayes_fusion_enabled": true, - "conf_threshold": 0.9, - "conf_temp": 2.0, - "anytime_fracs": "0.25,0.5,1.0", - "anytime_weights": "0.3,0.3,1.0", - "ic_loss_weight": 0.1, - "ic_layer": 0, - "mean_exit_tick": 7.478571428571429, - "mean_exit_ms": 373.92857142857144, - "ticks_saved_frac": 0.9252142857142858, - "exit_tick_histogram": { - "counts": [ - 1079, - 270, - 43, - 7, - 1, - 0, - 0, - 0, - 0, - 0 - ], - "edges": [ - 0.0, - 10.0, - 20.0, - 30.0, - 40.0, - 50.0, - 60.0, - 70.0, - 80.0, - 90.0, - 100.0 - ] + "pruning_enabled": true, + "balance_enabled": true, + "rounds": 5, + "final_sparsity": 0.9918981481481481, + "nnz_per_layer": { + "layer0": 640, + "layer1": 5120, + "layer2": 4096 }, - "acc_fc_only": 0.9271428571428572, - "acc_ic_only": 0.9357142857142857, - "acc_fused": 0.9314285714285714, - "acc_early_exit": 0.3657142857142857, - "syn_ops_saved_frac": 0.5258441549783321, - "syn_per_inference_early_exit": 1629213.7150974027, - "anytime_acc": { - "0.25": 0.6142857142857143, - "0.5": 0.825, - "0.75": 0.89, - "1.0": 0.9271428571428572 + "pe_stats_per_layer": { + "layer0": { + "mean_nnz": 2.0, + "max_nnz": 2.0, + "min_nnz": 2.0, + "total_nnz": 640, + "numel": 921600, + "sparsity": 0.9993055555555556, + "utilization": 1.0 + }, + "layer1": { + "mean_nnz": 10.0, + "max_nnz": 10.0, + "min_nnz": 10.0, + "total_nnz": 5120, + "numel": 163840, + "sparsity": 0.96875, + "utilization": 1.0 + }, + "layer2": { + "mean_nnz": 16.0, + "max_nnz": 16.0, + "min_nnz": 16.0, + "total_nnz": 4096, + "numel": 131072, + "sparsity": 0.96875, + "utilization": 1.0 + } }, - "pareto": [ - { - "conf": 0.5, - "mean_ticks": 4.1292857142857144, - "acc": 0.23142857142857143, - "syn_ops": 1500398.4293831168 - }, - { - "conf": 0.7, - "mean_ticks": 4.82, - "acc": 0.26357142857142857, - "syn_ops": 1529523.75487013 - }, - { - "conf": 0.8, - "mean_ticks": 5.501428571428572, - "acc": 0.29214285714285715, - "syn_ops": 1558365.8676948054 - }, - { - "conf": 0.9, - "mean_ticks": 7.478571428571429, - "acc": 0.3657142857142857, - "syn_ops": 1629213.413961039 - }, - { - "conf": 0.95, - "mean_ticks": 9.495714285714286, - "acc": 0.42928571428571427, - "syn_ops": 1689368.923701299 - }, - { - "conf": 0.99, - "mean_ticks": 13.76, - "acc": 0.5578571428571428, - "syn_ops": 1796731.148538961 + "pe_utilization_final": 1.0, + "pe_utilization_lth": 0.5554385964912281, + "pe_utilization_uticket": 1.0, + "latency_proxy_dense": 3712, + "latency_proxy_final": 28, + "latency_reduction_frac": 0.9924568965517241, + "rounds_history": [ + { + "round": 1, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 36.0, + "max_nnz": 46.0, + "min_nnz": 27.0, + "total_nnz": 11520, + "numel": 921600, + "sparsity": 0.9875, + "utilization": 0.782608695652174 + }, + "uticket": { + "mean_nnz": 36.0, + "max_nnz": 36.0, + "min_nnz": 36.0, + "total_nnz": 11520, + "numel": 921600, + "sparsity": 0.9875, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 160.0, + "max_nnz": 189.0, + "min_nnz": 123.0, + "total_nnz": 81920, + "numel": 163840, + "sparsity": 0.5, + "utilization": 0.8465608465608465 + }, + "uticket": { + "mean_nnz": 160.0, + "max_nnz": 160.0, + "min_nnz": 160.0, + "total_nnz": 81920, + "numel": 163840, + "sparsity": 0.5, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 256.0, + "max_nnz": 292.0, + "min_nnz": 197.0, + "total_nnz": 65536, + "numel": 131072, + "sparsity": 0.5, + "utilization": 0.8767123287671232 + }, + "uticket": { + "mean_nnz": 256.0, + "max_nnz": 256.0, + "min_nnz": 256.0, + "total_nnz": 65536, + "numel": 131072, + "sparsity": 0.5, + "utilization": 1.0 + } + } + }, + "latency_proxy": 452, + "sparsity": 0.8693181818181819 + }, + { + "round": 2, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 18.0, + "max_nnz": 27.0, + "min_nnz": 9.0, + "total_nnz": 5760, + "numel": 921600, + "sparsity": 0.99375, + "utilization": 0.6666666666666666 + }, + "uticket": { + "mean_nnz": 18.0, + "max_nnz": 18.0, + "min_nnz": 18.0, + "total_nnz": 5760, + "numel": 921600, + "sparsity": 0.99375, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 80.0, + "max_nnz": 110.0, + "min_nnz": 32.0, + "total_nnz": 40960, + "numel": 163840, + "sparsity": 0.75, + "utilization": 0.7272727272727273 + }, + "uticket": { + "mean_nnz": 80.0, + "max_nnz": 80.0, + "min_nnz": 80.0, + "total_nnz": 40960, + "numel": 163840, + "sparsity": 0.75, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 128.0, + "max_nnz": 161.0, + "min_nnz": 90.0, + "total_nnz": 32768, + "numel": 131072, + "sparsity": 0.75, + "utilization": 0.7950310559006211 + }, + "uticket": { + "mean_nnz": 128.0, + "max_nnz": 128.0, + "min_nnz": 128.0, + "total_nnz": 32768, + "numel": 131072, + "sparsity": 0.75, + "utilization": 1.0 + } + } + }, + "latency_proxy": 226, + "sparsity": 0.9346590909090909 + }, + { + "round": 3, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 9.0, + "max_nnz": 17.0, + "min_nnz": 3.0, + "total_nnz": 2880, + "numel": 921600, + "sparsity": 0.996875, + "utilization": 0.5294117647058824 + }, + "uticket": { + "mean_nnz": 9.0, + "max_nnz": 9.0, + "min_nnz": 9.0, + "total_nnz": 2880, + "numel": 921600, + "sparsity": 0.996875, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 40.0, + "max_nnz": 63.0, + "min_nnz": 3.0, + "total_nnz": 20480, + "numel": 163840, + "sparsity": 0.875, + "utilization": 0.6349206349206349 + }, + "uticket": { + "mean_nnz": 40.0, + "max_nnz": 40.0, + "min_nnz": 40.0, + "total_nnz": 20480, + "numel": 163840, + "sparsity": 0.875, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 64.0, + "max_nnz": 81.0, + "min_nnz": 40.0, + "total_nnz": 16384, + "numel": 131072, + "sparsity": 0.875, + "utilization": 0.7901234567901234 + }, + "uticket": { + "mean_nnz": 64.0, + "max_nnz": 64.0, + "min_nnz": 64.0, + "total_nnz": 16384, + "numel": 131072, + "sparsity": 0.875, + "utilization": 1.0 + } + } + }, + "latency_proxy": 113, + "sparsity": 0.9673295454545454 + }, + { + "round": 4, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 4.5, + "max_nnz": 9.0, + "min_nnz": 0.0, + "total_nnz": 1440, + "numel": 921600, + "sparsity": 0.9984375, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 4.0, + "max_nnz": 4.0, + "min_nnz": 4.0, + "total_nnz": 1280, + "numel": 921600, + "sparsity": 0.9986111111111111, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 20.0, + "max_nnz": 34.0, + "min_nnz": 0.0, + "total_nnz": 10240, + "numel": 163840, + "sparsity": 0.9375, + "utilization": 0.5882352941176471 + }, + "uticket": { + "mean_nnz": 20.0, + "max_nnz": 20.0, + "min_nnz": 20.0, + "total_nnz": 10240, + "numel": 163840, + "sparsity": 0.9375, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 32.0, + "max_nnz": 44.0, + "min_nnz": 20.0, + "total_nnz": 8192, + "numel": 131072, + "sparsity": 0.9375, + "utilization": 0.7272727272727273 + }, + "uticket": { + "mean_nnz": 32.0, + "max_nnz": 32.0, + "min_nnz": 32.0, + "total_nnz": 8192, + "numel": 131072, + "sparsity": 0.9375, + "utilization": 1.0 + } + } + }, + "latency_proxy": 56, + "sparsity": 0.9837962962962963 + }, + { + "round": 5, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 2.0, + "max_nnz": 4.0, + "min_nnz": 0.0, + "total_nnz": 640, + "numel": 921600, + "sparsity": 0.9993055555555556, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 2.0, + "max_nnz": 2.0, + "min_nnz": 2.0, + "total_nnz": 640, + "numel": 921600, + "sparsity": 0.9993055555555556, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 10.0, + "max_nnz": 19.0, + "min_nnz": 0.0, + "total_nnz": 5120, + "numel": 163840, + "sparsity": 0.96875, + "utilization": 0.5263157894736842 + }, + "uticket": { + "mean_nnz": 10.0, + "max_nnz": 10.0, + "min_nnz": 10.0, + "total_nnz": 5120, + "numel": 163840, + "sparsity": 0.96875, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 16.0, + "max_nnz": 25.0, + "min_nnz": 5.0, + "total_nnz": 4096, + "numel": 131072, + "sparsity": 0.96875, + "utilization": 0.64 + }, + "uticket": { + "mean_nnz": 16.0, + "max_nnz": 16.0, + "min_nnz": 16.0, + "total_nnz": 4096, + "numel": 131072, + "sparsity": 0.96875, + "utilization": 1.0 + } + } + }, + "latency_proxy": 28, + "sparsity": 0.9918981481481481 } - ] + ], + "latency_proxy_lth": 48, + "latency_reduction_vs_lth_frac": 0.41666666666666663 }, "config": { "method": "harmonic_matched", @@ -1825,163 +3173,455 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "baseline": false, - "bench": true, - "bench_warmup": 10, - "bench_iters": 50, - "bench_train_steps": 20, - "anytime_fracs": "0.25,0.5,1.0", - "anytime_weights": "0.3,0.3,1.0", - "ic_layer": 0, - "ic_loss_weight": 0.1, - "bayes_fusion": true, - "fusion_prior_weight": 1.0, - "conf_threshold": 0.9, - "conf_temp": 2.0, - "conf_sweep": "0.5,0.7,0.8,0.9,0.95,0.99", - "hub_dir": "/kaggle/working/hub/opt02_seenn_bayes/ic_loss_0.1", + "hub_dir": "/kaggle/working/hub/opt06_uticket_prune/rewind_epoch_0", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, "no_resume": true, "selftest": false, + "baseline": false, + "bench": true, + "bench_warmup": 10, + "bench_iters": 50, + "bench_train_steps": 20, + "prune_rounds": 5, + "prune_rate": 0.5, + "rewind_epoch": 0, + "balance": true, + "balance_iters": 2, + "prune_readout": false, "sampling_rate": 250.0 } }, "zip": { - "name": "ic_loss_0.1.zip", + "name": "rewind_epoch_0.zip", "files": 81, - "bytes": 12641458, - "MB": 12.6, - "sha256": "eee45bc432ffd53ea313c17e39d1034457326176b0a2447f42b1ecf6ed5b52a5" + "bytes": 8131993, + "MB": 8.1, + "sha256": "2759d4be6c9db943f9a4bdcb3c0501243b30044c7b1df1e065f9089899f000db" } }, { - "key": "opt02_seenn_bayes/ic_loss_0.6", - "tag": "ic_loss_0.6", - "script": "opt02_seenn_bayes.py", - "hf_folder": "applied/opt02_seenn_bayes", - "phase": 7, + "key": "opt06_uticket_prune/rewind_epoch_5", + "tag": "rewind_epoch_5", + "script": "opt06_uticket_prune.py", + "hf_folder": "applied/opt06_uticket_prune", + "phase": 9, "status": "ok", - "duration_s": 613.6, + "duration_s": 321.7, "optimization_summary": { - "optimization": "seenn_early_exit_bayesian_fusion", - "research_section": "2. Temporal-dimension reduction", - "reference": "SEENN: Towards Temporal Spiking Early-Exit Neural Networks, arXiv:2304.01230 (NeurIPS 2023); BayesianSpikeFusion, Front. Neurosci. 2024 (PMC11330889)", - "script": "opt02_seenn_bayes.py", + "optimization": "uticket_workload_balanced_pruning", + "research_section": "3. Spike sparsity & MAC->AC restructuring", + "reference": "u-Ticket: Workload-Balanced Pruning for Sparse SNNs, arXiv:2302.06746", + "script": "opt06_uticket_prune.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.9278571428571428, - "test_target_acc": 0.9278571428571428, - "best_val_acc": 0.93 + "test_acc": 0.77, + "test_target_acc": 0.77, + "best_val_acc": 0.9257142857142857 }, "efficiency": { "spikes_per_inference": null, - "syn_ops_per_inference": 3881920.3644480524, - "dense_macs_per_inference": 123955200.0, - "event_fraction_of_dense": 0.03131712396452954, - "syn_ops_per_inference_dense_fanout": 65458538.15097403, - "weight_sparsity": 0.7324707846410685, - "effective_ticks": 8.4, + "syn_ops_per_inference": 216033.23041801946, + "dense_macs_per_inference": 122675200.0, + "event_fraction_of_dense": 0.001761017959767088, + "syn_ops_per_inference_dense_fanout": 65245333.36607143, + "weight_sparsity": 0.983618530884808, + "effective_ticks": 13.376428571428571, "num_ticks": 100, "membrane_bytes_per_inference": 451200, "weight_bytes": 4907008, "input_raster_bytes": null, - "train_ms_per_step": 187.2155785560608, - "infer_ms_per_sample": 0.6591762280344364, - "peak_vram_mb": 10915.289088, - "syn_ops_per_inference_early_exit": 2075315.3977272727 + "train_ms_per_step": 95.69244695001089, + "infer_ms_per_sample": 0.6242595653125704, + "peak_vram_mb": 2837.865472, + "weight_bytes_nonzero": 39755 }, "optimization_specific": { - "seenn_enabled": true, - "bayes_fusion_enabled": true, - "conf_threshold": 0.9, - "conf_temp": 2.0, - "anytime_fracs": "0.25,0.5,1.0", - "anytime_weights": "0.3,0.3,1.0", - "ic_loss_weight": 0.6, - "ic_layer": 0, - "mean_exit_tick": 8.4, - "mean_exit_ms": 420.00000000000006, - "ticks_saved_frac": 0.916, - "exit_tick_histogram": { - "counts": [ - 996, - 350, - 41, - 11, - 1, - 1, - 0, - 0, - 0, - 0 - ], - "edges": [ - 0.0, - 10.0, - 20.0, - 30.0, - 40.0, - 50.0, - 60.0, - 70.0, - 80.0, - 90.0, - 100.0 - ] + "pruning_enabled": true, + "balance_enabled": true, + "rounds": 5, + "final_sparsity": 0.9918981481481481, + "nnz_per_layer": { + "layer0": 640, + "layer1": 5120, + "layer2": 4096 }, - "acc_fc_only": 0.9278571428571428, - "acc_ic_only": 0.9364285714285714, - "acc_fused": 0.935, - "acc_early_exit": 0.3942857142857143, - "syn_ops_saved_frac": 0.4653894972360285, - "syn_per_inference_early_exit": 2075315.3977272727, - "anytime_acc": { - "0.25": 0.6278571428571429, - "0.5": 0.8207142857142857, - "0.75": 0.8907142857142857, - "1.0": 0.9278571428571428 + "pe_stats_per_layer": { + "layer0": { + "mean_nnz": 2.0, + "max_nnz": 2.0, + "min_nnz": 2.0, + "total_nnz": 640, + "numel": 921600, + "sparsity": 0.9993055555555556, + "utilization": 1.0 + }, + "layer1": { + "mean_nnz": 10.0, + "max_nnz": 10.0, + "min_nnz": 10.0, + "total_nnz": 5120, + "numel": 163840, + "sparsity": 0.96875, + "utilization": 1.0 + }, + "layer2": { + "mean_nnz": 16.0, + "max_nnz": 16.0, + "min_nnz": 16.0, + "total_nnz": 4096, + "numel": 131072, + "sparsity": 0.96875, + "utilization": 1.0 + } }, - "pareto": [ - { - "conf": 0.5, - "mean_ticks": 4.551428571428572, - "acc": 0.24857142857142858, - "syn_ops": 1938717.4821428573 - }, - { - "conf": 0.7, - "mean_ticks": 5.35, - "acc": 0.2757142857142857, - "syn_ops": 1970675.2102272727 - }, - { - "conf": 0.8, - "mean_ticks": 6.325714285714286, - "acc": 0.31785714285714284, - "syn_ops": 2008171.7573051946 - }, - { - "conf": 0.9, - "mean_ticks": 8.4, - "acc": 0.3942857142857143, - "syn_ops": 2075315.4375 - }, - { - "conf": 0.95, - "mean_ticks": 10.375714285714286, - "acc": 0.46214285714285713, - "syn_ops": 2131283.1087662335 - }, - { - "conf": 0.99, - "mean_ticks": 14.411428571428571, - "acc": 0.5742857142857143, - "syn_ops": 2220022.9334415584 + "pe_utilization_final": 1.0, + "pe_utilization_lth": 0.5739893211289092, + "pe_utilization_uticket": 1.0, + "latency_proxy_dense": 3712, + "latency_proxy_final": 28, + "latency_reduction_frac": 0.9924568965517241, + "rounds_history": [ + { + "round": 1, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 36.0, + "max_nnz": 46.0, + "min_nnz": 24.0, + "total_nnz": 11520, + "numel": 921600, + "sparsity": 0.9875, + "utilization": 0.782608695652174 + }, + "uticket": { + "mean_nnz": 36.0, + "max_nnz": 36.0, + "min_nnz": 36.0, + "total_nnz": 11520, + "numel": 921600, + "sparsity": 0.9875, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 160.0, + "max_nnz": 189.0, + "min_nnz": 126.0, + "total_nnz": 81920, + "numel": 163840, + "sparsity": 0.5, + "utilization": 0.8465608465608465 + }, + "uticket": { + "mean_nnz": 160.0, + "max_nnz": 160.0, + "min_nnz": 160.0, + "total_nnz": 81920, + "numel": 163840, + "sparsity": 0.5, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 256.0, + "max_nnz": 288.0, + "min_nnz": 214.0, + "total_nnz": 65536, + "numel": 131072, + "sparsity": 0.5, + "utilization": 0.8888888888888888 + }, + "uticket": { + "mean_nnz": 256.0, + "max_nnz": 256.0, + "min_nnz": 256.0, + "total_nnz": 65536, + "numel": 131072, + "sparsity": 0.5, + "utilization": 1.0 + } + } + }, + "latency_proxy": 452, + "sparsity": 0.8693181818181819 + }, + { + "round": 2, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 18.0, + "max_nnz": 28.0, + "min_nnz": 10.0, + "total_nnz": 5760, + "numel": 921600, + "sparsity": 0.99375, + "utilization": 0.6428571428571429 + }, + "uticket": { + "mean_nnz": 18.0, + "max_nnz": 18.0, + "min_nnz": 18.0, + "total_nnz": 5760, + "numel": 921600, + "sparsity": 0.99375, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 80.0, + "max_nnz": 103.0, + "min_nnz": 31.0, + "total_nnz": 40960, + "numel": 163840, + "sparsity": 0.75, + "utilization": 0.7766990291262136 + }, + "uticket": { + "mean_nnz": 80.0, + "max_nnz": 80.0, + "min_nnz": 80.0, + "total_nnz": 40960, + "numel": 163840, + "sparsity": 0.75, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 128.0, + "max_nnz": 155.0, + "min_nnz": 99.0, + "total_nnz": 32768, + "numel": 131072, + "sparsity": 0.75, + "utilization": 0.8258064516129032 + }, + "uticket": { + "mean_nnz": 128.0, + "max_nnz": 128.0, + "min_nnz": 128.0, + "total_nnz": 32768, + "numel": 131072, + "sparsity": 0.75, + "utilization": 1.0 + } + } + }, + "latency_proxy": 226, + "sparsity": 0.9346590909090909 + }, + { + "round": 3, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 9.0, + "max_nnz": 17.0, + "min_nnz": 2.0, + "total_nnz": 2880, + "numel": 921600, + "sparsity": 0.996875, + "utilization": 0.5294117647058824 + }, + "uticket": { + "mean_nnz": 9.0, + "max_nnz": 9.0, + "min_nnz": 9.0, + "total_nnz": 2880, + "numel": 921600, + "sparsity": 0.996875, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 40.0, + "max_nnz": 63.0, + "min_nnz": 0.0, + "total_nnz": 20480, + "numel": 163840, + "sparsity": 0.875, + "utilization": 0.6349206349206349 + }, + "uticket": { + "mean_nnz": 40.0, + "max_nnz": 40.0, + "min_nnz": 40.0, + "total_nnz": 20480, + "numel": 163840, + "sparsity": 0.875, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 64.0, + "max_nnz": 83.0, + "min_nnz": 44.0, + "total_nnz": 16384, + "numel": 131072, + "sparsity": 0.875, + "utilization": 0.7710843373493976 + }, + "uticket": { + "mean_nnz": 64.0, + "max_nnz": 64.0, + "min_nnz": 64.0, + "total_nnz": 16384, + "numel": 131072, + "sparsity": 0.875, + "utilization": 1.0 + } + } + }, + "latency_proxy": 113, + "sparsity": 0.9673295454545454 + }, + { + "round": 4, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 4.5, + "max_nnz": 9.0, + "min_nnz": 0.0, + "total_nnz": 1440, + "numel": 921600, + "sparsity": 0.9984375, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 4.0, + "max_nnz": 4.0, + "min_nnz": 4.0, + "total_nnz": 1280, + "numel": 921600, + "sparsity": 0.9986111111111111, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 20.0, + "max_nnz": 33.0, + "min_nnz": 0.0, + "total_nnz": 10240, + "numel": 163840, + "sparsity": 0.9375, + "utilization": 0.6060606060606061 + }, + "uticket": { + "mean_nnz": 20.0, + "max_nnz": 20.0, + "min_nnz": 20.0, + "total_nnz": 10240, + "numel": 163840, + "sparsity": 0.9375, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 32.0, + "max_nnz": 43.0, + "min_nnz": 19.0, + "total_nnz": 8192, + "numel": 131072, + "sparsity": 0.9375, + "utilization": 0.7441860465116279 + }, + "uticket": { + "mean_nnz": 32.0, + "max_nnz": 32.0, + "min_nnz": 32.0, + "total_nnz": 8192, + "numel": 131072, + "sparsity": 0.9375, + "utilization": 1.0 + } + } + }, + "latency_proxy": 56, + "sparsity": 