add run u_headroom_4.0

applied/opt09_srif_softreset/u_headroom_4.0.summary.json

@@ -0,0 +1,403 @@
+{
+  "key": "opt09_srif_softreset/u_headroom_4.0",
+  "tag": "u_headroom_4.0",
+  "script": "opt09_srif_softreset.py",
+  "script_sha256_16": "dd62e980e1fe3fd1",
+  "hf_folder": "applied/opt09_srif_softreset",
+  "priority": 2,
+  "note": "wide headroom, coarser grid step at the same bit count",
+  "phase": 10,
+  "n_phases": 10,
+  "method": "harmonic_matched",
+  "command": [
+    "/usr/bin/python3",
+    "-m",
+    "torch.distributed.run",
+    "--standalone",
+    "--nproc_per_node=2",
+    "/kaggle/working/scripts/opt09_srif_softreset.py",
+    "--method",
+    "harmonic_matched",
+    "--data-root",
+    "/kaggle/working/data",
+    "--epochs",
+    "60",
+    "--seed",
+    "1234",
+    "--batch-size",
+    "64",
+    "--event-eval",
+    "--no-resume",
+    "--decision-margin",
+    "2.0",
+    "--hub-dir",
+    "/kaggle/working/hub/opt09_srif_softreset/u_headroom_4.0",
+    "--keep-epoch-weights",
+    "1",
+    "--resplit",
+    "block",
+    "--drop-rest",
+    "--bench",
+    "--u-bits",
+    "4",
+    "--u-headroom",
+    "4.0"
+  ],
+  "flags": [
+    "--u-bits",
+    "4",
+    "--u-headroom",
+    "4.0"
+  ],
+  "epochs_requested": 60,
+  "gpus": 2,
+  "relative_epoch_cost": 1.0,
+  "exit_code": 0,
+  "status": "ok",
+  "duration_s": 327.3,
+  "started_utc": "2026-08-23T05:53:52.459741+00:00",
+  "finished_utc": "2026-08-23T05:59:19.805759+00:00",
+  "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": 0.8307142857142857,
+      "test_target_acc": 0.8307142857142857,
+      "best_val_acc": 0.9214285714285714
+    },
+    "efficiency": {
+      "spikes_per_inference": 1344.265625,
+      "syn_ops_per_inference": 1891052.4415584416,
+      "dense_macs_per_inference": 122675200.0,
+      "event_fraction_of_dense": 0.015415116026372418,
+      "syn_ops_per_inference_dense_fanout": 63467670.228084415,
+      "weight_sparsity": 0.7324707846410685,
+      "effective_ticks": 30.065714285714286,
+      "num_ticks": 100,
+      "membrane_bytes_per_inference": 0,
+      "weight_bytes": 4907008,
+      "input_raster_bytes": null,
+      "train_ms_per_step": 89.94520204996661,
+      "infer_ms_per_sample": 0.5865823306251627,
+      "peak_vram_mb": 2821.165568,
+      "discarded_charge_per_reset": -0.30676767663002447
+    },
+    "optimization_specific": {
+      "reset_mode": "srif",
+      "learn_reset": true,
+      "lam_init": 0.9,
+      "v_reset_init": 0.0,
+      "u_bits": 4,
+      "u_scale": 0.5,
+      "u_headroom_requested": 4.0,
+      "u_headroom_actual": 3.5,
+      "u_threshold_level": 2,
+      "u_levels_above_threshold": 5,
+      "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": [
+          0.5882294178009033,
+          0.5828343033790588,
+          0.5932866930961609,
+          0.8433560729026794
+        ],
+        "learned_v_reset_per_layer": [
+          0.09568826854228973,
+          0.24902032315731049,
+          0.11872328817844391,
+          0.9482309222221375
+        ]
+      },
+      "reset_stats": {
+        "reset_mode": "srif",
+        "learn_reset": true,
+        "discarded_charge_per_reset": -0.30676767663002447,
+        "spikes_per_inference": 1344.265625,
+        "syn_per_inference": 1978610.34375,
+        "lam_per_site": [
+          0.5882294178009033,
+          0.5828343033790588,
+          0.5932866930961609,
+          0.8433560729026794
+        ],
+        "v_reset_per_site": [
+          0.09568826854228973,
+          0.24902032315731049,
+          0.11872328817844391,
+          0.9482309222221375
+        ],
+        "u_bits": 4,
+        "membrane_saturation_frac": 0.11609094498005319,
+        "batches": 4,
+        "grid_headroom": 3.5,