0.9837962962962963 + }, + { + "round": 5, + "layers": { + "layer0": { + "lth": { + "mean_nnz": 2.0, + "max_nnz": 4.0, + "min_nnz": 0.0, + "total_nnz": 640, + "numel": 921600, + "sparsity": 0.9993055555555556, + "utilization": 0.5 + }, + "uticket": { + "mean_nnz": 2.0, + "max_nnz": 2.0, + "min_nnz": 2.0, + "total_nnz": 640, + "numel": 921600, + "sparsity": 0.9993055555555556, + "utilization": 1.0 + } + }, + "layer1": { + "lth": { + "mean_nnz": 10.0, + "max_nnz": 19.0, + "min_nnz": 0.0, + "total_nnz": 5120, + "numel": 163840, + "sparsity": 0.96875, + "utilization": 0.5263157894736842 + }, + "uticket": { + "mean_nnz": 10.0, + "max_nnz": 10.0, + "min_nnz": 10.0, + "total_nnz": 5120, + "numel": 163840, + "sparsity": 0.96875, + "utilization": 1.0 + } + }, + "layer2": { + "lth": { + "mean_nnz": 16.0, + "max_nnz": 23.0, + "min_nnz": 7.0, + "total_nnz": 4096, + "numel": 131072, + "sparsity": 0.96875, + "utilization": 0.6956521739130435 + }, + "uticket": { + "mean_nnz": 16.0, + "max_nnz": 16.0, + "min_nnz": 16.0, + "total_nnz": 4096, + "numel": 131072, + "sparsity": 0.96875, + "utilization": 1.0 + } + } + }, + "latency_proxy": 28, + "sparsity": 0.9918981481481481 } - ] + ], + "latency_proxy_lth": 46, + "latency_reduction_vs_lth_frac": 0.3913043478260869 }, "config": { "method": "harmonic_matched", @@ -2021,167 +3661,172 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "baseline": false, - "bench": true, - "bench_warmup": 10, - "bench_iters": 50, - "bench_train_steps": 20, - "anytime_fracs": "0.25,0.5,1.0", - "anytime_weights": "0.3,0.3,1.0", - "ic_layer": 0, - "ic_loss_weight": 0.6, - "bayes_fusion": true, - "fusion_prior_weight": 1.0, - "conf_threshold": 0.9, - "conf_temp": 2.0, - "conf_sweep": "0.5,0.7,0.8,0.9,0.95,0.99", - "hub_dir": "/kaggle/working/hub/opt02_seenn_bayes/ic_loss_0.6", + "hub_dir": "/kaggle/working/hub/opt06_uticket_prune/rewind_epoch_5", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, "no_resume": true, "selftest": false, + "baseline": false, + "bench": true, + "bench_warmup": 10, + "bench_iters": 50, + "bench_train_steps": 20, + "prune_rounds": 5, + "prune_rate": 0.5, + "rewind_epoch": 5, + "balance": true, + "balance_iters": 2, + "prune_readout": false, "sampling_rate": 250.0 } }, "zip": { - "name": "ic_loss_0.6.zip", + "name": "rewind_epoch_5.zip", "files": 81, - "bytes": 12634970, - "MB": 12.6, - "sha256": "8b06be6ffe12ced55b0fa84f116fc4476be96b97264b7c85d13026df05393ec7" + "bytes": 8032799, + "MB": 8.0, + "sha256": "1d7273c1361cb8ae29ebdbf7f2044317537adba581891a63ec662bd9b3b8d948" } }, { - "key": "opt02_seenn_bayes/no_bayes_fusion", - "tag": "no_bayes_fusion", - "script": "opt02_seenn_bayes.py", - "hf_folder": "applied/opt02_seenn_bayes", - "phase": 7, + "key": "opt07_sfa_integer_train/hard_reset_D1", + "tag": "hard_reset_D1", + "script": "opt07_sfa_integer_train.py", + "hf_folder": "applied/opt07_sfa_integer_train", + "phase": 9, "status": "ok", - "duration_s": 605.0, + "duration_s": 333.6, "optimization_summary": { - "optimization": "seenn_early_exit_bayesian_fusion", - "research_section": "2. Temporal-dimension reduction", - "reference": "SEENN: Towards Temporal Spiking Early-Exit Neural Networks, arXiv:2304.01230 (NeurIPS 2023); BayesianSpikeFusion, Front. Neurosci. 2024 (PMC11330889)", - "script": "opt02_seenn_bayes.py", + "optimization": "sfa_integer_train", + "research_section": "3. Training-Time Approximation for Fewer Timesteps (SFA) / spike-storm suppression", + "reference": "Yao et al., Scaling Spike-driven Transformer with Efficient Spike Firing Approximation Training, arXiv:2411.16061", + "script": "opt07_sfa_integer_train.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.9257142857142857, - "test_target_acc": 0.9257142857142857, - "best_val_acc": 0.9257142857142857 + "test_acc": 0.9457142857142857, + "test_target_acc": 0.9457142857142857, + "best_val_acc": 0.9464285714285714 }, "efficiency": { "spikes_per_inference": null, - "syn_ops_per_inference": 3621647.9464285714, - "dense_macs_per_inference": 123955200.0, - "event_fraction_of_dense": 0.029217394239439504, - "syn_ops_per_inference_dense_fanout": 65198265.73295455, + "syn_ops_per_inference": 3248166.6850649347, + "dense_macs_per_inference": 122675200.0, + "event_fraction_of_dense": 0.02647777778283577, + "syn_ops_per_inference_dense_fanout": 64824784.47159091, "weight_sparsity": 0.7324707846410685, - "effective_ticks": 10.228571428571428, + "effective_ticks": 12.107142857142858, "num_ticks": 100, "membrane_bytes_per_inference": 451200, "weight_bytes": 4907008, "input_raster_bytes": null, - "train_ms_per_step": 184.58738327026367, - "infer_ms_per_sample": 0.6467816068898493, - "peak_vram_mb": 10914.273792, - "syn_ops_per_inference_early_exit": 1852356.0259740262 + "train_ms_per_step": 93.90855915000884, + "infer_ms_per_sample": 0.6246371478124502, + "peak_vram_mb": 2828.755968 }, "optimization_specific": { - "seenn_enabled": true, - "bayes_fusion_enabled": false, - "conf_threshold": 0.9, - "conf_temp": 2.0, - "anytime_fracs": "0.25,0.5,1.0", - "anytime_weights": "0.3,0.3,1.0", - "ic_loss_weight": 0.3, - "ic_layer": 0, - "mean_exit_tick": 10.228571428571428, - "mean_exit_ms": 511.42857142857144, - "ticks_saved_frac": 0.8977142857142857, - "exit_tick_histogram": { - "counts": [ - 784, - 517, - 79, - 19, - 1, - 0, - 0, - 0, - 0, - 0 - ], - "edges": [ - 0.0, - 10.0, - 20.0, - 30.0, - 40.0, - 50.0, - 60.0, - 70.0, - 80.0, - 90.0, - 100.0 - ] - }, - "acc_fc_only": 0.9257142857142857, - "acc_ic_only": 0.9342857142857143, - "acc_fused": 0.9257142857142857, - "acc_early_exit": 0.44, - "syn_ops_saved_frac": 0.48853227774370045, - "syn_per_inference_early_exit": 1852356.0259740262, - "anytime_acc": { - "0.25": 0.615, - "0.5": 0.8064285714285714, - "0.75": 0.8935714285714286, - "1.0": 0.9257142857142857 - }, - "pareto": [ + "sfa_levels_D": 1, + "soft_reset": false, + "storm_reg": 0.001, + "trained_T": 100, + "infer_mode": "graded", + "min_ticks_within_1pct_of_full": 100, + "speedup_at_min_ticks": 1.0, + "full_T_acc": 0.9457142857142857, + "t_sweep_rows": [ { - "conf": 0.5, - "mean_ticks": 4.534285714285715, - "acc": 0.24142857142857144, - "syn_ops": 1655985.486201299 + "T": 4, + "acc": 0.28214285714285714, + "syn_per_inference": 1865061.2053571427, + "dense_per_inference": 4907008.0, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 25.0 }, { - "conf": 0.7, - "mean_ticks": 5.685, - "acc": 0.2885714285714286, - "syn_ops": 1712321.3814935065 + "T": 8, + "acc": 0.5857142857142857, + "syn_per_inference": 2043636.288961039, + "dense_per_inference": 9814016.0, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 12.5 }, { - "conf": 0.8, - "mean_ticks": 6.976428571428571, - "acc": 0.3385714285714286, - "syn_ops": 1762251.6274350649 + "T": 16, + "acc": 0.8035714285714286, + "syn_per_inference": 2325766.0089285714, + "dense_per_inference": 19628032.0, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 6.25 }, { - "conf": 0.9, - "mean_ticks": 10.228571428571428, - "acc": 0.44, - "syn_ops": 1852356.7021103897 + "T": 25, + "acc": 0.8757142857142857, + "syn_per_inference": 2540783.9699675324, + "dense_per_inference": 30668800.0, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 4.0 }, { - "conf": 0.95, - "mean_ticks": 10.916428571428572, - "acc": 0.4614285714285714, - "syn_ops": 1872607.6915584416 + "T": 50, + "acc": 0.93, + "syn_per_inference": 2887393.0584415584, + "dense_per_inference": 61337600.0, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 2.0 }, { - "conf": 0.99, - "mean_ticks": 15.165714285714285, - "acc": 0.5571428571428572, - "syn_ops": 1963321.9383116884 + "T": 100, + "acc": 0.9457142857142857, + "syn_per_inference": 3248166.713474026, + "dense_per_inference": 122675200.0, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 1.0 } - ] - }, - "config": { - "method": "harmonic_matched", - "data_root": "/kaggle/working/data", + ], + "storm_ratio_peak_over_mean": 4.402592156354595, + "tick_activity_cv": 0.7492657914512895, + "mean_integer_activation": 0.07816793096872668, + "frac_activations_gt1": 0.0, + "max_activation": 1.0, + "per_layer_activity": [ + { + "layer": 0, + "mean_integer_activation": 0.064309815550223, + "peak_tick_activation": 0.28312988951802254, + "storm_ratio": 4.402592156354595, + "tick_activity_cv": 0.7492657914512895, + "frac_activations_gt1": 0.0, + "max_activation": 1.0 + }, + { + "layer": 1, + "mean_integer_activation": 0.036434059143066404, + "peak_tick_activation": 0.07265853881835938, + "storm_ratio": 1.9942477046833988, + "tick_activity_cv": 0.22539867053210047, + "frac_activations_gt1": 0.0, + "max_activation": 1.0 + }, + { + "layer": 2, + "mean_integer_activation": 0.13375991821289063, + "peak_tick_activation": 0.158294677734375, + "storm_ratio": 1.183423852595627, + "tick_activity_cv": 0.12266443917391988, + "frac_activations_gt1": 0.0, + "max_activation": 1.0 + } + ], + "expanded_vs_graded_agreement": null, + "expanded_acc": null, + "graded_acc": null, + "syn_ratio_expanded_over_graded": null, + "expansion_caveat": "exact only at beta=1 (IAF); with beta<1 the expanded mode applies decay per sub-tick and drifts" + }, + "config": { + "method": "harmonic_matched", + "data_root": "/kaggle/working/data", "subjects": "", "blocks": "", "limit": 0, @@ -2217,297 +3862,167 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "baseline": false, - "bench": true, - "bench_warmup": 10, - "bench_iters": 50, - "bench_train_steps": 20, - "anytime_fracs": "0.25,0.5,1.0", - "anytime_weights": "0.3,0.3,1.0", - "ic_layer": 0, - "ic_loss_weight": 0.3, - "bayes_fusion": false, - "fusion_prior_weight": 1.0, - "conf_threshold": 0.9, - "conf_temp": 2.0, - "conf_sweep": "0.5,0.7,0.8,0.9,0.95,0.99", - "hub_dir": "/kaggle/working/hub/opt02_seenn_bayes/no_bayes_fusion", + "hub_dir": "/kaggle/working/hub/opt07_sfa_integer_train/hard_reset_D1", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, "no_resume": true, "selftest": false, + "baseline": false, + "bench": true, + "bench_warmup": 10, + "bench_iters": 50, + "bench_train_steps": 20, + "sfa_levels": 4, + "soft_reset": false, + "storm_reg": 0.001, + "infer_mode": "graded", + "t_sweep": "4,8,16,25,50,100", "sampling_rate": 250.0 } }, "zip": { - "name": "no_bayes_fusion.zip", + "name": "hard_reset_D1.zip", "files": 81, - "bytes": 12641798, - "MB": 12.6, - "sha256": "e2e83c0d51a17b32b845d07b1c1cd99dd318c4b15f891615c90b53a342810c59" + "bytes": 12247474, + "MB": 12.2, + "sha256": "14ec9f04c11682c9afbfa8aa48edc10a0d1b46c0b71aa2f5fc9f393b5519c9ad" } }, { - "key": "opt03_ftbc_threshold_balance/calib_batches_32", - "tag": "calib_batches_32", - "script": "opt03_ftbc_threshold_balance.py", - "hf_folder": "applied/opt03_ftbc_threshold_balance", - "phase": 7, + "key": "opt07_sfa_integer_train/infer_expanded", + "tag": "infer_expanded", + "script": "opt07_sfa_integer_train.py", + "hf_folder": "applied/opt07_sfa_integer_train", + "phase": 9, "status": "ok", - "duration_s": 399.1, + "duration_s": 361.5, "optimization_summary": { - "optimization": "ftbc_threshold_balance", - "research_section": "2. Temporal-Dimension Reduction (threshold balancing + FTBC)", - "reference": "Rueckauer et al. arXiv:1612.04052; Sengupta et al. arXiv:1802.02627; FTBC: Forward Temporal Bias Correction, ECCV 2024", - "script": "opt03_ftbc_threshold_balance.py", + "optimization": "sfa_integer_train", + "research_section": "3. Training-Time Approximation for Fewer Timesteps (SFA) / spike-storm suppression", + "reference": "Yao et al., Scaling Spike-driven Transformer with Efficient Spike Firing Approximation Training, arXiv:2411.16061", + "script": "opt07_sfa_integer_train.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.9378571428571428, - "test_target_acc": 0.9378571428571428, - "best_val_acc": 0.9478571428571428 + "test_acc": 0.9264285714285714, + "test_target_acc": 0.9264285714285714, + "best_val_acc": 0.93 }, "efficiency": { "spikes_per_inference": null, - "syn_ops_per_inference": 3261141.8693181816, + "syn_ops_per_inference": 5329171.874188311, "dense_macs_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026583546383606316, - "syn_ops_per_inference_dense_fanout": 64837759.65584416, + "event_fraction_of_dense": 0.043441313926435916, + "syn_ops_per_inference_dense_fanout": 66905789.66071429, "weight_sparsity": 0.7324707846410685, - "effective_ticks": 12.063571428571429, + "effective_ticks": 4.7178571428571425, "num_ticks": 100, "membrane_bytes_per_inference": 451200, "weight_bytes": 4907008, "input_raster_bytes": null, - "train_ms_per_step": 96.95494634997885, - "infer_ms_per_sample": 0.6346498156250391, - "peak_vram_mb": 2824.29952 + "train_ms_per_step": 98.45181450000382, + "infer_ms_per_sample": 0.660058239062522, + "peak_vram_mb": 2832.446464 }, "optimization_specific": { - "calibrate_thresh": true, - "ftbc": true, - "calib_percentile": 99.9, - "calib_batches": 32, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 100, + "sfa_levels_D": 4, + "soft_reset": true, + "storm_reg": 0.001, "trained_T": 100, - "thresh_before": [ - 1.0, - 1.0, - 1.0, - 1.0 - ], - "thresh_after": [ - 2.3753418922424316, - 0.8675889372825623, - 1.0950974225997925, - 2.8038883209228516 - ], - "min_ticks_within_1pct_of_full": null, - "min_ticks_within_1pct_per_arm": { - "raw": 100, - "thresh": 50, - "ftbc": 100, - "both": null - }, - "speedup_at_min_ticks": null, - "full_T_acc": 0.9442857142857143, + "infer_mode": "expanded", + "min_ticks_within_1pct_of_full": 100, + "speedup_at_min_ticks": 1.0, + "full_T_acc": 0.9264285714285714, "t_sweep_rows": [ { "T": 4, - "arm": "raw", - "acc": 0.27, - "syn_per_inference": 1859619.69237013, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.37897221532349856 - }, - { - "T": 4, - "arm": "thresh", - "acc": 0.4235714285714286, - "syn_per_inference": 1819867.7662337662, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.3708711634938778 - }, - { - "T": 4, - "arm": "ftbc", - "acc": 0.17, - "syn_per_inference": 1981539.6152597403, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.4038182972719303 - }, - { - "T": 4, - "arm": "both", - "acc": 0.23, - "syn_per_inference": 1951891.856331169, + "acc": 0.16, + "syn_per_inference": 2778669.6290584416, "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.39777637540659583 - }, - { - "T": 8, - "arm": "raw", - "acc": 0.605, - "syn_per_inference": 2038958.4732142857, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.20775984807995887 - }, - { - "T": 8, - "arm": "thresh", - "acc": 0.7764285714285715, - "syn_per_inference": 1942887.4034090908, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.1979706782023884 - }, - { - "T": 8, - "arm": "ftbc", - "acc": 0.315, - "syn_per_inference": 2370014.435064935, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.24149282363763572 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 25.0 }, { "T": 8, - "arm": "both", - "acc": 0.4764285714285714, - "syn_per_inference": 2295570.1801948054, + "acc": 0.2892857142857143, + "syn_per_inference": 3495133.1233766233, "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.2339073199182481 - }, - { - "T": 16, - "arm": "raw", - "acc": 0.8192857142857143, - "syn_per_inference": 2325345.020292208, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.11847061489874319 - }, - { - "T": 16, - "arm": "thresh", - "acc": 0.8992857142857142, - "syn_per_inference": 2113511.4650974027, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.10767821578329415 - }, - { - "T": 16, - "arm": "ftbc", - "acc": 0.6028571428571429, - "syn_per_inference": 2770224.811688312, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.14113614710269026 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 12.5 }, { "T": 16, - "arm": "both", - "acc": 0.6885714285714286, - "syn_per_inference": 2556422.022727273, + "acc": 0.5628571428571428, + "syn_per_inference": 4105260.607142857, "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.13024342036569295 - }, - { - "T": 25, - "arm": "raw", - "acc": 0.8892857142857142, - "syn_per_inference": 2546046.7800324676, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.08301748943657618 - }, - { - "T": 25, - "arm": "thresh", - "acc": 0.9214285714285714, - "syn_per_inference": 2230110.1696428573, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.07271592529355102 - }, - { - "T": 25, - "arm": "ftbc", - "acc": 0.7728571428571429, - "syn_per_inference": 3055456.0194805195, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.09962750480881284 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 6.25 }, { "T": 25, - "arm": "both", - "acc": 0.8035714285714286, - "syn_per_inference": 2718618.5576298703, + "acc": 0.765, + "syn_per_inference": 4387251.887987013, "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.08864443857046478 - }, - { - "T": 50, - "arm": "raw", - "acc": 0.9285714285714286, - "syn_per_inference": 2902600.2816558443, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.047321712646987237 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 4.0 }, { "T": 50, - "arm": "thresh", - "acc": 0.9428571428571428, - "syn_per_inference": 2409330.423701299, + "acc": 0.8835714285714286, + "syn_per_inference": 4760043.469155844, "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.03927982874617362 - }, - { - "T": 50, - "arm": "ftbc", - "acc": 0.9007142857142857, - "syn_per_inference": 3497088.7767857146, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.05701378561902837 - }, - { - "T": 50, - "arm": "both", - "acc": 0.8971428571428571, - "syn_per_inference": 2966595.8733766233, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.04836504645399597 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 2.0 }, { "T": 100, - "arm": "raw", - "acc": 0.9442857142857143, - "syn_per_inference": 3261141.1704545454, + "acc": 0.9264285714285714, + "syn_per_inference": 5327084.303571429, "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026583540686744716 - }, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 1.0 + } + ], + "storm_ratio_peak_over_mean": 3.6731897953532338, + "tick_activity_cv": 0.5687552774544564, + "mean_integer_activation": 0.10272448336705566, + "frac_activations_gt1": 0.021054688608273864, + "max_activation": 4.0, + "per_layer_activity": [ { - "T": 100, - "arm": "thresh", - "acc": 0.9378571428571428, - "syn_per_inference": 2675097.6079545454, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.021806343971353177 + "layer": 0, + "mean_integer_activation": 0.18541565285995604, + "peak_tick_activation": 0.640545666217804, + "storm_ratio": 3.454647201234981, + "tick_activity_cv": 0.5687552774544564, + "frac_activations_gt1": 0.021054688608273864, + "max_activation": 4.0 }, { - "T": 100, - "arm": "ftbc", - "acc": 0.935, - "syn_per_inference": 