+        "grid_levels_above_thresh": 5
+      },
+      "reset_x_bits_sweep": {
+        "rows": [
+          {
+            "reset_mode": "hard",
+            "u_bits": 0,
+            "acc": 0.645,
+            "target_acc": 0.645,
+            "grid_headroom": 0.0,
+            "grid_levels_above_thresh": null,
+            "discarded_charge_per_reset": 0.11399426552829378,
+            "spikes_per_inference": 1360.8203125,
+            "membrane_saturation_frac": null,
+            "membrane_memory_reduction_x": 1.0,
+            "acc_drop_vs_fp32": 0.0
+          },
+          {
+            "reset_mode": "hard",
+            "u_bits": 4,
+            "acc": 0.8457142857142858,
+            "target_acc": 0.8457142857142858,
+            "grid_headroom": 3.5,
+            "grid_levels_above_thresh": 5,
+            "discarded_charge_per_reset": 0.032880633416589214,
+            "spikes_per_inference": 788.109375,
+            "membrane_saturation_frac": 0.07383775210549645,
+            "membrane_memory_reduction_x": 8.0,
+            "acc_drop_vs_fp32": -0.20071428571428573
+          },
+          {
+            "reset_mode": "hard",
+            "u_bits": 2,
+            "acc": 0.21357142857142858,
+            "target_acc": 0.21357142857142858,
+            "grid_headroom": 1.0,
+            "grid_levels_above_thresh": 0,
+            "discarded_charge_per_reset": 0.0,
+            "spikes_per_inference": 152.6796875,
+            "membrane_saturation_frac": 0.04726756427304965,
+            "membrane_memory_reduction_x": 16.0,
+            "acc_drop_vs_fp32": 0.43142857142857144
+          },
+          {
+            "reset_mode": "soft",
+            "u_bits": 0,
+            "acc": 0.7114285714285714,
+            "target_acc": 0.7114285714285714,
+            "grid_headroom": 0.0,
+            "grid_levels_above_thresh": null,
+            "discarded_charge_per_reset": 0.0,
+            "spikes_per_inference": 1499.1796875,
+            "membrane_saturation_frac": null,
+            "membrane_memory_reduction_x": 1.0,
+            "acc_drop_vs_fp32": 0.0
+          },
+          {
+            "reset_mode": "soft",
+            "u_bits": 4,
+            "acc": 0.8435714285714285,
+            "target_acc": 0.8435714285714285,
+            "grid_headroom": 3.5,
+            "grid_levels_above_thresh": 5,
+            "discarded_charge_per_reset": 0.0,
+            "spikes_per_inference": 821.4296875,
+            "membrane_saturation_frac": 0.07612069204343971,
+            "membrane_memory_reduction_x": 8.0,
+            "acc_drop_vs_fp32": -0.13214285714285712
+          },
+          {
+            "reset_mode": "soft",
+            "u_bits": 2,
+            "acc": 0.21357142857142858,
+            "target_acc": 0.21357142857142858,
+            "grid_headroom": 1.0,
+            "grid_levels_above_thresh": 0,
+            "discarded_charge_per_reset": 0.0,
+            "spikes_per_inference": 152.6796875,
+            "membrane_saturation_frac": 0.04726756427304965,
+            "membrane_memory_reduction_x": 16.0,
+            "acc_drop_vs_fp32": 0.49785714285714283
+          },
+          {
+            "reset_mode": "trained",
+            "u_bits": 0,
+            "acc": 0.6678571428571428,
+            "target_acc": 0.6678571428571428,
+            "grid_headroom": 0.0,
+            "grid_levels_above_thresh": null,
+            "discarded_charge_per_reset": -0.3709850527872548,
+            "spikes_per_inference": 2268.5703125,
+            "membrane_saturation_frac": null,
+            "membrane_memory_reduction_x": 1.0,
+            "acc_drop_vs_fp32": 0.0
+          },
+          {
+            "reset_mode": "trained",
+            "u_bits": 4,
+            "acc": 0.8307142857142857,
+            "target_acc": 0.8307142857142857,