3939535.5738636362, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.03211354514900841 + "layer": 1, + "mean_integer_activation": 0.04947120666503906, + "peak_tick_activation": 0.12639999389648438, + "storm_ratio": 2.555021444136521, + "tick_activity_cv": 0.3229986629594537, + "frac_activations_gt1": 0.00117851253162371, + "max_activation": 4.0 }, { - "T": 100, - "arm": "both", - "acc": 0.9235714285714286, - "syn_per_inference": 3268745.5649350653, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026645528720842235 + "layer": 2, + "mean_integer_activation": 0.07328659057617187, + "peak_tick_activation": 0.269195556640625, + "storm_ratio": 3.6731897953532338, + "tick_activity_cv": 0.5266420152224381, + "frac_activations_gt1": 0.0013648986277985387, + "max_activation": 4.0 } ], - "calibration_reference": "same network at full T (NOT a source ANN)" + "expanded_vs_graded_agreement": 1.0, + "expanded_acc": 0.9140625, + "graded_acc": 0.9140625, + "syn_ratio_expanded_over_graded": 1.0003622535345824, + "expansion_caveat": "exact only at beta=1 (IAF); with beta<1 the expanded mode applies decay per sub-tick and drifts" }, "config": { "method": "harmonic_matched", @@ -2547,7 +4062,7 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "hub_dir": "/kaggle/working/hub/opt03_ftbc_threshold_balance/calib_batches_32", + "hub_dir": "/kaggle/working/hub/opt07_sfa_integer_train/infer_expanded", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, @@ -2558,285 +4073,156 @@ "bench_warmup": 10, "bench_iters": 50, "bench_train_steps": 20, - "calibrate_thresh": true, - "calib_percentile": 99.9, - "calib_batches": 32, - "ftbc": true, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 0, - "ftbc_lr": 1.0, + "sfa_levels": 4, + "soft_reset": true, + "storm_reg": 0.001, + "infer_mode": "expanded", "t_sweep": "4,8,16,25,50,100", "sampling_rate": 250.0 } }, "zip": { - "name": "calib_batches_32.zip", + "name": "infer_expanded.zip", "files": 81, - "bytes": 12248653, - "MB": 12.2, - "sha256": "9359402eddfde164b345f308da096b3218c7dd55bd14b4dbd1e9a946ac98d5e9" + "bytes": 12257806, + "MB": 12.3, + "sha256": "ce5e194365b062cb3866b982931e820ee73fb98483e8b12a4970e99e56ca8bcf" } }, { - "key": "opt03_ftbc_threshold_balance/ftbc_lr_0.5", - "tag": "ftbc_lr_0.5", - "script": "opt03_ftbc_threshold_balance.py", - "hf_folder": "applied/opt03_ftbc_threshold_balance", - "phase": 7, + "key": "opt07_sfa_integer_train/storm_reg_1e-2", + "tag": "storm_reg_1e-2", + "script": "opt07_sfa_integer_train.py", + "hf_folder": "applied/opt07_sfa_integer_train", + "phase": 9, "status": "ok", - "duration_s": 386.3, + "duration_s": 352.7, "optimization_summary": { - "optimization": "ftbc_threshold_balance", - "research_section": "2. Temporal-Dimension Reduction (threshold balancing + FTBC)", - "reference": "Rueckauer et al. arXiv:1612.04052; Sengupta et al. arXiv:1802.02627; FTBC: Forward Temporal Bias Correction, ECCV 2024", - "script": "opt03_ftbc_threshold_balance.py", + "optimization": "sfa_integer_train", + "research_section": "3. Training-Time Approximation for Fewer Timesteps (SFA) / spike-storm suppression", + "reference": "Yao et al., Scaling Spike-driven Transformer with Efficient Spike Firing Approximation Training, arXiv:2411.16061", + "script": "opt07_sfa_integer_train.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.9328571428571428, - "test_target_acc": 0.9328571428571428, - "best_val_acc": 0.945 + "test_acc": 0.9271428571428572, + "test_target_acc": 0.9271428571428572, + "best_val_acc": 0.9257142857142857 }, "efficiency": { "spikes_per_inference": null, - "syn_ops_per_inference": 3251424.406655844, + "syn_ops_per_inference": 5481484.731331169, "dense_macs_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026504333448454487, - "syn_ops_per_inference_dense_fanout": 64828042.19318182, + "event_fraction_of_dense": 0.04468290845526372, + "syn_ops_per_inference_dense_fanout": 67058102.51785714, "weight_sparsity": 0.7324707846410685, - "effective_ticks": 12.322857142857142, + "effective_ticks": 3.505714285714286, "num_ticks": 100, "membrane_bytes_per_inference": 451200, "weight_bytes": 4907008, "input_raster_bytes": null, - "train_ms_per_step": 96.83538014996884, - "infer_ms_per_sample": 0.6346282890626753, - "peak_vram_mb": 2825.357824 + "train_ms_per_step": 98.8862052499826, + "infer_ms_per_sample": 0.6657202099999893, + "peak_vram_mb": 2833.100288 }, "optimization_specific": { - "calibrate_thresh": true, - "ftbc": true, - "calib_percentile": 99.9, - "calib_batches": 8, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 100, + "sfa_levels_D": 4, + "soft_reset": true, + "storm_reg": 0.01, "trained_T": 100, - "thresh_before": [ - 1.0, - 1.0, - 1.0, - 1.0 - ], - "thresh_after": [ - 2.332695722579956, - 0.859599232673645, - 1.0915957689285278, - 2.6890203952789307 - ], - "min_ticks_within_1pct_of_full": null, - "min_ticks_within_1pct_per_arm": { - "raw": 100, - "thresh": null, - "ftbc": null, - "both": null - }, - "speedup_at_min_ticks": null, - "full_T_acc": 0.95, + "infer_mode": "graded", + "min_ticks_within_1pct_of_full": 100, + "speedup_at_min_ticks": 1.0, + "full_T_acc": 0.9271428571428572, "t_sweep_rows": [ { "T": 4, - "arm": "raw", - "acc": 0.2671428571428571, - "syn_per_inference": 1861908.2905844154, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.3794386091452093 - }, - { - "T": 4, - "arm": "thresh", - "acc": 0.4085714285714286, - "syn_per_inference": 1822856.6185064935, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.37148026220998487 - }, - { - "T": 4, - "arm": "ftbc", - "acc": 0.31785714285714284, - "syn_per_inference": 1921884.1883116884, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.3916611076060378 - }, - { - "T": 4, - "arm": "both", - "acc": 0.43785714285714283, - "syn_per_inference": 1887901.7792207792, + "acc": 0.21571428571428572, + "syn_per_inference": 2678610.4642857146, "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.3847358266423815 - }, - { - "T": 8, - "arm": "raw", - "acc": 0.5807142857142857, - "syn_per_inference": 2040732.888798701, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.20794065230775058 - }, - { - "T": 8, - "arm": "thresh", - "acc": 0.7564285714285715, - "syn_per_inference": 1946398.6477272727, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.19832845674260902 - }, - { - "T": 8, - "arm": "ftbc", - "acc": 0.5457142857142857, - "syn_per_inference": 2210247.6266233767, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.22521337102195235 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 25.0 }, { "T": 8, - "arm": "both", - "acc": 0.7142857142857143, - "syn_per_inference": 2123275.1217532465, + "acc": 0.3557142857142857, + "syn_per_inference": 3344226.649350649, "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.2163513001969068 - }, - { - "T": 16, - "arm": "raw", - "acc": 0.8042857142857143, - "syn_per_inference": 2325398.8141233767, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.11847335556225794 - }, - { - "T": 16, - "arm": "thresh", - "acc": 0.8992857142857142, - "syn_per_inference": 2118264.750811688, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.10792038401056653 - }, - { - "T": 16, - "arm": "ftbc", - "acc": 0.7342857142857143, - "syn_per_inference": 2572779.288961039, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.1310767828868956 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 12.5 }, { "T": 16, - "arm": "both", - "acc": 0.8478571428571429, - "syn_per_inference": 2357532.7719155843, + "acc": 0.6378571428571429, + "syn_per_inference": 3941638.6444805195, "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.12011050175155534 - }, - { - "T": 25, - "arm": "raw", - "acc": 0.8764285714285714, - "syn_per_inference": 2543208.9926948054, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.08292495932983375 - }, - { - "T": 25, - "arm": "thresh", - "acc": 0.9242857142857143, - "syn_per_inference": 2236140.2930194805, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.07291254607351708 - }, - { - "T": 25, - "arm": "ftbc", - "acc": 0.8357142857142857, - "syn_per_inference": 2838940.3084415584, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.092567701000416 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 6.25 }, { "T": 25, - "arm": "both", - "acc": 0.895, - "syn_per_inference": 2509213.547077922, + "acc": 0.7914285714285715, + "syn_per_inference": 4239864.0032467535, "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.0818164893011113 - }, - { - "T": 50, - "arm": "raw", - "acc": 0.9307142857142857, - "syn_per_inference": 2895740.404220779, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.04720987459927971 - }, - { - "T": 50, - "arm": "thresh", - "acc": 0.9392857142857143, - "syn_per_inference": 2418345.2930194805, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.03942680008705069 - }, - { - "T": 50, - "arm": "ftbc", - "acc": 0.9214285714285714, - "syn_per_inference": 3263736.2483766233, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.053209389483393925 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 4.0 }, { "T": 50, - "arm": "both", - "acc": 0.9307142857142857, - "syn_per_inference": 2744258.13474026, + "acc": 0.8857142857142857, + "syn_per_inference": 4694799.851461039, "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.044740226789771034 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 2.0 }, { "T": 100, - "arm": "raw", - "acc": 0.95, - "syn_per_inference": 3251424.2077922076, + "acc": 0.9271428571428572, + "syn_per_inference": 5481484.731331169, "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026504331827396307 - }, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 1.0 + } + ], + "storm_ratio_peak_over_mean": 3.003345709243902, + "tick_activity_cv": 0.46190065268000957, + "mean_integer_activation": 0.10574525063236555, + "frac_activations_gt1": 0.020694764563813806, + "max_activation": 4.0, + "per_layer_activity": [ { - "T": 100, - "arm": "thresh", - "acc": 0.9328571428571428, - "syn_per_inference": 2685684.0251623373, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.0218926402823255 + "layer": 0, + "mean_integer_activation": 0.19540179803967475, + "peak_tick_activation": 0.5868591517210007, + "storm_ratio": 3.003345709243902, + "tick_activity_cv": 0.46190065268000957, + "frac_activations_gt1": 0.020694764563813806, + "max_activation": 4.0 }, { - "T": 100, - "arm": "ftbc", - "acc": 0.9378571428571428, - "syn_per_inference": 3688635.3920454546, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.030068305509552497 + "layer": 1, + "mean_integer_activation": 0.04897125244140625, + "peak_tick_activation": 0.11265945434570312, + "storm_ratio": 2.3005222192448387, + "tick_activity_cv": 0.26290936311848917, + "frac_activations_gt1": 0.0010063552690553479, + "max_activation": 4.0 }, { - "T": 100, - "arm": "both", - "acc": 0.925, - "syn_per_inference": 3041737.727272727, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.024795050077544012 + "layer": 2, + "mean_integer_activation": 0.07286270141601563, + "peak_tick_activation": 0.21881866455078125, + "storm_ratio": 3.0031643117569575, + "tick_activity_cv": 0.36416396678259205, + "frac_activations_gt1": 0.0011422729294281453, + "max_activation": 4.0 } ], - "calibration_reference": "same network at full T (NOT a source ANN)" + "expanded_vs_graded_agreement": 0.9921875, + "expanded_acc": 0.90625, + "graded_acc": 0.90625, + "syn_ratio_expanded_over_graded": 1.0003618800383918, + "expansion_caveat": "exact only at beta=1 (IAF); with beta<1 the expanded mode applies decay per sub-tick and drifts" }, "config": { "method": "harmonic_matched", @@ -2876,7 +4262,7 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "hub_dir": "/kaggle/working/hub/opt03_ftbc_threshold_balance/ftbc_lr_0.5", + "hub_dir": "/kaggle/working/hub/opt07_sfa_integer_train/storm_reg_1e-2", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, @@ -2887,285 +4273,971 @@ "bench_warmup": 10, "bench_iters": 50, "bench_train_steps": 20, - "calibrate_thresh": true, - "calib_percentile": 99.9, - "calib_batches": 8, - "ftbc": true, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 0, - "ftbc_lr": 0.5, + "sfa_levels": 4, + "soft_reset": true, + "storm_reg": 0.01, + "infer_mode": "graded", "t_sweep": "4,8,16,25,50,100", "sampling_rate": 250.0 } }, "zip": { - "name": "ftbc_lr_0.5.zip", + "name": "storm_reg_1e-2.zip", "files": 81, - "bytes": 12251677, + "bytes": 12255265, "MB": 12.3, - "sha256": "77804b842073b38ff69c86ae0358cafc315ed019d7dc45240ea65afc1daf7137" + "sha256": "c2a25d9ee33de9bba942a4027d5d7df8d62fa45eb05f8d3d39e831de2f724632" } }, { - "key": "opt03_ftbc_threshold_balance/ftbc_lr_2.0", - "tag": "ftbc_lr_2.0", - "script": "opt03_ftbc_threshold_balance.py", - "hf_folder": "applied/opt03_ftbc_threshold_balance", - "phase": 7, + "key": "opt07_sfa_integer_train/t_sweep_ultra_low", + "tag": "t_sweep_ultra_low", + "script": "opt07_sfa_integer_train.py", + "hf_folder": "applied/opt07_sfa_integer_train", + "phase": 9, "status": "ok", - "duration_s": 387.4, + "duration_s": 351.0, "optimization_summary": { - "optimization": "ftbc_threshold_balance", - "research_section": "2. Temporal-Dimension Reduction (threshold balancing + FTBC)", - "reference": "Rueckauer et al. arXiv:1612.04052; Sengupta et al. arXiv:1802.02627; FTBC: Forward Temporal Bias Correction, ECCV 2024", - "script": "opt03_ftbc_threshold_balance.py", + "optimization": "sfa_integer_train", + "research_section": "3. Training-Time Approximation for Fewer Timesteps (SFA) / spike-storm suppression", + "reference": "Yao et al., Scaling Spike-driven Transformer with Efficient Spike Firing Approximation Training, arXiv:2411.16061", + "script": "opt07_sfa_integer_train.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { "test_acc": 0.925, "test_target_acc": 0.925, - "best_val_acc": 0.9442857142857143 + "best_val_acc": 0.9285714285714286 }, "efficiency": { "spikes_per_inference": null, - "syn_ops_per_inference": 3270530.453733766, + "syn_ops_per_inference": 5325856.609577922, "dense_macs_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026660078432590825, - "syn_ops_per_inference_dense_fanout": 64847148.240259744, + "event_fraction_of_dense": 0.04341428919274574, + "syn_ops_per_inference_dense_fanout": 66902474.39610389, "weight_sparsity": 0.7324707846410685, - "effective_ticks": 12.499285714285714, + "effective_ticks": 3.922142857142857, "num_ticks": 100, "membrane_bytes_per_inference": 451200, "weight_bytes": 4907008, "input_raster_bytes": null, - "train_ms_per_step": 97.14889454999138, - "infer_ms_per_sample": 0.6363494856253737, - "peak_vram_mb": 2825.317888 + "train_ms_per_step": 98.89879284999097, + "infer_ms_per_sample": 0.6661157987500133, + "peak_vram_mb": 2832.45568 }, "optimization_specific": { - "calibrate_thresh": true, - "ftbc": true, - "calib_percentile": 99.9, - "calib_batches": 8, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 100, + "sfa_levels_D": 4, + "soft_reset": true, + "storm_reg": 0.001, "trained_T": 100, - "thresh_before": [ - 1.0, - 1.0, - 1.0, - 1.0 - ], - "thresh_after": [ - 2.366356134414673, - 0.8545898795127869, - 1.1143375635147095, - 2.678464412689209 - ], - "min_ticks_within_1pct_of_full": null, - "min_ticks_within_1pct_per_arm": { - "raw": 100, - "thresh": 50, - "ftbc": null, - "both": null - }, - "speedup_at_min_ticks": null, - "full_T_acc": 0.9485714285714286, + "infer_mode": "graded", + "min_ticks_within_1pct_of_full": 100, + "speedup_at_min_ticks": 1.0, + "full_T_acc": 0.925, "t_sweep_rows": [ { - "T": 4, - "arm": "raw", - "acc": 0.25285714285714284, - "syn_per_inference": 1866740.6030844154, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.3804233869364826 - }, - { - "T": 4, - "arm": "thresh", - "acc": 0.3921428571428571, - "syn_per_inference": 1825928.9253246754, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.3721063681421908 + "T": 2, + "acc": 0.07642857142857143, + "syn_per_inference": 2239954.340909091, + "dense_per_inference": 2453504.0, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 50.0 }, { - "T": 4, - "arm": "ftbc", - "acc": 0.05071428571428571, - "syn_per_inference": 2045034.4821428573, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.41675792705918907 + "T": 3, + "acc": 0.1357142857142857, + "syn_per_inference": 2528104.3733766233, + "dense_per_inference": 3680256.0, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 33.333333333333336 }, { "T": 4, - "arm": "both", - "acc": 0.030714285714285715, - "syn_per_inference": 2018497.18262987, + "acc": 0.1657142857142857, + "syn_per_inference": 2779257.885551948, "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.4113498862504137 - }, - { - "T": 8, - "arm": "raw", - "acc": 0.5771428571428572, - "syn_per_inference": 2048153.543831169, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.20869678058718968 - }, - { - "T": 8, - "arm": "thresh", - "acc": 0.7385714285714285, - "syn_per_inference": 1950252.5267857143, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.19872114807900398 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 25.0 }, { - "T": 8, - "arm": "ftbc", - "acc": 0.05285714285714286, - "syn_per_inference": 2452784.042207792, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.2499266398391639 + "T": 6, + "acc": 0.23142857142857143, + "syn_per_inference": 3189504.724837662, + "dense_per_inference": 7360512.0, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 16.666666666666668 }, { "T": 8, - "arm": "both", - "acc": 0.06642857142857143, - "syn_per_inference": 2399304.793831169, + "acc": 0.30642857142857144, + "syn_per_inference": 3490675.0357142854, "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.2444773672501827 - }, - { - "T": 16, - "arm": "raw", - "acc": 0.7892857142857143, - "syn_per_inference": 2337609.4261363638, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.11909545624015509 - }, - { - "T": 16, - "arm": "thresh", - "acc": 0.9014285714285715, - "syn_per_inference": 2121910.1517857146, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.10810610823263965 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 12.5 }, { - "T": 16, - "arm": "ftbc", - "acc": 0.4092857142857143, - "syn_per_inference": 2811218.702922078, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.14322468513002617 + "T": 12, + "acc": 0.4342857142857143, + "syn_per_inference": 3871810.025162338, + "dense_per_inference": 14721024.0, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 8.333333333333334 }, { "T": 16, - "arm": "both", - "acc": 0.3514285714285714, - "syn_per_inference": 2648069.6298701297, + "acc": 0.5678571428571428, + "syn_per_inference": 4094637.9375, "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.13491264075125461 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 6.25 }, { "T": 25, - "arm": "raw", - "acc": 0.8792857142857143, - "syn_per_inference": 2558687.908279221, + "acc": 0.7721428571428571, + "syn_per_inference": 4374663.069805195, "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.08342967146674213 + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 4.0 }, { - "T": 25, - "arm": "thresh", + "T": 100, "acc": 0.925, - "syn_per_inference": 2239377.7564935065, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.07301810819117495 - }, + "syn_per_inference": 5325856.666396104, + "dense_per_inference": 122675200.0, + "infer_mode": "graded", + "ticks_speedup_vs_trained_T": 1.0 + } + ], + "storm_ratio_peak_over_mean": 3.6839613211418474, + "tick_activity_cv": 0.5593002627242345, + "mean_integer_activation": 0.10257062120983997, + "frac_activations_gt1": 0.020979859633371234, + "max_activation": 4.0, + "per_layer_activity": [ { - "T": 25, - "arm": "ftbc", - "acc": 0.685, - "syn_per_inference": 3072336.8579545454, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.10017792864261221 + "layer": 0, + "mean_integer_activation": 0.18541919242590665, + "peak_tick_activation": 0.6322021558880806, + "storm_ratio": 3.409583159201322, + "tick_activity_cv": 0.5593002627242345, + "frac_activations_gt1": 0.020979859633371234, + "max_activation": 4.0 }, { - "T": 25, - "arm": "both", - "acc": 0.5571428571428572, - "syn_per_inference": 2806018.2516233767, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.0914942303456078 + "layer": 1, + "mean_integer_activation": 0.049783668518066405, + "peak_tick_activation": 0.12538528442382812, + "storm_ratio": 2.518602749781809, + "tick_activity_cv": 0.3183053742457197, + "frac_activations_gt1": 0.0011735152947949246, + "max_activation": 4.0 }, { - "T": 50, - "arm": "raw", - "acc": 0.9328571428571428, - "syn_per_inference": 2913571.6136363638, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.047500580616723895 + "layer": 2, + "mean_integer_activation": 0.07250900268554687, + "peak_tick_activation": 0.267120361328125, + "storm_ratio": 3.6839613211418474, + "tick_activity_cv": 0.5338615572684136, + "frac_activations_gt1": 0.0013800048400298692, + "max_activation": 4.0 + } + ], + "expanded_vs_graded_agreement": 1.0, + "expanded_acc": 0.90234375, + "graded_acc": 0.90234375, + "syn_ratio_expanded_over_graded": 1.0003646028320816, + "expansion_caveat": "exact only at beta=1 (IAF); with beta<1 the expanded mode applies decay per sub-tick and drifts" + }, + "config": { + "method": "harmonic_matched", + "data_root": "/kaggle/working/data", + "subjects": "", + "blocks": "", + "limit": 0, + "drop_rest": true, + "resplit": "block", + "test_blocks": "6", + "class_weight": false, + "epochs": 60, + "batch_size": 64, + "lr": 0.002, + "weight_decay": 0.0005, + "grad_clip": 5.0, + "num_bins": 100, + "window_sec": 5.0, + "hidden": "512,256", + "freq_groups": 8, + "dropout": 0.3, + "spike_drop": 0.1, + "time_jitter": 2, + "label_smoothing": 0.05, + "no_batchnorm": false, + "binary_input": false, + "beta": 1.0, + "thresh": 1.0, + "window": 0.5, + "gain": 1.0, + "alpha": 0.9, + "readout": "spikecount", + "fake_quant": 0, + "amp": false, + "spike_reg": 0.0, + "event_eval": true, + "decision_margin": 2.0, + "export": false, + "seed": 1234, + "hub_dir": "/kaggle/working/hub/opt07_sfa_integer_train/t_sweep_ultra_low", + "repo_id": "UWU-R-13/SSVEP-SNN", + "push_hf": false, + "keep_epoch_weights": 1, + "no_resume": true, + "selftest": false, + "baseline": false, + "bench": true, + "bench_warmup": 10, + "bench_iters": 50, + "bench_train_steps": 20, + "sfa_levels": 4, + "soft_reset": true, + "storm_reg": 0.001, + "infer_mode": "graded", + "t_sweep": "2,3,4,6,8,12,16,25", + "sampling_rate": 250.0 + } + }, + "zip": { + "name": "t_sweep_ultra_low.zip", + "files": 81, + "bytes": 12258005, + "MB": 12.3, + "sha256": "9b40b0bf29d963e1cd4165e4b8d4e3a1993bf7be7ddbb241165fc0976c82fe13" + } + }, + { + "key": "opt08_mint_quant/bits_sweep_fine", + "tag": "bits_sweep_fine", + "script": "opt08_mint_quant.py", + "hf_folder": "applied/opt08_mint_quant", + "phase": 9, + "status": "ok", + "duration_s": 323.5, + "optimization_summary": { + "optimization": "mint_quant", + "research_section": "4. Membrane-Potential Memory-Wall Optimizations (multiplier-less integer quantization)", + "reference": "Yin et al., MINT: Multiplier-less INTeger Quantization for Energy Efficient Spiking Neural Networks, arXiv:2305.09850 (ASP-DAC 2024)", + "script": "opt08_mint_quant.py", + "baseline_mode": false, + "method": "harmonic_matched", + "accuracy": { + "test_acc": 0.22428571428571428, + "test_target_acc": 0.22428571428571428, + "best_val_acc": 0.49357142857142855 + }, + "efficiency": { + "spikes_per_inference": null, + "syn_ops_per_inference": 737711.1282467532, + "dense_macs_per_inference": 122675200.0, + "event_fraction_of_dense": 0.006013531082458013, + "syn_ops_per_inference_dense_fanout": 63661272.290584415, + "weight_sparsity": 0.9789982001252087, + "effective_ticks": 27.373571428571427, + "num_ticks": 100, + "membrane_bytes_per_inference": 0, + "weight_bytes": 613376, + "input_raster_bytes": null, + "train_ms_per_step": 88.68454749999728, + "infer_ms_per_sample": 0.6045492706249433, + "peak_vram_mb": 2835.119104 + }, + "optimization_specific": { + "w_bits": 4, + "u_bits": 2, + "leak_shift": 0, + "beta_eff": 1.0, + "qat": true, + "batchnorm": true, + "membrane_memory_reduction_x": 16.0, + "weight_memory_reduction_x": 8.0, + "membrane_bytes_per_inference": 28200.0, + "working_set_bytes": 613658.0, + "baseline_working_set_bytes": 4911520.0, + "l2_cache_fits": true, + "baseline_l2_cache_fits": false, + "l2_budget_bytes": 2097152, + "membrane_saturation_frac": 0.018447992852393617, + "acc_vs_ubits": { + "0": 0.18428571428571427, + "2": 0.22428571428571428, + "3": 0.21571428571428572, + "4": 0.21428571428571427, + "6": 0.21571428571428572, + "8": 0.21214285714285713 + }, + "bits_sweep_rows": [ + { + "u_bits": 0, + "acc": 0.18428571428571427, + "target_acc": 0.18428571428571427, + "membrane_bytes_per_neuron_per_tick": 4.0, + "membrane_memory_reduction_x": 1.0, + "membrane_saturation_frac": null, + "acc_drop_vs_fp32": 0.0 + }, + { + "u_bits": 2, + "acc": 0.22428571428571428, + "target_acc": 0.22428571428571428, + "membrane_bytes_per_neuron_per_tick": 0.25, + "membrane_memory_reduction_x": 16.0, + "membrane_saturation_frac": 0.016339829898049645, + "acc_drop_vs_fp32": -0.04000000000000001 + }, + { + "u_bits": 3, + "acc": 0.21571428571428572, + "target_acc": 0.21571428571428572, + "membrane_bytes_per_neuron_per_tick": 0.375, + "membrane_memory_reduction_x": 10.666666666666666, + "membrane_saturation_frac": 0.07418640569592198, + "acc_drop_vs_fp32": -0.031428571428571445 + }, + { + "u_bits": 4, + "acc": 0.21428571428571427, + "target_acc": 0.21428571428571427, + "membrane_bytes_per_neuron_per_tick": 0.5, + "membrane_memory_reduction_x": 8.0, + "membrane_saturation_frac": 0.1724607297207447, + "acc_drop_vs_fp32": -0.03 + }, + { + "u_bits": 6, + "acc": 0.21571428571428572, + "target_acc": 0.21571428571428572, + "membrane_bytes_per_neuron_per_tick": 0.75, + "membrane_memory_reduction_x": 5.333333333333333, + "membrane_saturation_frac": 0.3710663923980496, + "acc_drop_vs_fp32": -0.031428571428571445 + }, + { + "u_bits": 8, + "acc": 0.21214285714285713, + "target_acc": 0.21214285714285713, + "membrane_bytes_per_neuron_per_tick": 1.0, + "membrane_memory_reduction_x": 4.0, + "membrane_saturation_frac": 0.37882293051861704, + "acc_drop_vs_fp32": -0.027857142857142858 + } + ], + "fp32_membrane_acc": 0.18428571428571427, + "per_layer_scales": [ + { + "layer": 0, + "weight_scale": 0.03549237549304962, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 28.17506537976945 + }, + { + "layer": 1, + "weight_scale": 0.03008062924657549, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 33.24398541675601 + }, + { + "layer": 2, + "weight_scale": 0.027376641120229448, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 36.52749055694305 + }, + { + "layer": 3, + "weight_scale": 0.02388627827167511, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 41.865040196982996 + } + ], + "shared_scale_note": "weight grid and membrane grid are the same grid up to a per-layer positive scalar that BatchNorm absorbs exactly; exact only with --batchnorm on", + "multiplier_less_compare": true, + "wall_clock_caveat": "this reference kernel computes in fp32 registers, so the traffic win does not appear in train_ms_per_step; a fixed-point core is where it becomes energy" + }, + "config": { + "method": "harmonic_matched", + "data_root": "/kaggle/working/data", + "subjects": "", + "blocks": "", + "limit": 0, + "drop_rest": true, + "resplit": "block", + "test_blocks": "6", + "class_weight": false, + "epochs": 60, + "batch_size": 64, + "lr": 0.002, + "weight_decay": 0.0005, + "grad_clip": 5.0, + "num_bins": 100, + "window_sec": 5.0, + "hidden": "512,256", + "freq_groups": 8, + "dropout": 0.3, + "spike_drop": 0.1, + "time_jitter": 2, + "label_smoothing": 0.05, + "no_batchnorm": false, + "binary_input": false, + "beta": 1.0, + "thresh": 1.0, + "window": 0.5, + "gain": 1.0, + "alpha": 0.9, + "readout": "spikecount", + "fake_quant": 0, + "amp": false, + "spike_reg": 0.0, + "event_eval": true, + "decision_margin": 2.0, + "export": false, + "seed": 1234, + "hub_dir": "/kaggle/working/hub/opt08_mint_quant/bits_sweep_fine", + "repo_id": "UWU-R-13/SSVEP-SNN", + "push_hf": false, + "keep_epoch_weights": 1, + "no_resume": true, + "selftest": false, + "baseline": false, + "bench": true, + "bench_warmup": 10, + "bench_iters": 50, + "bench_train_steps": 20, + "w_bits": 4, + "u_bits": 2, + "leak_shift": 0, + "qat": true, + "bits_sweep": "0,2,3,4,6,8", + "l2_bytes": 2097152, + "sampling_rate": 250.0 + } + }, + "zip": { + "name": "bits_sweep_fine.zip", + "files": 81, + "bytes": 12373116, + "MB": 12.4, + "sha256": "2fe3f4aad883a9ccc3ebe2dd4259aaa6e4053f00e53278a241e2f51f70d48e6c" + } + }, + { + "key": "opt08_mint_quant/l2_jetson_nano", + "tag": "l2_jetson_nano", + "script": "opt08_mint_quant.py", + "hf_folder": "applied/opt08_mint_quant", + "phase": 9, + "status": "ok", + "duration_s": 321.2, + "optimization_summary": { + "optimization": "mint_quant", + "research_section": "4. Membrane-Potential Memory-Wall Optimizations (multiplier-less integer quantization)", + "reference": "Yin et al., MINT: Multiplier-less INTeger Quantization for Energy Efficient Spiking Neural Networks, arXiv:2305.09850 (ASP-DAC 2024)", + "script": "opt08_mint_quant.py", + "baseline_mode": false, + "method": "harmonic_matched", + "accuracy": { + "test_acc": 0.22642857142857142, + "test_target_acc": 0.22642857142857142, + "best_val_acc": 0.48142857142857143 + }, + "efficiency": { + "spikes_per_inference": null, + "syn_ops_per_inference": 819585.7610592532, + "dense_macs_per_inference": 122675200.0, + "event_fraction_of_dense": 0.006680940899703064, + "syn_ops_per_inference_dense_fanout": 63636744.434253246, + "weight_sparsity": 0.9802144198664441, + "effective_ticks": 35.24857142857143, + "num_ticks": 100, + "membrane_bytes_per_inference": 0, + "weight_bytes": 613376, + "input_raster_bytes": null, + "train_ms_per_step": 88.0644066500281, + "infer_ms_per_sample": 0.6020276593750395, + "peak_vram_mb": 2835.111424 + }, + "optimization_specific": { + "w_bits": 4, + "u_bits": 2, + "leak_shift": 0, + "beta_eff": 1.0, + "qat": true, + "batchnorm": true, + "membrane_memory_reduction_x": 16.0, + "weight_memory_reduction_x": 8.0, + "membrane_bytes_per_inference": 28200.0, + "working_set_bytes": 613658.0, + "baseline_working_set_bytes": 4911520.0, + "l2_cache_fits": false, + "baseline_l2_cache_fits": false, + "l2_budget_bytes": 262144, + "membrane_saturation_frac": 0.019173973570478724, + "acc_vs_ubits": { + "2": 0.22642857142857142, + "4": 0.2057142857142857, + "8": 0.20642857142857143, + "0": 0.155 + }, + "bits_sweep_rows": [ + { + "u_bits": 2, + "acc": 0.22642857142857142, + "target_acc": 0.22642857142857142, + "membrane_bytes_per_neuron_per_tick": 0.25, + "membrane_memory_reduction_x": 16.0, + "membrane_saturation_frac": 0.01658403978280142, + "acc_drop_vs_fp32": -0.07142857142857142 + }, + { + "u_bits": 4, + "acc": 0.2057142857142857, + "target_acc": 0.2057142857142857, + "membrane_bytes_per_neuron_per_tick": 0.5, + "membrane_memory_reduction_x": 8.0, + "membrane_saturation_frac": 0.18061516511524822, + "acc_drop_vs_fp32": -0.05071428571428571 + }, + { + "u_bits": 8, + "acc": 0.20642857142857143, + "target_acc": 0.20642857142857143, + "membrane_bytes_per_neuron_per_tick": 1.0, + "membrane_memory_reduction_x": 4.0, + "membrane_saturation_frac": 0.39453422816932626, + "acc_drop_vs_fp32": -0.051428571428571435 + }, + { + "u_bits": 0, + "acc": 0.155, + "target_acc": 0.155, + "membrane_bytes_per_neuron_per_tick": 4.0, + "membrane_memory_reduction_x": 1.0, + "membrane_saturation_frac": null, + "acc_drop_vs_fp32": 0.0 + } + ], + "fp32_membrane_acc": 0.155, + "per_layer_scales": [ + { + "layer": 0, + "weight_scale": 0.02853875287941524, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 35.040073552803754 + }, + { + "layer": 1, + "weight_scale": 0.03575079781668527, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 27.97140374677987 + }, + { + "layer": 2, + "weight_scale": 0.03166966778891427, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 31.57595484313992 + }, + { + "layer": 3, + "weight_scale": 0.022976822086742947, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 43.52211964843367 + } + ], + "shared_scale_note": "weight grid and membrane grid are the same grid up to a per-layer positive scalar that BatchNorm absorbs exactly; exact only with --batchnorm on", + "multiplier_less_compare": true, + "wall_clock_caveat": "this reference kernel computes in fp32 registers, so the traffic win does not appear in train_ms_per_step; a fixed-point core is where it becomes energy" + }, + "config": { + "method": "harmonic_matched", + "data_root": "/kaggle/working/data", + "subjects": "", + "blocks": "", + "limit": 0, + "drop_rest": true, + "resplit": "block", + "test_blocks": "6", + "class_weight": false, + "epochs": 60, + "batch_size": 64, + "lr": 0.002, + "weight_decay": 0.0005, + "grad_clip": 5.0, + "num_bins": 100, + "window_sec": 5.0, + "hidden": "512,256", + "freq_groups": 8, + "dropout": 0.3, + "spike_drop": 0.1, + "time_jitter": 2, + "label_smoothing": 0.05, + "no_batchnorm": false, + "binary_input": false, + "beta": 1.0, + "thresh": 1.0, + "window": 0.5, + "gain": 1.0, + "alpha": 0.9, + "readout": "spikecount", + "fake_quant": 0, + "amp": false, + "spike_reg": 0.0, + "event_eval": true, + "decision_margin": 2.0, + "export": false, + "seed": 1234, + "hub_dir": "/kaggle/working/hub/opt08_mint_quant/l2_jetson_nano", + "repo_id": "UWU-R-13/SSVEP-SNN", + "push_hf": false, + "keep_epoch_weights": 1, + "no_resume": true, + "selftest": false, + "baseline": false, + "bench": true, + "bench_warmup": 10, + "bench_iters": 50, + "bench_train_steps": 20, + "w_bits": 4, + "u_bits": 2, + "leak_shift": 0, + "qat": true, + "bits_sweep": "2,4,8,0", + "l2_bytes": 262144, + "sampling_rate": 250.0 + } + }, + "zip": { + "name": "l2_jetson_nano.zip", + "files": 81, + "bytes": 12361956, + "MB": 12.4, + "sha256": "d29263c2d6d2174c1049eacdc639e439ab6721dbf1819649dee4622ebabc2743" + } + }, + { + "key": "opt08_mint_quant/leak_shift_2", + "tag": "leak_shift_2", + "script": "opt08_mint_quant.py", + "hf_folder": "applied/opt08_mint_quant", + "phase": 9, + "status": "ok", + "duration_s": 308.1, + "optimization_summary": { + "optimization": "mint_quant", + "research_section": "4. Membrane-Potential Memory-Wall Optimizations (multiplier-less integer quantization)", + "reference": "Yin et al., MINT: Multiplier-less INTeger Quantization for Energy Efficient Spiking Neural Networks, arXiv:2305.09850 (ASP-DAC 2024)", + "script": "opt08_mint_quant.py", + "baseline_mode": false, + "method": "harmonic_matched", + "accuracy": { + "test_acc": 0.025, + "test_target_acc": 0.025, + "best_val_acc": 0.6871428571428572 + }, + "efficiency": { + "spikes_per_inference": null, + "syn_ops_per_inference": 8857.618810876624, + "dense_macs_per_inference": 122675200.0, + "event_fraction_of_dense": 7.22038261268506e-05, + "syn_ops_per_inference_dense_fanout": 63155950.40909091, + "weight_sparsity": 0.9982718593489148, + "effective_ticks": 100.0, + "num_ticks": 100, + "membrane_bytes_per_inference": 0, + "weight_bytes": 613376, + "input_raster_bytes": null, + "train_ms_per_step": 83.6030630499863, + "infer_ms_per_sample": 0.5862654068749862, + "peak_vram_mb": 2832.311808 + }, + "optimization_specific": { + "w_bits": 4, + "u_bits": 2, + "leak_shift": 2, + "beta_eff": 0.75, + "qat": true, + "batchnorm": true, + "membrane_memory_reduction_x": 16.0, + "weight_memory_reduction_x": 8.0, + "membrane_bytes_per_inference": 28200.0, + "working_set_bytes": 613658.0, + "baseline_working_set_bytes": 4911520.0, + "l2_cache_fits": true, + "baseline_l2_cache_fits": false, + "l2_budget_bytes": 2097152, + "membrane_saturation_frac": 0.0, + "acc_vs_ubits": { + "2": 0.025, + "4": 0.025714285714285714, + "8": 0.03, + "0": 0.024285714285714285 + }, + "bits_sweep_rows": [ + { + "u_bits": 2, + "acc": 0.025, + "target_acc": 0.025, + "membrane_bytes_per_neuron_per_tick": 0.25, + "membrane_memory_reduction_x": 16.0, + "membrane_saturation_frac": 0.0, + "acc_drop_vs_fp32": -0.0007142857142857159 + }, + { + "u_bits": 4, + "acc": 0.025714285714285714, + "target_acc": 0.025714285714285714, + "membrane_bytes_per_neuron_per_tick": 0.5, + "membrane_memory_reduction_x": 8.0, + "membrane_saturation_frac": 3.0197251773049646e-05, + "acc_drop_vs_fp32": -0.0014285714285714284 + }, + { + "u_bits": 8, + "acc": 0.03, + "target_acc": 0.03, + "membrane_bytes_per_neuron_per_tick": 1.0, + "membrane_memory_reduction_x": 4.0, + "membrane_saturation_frac": 0.0005447279476950355, + "acc_drop_vs_fp32": -0.005714285714285713 + }, + { + "u_bits": 0, + "acc": 0.024285714285714285, + "target_acc": 0.024285714285714285, + "membrane_bytes_per_neuron_per_tick": 4.0, + "membrane_memory_reduction_x": 1.0, + "membrane_saturation_frac": null, + "acc_drop_vs_fp32": 0.0 + } + ], + "fp32_membrane_acc": 0.024285714285714285, + "per_layer_scales": [ + { + "layer": 0, + "weight_scale": 0.005439152142831257, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 183.8521838956075 + }, + { + "layer": 1, + "weight_scale": 0.0023113220397915158, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 432.65282067322875 + }, + { + "layer": 2, + "weight_scale": 0.008123405277729034, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 123.10108455890783 }, { - "T": 50, - "arm": "thresh", - "acc": 0.9407142857142857, - "syn_per_inference": 