+            "grid_headroom": 3.5,
+            "grid_levels_above_thresh": 5,
+            "discarded_charge_per_reset": -0.32191684826040434,
+            "spikes_per_inference": 1345.703125,
+            "membrane_saturation_frac": 0.11365830008865249,
+            "membrane_memory_reduction_x": 8.0,
+            "acc_drop_vs_fp32": -0.16285714285714292
+          },
+          {
+            "reset_mode": "trained",
+            "u_bits": 2,
+            "acc": 0.16142857142857142,
+            "target_acc": 0.16142857142857142,
+            "grid_headroom": 1.0,
+            "grid_levels_above_thresh": 0,
+            "discarded_charge_per_reset": -0.7320368063795608,
+            "spikes_per_inference": 868.6796875,
+            "membrane_saturation_frac": 0.06705091976950354,
+            "membrane_memory_reduction_x": 16.0,
+            "acc_drop_vs_fp32": 0.5064285714285715
+          }
+        ],
+        "by_mode": {
+          "hard": {
+            "fp32_acc": 0.645,
+            "acc_by_bits": {
+              "0": 0.645,
+              "4": 0.8457142857142858,
+              "2": 0.21357142857142858
+            }
+          },
+          "soft": {
+            "fp32_acc": 0.7114285714285714,
+            "acc_by_bits": {
+              "0": 0.7114285714285714,
+              "4": 0.8435714285714285,
+              "2": 0.21357142857142858
+            }
+          },
+          "trained": {
+            "fp32_acc": 0.6678571428571428,
+            "acc_by_bits": {
+              "0": 0.6678571428571428,
+              "4": 0.8307142857142857,
+              "2": 0.16142857142857142
+            }
+          }
+        }
+      },
+      "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",
+      "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/opt09_srif_softreset/u_headroom_4.0",
+      "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": 4.0,
+      "reset_sweep": "hard,soft,trained",
+      "bits_sweep": "0,4,2",
+      "baseline": false,
+      "bench": true,
+      "bench_warmup": 10,
+      "bench_iters": 50,
+      "bench_train_steps": 20,
+      "sampling_rate": 250.0
+    }
+  },
+  "final_summary": {
+    "final_test_acc": 0.8307142857142857,
+    "final_test_target_acc": 0.8307142857142857,
+    "final_test_rest_recall": NaN,
+    "best_acc": 0.9214285714285714,
+    "best_epoch": 8,
+    "resplit": "block",
+    "class_weight": false,
+    "drop_rest": true,
+    "epochs": 60,
+    "method": "harmonic_matched",
+    "n_in": 2880,
+    "n_classes": 40,
+    "num_bins": 100
+  },
+  "event_driven_summary": {
+    "acc": 0.8307142857142857,
+    "anytime_acc": {
+      "0.25": 0.33714285714285713,
+      "0.5": 0.6392857142857142,
+      "0.75": 0.7664285714285715,
+      "1.0": 0.8307142857142857
+    },
+    "syn_per_inference": 1891052.4415584416,
+    "dense_per_inference": 122675200.0,
+    "event_fraction_of_dense": 0.015415116026372418,
+    "syn_per_inference_dense_fanout": 63467670.228084415,
+    "weight_sparsity": 0.7324707846410685,
+    "num_ticks": 100,
+    "dt_ms": 50.0,
+    "margin_decision_acc": 0.4207142857142857,
+    "mean_decision_tick": 30.065714285714286,
+    "mean_decision_ms": 1503.2857142857144,
+    "dense_forward_test_acc": 0.8307142857142857
+  },
+  "extra_sweep_file_present": true,
+  "selftest": {
+    "mode": "gate",
+    "passed": true
+  },
+  "driver_version": 1,
+  "zip": {
+    "name": "u_headroom_4.0.zip",
+    "files": 81,
+    "bytes": 12340512,
+    "MB": 12.3,
+    "sha256": "3394974815eca792e339c4516e7ced59d4a7259d00e7f9b0a8906fdcb18cea8d"
+  }
+}

applied/opt09_srif_softreset/u_headroom_4.0.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3394974815eca792e339c4516e7ced59d4a7259d00e7f9b0a8906fdcb18cea8d
+size 12340512