2421276.7459415584, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.039474592190459984 + "layer": 3, + "weight_scale": 0.016667025429861888, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 59.99870848029819 + } + ], + "shared_scale_note": "weight grid and membrane grid are the same grid up to a per-layer positive scalar that BatchNorm absorbs exactly; exact only with --batchnorm on", + "multiplier_less_compare": true, + "wall_clock_caveat": "this reference kernel computes in fp32 registers, so the traffic win does not appear in train_ms_per_step; a fixed-point core is where it becomes energy" + }, + "config": { + "method": "harmonic_matched", + "data_root": "/kaggle/working/data", + "subjects": "", + "blocks": "", + "limit": 0, + "drop_rest": true, + "resplit": "block", + "test_blocks": "6", + "class_weight": false, + "epochs": 60, + "batch_size": 64, + "lr": 0.002, + "weight_decay": 0.0005, + "grad_clip": 5.0, + "num_bins": 100, + "window_sec": 5.0, + "hidden": "512,256", + "freq_groups": 8, + "dropout": 0.3, + "spike_drop": 0.1, + "time_jitter": 2, + "label_smoothing": 0.05, + "no_batchnorm": false, + "binary_input": false, + "beta": 1.0, + "thresh": 1.0, + "window": 0.5, + "gain": 1.0, + "alpha": 0.9, + "readout": "spikecount", + "fake_quant": 0, + "amp": false, + "spike_reg": 0.0, + "event_eval": true, + "decision_margin": 2.0, + "export": false, + "seed": 1234, + "hub_dir": "/kaggle/working/hub/opt08_mint_quant/leak_shift_2", + "repo_id": "UWU-R-13/SSVEP-SNN", + "push_hf": false, + "keep_epoch_weights": 1, + "no_resume": true, + "selftest": false, + "baseline": false, + "bench": true, + "bench_warmup": 10, + "bench_iters": 50, + "bench_train_steps": 20, + "w_bits": 4, + "u_bits": 2, + "leak_shift": 2, + "qat": true, + "bits_sweep": "2,4,8,0", + "l2_bytes": 2097152, + "sampling_rate": 250.0 + } + }, + "zip": { + "name": "leak_shift_2.zip", + "files": 81, + "bytes": 12398575, + "MB": 12.4, + "sha256": "f3265cdd1b2d111f678c285a1c19cf7387d1102687a0b4cf493deb9a791519cc" + } + }, + { + "key": "opt08_mint_quant/no_qat", + "tag": "no_qat", + "script": "opt08_mint_quant.py", + "hf_folder": "applied/opt08_mint_quant", + "phase": 9, + "status": "ok", + "duration_s": 328.8, + "optimization_summary": { + "optimization": "mint_quant", + "research_section": "4. Membrane-Potential Memory-Wall Optimizations (multiplier-less integer quantization)", + "reference": "Yin et al., MINT: Multiplier-less INTeger Quantization for Energy Efficient Spiking Neural Networks, arXiv:2305.09850 (ASP-DAC 2024)", + "script": "opt08_mint_quant.py", + "baseline_mode": false, + "method": "harmonic_matched", + "accuracy": { + "test_acc": 0.8814285714285715, + "test_target_acc": 0.8814285714285715, + "best_val_acc": 0.9071428571428571 + }, + "efficiency": { + "spikes_per_inference": null, + "syn_ops_per_inference": 2369107.872970779, + "dense_macs_per_inference": 122675200.0, + "event_fraction_of_dense": 0.01931203595323895, + "syn_ops_per_inference_dense_fanout": 65155147.9375, + "weight_sparsity": 0.8282806956907346, + "effective_ticks": 10.586428571428572, + "num_ticks": 100, + "membrane_bytes_per_inference": 0, + "weight_bytes": 613376, + "input_raster_bytes": null, + "train_ms_per_step": 93.63406789998407, + "infer_ms_per_sample": 0.6351243703124965, + "peak_vram_mb": 2828.762112 + }, + "optimization_specific": { + "w_bits": 4, + "u_bits": 2, + "leak_shift": 0, + "beta_eff": 1.0, + "qat": false, + "batchnorm": true, + "membrane_memory_reduction_x": 16.0, + "weight_memory_reduction_x": 8.0, + "membrane_bytes_per_inference": 28200.0, + "working_set_bytes": 613658.0, + "baseline_working_set_bytes": 4911520.0, + "l2_cache_fits": true, + "baseline_l2_cache_fits": false, + "l2_budget_bytes": 2097152, + "membrane_saturation_frac": 0.10079070395611703, + "acc_vs_ubits": { + "2": 0.8814285714285715, + "4": 0.9357142857142857, + "8": 0.9421428571428572, + "0": 0.9278571428571428 + }, + "bits_sweep_rows": [ + { + "u_bits": 2, + "acc": 0.8814285714285715, + "target_acc": 0.8814285714285715, + "membrane_bytes_per_neuron_per_tick": 0.25, + "membrane_memory_reduction_x": 16.0, + "membrane_saturation_frac": 0.09545739140070922, + "acc_drop_vs_fp32": 0.046428571428571375 }, { - "T": 50, - "arm": "ftbc", - "acc": 0.8578571428571429, - "syn_per_inference": 3487039.4261363638, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.05684994890795146 + "u_bits": 4, + "acc": 0.9357142857142857, + "target_acc": 0.9357142857142857, + "membrane_bytes_per_neuron_per_tick": 0.5, + "membrane_memory_reduction_x": 8.0, + "membrane_saturation_frac": 0.2861436862810284, + "acc_drop_vs_fp32": -0.007857142857142896 }, { - "T": 50, - "arm": "both", - "acc": 0.7778571428571428, - "syn_per_inference": 3057650.978896104, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.04984953729679844 - }, + "u_bits": 8, + "acc": 0.9421428571428572, + "target_acc": 0.9421428571428572, + "membrane_bytes_per_neuron_per_tick": 1.0, + "membrane_memory_reduction_x": 4.0, + "membrane_saturation_frac": 0.3314320146276596, + "acc_drop_vs_fp32": -0.014285714285714346 + }, + { + "u_bits": 0, + "acc": 0.9278571428571428, + "target_acc": 0.9278571428571428, + "membrane_bytes_per_neuron_per_tick": 4.0, + "membrane_memory_reduction_x": 1.0, + "membrane_saturation_frac": null, + "acc_drop_vs_fp32": 0.0 + } + ], + "fp32_membrane_acc": 0.9278571428571428, + "per_layer_scales": [ { - "T": 100, - "arm": "raw", - "acc": 0.9485714285714286, - "syn_per_inference": 3270530.1485389615, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.02666007594476277 + "layer": 0, + "weight_scale": 0.03091752529144287, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 32.34411520888358 }, { - "T": 100, - "arm": "thresh", - "acc": 0.925, - "syn_per_inference": 2694214.478896104, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.02196217718737042 + "layer": 1, + "weight_scale": 0.01778849107878549, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 56.21612286117947 }, { - "T": 100, - "arm": "ftbc", - "acc": 0.9164285714285715, - "syn_per_inference": 3933733.825487013, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.03206625157723006 + "layer": 2, + "weight_scale": 0.01220501959323883, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 81.93350222509812 }, { - "T": 100, - "arm": "both", - "acc": 0.8485714285714285, - "syn_per_inference": 3378956.2589285714, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.02754392296836338 + "layer": 3, + "weight_scale": 0.018680229783058167, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 53.53253207339769 } ], - "calibration_reference": "same network at full T (NOT a source ANN)" + "shared_scale_note": "weight grid and membrane grid are the same grid up to a per-layer positive scalar that BatchNorm absorbs exactly; exact only with --batchnorm on", + "multiplier_less_compare": true, + "wall_clock_caveat": "this reference kernel computes in fp32 registers, so the traffic win does not appear in train_ms_per_step; a fixed-point core is where it becomes energy" }, "config": { "method": "harmonic_matched", @@ -3205,7 +5277,7 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "hub_dir": "/kaggle/working/hub/opt03_ftbc_threshold_balance/ftbc_lr_2.0", + "hub_dir": "/kaggle/working/hub/opt08_mint_quant/no_qat", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, @@ -3216,182 +5288,149 @@ "bench_warmup": 10, "bench_iters": 50, "bench_train_steps": 20, - "calibrate_thresh": true, - "calib_percentile": 99.9, - "calib_batches": 8, - "ftbc": true, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 0, - "ftbc_lr": 2.0, - "t_sweep": "4,8,16,25,50,100", + "w_bits": 4, + "u_bits": 2, + "leak_shift": 0, + "qat": false, + "bits_sweep": "2,4,8,0", + "l2_bytes": 2097152, "sampling_rate": 250.0 } }, "zip": { - "name": "ftbc_lr_2.0.zip", + "name": "no_qat.zip", "files": 81, - "bytes": 12249793, + "bytes": 12241785, "MB": 12.2, - "sha256": "8da510a7f6f6ea93fcb881624cd1c8187170ac3da8af0683fcf18c123a1fac73" + "sha256": "23aa8a4da5964260deb080b490dd81651bc1ec81a1e96263826a88013a927a7d" } }, { - "key": "opt03_ftbc_threshold_balance/ftbc_only", - "tag": "ftbc_only", - "script": "opt03_ftbc_threshold_balance.py", - "hf_folder": "applied/opt03_ftbc_threshold_balance", - "phase": 7, + "key": "opt08_mint_quant/w_bits_2", + "tag": "w_bits_2", + "script": "opt08_mint_quant.py", + "hf_folder": "applied/opt08_mint_quant", + "phase": 10, "status": "ok", - "duration_s": 362.0, + "duration_s": 313.7, "optimization_summary": { - "optimization": "ftbc_threshold_balance", - "research_section": "2. Temporal-Dimension Reduction (threshold balancing + FTBC)", - "reference": "Rueckauer et al. arXiv:1612.04052; Sengupta et al. arXiv:1802.02627; FTBC: Forward Temporal Bias Correction, ECCV 2024", - "script": "opt03_ftbc_threshold_balance.py", + "optimization": "mint_quant", + "research_section": "4. Membrane-Potential Memory-Wall Optimizations (multiplier-less integer quantization)", + "reference": "Yin et al., MINT: Multiplier-less INTeger Quantization for Energy Efficient Spiking Neural Networks, arXiv:2305.09850 (ASP-DAC 2024)", + "script": "opt08_mint_quant.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.9414285714285714, - "test_target_acc": 0.9414285714285714, - "best_val_acc": 0.9421428571428572 + "test_acc": 0.03214285714285714, + "test_target_acc": 0.03214285714285714, + "best_val_acc": 0.23142857142857143 }, "efficiency": { "spikes_per_inference": null, - "syn_ops_per_inference": 3249375.69724026, + "syn_ops_per_inference": 64874.41061282468, "dense_macs_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026487633174759526, - "syn_ops_per_inference_dense_fanout": 64825993.48376623, - "weight_sparsity": 0.7324707846410685, - "effective_ticks": 12.075714285714286, + "event_fraction_of_dense": 0.0005288306896000551, + "syn_ops_per_inference_dense_fanout": 63605566.99918831, + "weight_sparsity": 0.9991204416214524, + "effective_ticks": 84.10142857142857, "num_ticks": 100, - "membrane_bytes_per_inference": 451200, - "weight_bytes": 4907008, + "membrane_bytes_per_inference": 0, + "weight_bytes": 306688, "input_raster_bytes": null, - "train_ms_per_step": 101.28010819998963, - "infer_ms_per_sample": 0.656100417812695, - "peak_vram_mb": 2828.060672 + "train_ms_per_step": 85.21973990000333, + "infer_ms_per_sample": 0.5962876021874308, + "peak_vram_mb": 2835.795968 }, "optimization_specific": { - "calibrate_thresh": false, - "ftbc": true, - "calib_percentile": 99.9, - "calib_batches": 8, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 100, - "trained_T": 100, - "thresh_before": [ - 1.0, - 1.0, - 1.0, - 1.0 - ], - "thresh_after": null, - "min_ticks_within_1pct_of_full": null, - "min_ticks_within_1pct_per_arm": { - "raw": 100, - "ftbc": 100 + "w_bits": 2, + "u_bits": 2, + "leak_shift": 0, + "beta_eff": 1.0, + "qat": true, + "batchnorm": true, + "membrane_memory_reduction_x": 16.0, + "weight_memory_reduction_x": 16.0, + "membrane_bytes_per_inference": 28200.0, + "working_set_bytes": 306970.0, + "baseline_working_set_bytes": 4911520.0, + "l2_cache_fits": true, + "baseline_l2_cache_fits": false, + "l2_budget_bytes": 2097152, + "membrane_saturation_frac": 0.008376689937943263, + "acc_vs_ubits": { + "2": 0.03214285714285714, + "4": 0.03785714285714286, + "8": 0.03357142857142857, + "0": 0.03428571428571429 }, - "speedup_at_min_ticks": null, - "full_T_acc": 0.9414285714285714, - "t_sweep_rows": [ - { - "T": 4, - "arm": "raw", - "acc": 0.285, - "syn_per_inference": 1863959.37987013, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.3798566009817245 - }, - { - "T": 4, - "arm": "ftbc", - "acc": 0.18, - "syn_per_inference": 1988746.081168831, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.4052869041927038 - }, - { - "T": 8, - "arm": "raw", - "acc": 0.6028571428571429, - "syn_per_inference": 2041225.4082792208, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.20799083762235773 - }, - { - "T": 8, - "arm": "ftbc", - "acc": 0.375, - "syn_per_inference": 2398680.064935065, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.24441371044586283 - }, - { - "T": 16, - "arm": "raw", - "acc": 0.8192857142857143, - "syn_per_inference": 2324367.765422078, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.11842082616444062 - }, - { - "T": 16, - "arm": "ftbc", - "acc": 0.6142857142857143, - "syn_per_inference": 2803018.7873376627, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.1428069195799998 - }, - { - "T": 25, - "arm": "raw", - "acc": 0.8821428571428571, - "syn_per_inference": 2541577.2775974027, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.08287175493000713 - }, + "bits_sweep_rows": [ { - "T": 25, - "arm": "ftbc", - "acc": 0.78, - "syn_per_inference": 3086706.7922077924, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.10064648086028122 - }, + "u_bits": 2, + "acc": 0.03214285714285714, + "target_acc": 0.03214285714285714, + "membrane_bytes_per_neuron_per_tick": 0.25, + "membrane_memory_reduction_x": 16.0, + "membrane_saturation_frac": 0.007398326684397164, + "acc_drop_vs_fp32": 0.0021428571428571477 + }, + { + "u_bits": 4, + "acc": 0.03785714285714286, + "target_acc": 0.03785714285714286, + "membrane_bytes_per_neuron_per_tick": 0.5, + "membrane_memory_reduction_x": 8.0, + "membrane_saturation_frac": 0.05461069093528369, + "acc_drop_vs_fp32": -0.0035714285714285726 + }, + { + "u_bits": 8, + "acc": 0.03357142857142857, + "target_acc": 0.03357142857142857, + "membrane_bytes_per_neuron_per_tick": 1.0, + "membrane_memory_reduction_x": 4.0, + "membrane_saturation_frac": 0.1291535073138298, + "acc_drop_vs_fp32": 0.0007142857142857159 + }, + { + "u_bits": 0, + "acc": 0.03428571428571429, + "target_acc": 0.03428571428571429, + "membrane_bytes_per_neuron_per_tick": 4.0, + "membrane_memory_reduction_x": 1.0, + "membrane_saturation_frac": null, + "acc_drop_vs_fp32": 0.0 + } + ], + "fp32_membrane_acc": 0.03428571428571429, + "per_layer_scales": [ { - "T": 50, - "arm": "raw", - "acc": 0.9257142857142857, - "syn_per_inference": 2894576.2873376627, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.04719089575297473 + "layer": 0, + "weight_scale": 0.1566811352968216, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 6.382389290871355 }, { - "T": 50, - "arm": "ftbc", - "acc": 0.9128571428571428, - "syn_per_inference": 3537494.286525974, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.057672525278556284 + "layer": 1, + "weight_scale": 0.1379482001066208, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 7.249097844169735 }, { - "T": 100, - "arm": "raw", - "acc": 0.9414285714285714, - "syn_per_inference": 3249375.612012987, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026487632480020305 + "layer": 2, + "weight_scale": 0.07789017260074615, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 12.83859011490264 }, { - "T": 100, - "arm": "ftbc", - "acc": 0.9385714285714286, - "syn_per_inference": 3966864.5608766233, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.032336320306603314 + "layer": 3, + "weight_scale": 0.1205046996474266, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 8.298431537739242 } ], - "calibration_reference": "same network at full T (NOT a source ANN)" + "shared_scale_note": "weight grid and membrane grid are the same grid up to a per-layer positive scalar that BatchNorm absorbs exactly; exact only with --batchnorm on", + "multiplier_less_compare": true, + "wall_clock_caveat": "this reference kernel computes in fp32 registers, so the traffic win does not appear in train_ms_per_step; a fixed-point core is where it becomes energy" }, "config": { "method": "harmonic_matched", @@ -3431,7 +5470,7 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "hub_dir": "/kaggle/working/hub/opt03_ftbc_threshold_balance/ftbc_only", + "hub_dir": "/kaggle/working/hub/opt08_mint_quant/w_bits_2", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, @@ -3442,285 +5481,458 @@ "bench_warmup": 10, "bench_iters": 50, "bench_train_steps": 20, - "calibrate_thresh": false, - "calib_percentile": 99.9, - "calib_batches": 8, - "ftbc": true, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 0, - "ftbc_lr": 1.0, - "t_sweep": "4,8,16,25,50,100", + "w_bits": 2, + "u_bits": 2, + "leak_shift": 0, + "qat": true, + "bits_sweep": "2,4,8,0", + "l2_bytes": 2097152, "sampling_rate": 250.0 } }, "zip": { - "name": "ftbc_only.zip", + "name": "w_bits_2.zip", "files": 81, - "bytes": 12244861, - "MB": 12.2, - "sha256": "736f330397784874481ef351ef9ef2dc0fd23464b032e5ef4d4e65a8729b5ab0" + "bytes": 12361927, + "MB": 12.4, + "sha256": "2bbbd0d38983d52dddfea5a1ba1805c017e7f9fd0cf0c7815896dc8b4f1b95ce" } }, { - "key": "opt03_ftbc_threshold_balance/ftbc_ref_ticks_25", - "tag": "ftbc_ref_ticks_25", - "script": "opt03_ftbc_threshold_balance.py", - "hf_folder": "applied/opt03_ftbc_threshold_balance", - "phase": 7, + "key": "opt08_mint_quant/w_bits_8", + "tag": "w_bits_8", + "script": "opt08_mint_quant.py", + "hf_folder": "applied/opt08_mint_quant", + "phase": 10, "status": "ok", - "duration_s": 386.5, + "duration_s": 321.1, "optimization_summary": { - "optimization": "ftbc_threshold_balance", - "research_section": "2. Temporal-Dimension Reduction (threshold balancing + FTBC)", - "reference": "Rueckauer et al. arXiv:1612.04052; Sengupta et al. arXiv:1802.02627; FTBC: Forward Temporal Bias Correction, ECCV 2024", - "script": "opt03_ftbc_threshold_balance.py", + "optimization": "mint_quant", + "research_section": "4. Membrane-Potential Memory-Wall Optimizations (multiplier-less integer quantization)", + "reference": "Yin et al., MINT: Multiplier-less INTeger Quantization for Energy Efficient Spiking Neural Networks, arXiv:2305.09850 (ASP-DAC 2024)", + "script": "opt08_mint_quant.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.9357142857142857, - "test_target_acc": 0.9357142857142857, - "best_val_acc": 0.945 + "test_acc": 0.17714285714285713, + "test_target_acc": 0.17714285714285713, + "best_val_acc": 0.5 }, "efficiency": { "spikes_per_inference": null, - "syn_ops_per_inference": 3264603.7832792206, + "syn_ops_per_inference": 1656847.0116680192, "dense_macs_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026611766545147027, - "syn_ops_per_inference_dense_fanout": 64841221.5698052, - "weight_sparsity": 0.7324707846410685, - "effective_ticks": 12.280714285714286, + "event_fraction_of_dense": 0.013505965441002087, + "syn_ops_per_inference_dense_fanout": 63575177.9935065, + "weight_sparsity": 0.9049547096723706, + "effective_ticks": 34.494285714285716, "num_ticks": 100, - "membrane_bytes_per_inference": 451200, - "weight_bytes": 4907008, + "membrane_bytes_per_inference": 0, + "weight_bytes": 1226752, "input_raster_bytes": null, - "train_ms_per_step": 97.85671250001542, - "infer_ms_per_sample": 0.6403399928126419, - "peak_vram_mb": 2825.387008 + "train_ms_per_step": 88.07029660001717, + "infer_ms_per_sample": 0.5999964265623703, + "peak_vram_mb": 2834.582528 }, "optimization_specific": { - "calibrate_thresh": true, - "ftbc": true, - "calib_percentile": 99.9, - "calib_batches": 8, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 25, - "trained_T": 100, - "thresh_before": [ - 1.0, - 1.0, - 1.0, - 1.0 - ], - "thresh_after": [ - 2.3277933597564697, - 0.8438996076583862, - 1.0949429273605347, - 2.7184956073760986 - ], - "min_ticks_within_1pct_of_full": null, - "min_ticks_within_1pct_per_arm": { - "raw": 100, - "thresh": 50, - "ftbc": 100, - "both": null + "w_bits": 8, + "u_bits": 2, + "leak_shift": 0, + "beta_eff": 1.0, + "qat": true, + "batchnorm": true, + "membrane_memory_reduction_x": 16.0, + "weight_memory_reduction_x": 4.0, + "membrane_bytes_per_inference": 28200.0, + "working_set_bytes": 1227034.0, + "baseline_working_set_bytes": 4911520.0, + "l2_cache_fits": true, + "baseline_l2_cache_fits": false, + "l2_budget_bytes": 2097152, + "membrane_saturation_frac": 0.017583977449024825, + "acc_vs_ubits": { + "2": 0.17714285714285713, + "4": 0.19642857142857142, + "8": 0.19642857142857142, + "0": 0.16357142857142856 }, - "speedup_at_min_ticks": null, - "full_T_acc": 0.9471428571428572, - "t_sweep_rows": [ - { - "T": 4, - "arm": "raw", - "acc": 0.29928571428571427, - "syn_per_inference": 1866356.1696428573, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.38034504317964374 - }, - { - "T": 4, - "arm": "thresh", - "acc": 0.4421428571428571, - "syn_per_inference": 1826142.4926948054, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.37214989107309493 - }, - { - "T": 4, - "arm": "ftbc", - "acc": 0.1457142857142857, - "syn_per_inference": 1930750.6753246754, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.39346801051163466 - }, - { - "T": 4, - "arm": "both", - "acc": 0.23785714285714285, - "syn_per_inference": 1891963.7873376624, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.385563623971606 - }, - { - "T": 8, - "arm": "raw", - "acc": 0.6071428571428571, - "syn_per_inference": 2047413.7930194808, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.2086214036149402 - }, - { - "T": 8, - "arm": "thresh", - "acc": 0.7614285714285715, - "syn_per_inference": 1950466.1079545454, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.19874291095047586 - }, - { - "T": 8, - "arm": "ftbc", - "acc": 0.22642857142857142, - "syn_per_inference": 2217957.851461039, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.2259990050414671 - }, - { - "T": 8, - "arm": "both", - "acc": 0.4878571428571429, - "syn_per_inference": 2107893.627435065, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.21478400151732632 - }, - { - "T": 16, - "arm": "raw", - "acc": 0.795, - "syn_per_inference": 2334632.549512987, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.11894379169103592 - }, - { - "T": 16, - "arm": "thresh", - "acc": 0.8978571428571429, - "syn_per_inference": 2122709.396103896, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.1081468277667316 - }, - { - "T": 16, - "arm": "ftbc", - "acc": 0.49642857142857144, - "syn_per_inference": 2581577.3790584416, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.13152502395851207 - }, - { - "T": 16, - "arm": "both", - "acc": 0.8085714285714286, - "syn_per_inference": 2327981.189935065, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.118604921264397 - }, - { - "T": 25, - "arm": "raw", - "acc": 0.8828571428571429, - "syn_per_inference": 2554239.542207792, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.08328462614147902 - }, - { - "T": 25, - "arm": "thresh", - "acc": 0.9221428571428572, - "syn_per_inference": 2240504.061688312, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.07305483297971593 - }, - { - "T": 25, - "arm": "ftbc", - "acc": 0.7678571428571429, - "syn_per_inference": 2843913.9107142854, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.09272987240173354 - }, - { - "T": 25, - "arm": "both", - "acc": 0.8828571428571429, - "syn_per_inference": 2468573.5324675324, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.08049136361603755 - }, - { - "T": 50, - "arm": "raw", - "acc": 0.9335714285714286, - "syn_per_inference": 2908471.4699675324, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.04741743188464388 - }, - { - "T": 50, - "arm": "thresh", - "acc": 0.9385714285714286, - "syn_per_inference": 2421485.2816558443, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.039477991992771876 - }, - { - "T": 50, - "arm": "ftbc", - "acc": 0.9185714285714286, - "syn_per_inference": 3242881.149350649, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.05286938434745815 - }, - { - "T": 50, - "arm": "both", - "acc": 0.9285714285714286, - "syn_per_inference": 2676649.548701299, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.043637989564334095 - }, + "bits_sweep_rows": [ + { + "u_bits": 2, + "acc": 0.17714285714285713, + "target_acc": 0.17714285714285713, + "membrane_bytes_per_neuron_per_tick": 0.25, + "membrane_memory_reduction_x": 16.0, + "membrane_saturation_frac": 0.015716422872340426, + "acc_drop_vs_fp32": -0.013571428571428568 + }, + { + "u_bits": 4, + "acc": 0.19642857142857142, + "target_acc": 0.19642857142857142, + "membrane_bytes_per_neuron_per_tick": 0.5, + "membrane_memory_reduction_x": 8.0, + "membrane_saturation_frac": 0.1619908715647163, + "acc_drop_vs_fp32": -0.03285714285714286 + }, + { + "u_bits": 8, + "acc": 0.19642857142857142, + "target_acc": 0.19642857142857142, + "membrane_bytes_per_neuron_per_tick": 1.0, + "membrane_memory_reduction_x": 4.0, + "membrane_saturation_frac": 0.3510126468306738, + "acc_drop_vs_fp32": -0.03285714285714286 + }, + { + "u_bits": 0, + "acc": 0.16357142857142856, + "target_acc": 0.16357142857142856, + "membrane_bytes_per_neuron_per_tick": 4.0, + "membrane_memory_reduction_x": 1.0, + "membrane_saturation_frac": null, + "acc_drop_vs_fp32": 0.0 + } + ], + "fp32_membrane_acc": 0.16357142857142856, + "per_layer_scales": [ { - "T": 100, - "arm": "raw", - "acc": 0.9471428571428572, - "syn_per_inference": 3264602.8230519476, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026611758717751816 + "layer": 0, + "weight_scale": 0.0016182805374851377, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 617.9398298604231 }, { - "T": 100, - "arm": "thresh", - "acc": 0.9357142857142857, - "syn_per_inference": 2689589.7467532465, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.02192447818917961 + "layer": 1, + "weight_scale": 0.0015527149589042964, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 644.0332105164167 }, { - "T": 100, - "arm": "ftbc", - "acc": 0.9428571428571428, - "syn_per_inference": 3643597.5600649347, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.029701174810107787 + "layer": 2, + "weight_scale": 0.0011796849215124537, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 847.6839720201876 }, { - "T": 100, - "arm": "both", - "acc": 0.9314285714285714, - "syn_per_inference": 2949095.00974026, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.024039863067190925 + "layer": 3, + "weight_scale": 0.0008578119198168356, + "membrane_scale": 1.0, + "shared_scale_gauge_factor": 1165.7567083161132 + } + ], + "shared_scale_note": "weight grid and membrane grid are the same grid up to a per-layer positive scalar that BatchNorm absorbs exactly; exact only with --batchnorm on", + "multiplier_less_compare": true, + "wall_clock_caveat": "this reference kernel computes in fp32 registers, so the traffic win does not appear in train_ms_per_step; a fixed-point core is where it becomes energy" + }, + "config": { + "method": "harmonic_matched", + "data_root": "/kaggle/working/data", + "subjects": "", + "blocks": "", + "limit": 0, + "drop_rest": true, + "resplit": "block", + "test_blocks": "6", + "class_weight": false, + "epochs": 60, + "batch_size": 64, + "lr": 0.002, + "weight_decay": 0.0005, + "grad_clip": 5.0, + "num_bins": 100, + "window_sec": 5.0, + "hidden": "512,256", + "freq_groups": 8, + "dropout": 0.3, + "spike_drop": 0.1, + "time_jitter": 2, + "label_smoothing": 0.05, + "no_batchnorm": false, + "binary_input": false, + "beta": 1.0, + "thresh": 1.0, + "window": 0.5, + "gain": 1.0, + "alpha": 0.9, + "readout": "spikecount", + "fake_quant": 0, + "amp": false, + "spike_reg": 0.0, + "event_eval": true, + "decision_margin": 2.0, + "export": false, + "seed": 1234, + "hub_dir": "/kaggle/working/hub/opt08_mint_quant/w_bits_8", + "repo_id": "UWU-R-13/SSVEP-SNN", + "push_hf": false, + "keep_epoch_weights": 1, + "no_resume": true, + "selftest": false, + "baseline": false, + "bench": true, + "bench_warmup": 10, + "bench_iters": 50, + "bench_train_steps": 20, + "w_bits": 8, + "u_bits": 2, + "leak_shift": 0, + "qat": true, + "bits_sweep": "2,4,8,0", + "l2_bytes": 2097152, + "sampling_rate": 250.0 + } + }, + "zip": { + "name": "w_bits_8.zip", + "files": 81, + "bytes": 12364242, + "MB": 12.4, + "sha256": "56ab4c27abf4fa9bb3d03b40fdebdc1a39a01711c51366a348fc10f4df048451" + } + }, + { + "key": "opt09_srif_softreset/no_learn_reset", + "tag": "no_learn_reset", + "script": "opt09_srif_softreset.py", + "hf_folder": "applied/opt09_srif_softreset", + "phase": 10, + "status": "ok", + "duration_s": 338.9, + "optimization_summary": { + "optimization": "srif_softreset", + "research_section": "4. Membrane-Potential Memory Wall -- Soft Reset Integrate-and-Fire (SRIF), dynamic reset potential", + "reference": "Guo et al., 'Reducing Information Loss for Spiking Neural Networks' (SRIF), ECCV 2022; research.md S4 (stacks with MINT / SpQuant-SNN membrane quantization)", + "script": "opt09_srif_softreset.py", + "baseline_mode": false, + "method": "harmonic_matched", + "accuracy": { + "test_acc": 0.9157142857142857, + "test_target_acc": 0.9157142857142857, + "best_val_acc": 0.9135714285714286 + }, + "efficiency": { + "spikes_per_inference": 6883.3984375, + "syn_ops_per_inference": 3723144.671266234, + "dense_macs_per_inference": 122675200.0, + "event_fraction_of_dense": 0.03034961158625569, + "syn_ops_per_inference_dense_fanout": 65299762.45779221, + "weight_sparsity": 0.7324707846410685, + "effective_ticks": 12.082142857142857, + "num_ticks": 100, + "membrane_bytes_per_inference": 451200, + "weight_bytes": 4907008, + "input_raster_bytes": null, + "train_ms_per_step": 95.2910739500112, + "infer_ms_per_sample": 0.6323959431250614, + "peak_vram_mb": 2829.398528, + "discarded_charge_per_reset": 0.038193207689203236 + }, + "optimization_specific": { + "reset_mode": "srif", + "learn_reset": false, + "lam_init": 0.9, + "v_reset_init": 0.0, + "u_bits": 0, + "u_scale": 1.0, + "u_headroom_requested": 2.0, + "u_headroom_actual": 0.0, + "u_threshold_level": 0, + "u_levels_above_threshold": null, + "reset_family": { + "reset_mode": "srif", + "learn_reset": false, + "num_spiking_sites": 4, + "extra_params": 0, + "extra_params_frac_of_weights": 0.0, + "extra_flops_per_tick_per_neuron": 0, + "learned_lam_per_layer": [ + 0.8999999761581421, + 0.8999999761581421, + 0.8999999761581421, + 0.8999999761581421 + ], + "learned_v_reset_per_layer": [ + 0.0, + 0.0, + 0.0, + 0.0 + ] + }, + "reset_stats": { + "reset_mode": "srif", + "learn_reset": false, + "discarded_charge_per_reset": 0.038193207689203236, + "spikes_per_inference": 6883.3984375, + "syn_per_inference": 3811349.5, + "lam_per_site": [ + 0.8999999761581421, + 0.8999999761581421, + 0.8999999761581421, + 0.8999999761581421 + ], + "v_reset_per_site": [ + 0.0, + 0.0, + 0.0, + 0.0 + ], + "u_bits": 0, + "membrane_saturation_frac": null, + "batches": 4, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null + }, + "reset_x_bits_sweep": { + "rows": [ + { + "reset_mode": "hard", + "u_bits": 0, + "acc": 0.9264285714285714, + "target_acc": 0.9264285714285714, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null, + "discarded_charge_per_reset": 0.22403764747430177, + "spikes_per_inference": 5916.7890625, + "membrane_saturation_frac": null, + "membrane_memory_reduction_x": 1.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "hard", + "u_bits": 4, + "acc": 0.94, + "target_acc": 0.94, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3, + "discarded_charge_per_reset": 0.11177482137016213, + "spikes_per_inference": 6947.1171875, + "membrane_saturation_frac": 0.2622705424423759, + "membrane_memory_reduction_x": 8.0, + "acc_drop_vs_fp32": -0.013571428571428568 + }, + { + "reset_mode": "hard", + "u_bits": 2, + "acc": 0.8428571428571429, + "target_acc": 0.8428571428571429, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 6449.640625, + "membrane_saturation_frac": 0.3889967032358156, + "membrane_memory_reduction_x": 16.0, + "acc_drop_vs_fp32": 0.08357142857142852 + }, + { + "reset_mode": "soft", + "u_bits": 0, + "acc": 0.9014285714285715, + "target_acc": 0.9014285714285715, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 7159.375, + "membrane_saturation_frac": null, + "membrane_memory_reduction_x": 1.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "soft", + "u_bits": 4, + "acc": 0.9421428571428572, + "target_acc": 0.9421428571428572, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 7629.8828125, + "membrane_saturation_frac": 0.30001184341755316, + "membrane_memory_reduction_x": 8.0, + "acc_drop_vs_fp32": -0.0407142857142857 + }, + { + "reset_mode": "soft", + "u_bits": 2, + "acc": 0.8428571428571429, + "target_acc": 0.8428571428571429, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 6449.640625, + "membrane_saturation_frac": 0.3889967032358156, + "membrane_memory_reduction_x": 16.0, + "acc_drop_vs_fp32": 0.05857142857142861 + }, + { + "reset_mode": "trained", + "u_bits": 0, + "acc": 0.9157142857142857, + "target_acc": 0.9157142857142857, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null, + "discarded_charge_per_reset": 0.03844964373905227, + "spikes_per_inference": 6912.2421875, + "membrane_saturation_frac": null, + "membrane_memory_reduction_x": 1.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "trained", + "u_bits": 4, + "acc": 0.9428571428571428, + "target_acc": 0.9428571428571428, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3, + "discarded_charge_per_reset": 0.013396262930022496, + "spikes_per_inference": 7548.390625, + "membrane_saturation_frac": 0.2957444730718085, + "membrane_memory_reduction_x": 8.0, + "acc_drop_vs_fp32": -0.027142857142857135 + }, + { + "reset_mode": "trained", + "u_bits": 2, + "acc": 0.8428571428571429, + "target_acc": 0.8428571428571429, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 6449.640625, + "membrane_saturation_frac": 0.3889967032358156, + "membrane_memory_reduction_x": 16.0, + "acc_drop_vs_fp32": 0.07285714285714284 + } + ], + "by_mode": { + "hard": { + "fp32_acc": 0.9264285714285714, + "acc_by_bits": { + "0": 0.9264285714285714, + "4": 0.94, + "2": 0.8428571428571429 + } + }, + "soft": { + "fp32_acc": 0.9014285714285715, + "acc_by_bits": { + "0": 0.9014285714285715, + "4": 0.9421428571428572, + "2": 0.8428571428571429 + } + }, + "trained": { + "fp32_acc": 0.9157142857142857, + "acc_by_bits": { + "0": 0.9157142857142857, + "4": 0.9428571428571428, + "2": 0.8428571428571429 + } + } } - ], - "calibration_reference": "same network at full T (NOT a source ANN)" + }, + "mechanism": "V = u - s*((1-lam)*u + lam*thresh) + s*v_r; lam=0,v_r=0 is the baseline hard reset exactly, lam=1,v_r=0 discards zero charge", + "cost": "2 scalars per spiking layer; identical FMA count per tick, so the reset change is free at inference time", + "caveat": "the sweep swaps the reset on ALREADY-TRAINED weights, which under-states what training with each reset would give; and a charge-conserving reset can raise the spike count, so read spikes_per_inference next to the accuracy" }, "config": { "method": "harmonic_matched", @@ -3760,360 +5972,362 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "hub_dir": "/kaggle/working/hub/opt03_ftbc_threshold_balance/ftbc_ref_ticks_25", + "hub_dir": "/kaggle/working/hub/opt09_srif_softreset/no_learn_reset", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, "no_resume": true, "selftest": false, + "reset_mode": "srif", + "learn_reset": false, + "lam_init": 0.9, + "v_reset_init": 0.0, + "u_bits": 0, + "u_headroom": 2.0, + "reset_sweep": "hard,soft,trained", + "bits_sweep": "0,4,2", "baseline": false, "bench": true, "bench_warmup": 10, "bench_iters": 50, "bench_train_steps": 20, - "calibrate_thresh": true, - "calib_percentile": 99.9, - "calib_batches": 8, - "ftbc": true, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 25, - "ftbc_lr": 1.0, - "t_sweep": "4,8,16,25,50,100", "sampling_rate": 250.0 } }, "zip": { - "name": "ftbc_ref_ticks_25.zip", + "name": "no_learn_reset.zip", "files": 81, - "bytes": 12245889, - "MB": 12.2, - "sha256": "cf78b5486d0a886d731d0302948dd4676970143adfcc5145211d7413c97f33c7" + "bytes": 12252281, + "MB": 12.3, + "sha256": "600844de2ff500a224c3e08c679649a712a823e9b59de96b2ebad2d66308f326" } }, { - "key": "opt03_ftbc_threshold_balance/t_sweep_ultra_low", - "tag": "t_sweep_ultra_low", - "script": "opt03_ftbc_threshold_balance.py", - "hf_folder": "applied/opt03_ftbc_threshold_balance", - "phase": 8, + "key": "opt09_srif_softreset/reset_x_bits_full", + "tag": "reset_x_bits_full", + "script": "opt09_srif_softreset.py", + "hf_folder": "applied/opt09_srif_softreset", + "phase": 10, "status": "ok", - "duration_s": 361.4, + "duration_s": 352.8, "optimization_summary": { - "optimization": "ftbc_threshold_balance", - "research_section": "2. Temporal-Dimension Reduction (threshold balancing + FTBC)", - "reference": "Rueckauer et al. arXiv:1612.04052; Sengupta et al. arXiv:1802.02627; FTBC: Forward Temporal Bias Correction, ECCV 2024", - "script": "opt03_ftbc_threshold_balance.py", + "optimization": "srif_softreset", + "research_section": "4. Membrane-Potential Memory Wall -- Soft Reset Integrate-and-Fire (SRIF), dynamic reset potential", + "reference": "Guo et al., 'Reducing Information Loss for Spiking Neural Networks' (SRIF), ECCV 2022; research.md S4 (stacks with MINT / SpQuant-SNN membrane quantization)", + "script": "opt09_srif_softreset.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.9242857142857143, - "test_target_acc": 0.9242857142857143, - "best_val_acc": 0.9464285714285714 + "test_acc": 0.8478571428571429, + "test_target_acc": 0.8478571428571429, + "best_val_acc": 0.8692857142857143 }, "efficiency": { - "spikes_per_inference": null, - "syn_ops_per_inference": 3280250.712662338, + "spikes_per_inference": 9121.73828125, + "syn_ops_per_inference": 4522878.1777597405, "dense_macs_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026739314161805628, - "syn_ops_per_inference_dense_fanout": 64856868.49918831, + "event_fraction_of_dense": 0.03686872471175707, + "syn_ops_per_inference_dense_fanout": 66099495.96428571, "weight_sparsity": 0.7324707846410685, - "effective_ticks": 12.063571428571429, + "effective_ticks": 15.272142857142857, "num_ticks": 100, "membrane_bytes_per_inference": 451200, "weight_bytes": 4907008, "input_raster_bytes": null, - "train_ms_per_step": 98.6024443000133, - "infer_ms_per_sample": 0.6513143553127065, - "peak_vram_mb": 2825.365504 + "train_ms_per_step": 97.55348075000256, + "infer_ms_per_sample": 0.6508080149998818, + "peak_vram_mb": 2830.13888, + "discarded_charge_per_reset": -0.40509077622122347 }, "optimization_specific": { - "calibrate_thresh": true, - "ftbc": true, - "calib_percentile": 99.9, - "calib_batches": 8, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 100, - "trained_T": 100, - "thresh_before": [ - 1.0, - 1.0, - 1.0, - 1.0 - ], - "thresh_after": [ - 2.332289695739746, - 0.8534926176071167, - 1.113392949104309, - 2.6782219409942627 - ], - "min_ticks_within_1pct_of_full": null, - "min_ticks_within_1pct_per_arm": { - "raw": 25, - "thresh": 16, - "ftbc": null, - "both": null + "reset_mode": "srif", + "learn_reset": true, + "lam_init": 0.9, + "v_reset_init": 0.0, + "u_bits": 0, + "u_scale": 1.0, + "u_headroom_requested": 2.0, + "u_headroom_actual": 0.0, + "u_threshold_level": 0, + "u_levels_above_threshold": null, + "reset_family": { + "reset_mode": "srif", + "learn_reset": true, + "num_spiking_sites": 4, + "extra_params": 8, + "extra_params_frac_of_weights": 6.5212854757929885e-06, + "extra_flops_per_tick_per_neuron": 0, + "learned_lam_per_layer": [ + 0.7039555907249451, + 0.7162760496139526, + 0.6912680864334106, + 0.8535905480384827 + ], + "learned_v_reset_per_layer": [ + 0.4652628004550934, + 0.5016207695007324, + 0.4343704283237457, + 0.916726291179657 + ] }, - "speedup_at_min_ticks": null, - "full_T_acc": 0.8828571428571429, - "t_sweep_rows": [ - { - "T": 2, - "arm": "raw", - "acc": 0.02642857142857143, - "syn_per_inference": 1752319.4204545454, - "dense_per_inference": 2453504.0, - "event_fraction_of_dense": 0.7142109490975134 - }, - { - "T": 2, - "arm": "thresh", - "acc": 0.03214285714285714, - "syn_per_inference": 1738330.94237013, - "dense_per_inference": 2453504.0, - "event_fraction_of_dense": 0.7085095204124917 - }, - { - "T": 2, - "arm": "ftbc", - "acc": 0.04428571428571428, - "syn_per_inference": 1793459.4740259738, - "dense_per_inference": 2453504.0, - "event_fraction_of_dense": 0.7309788262118072 - }, - { - "T": 2, - "arm": "both", - "acc": 0.07714285714285714, - "syn_per_inference": 1784578.0275974027, - "dense_per_inference": 2453504.0, - "event_fraction_of_dense": 0.72735892323689 - }, - { - "T": 3, - "arm": "raw", - "acc": 0.14642857142857144, - "syn_per_inference": 1817622.8660714284, - "dense_per_inference": 3680256.0, - "event_fraction_of_dense": 0.49388489987420126 - }, - { - "T": 3, - "arm": "thresh", - "acc": 0.22714285714285715, - "syn_per_inference": 1790499.9334415584, - "dense_per_inference": 3680256.0, - "event_fraction_of_dense": 0.4865150504316978 - }, - { - "T": 3, - "arm": "ftbc", - "acc": 0.08428571428571428, - "syn_per_inference": 1893111.2353896103, - "dense_per_inference": 3680256.0, - "event_fraction_of_dense": 0.5143966168086161 - }, - { - "T": 3, - "arm": "both", - "acc": 0.14714285714285713, - "syn_per_inference": 1874461.9123376624, - "dense_per_inference": 3680256.0, - "event_fraction_of_dense": 0.5093292184939479 - }, - { - "T": 4, - "arm": "raw", - "acc": 0.26785714285714285, - "syn_per_inference": 1874738.8027597403, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.3820533414169572 - }, - { - "T": 4, - "arm": "thresh", - "acc": 0.3871428571428571, - "syn_per_inference": 1833943.452922078, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.37373965009269966 - }, - { - "T": 4, - "arm": "ftbc", - "acc": 0.135, - "syn_per_inference": 1990891.393668831, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.4057240977941815 - }, - { - "T": 4, - "arm": "both", - "acc": 0.21142857142857144, - "syn_per_inference": 1961136.798701299, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.39966040379418555 - }, - { - "T": 6, - "arm": "raw", - "acc": 0.4764285714285714, - "syn_per_inference": 1969987.3693181819, - "dense_per_inference": 7360512.0, - "event_fraction_of_dense": 0.26764270872979784 - }, - { - "T": 6, - "arm": "thresh", - "acc": 0.6278571428571429, - "syn_per_inference": 1902471.609577922, - "dense_per_inference": 7360512.0, - "event_fraction_of_dense": 0.25847000990935437 - }, - { - "T": 6, - "arm": "ftbc", - "acc": 0.255, - "syn_per_inference": 2175123.340909091, - "dense_per_inference": 7360512.0, - "event_fraction_of_dense": 0.29551250523184946 - }, - { - "T": 6, - "arm": "both", - "acc": 0.3507142857142857, - "syn_per_inference": 2125190.585227273, - "dense_per_inference": 7360512.0, - "event_fraction_of_dense": 0.28872863534863785 - }, - { - "T": 8, - "arm": "raw", - "acc": 0.6135714285714285, - "syn_per_inference": 2061354.8465909092, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.21004192846138717 - }, - { - "T": 8, - "arm": "thresh", - "acc": 0.7464285714285714, - "syn_per_inference": 1962566.861201299, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.19997591823788538 - }, - { - "T": 8, - "arm": "ftbc", - "acc": 0.32357142857142857, - "syn_per_inference": 2363100.6761363638, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.2407883455800728 - }, - { - "T": 8, - "arm": "both", - "acc": 0.4635714285714286, - "syn_per_inference": 2291690.112012987, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.2335119600388859 - }, - { - "T": 12, - "arm": "raw", - "acc": 0.7221428571428572, - "syn_per_inference": 2225777.935064935, - "dense_per_inference": 14721024.0, - "event_fraction_of_dense": 0.15119722208624448 - }, - { - "T": 12, - "arm": "thresh", - "acc": 0.8571428571428571, - "syn_per_inference": 2062856.625, - "dense_per_inference": 14721024.0, - "event_fraction_of_dense": 0.1401299682005817 - }, - { - "T": 12, - "arm": "ftbc", - "acc": 0.47714285714285715, - "syn_per_inference": 2584647.832792208, - "dense_per_inference": 14721024.0, - "event_fraction_of_dense": 0.17557527470862136 - }, - { - "T": 12, - "arm": "both", - "acc": 0.5707142857142857, - "syn_per_inference": 2447611.5698051946, - "dense_per_inference": 14721024.0, - "event_fraction_of_dense": 0.1662663935474322 - }, - { - "T": 16, - "arm": "raw", - "acc": 0.8064285714285714, - "syn_per_inference": 2357961.4675324676, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.12013234273983595 - }, - { - "T": 16, - "arm": "thresh", - "acc": 0.8921428571428571, - "syn_per_inference": 2139811.6217532465, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.10901814413962879 - }, - { - "T": 16, - "arm": "ftbc", - "acc": 0.6371428571428571, - "syn_per_inference": 2753975.573051948, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.14030828832212766 - }, - { - "T": 16, - "arm": "both", - "acc": 0.6585714285714286, - "syn_per_inference": 2552953.403409091, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.13006670273459361 - }, - { - "T": 25, - "arm": "raw", - "acc": 0.8828571428571429, - "syn_per_inference": 2582600.840909091, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.08420938676795606 - }, - { - "T": 25, - "arm": "thresh", - "acc": 0.9257142857142857, - "syn_per_inference": 2259412.2305194805, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.07367136081357864 - }, - { - "T": 25, - "arm": "ftbc", - "acc": 0.8107142857142857, - "syn_per_inference": 3033411.4326298703, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.09890870958856787 - }, - { - "T": 25, - "arm": "both", - "acc": 0.765, - "syn_per_inference": 2715560.4196428573, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.08854472361627638 + "reset_stats": { + "reset_mode": "srif", + "learn_reset": true, + "discarded_charge_per_reset": -0.40509077622122347, + "spikes_per_inference": 9121.73828125, + "syn_per_inference": 4631522.0, + "lam_per_site": [ + 0.7039555907249451, + 0.7162760496139526, + 0.6912680864334106, + 0.8535905480384827 + ], + "v_reset_per_site": [ + 0.4652628004550934, + 0.5016207695007324, + 0.4343704283237457, + 0.916726291179657 + ], + "u_bits": 0, + "membrane_saturation_frac": null, + "batches": 4, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null + }, + "reset_x_bits_sweep": { + "rows": [ + { + "reset_mode": "hard", + "u_bits": 0, + "acc": 0.7407142857142857, + "target_acc": 0.7407142857142858, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null, + "discarded_charge_per_reset": 0.15744831227474765, + "spikes_per_inference": 6255.1328125, + "membrane_saturation_frac": null, + "membrane_memory_reduction_x": 1.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "hard", + "u_bits": 8, + "acc": 0.8028571428571428, + "target_acc": 0.8028571428571428, + "grid_headroom": 1.984375, + "grid_levels_above_thresh": 63, + "discarded_charge_per_reset": 0.1512959886308234, + "spikes_per_inference": 6588.09375, + "membrane_saturation_frac": 0.10686842032358157, + "membrane_memory_reduction_x": 4.0, + "acc_drop_vs_fp32": -0.062142857142857166 + }, + { + "reset_mode": "hard", + "u_bits": 4, + "acc": 0.8342857142857143, + "target_acc": 0.8342857142857143, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3, + "discarded_charge_per_reset": 0.05941473367507645, + "spikes_per_inference": 6960.21875, + "membrane_saturation_frac": 0.16326490469858157, + "membrane_memory_reduction_x": 8.0, + "acc_drop_vs_fp32": -0.09357142857142864 + }, + { + "reset_mode": "hard", + "u_bits": 3, + "acc": 0.8107142857142857, + "target_acc": 0.8107142857142857, + "grid_headroom": 1.5, + "grid_levels_above_thresh": 1, + "discarded_charge_per_reset": 0.01613670709751149, + "spikes_per_inference": 7805.390625, + "membrane_saturation_frac": 0.18435373725620569, + "membrane_memory_reduction_x": 10.666666666666666, + "acc_drop_vs_fp32": -0.07000000000000006 + }, + { + "reset_mode": "hard", + "u_bits": 2, + "acc": 0.4085714285714286, + "target_acc": 0.4085714285714286, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 2916.671875, + "membrane_saturation_frac": 0.1515874335106383, + "membrane_memory_reduction_x": 16.0, + "acc_drop_vs_fp32": 0.3321428571428571 + }, + { + "reset_mode": "soft", + "u_bits": 0, + "acc": 0.8307142857142857, + "target_acc": 0.8307142857142857, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 7030.25, + "membrane_saturation_frac": null, + "membrane_memory_reduction_x": 1.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "soft", + "u_bits": 8, + "acc": 0.87, + "target_acc": 0.87, + "grid_headroom": 1.984375, + "grid_levels_above_thresh": 63, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 7334.640625, + "membrane_saturation_frac": 0.1289720467641844, + "membrane_memory_reduction_x": 4.0, + "acc_drop_vs_fp32": -0.03928571428571426 + }, + { + "reset_mode": "soft", + "u_bits": 4, + "acc": 0.8657142857142857, + "target_acc": 0.8657142857142858, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 7372.984375, + "membrane_saturation_frac": 0.180172733820922, + "membrane_memory_reduction_x": 8.0, + "acc_drop_vs_fp32": -0.03499999999999992 + }, + { + "reset_mode": "soft", + "u_bits": 3, + "acc": 0.8271428571428572, + "target_acc": 0.8271428571428572, + "grid_headroom": 1.5, + "grid_levels_above_thresh": 1, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 7913.796875, + "membrane_saturation_frac": 0.1904026761968085, + "membrane_memory_reduction_x": 10.666666666666666, + "acc_drop_vs_fp32": 0.0035714285714285587 + }, + { + "reset_mode": "soft", + "u_bits": 2, + "acc": 0.4085714285714286, + "target_acc": 0.4085714285714286, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 2916.6796875, + "membrane_saturation_frac": 0.15158750277039007, + "membrane_memory_reduction_x": 16.0, + "acc_drop_vs_fp32": 0.42214285714285715 + }, + { + "reset_mode": "trained", + "u_bits": 0, + "acc": 0.8478571428571429, + "target_acc": 0.8478571428571429, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null, + "discarded_charge_per_reset": -0.4054994834959055, + "spikes_per_inference": 9098.09375, + "membrane_saturation_frac": null, + "membrane_memory_reduction_x": 1.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "trained", + "u_bits": 8, + "acc": 0.8957142857142857, + "target_acc": 0.8957142857142857, + "grid_headroom": 1.984375, + "grid_levels_above_thresh": 63, + "discarded_charge_per_reset": -0.41906449639349935, + "spikes_per_inference": 9753.5703125, + "membrane_saturation_frac": 0.23588444703014183, + "membrane_memory_reduction_x": 4.0, + "acc_drop_vs_fp32": -0.04785714285714282 + }, + { + "reset_mode": "trained", + "u_bits": 4, + "acc": 0.8957142857142857, + "target_acc": 0.8957142857142857, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3, + "discarded_charge_per_reset": -0.4484818391381923, + "spikes_per_inference": 10158.6953125, + "membrane_saturation_frac": 0.3265235067597517, + "membrane_memory_reduction_x": 8.0, + "acc_drop_vs_fp32": -0.04785714285714282 + }, + { + "reset_mode": "trained", + "u_bits": 3, + "acc": 0.8785714285714286, + "target_acc": 0.8785714285714286, + "grid_headroom": 1.5, + "grid_levels_above_thresh": 1, + "discarded_charge_per_reset": -0.45839456771794834, + "spikes_per_inference": 11573.875, + "membrane_saturation_frac": 0.3351565962987589, + "membrane_memory_reduction_x": 10.666666666666666, + "acc_drop_vs_fp32": -0.030714285714285694 + }, + { + "reset_mode": "trained", + "u_bits": 2, + "acc": 0.44357142857142856, + "target_acc": 0.44357142857142856, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0, + "discarded_charge_per_reset": -0.47925886040673965, + "spikes_per_inference": 12766.6015625, + "membrane_saturation_frac": 0.3268530446586879, + "membrane_memory_reduction_x": 16.0, + "acc_drop_vs_fp32": 0.4042857142857143 + } + ], + "by_mode": { + "hard": { + "fp32_acc": 0.7407142857142857, + "acc_by_bits": { + "0": 0.7407142857142857, + "8": 0.8028571428571428, + "4": 0.8342857142857143, + "3": 0.8107142857142857, + "2": 0.4085714285714286 + } + }, + "soft": { + "fp32_acc": 0.8307142857142857, + "acc_by_bits": { + "0": 0.8307142857142857, + "8": 0.87, + "4": 0.8657142857142857, + "3": 0.8271428571428572, + "2": 0.4085714285714286 + } + }, + "trained": { + "fp32_acc": 0.8478571428571429, + "acc_by_bits": { + "0": 0.8478571428571429, + "8": 0.8957142857142857, + "4": 0.8957142857142857, + "3": 0.8785714285714286, + "2": 0.44357142857142856 + } + } } - ], - "calibration_reference": "same network at full T (NOT a source ANN)" + }, + "mechanism": "V = u - s*((1-lam)*u + lam*thresh) + s*v_r; lam=0,v_r=0 is the baseline hard reset exactly, lam=1,v_r=0 discards zero charge", + "cost": "2 scalars per spiking layer; identical FMA count per tick, so the reset change is free at inference time", + "caveat": "the sweep swaps the reset on ALREADY-TRAINED weights, which under-states what training with each reset would give; and a charge-conserving reset can raise the spike count, so read spikes_per_inference next to the accuracy" }, "config": { "method": "harmonic_matched", @@ -4153,198 +6367,278 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "hub_dir": "/kaggle/working/hub/opt03_ftbc_threshold_balance/t_sweep_ultra_low", + "hub_dir": "/kaggle/working/hub/opt09_srif_softreset/reset_x_bits_full", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, "no_resume": true, "selftest": false, + "reset_mode": "srif", + "learn_reset": true, + "lam_init": 0.9, + "v_reset_init": 0.0, + "u_bits": 0, + "u_headroom": 2.0, + "reset_sweep": "hard,soft,trained", + "bits_sweep": "0,8,4,3,2", "baseline": false, "bench": true, "bench_warmup": 10, "bench_iters": 50, "bench_train_steps": 20, - "calibrate_thresh": true, - "calib_percentile": 99.9, - "calib_batches": 8, - "ftbc": true, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 0, - "ftbc_lr": 1.0, - "t_sweep": "2,3,4,6,8,12,16,25", "sampling_rate": 250.0 } }, "zip": { - "name": "t_sweep_ultra_low.zip", + "name": "reset_x_bits_full.zip", "files": 81, - "bytes": 12250607, + "bytes": 12269476, "MB": 12.3, - "sha256": "39c7a527158f3ac451c03f8cd69f8d779aa19f416260eb3eb5582179915c5f2a" + "sha256": "44447e048efd1105904179a9e495f1c420dc4b70d30bbf5a28d57dc17386d626" } }, { - "key": "opt03_ftbc_threshold_balance/thresh_only", - "tag": "thresh_only", - "script": "opt03_ftbc_threshold_balance.py", - "hf_folder": "applied/opt03_ftbc_threshold_balance", - "phase": 8, + "key": "opt09_srif_softreset/u_bits_2", + "tag": "u_bits_2", + "script": "opt09_srif_softreset.py", + "hf_folder": "applied/opt09_srif_softreset", + "phase": 10, "status": "ok", - "duration_s": 361.1, + "duration_s": 303.1, "optimization_summary": { - "optimization": "ftbc_threshold_balance", - "research_section": "2. Temporal-Dimension Reduction (threshold balancing + FTBC)", - "reference": "Rueckauer et al. arXiv:1612.04052; Sengupta et al. arXiv:1802.02627; FTBC: Forward Temporal Bias Correction, ECCV 2024", - "script": "opt03_ftbc_threshold_balance.py", + "optimization": "srif_softreset", + "research_section": "4. Membrane-Potential Memory Wall -- Soft Reset Integrate-and-Fire (SRIF), dynamic reset potential", + "reference": "Guo et al., 'Reducing Information Loss for Spiking Neural Networks' (SRIF), ECCV 2022; research.md S4 (stacks with MINT / SpQuant-SNN membrane quantization)", + "script": "opt09_srif_softreset.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.93, - "test_target_acc": 0.93, - "best_val_acc": 0.9442857142857143 + "test_acc": 0.025, + "test_target_acc": 0.025, + "best_val_acc": 0.025 }, "efficiency": { - "spikes_per_inference": null, - "syn_ops_per_inference": 3291572.5738636362, + "spikes_per_inference": 0.0, + "syn_ops_per_inference": 1578887.635551948, "dense_macs_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026831605523069343, - "syn_ops_per_inference_dense_fanout": 64868190.36038961, + "event_fraction_of_dense": 0.012870471257042566, + "syn_ops_per_inference_dense_fanout": 63155505.422077924, "weight_sparsity": 0.7324707846410685, - "effective_ticks": 12.039285714285715, + "effective_ticks": 100.0, "num_ticks": 100, - "membrane_bytes_per_inference": 451200, + "membrane_bytes_per_inference": 0, "weight_bytes": 4907008, "input_raster_bytes": null, - "train_ms_per_step": 98.61616005000542, - "infer_ms_per_sample": 0.648547162812747, - "peak_vram_mb": 2825.319424 + "train_ms_per_step": 82.91583395002817, + "infer_ms_per_sample": 0.5732730162498001, + "peak_vram_mb": 2820.341248, + "discarded_charge_per_reset": 0.0 }, "optimization_specific": { - "calibrate_thresh": true, - "ftbc": false, - "calib_percentile": 99.9, - "calib_batches": 8, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 100, - "trained_T": 100, - "thresh_before": [ - 1.0, - 1.0, - 1.0, - 1.0 - ], - "thresh_after": [ - 2.316941022872925, - 0.8509324193000793, - 1.1109048128128052, - 2.703218460083008 - ], - "min_ticks_within_1pct_of_full": 50, - "min_ticks_within_1pct_per_arm": { - "raw": 50, - "thresh": 50 + "reset_mode": "srif", + "learn_reset": true, + "lam_init": 0.9, + "v_reset_init": 0.0, + "u_bits": 2, + "u_scale": 1.0, + "u_headroom_requested": 2.0, + "u_headroom_actual": 1.0, + "u_threshold_level": 1, + "u_levels_above_threshold": 0, + "reset_family": { + "reset_mode": "srif", + "learn_reset": true, + "num_spiking_sites": 4, + "extra_params": 8, + "extra_params_frac_of_weights": 6.5212854757929885e-06, + "extra_flops_per_tick_per_neuron": 0, + "learned_lam_per_layer": [ + 0.5803266763687134, + 0.5803266763687134, + 0.5803266763687134, + 0.5803266763687134 + ], + "learned_v_reset_per_layer": [ + 0.0, + 0.0, + 0.0, + 0.0 + ] }, - "speedup_at_min_ticks": 2.0, - "full_T_acc": 0.9457142857142857, - "t_sweep_rows": [ - { - "T": 4, - "arm": "raw", - "acc": 0.2742857142857143, - "syn_per_inference": 1874088.702922078, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.38192085745979587 - }, - { - "T": 4, - "arm": "thresh", - "acc": 0.41285714285714287, - "syn_per_inference": 1834095.9163961038, - "dense_per_inference": 4907008.0, - "event_fraction_of_dense": 0.3737707206501607 - }, - { - "T": 8, - "arm": "raw", - "acc": 0.595, - "syn_per_inference": 2061721.9034090908, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.2100793297472809 - }, - { - "T": 8, - "arm": "thresh", - "acc": 0.7728571428571429, - "syn_per_inference": 1963701.1875, - "dense_per_inference": 9814016.0, - "event_fraction_of_dense": 0.20009150051314364 - }, - { - "T": 16, - "arm": "raw", - "acc": 0.8207142857142857, - "syn_per_inference": 2358720.708603896, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.1201710242068026 - }, - { - "T": 16, - "arm": "thresh", - "acc": 0.9035714285714286, - "syn_per_inference": 2142573.9935064935, - "dense_per_inference": 19628032.0, - "event_fraction_of_dense": 0.1091588801927006 - }, - { - "T": 25, - "arm": "raw", - "acc": 0.89, - "syn_per_inference": 2583865.2305194805, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.08425061399596595 - }, - { - "T": 25, - "arm": "thresh", - "acc": 0.9307142857142857, - "syn_per_inference": 2263356.4123376627, - "dense_per_inference": 30668800.0, - "event_fraction_of_dense": 0.07379996649160263 - }, - { - "T": 50, - "arm": "raw", - "acc": 0.9385714285714286, - "syn_per_inference": 2942934.289772727, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.047979286600270094 - }, - { - "T": 50, - "arm": "thresh", - "acc": 0.9421428571428572, - "syn_per_inference": 2448527.6655844157, - "dense_per_inference": 61337600.0, - "event_fraction_of_dense": 0.039918869756632405 - }, - { - "T": 100, - "arm": "raw", - "acc": 0.9457142857142857, - "syn_per_inference": 3291572.496753247, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.026831604894495768 - }, - { - "T": 100, - "arm": "thresh", - "acc": 0.93, - "syn_per_inference": 2714702.452922078, - "dense_per_inference": 122675200.0, - "event_fraction_of_dense": 0.02212918709667543 + "reset_stats": { + "reset_mode": "srif", + "learn_reset": true, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 0.0, + "syn_per_inference": 1664173.09375, + "lam_per_site": [ + 0.5803266763687134, + 0.5803266763687134, + 0.5803266763687134, + 0.5803266763687134 + ], + "v_reset_per_site": [ + 0.0, + 0.0, + 0.0, + 0.0 + ], + "u_bits": 2, + "membrane_saturation_frac": 0.0, + "batches": 4, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0 + }, + "reset_x_bits_sweep": { + "rows": [ + { + "reset_mode": "hard", + "u_bits": 0, + "acc": 0.025, + "target_acc": 0.025, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 0.0, + "membrane_saturation_frac": null, + "membrane_memory_reduction_x": 1.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "hard", + "u_bits": 4, + "acc": 0.025, + "target_acc": 0.025, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 0.0, + "membrane_saturation_frac": 0.0, + "membrane_memory_reduction_x": 8.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "hard", + "u_bits": 2, + "acc": 0.025, + "target_acc": 0.025, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 0.0, + "membrane_saturation_frac": 0.0, + "membrane_memory_reduction_x": 16.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "soft", + "u_bits": 0, + "acc": 0.025, + "target_acc": 0.025, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 0.0, + "membrane_saturation_frac": null, + "membrane_memory_reduction_x": 1.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "soft", + "u_bits": 4, + "acc": 0.025, + "target_acc": 0.025, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 0.0, + "membrane_saturation_frac": 0.0, + "membrane_memory_reduction_x": 8.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "soft", + "u_bits": 2, + "acc": 0.025, + "target_acc": 0.025, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 0.0, + "membrane_saturation_frac": 0.0, + "membrane_memory_reduction_x": 16.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "trained", + "u_bits": 0, + "acc": 0.025, + "target_acc": 0.025, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 0.0, + "membrane_saturation_frac": null, + "membrane_memory_reduction_x": 1.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "trained", + "u_bits": 4, + "acc": 0.025, + "target_acc": 0.025, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 0.0, + "membrane_saturation_frac": 0.0, + "membrane_memory_reduction_x": 8.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "trained", + "u_bits": 2, + "acc": 0.025, + "target_acc": 0.025, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 0.0, + "membrane_saturation_frac": 0.0, + "membrane_memory_reduction_x": 16.0, + "acc_drop_vs_fp32": 0.0 + } + ], + "by_mode": { + "hard": { + "fp32_acc": 0.025, + "acc_by_bits": { + "0": 0.025, + "4": 0.025, + "2": 0.025 + } + }, + "soft": { + "fp32_acc": 0.025, + "acc_by_bits": { + "0": 0.025, + "4": 0.025, + "2": 0.025 + } + }, + "trained": { + "fp32_acc": 0.025, + "acc_by_bits": { + "0": 0.025, + "4": 0.025, + "2": 0.025 + } + } } - ], - "calibration_reference": "same network at full T (NOT a source ANN)" + }, + "mechanism": "V = u - s*((1-lam)*u + lam*thresh) + s*v_r; lam=0,v_r=0 is the baseline hard reset exactly, lam=1,v_r=0 discards zero charge", + "cost": "2 scalars per spiking layer; identical FMA count per tick, so the reset change is free at inference time", + "caveat": "the sweep swaps the reset on ALREADY-TRAINED weights, which under-states what training with each reset would give; and a charge-conserving reset can raise the spike count, so read spikes_per_inference next to the accuracy" }, "config": { "method": "harmonic_matched", @@ -4384,137 +6678,278 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "hub_dir": "/kaggle/working/hub/opt03_ftbc_threshold_balance/thresh_only", + "hub_dir": "/kaggle/working/hub/opt09_srif_softreset/u_bits_2", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, "no_resume": true, "selftest": false, + "reset_mode": "srif", + "learn_reset": true, + "lam_init": 0.9, + "v_reset_init": 0.0, + "u_bits": 2, + "u_headroom": 2.0, + "reset_sweep": "hard,soft,trained", + "bits_sweep": "0,4,2", "baseline": false, "bench": true, "bench_warmup": 10, "bench_iters": 50, "bench_train_steps": 20, - "calibrate_thresh": true, - "calib_percentile": 99.9, - "calib_batches": 8, - "ftbc": false, - "ftbc_target_ticks": 8, - "ftbc_ref_ticks": 0, - "ftbc_lr": 1.0, - "t_sweep": "4,8,16,25,50,100", "sampling_rate": 250.0 } }, "zip": { - "name": "thresh_only.zip", + "name": "u_bits_2.zip", "files": 81, - "bytes": 12246575, - "MB": 12.2, - "sha256": "17a308e5444a23727992a34377bb7f0dc5f72c901454d5a8a2380dbb443a30b6" + "bytes": 10276157, + "MB": 10.3, + "sha256": "73e4f4b1f74a1efe218f865b59d03062b3c9d292f5b6a35747f4559d0736ef1f" } }, { - "key": "opt04_cudagraph_unroll/bench_batches_wide", - "tag": "bench_batches_wide", - "script": "opt04_cudagraph_unroll.py", - "hf_folder": "applied/opt04_cudagraph_unroll", - "phase": 8, + "key": "opt09_srif_softreset/u_bits_4", + "tag": "u_bits_4", + "script": "opt09_srif_softreset.py", + "hf_folder": "applied/opt09_srif_softreset", + "phase": 10, "status": "ok", - "duration_s": 639.6, + "duration_s": 345.7, "optimization_summary": { - "optimization": "cudagraph_static_unroll", - "research_section": "2. Temporal-dimension reduction / 5. OS & runtime", - "reference": "CUDA Graphs static capture/replay applied to the unrolled T-tick event-driven inference loop", - "script": "opt04_cudagraph_unroll.py", + "optimization": "srif_softreset", + "research_section": "4. Membrane-Potential Memory Wall -- Soft Reset Integrate-and-Fire (SRIF), dynamic reset potential", + "reference": "Guo et al., 'Reducing Information Loss for Spiking Neural Networks' (SRIF), ECCV 2022; research.md S4 (stacks with MINT / SpQuant-SNN membrane quantization)", + "script": "opt09_srif_softreset.py", "baseline_mode": false, "method": "harmonic_matched", "accuracy": { - "test_acc": 0.9492857142857143, - "test_target_acc": 0.9492857142857143, - "best_val_acc": 0.9428571428571428 + "test_acc": 0.92, + "test_target_acc": 0.92, + "best_val_acc": 0.935 }, "efficiency": { - "spikes_per_inference": null, - "syn_ops_per_inference": 3288914.2832792206, + "spikes_per_inference": 11915.01171875, + "syn_ops_per_inference": 4401361.720779221, "dense_macs_per_inference": 122675200.0, - "event_fraction_of_dense": 0.02680993618334611, - "syn_ops_per_inference_dense_fanout": 64865532.0698052, + "event_fraction_of_dense": 0.035878170329285954, + "syn_ops_per_inference_dense_fanout": 65977979.5073052, "weight_sparsity": 0.7324707846410685, - "effective_ticks": 12.153571428571428, + "effective_ticks": 21.25857142857143, "num_ticks": 100, - "membrane_bytes_per_inference": 451200, + "membrane_bytes_per_inference": 0, "weight_bytes": 4907008, "input_raster_bytes": null, - "train_ms_per_step": 96.87458276748657, - "infer_ms_per_sample": 0.6550618261098862, - "peak_vram_mb": 2804.85888 + "train_ms_per_step": 98.06833220000044, + "infer_ms_per_sample": 0.650274362812695, + "peak_vram_mb": 2834.694656, + "discarded_charge_per_reset": -0.7078402399200352 }, "optimization_specific": { - "graph_enabled": true, - "event_capacity_scan": { - "cap_scan": 98901, - "batch_size_scan": 64, - "n_batches_scanned": 20, - "density_per_sample": 1545.328125, - "per_layer_caps": [ - 98901, - 7935, - 3512, - 3383 + "reset_mode": "srif", + "learn_reset": true, + "lam_init": 0.9, + "v_reset_init": 0.0, + "u_bits": 4, + "u_scale": 0.25, + "u_headroom_requested": 2.0, + "u_headroom_actual": 1.75, + "u_threshold_level": 4, + "u_levels_above_threshold": 3, + "reset_family": { + "reset_mode": "srif", + "learn_reset": true, + "num_spiking_sites": 4, + "extra_params": 8, + "extra_params_frac_of_weights": 6.5212854757929885e-06, + "extra_flops_per_tick_per_neuron": 0, + "learned_lam_per_layer": [ + 0.5789579153060913, + 0.5431995391845703, + 0.5589428544044495, + 0.7961611747741699 + ], + "learned_v_reset_per_layer": [ + 0.6194214224815369, + 0.877713143825531, + 0.6152849197387695, + 0.9643499255180359 ] }, - "capacity": 98901, - "overflow_ticks_frac": 0.0, - "graph_vs_eager_pred_agreement": 1.0, - "graph_capture_ms": 246.45113945007324, - "graphed_acc": 0.9492857142857143, - "kernel_launches_eager": 400, - "kernel_launches_graphed": 2, - "launch_reduction_x": 200.0, - "bench_batches": { - "1": { - "ms_per_sample_eager": 155.08618831634521, - "ms_per_sample_graphed": 32.538652420043945, - "throughput_eager_samples_per_s": 6.4480274539999485, - "throughput_graphed_samples_per_s": 30.732680231833935, - "speedup_x": 4.766214233900218, - "capacity": 1546 - }, - "4": { - "ms_per_sample_eager": 39.0606689453125, - "ms_per_sample_graphed": 15.353316068649292, - "throughput_eager_samples_per_s": 25.601200056252637, - "throughput_graphed_samples_per_s": 65.13250919402033, - "speedup_x": 2.5441193792051506, - "capacity": 6182 - }, - "16": { - "ms_per_sample_eager": 10.658027529716492, - "ms_per_sample_graphed": 11.284635066986084, - "throughput_eager_samples_per_s": 93.82599146153645, - "throughput_graphed_samples_per_s": 88.61606902340718, - "speedup_x": 0.9444725032267306, - "capacity": 24726 - }, - "64": { - "ms_per_sample_eager": 3.5885700583457947, - "ms_per_sample_graphed": 9.764026775956154, - "throughput_eager_samples_per_s": 278.66252678398735, - "throughput_graphed_samples_per_s": 102.41676133687925, - "speedup_x": 0.3675297232062721, - "capacity": 98901 - }, - "256": { - "ms_per_sample_eager": 1.2360872328281403, - "ms_per_sample_graphed": 9.808671623468399, - "throughput_eager_samples_per_s": 809.0043917952473, - "throughput_graphed_samples_per_s": 101.95060436189773, - "speedup_x": 0.1260198404308547, - "capacity": 395604 + "reset_stats": { + "reset_mode": "srif", + "learn_reset": true, + "discarded_charge_per_reset": -0.7078402399200352, + "spikes_per_inference": 11915.01171875, + "syn_per_inference": 4434214.6875, + "lam_per_site": [ + 0.5789579153060913, + 0.5431995391845703, + 0.5589428544044495, + 0.7961611747741699 + ], + "v_reset_per_site": [ + 0.6194214224815369, + 0.877713143825531, + 0.6152849197387695, + 0.9643499255180359 + ], + "u_bits": 4, + "membrane_saturation_frac": 0.2822270126883865, + "batches": 4, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3 + }, + "reset_x_bits_sweep": { + "rows": [ + { + "reset_mode": "hard", + "u_bits": 0, + "acc": 0.295, + "target_acc": 0.295, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null, + "discarded_charge_per_reset": 0.08765906069564103, + "spikes_per_inference": 4919.84375, + "membrane_saturation_frac": null, + "membrane_memory_reduction_x": 1.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "hard", + "u_bits": 4, + "acc": 0.7471428571428571, + "target_acc": 0.7471428571428571, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3, + "discarded_charge_per_reset": 0.013719326346275418, + "spikes_per_inference": 5178.40625, + "membrane_saturation_frac": 0.12054306571365248, + "membrane_memory_reduction_x": 8.0, + "acc_drop_vs_fp32": -0.4521428571428571 + }, + { + "reset_mode": "hard", + "u_bits": 2, + "acc": 0.04, + "target_acc": 0.04, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 428.6640625, + "membrane_saturation_frac": 0.06521574412677306, + "membrane_memory_reduction_x": 16.0, + "acc_drop_vs_fp32": 0.255 + }, + { + "reset_mode": "soft", + "u_bits": 0, + "acc": 0.43785714285714283, + "target_acc": 0.43785714285714283, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 5315.078125, + "membrane_saturation_frac": null, + "membrane_memory_reduction_x": 1.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "soft", + "u_bits": 4, + "acc": 0.7878571428571428, + "target_acc": 0.7878571428571428, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 5275.203125, + "membrane_saturation_frac": 0.1267951435062057, + "membrane_memory_reduction_x": 8.0, + "acc_drop_vs_fp32": -0.35 + }, + { + "reset_mode": "soft", + "u_bits": 2, + "acc": 0.04, + "target_acc": 0.04, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0, + "discarded_charge_per_reset": 0.0, + "spikes_per_inference": 428.6640625, + "membrane_saturation_frac": 0.06521574412677306, + "membrane_memory_reduction_x": 16.0, + "acc_drop_vs_fp32": 0.39785714285714285 + }, + { + "reset_mode": "trained", + "u_bits": 0, + "acc": 0.8771428571428571, + "target_acc": 0.8771428571428571, + "grid_headroom": 0.0, + "grid_levels_above_thresh": null, + "discarded_charge_per_reset": -0.672806146062752, + "spikes_per_inference": 10959.46875, + "membrane_saturation_frac": null, + "membrane_memory_reduction_x": 1.0, + "acc_drop_vs_fp32": 0.0 + }, + { + "reset_mode": "trained", + "u_bits": 4, + "acc": 0.92, + "target_acc": 0.92, + "grid_headroom": 1.75, + "grid_levels_above_thresh": 3, + "discarded_charge_per_reset": -0.70911303501342, + "spikes_per_inference": 11931.078125, + "membrane_saturation_frac": 0.2747938137189716, + "membrane_memory_reduction_x": 8.0, + "acc_drop_vs_fp32": -0.04285714285714293 + }, + { + "reset_mode": "trained", + "u_bits": 2, + "acc": 0.03785714285714286, + "target_acc": 0.03785714285714286, + "grid_headroom": 1.0, + "grid_levels_above_thresh": 0, + "discarded_charge_per_reset": -0.7804550254675015, + "spikes_per_inference": 6802.96875, + "membrane_saturation_frac": 0.12682423260195036, + "membrane_memory_reduction_x": 16.0, + "acc_drop_vs_fp32": 0.8392857142857143 + } + ], + "by_mode": { + "hard": { + "fp32_acc": 0.295, + "acc_by_bits": { + "0": 0.295, + "4": 0.7471428571428571, + "2": 0.04 + } + }, + "soft": { + "fp32_acc": 0.43785714285714283, + "acc_by_bits": { + "0": 0.43785714285714283, + "4": 0.7878571428571428, + "2": 0.04 + } + }, + "trained": { + "fp32_acc": 0.8771428571428571, + "acc_by_bits": { + "0": 0.8771428571428571, + "4": 0.92, + "2": 0.03785714285714286 + } + } } }, - "speedup_x_at_batch1": 4.766214233900218 + "mechanism": "V = u - s*((1-lam)*u + lam*thresh) + s*v_r; lam=0,v_r=0 is the baseline hard reset exactly, lam=1,v_r=0 discards zero charge", + "cost": "2 scalars per spiking layer; identical FMA count per tick, so the reset change is free at inference time", + "caveat": "the sweep swaps the reset on ALREADY-TRAINED weights, which under-states what training with each reset would give; and a charge-conserving reset can raise the spike count, so read spikes_per_inference next to the accuracy" }, "config": { "method": "harmonic_matched", @@ -4554,29 +6989,34 @@ "decision_margin": 2.0, "export": false, "seed": 1234, - "hub_dir": "/kaggle/working/hub/opt04_cudagraph_unroll/bench_batches_wide", + "hub_dir": "/kaggle/working/hub/opt09_srif_softreset/u_bits_4", "repo_id": "UWU-R-13/SSVEP-SNN", "push_hf": false, "keep_epoch_weights": 1, "no_resume": true, "selftest": false, + "reset_mode": "srif", + "learn_reset": true, + "lam_init": 0.9, + "v_reset_init": 0.0, + "u_bits": 4, + "u_headroom": 2.0, + "reset_sweep": "hard,soft,trained", + "bits_sweep": "0,4,2", "baseline": false, "bench": true, "bench_warmup": 10, "bench_iters": 50, "bench_train_steps": 20, - "event_capacity": 0, - "graph_warmup": 3, - "bench_batches": "1,4,16,64,256", "sampling_rate": 250.0 } }, "zip": { - "name": "bench_batches_wide.zip", - "files": 78, - "bytes": 7977178, - "MB": 8.0, - "sha256": "20c84fb916addc61aa295dfeead6cf4874dd253da092adb70e951ec451943fe1" + "name": "u_bits_4.zip", + "files": 81, + "bytes": 12262238, + "MB": 12.3, + "sha256": "e272a8770dc758e21ec340f0e670fb3661af3f7c3d82166c465f0ded0a39437c" } } ] \ No newline at end of file