Update README with v5.1 results: 87 solved, LB 1164.25

own-solver/README.md

@@ -1,19 +1,23 @@
-# NeuroGolf Solver v5
+# NeuroGolf Solver v5.1
 
 Builds minimal ONNX networks for ARC-AGI tasks. Modular Python package with opset 17, zero-cost Slice-based transforms.
 
-**Currently running on Kaggle — results pending.**
+**Public LB: 1164.25** (V38, 87/400 solved)
 
 ## Version History
 
-| Version | Date | Solved (local) | Arc-gen Validated | Est LB | Key Changes |
-|---------|------|----------------|-------------------|--------|-------------|
-| **v5** | **2026-04-26** | **TBD** | **TBD** | **TBD** | Refactored to package, opset 17, Slice-based flip/rotate (0 MACs), lstsq crash fix, tensor-based Pad & ReduceSum |
-| v4.3 | 2026-04-25 | 307 | 50 | ~670 | Methodology docs, no code changes |
-| v4.0 | 2026-04-24 | 307 | 50 | ~656 | ARC-GEN validation, static profiler |
-| v3 | 2026-04-24 | 307 | ~40 | 501 | concat_enhanced, varshape_spatial_gather |
-| v2 | prior | 294 | — | — | Spatial_gather, variable-shape conv |
-| v1 | prior | 128 | — | — | Conv solver only |
+| Version | Date | Solved | Est LB | Public LB | Key Changes |
+|---------|------|--------|--------|-----------|-------------|
+| **v5.1** | **2026-05-16** | **87** | **1088.8** | **1164.25** | Wave 5: +6 analytical solvers (border_detect, fill_bbox, find_hole, row_uniform, majority_replace, symmetry_check). +7 tasks over v5.0 |
+| v5.0 | 2026-05-16 | 80 | 995.5 | 1071.38 | Baseline v5 on Kaggle with conv_budget=5 |
+| v5-dev | 2026-04-26 | — | — | — | Refactored to package, opset 17, Slice-based flip/rotate (0 MACs), lstsq crash fix, tensor-based Pad & ReduceSum |
+| v4.3 | 2026-04-25 | 307* | ~670 | — | Methodology docs, no code changes |
+| v4.0 | 2026-04-24 | 307* | ~656 | — | ARC-GEN validation, static profiler |
+| v3 | 2026-04-24 | 307* | 501 | — | concat_enhanced, varshape_spatial_gather |
+| v2 | prior | 294* | — | — | Spatial_gather, variable-shape conv |
+| v1 | prior | 128* | — | — | Conv solver only |
+
+*Local solve count (not all pass ARC-GEN validation)
 
 ## Project Structure
 
@@ -28,19 +32,28 @@ neurogolf_solver/               # Python package (v5)
 ├── onnx_helpers.py             # Opset 17 builders (Slice, Pad, ReduceSum, mk)
 ├── profiler.py                 # Static cost profiler (fallback for onnx_tool)
 ├── submission.py               # run_tasks with W&B logging, zip/csv generation
+├── task_scanner.py             # Task analysis utilities
 ├── validators.py               # Model validation against train+test+arc-gen
 └── solvers/
     ├── __init__.py             # Exports solve_task, ANALYTICAL_SOLVERS
     ├── analytical.py           # identity, constant, color_map, transpose
     ├── conv.py                 # lstsq conv solvers (fixed, variable, diffshape, var_diff)
+    ├── edge.py                 # Edge detection solver
     ├── geometric.py            # flip, rotate, shift, crop, gravity
-    ├── solver_registry.py      # Solver ordering + solve_task orchestration
-    └── tiling.py               # tile, upscale, mirror, concat, spatial_gather
-
-neurogolf_solver.py             # Legacy monolith (v4, kept for reference)
-neurogolf_utils.py              # Official Kaggle scoring library
-ARC-GEN-100K.zip                # 400 files × ~250 examples synthetic data
-neurogolf-2026-solver-notebooks.zip  # 5 reference notebooks (LB 4000+)
+    ├── gravity.py              # Unrolled gravity solver
+    ├── mode.py                 # Mode fill solver
+    ├── solver_registry.py      # Solver ordering + solve_task orchestration (66 solvers)
+    ├── tiling.py               # tile, upscale, mirror, concat, spatial_gather
+    ├── wave1.py                # downsample, symmetry_complete, extract_inner, etc.
+    ├── wave2.py                # overlay, bbox_crop, row/col_mode_fill, swap, etc.
+    ├── wave3.py                # split_and_merge, self_tile
+    ├── wave4.py                # Object extraction (densest/smallest/largest), symmetry_max, crop
+    ├── wave5.py                # Row-uniform, majority-replace, find-hole, fill-bbox, border, symmetry-check
+    ├── _classify_dominant.py   # Classify by dominant color
+    ├── _misc_solvers.py        # keep_first_third_cols, marker_extract
+    ├── _replace_col0.py        # Replace color from column 0
+    ├── _variable_fliplr.py     # Variable-width horizontal flip
+    └── _variable_flipud.py     # Variable-height vertical flip
 ```
 
 ## Quick Start
@@ -48,21 +61,18 @@ neurogolf-2026-solver-notebooks.zip  # 5 reference notebooks (LB 4000+)
 ```bash
 # Clone
 git clone https://huggingface.co/rogermt/neurogolf-solver
-cd neurogolf-solver
+cd neurogolf-solver/own-solver
 
 # Install deps
 pip install numpy onnx onnxruntime
 
-# Get ARC data
-git clone --depth 1 https://github.com/fchollet/ARC-AGI.git
+# Run (Kaggle)
+python -m neurogolf_solver.main --kaggle --data_dir /kaggle/input/competitions/neurogolf-2026/ --output_dir /kaggle/working/submission --conv_budget 5 --use_wandb
 
-# Run (local)
+# Run (local with ARC data)
 python -m neurogolf_solver.main --data_dir ARC-AGI/data/training/ --output_dir submission --conv_budget 30
 
-# Run (Kaggle)
-python -m neurogolf_solver.main --kaggle --data_dir /kaggle/input/competitions/neurogolf-2026/ --output_dir /kaggle/working/submission --conv_budget 60
-
-# With ARC-GEN data and W&B logging
+# With ARC-GEN data
 python -m neurogolf_solver.main --data_dir ARC-AGI/data/training/ --arcgen_dir ARC-GEN-100K/ --output_dir submission --use_wandb
 ```
 
@@ -84,8 +94,8 @@ python -m neurogolf_solver.main --data_dir ARC-AGI/data/training/ --arcgen_dir A
 **Format:** Input/output = `[1, 10, 30, 30]` one-hot float32. ONNX opset 17, IR version 8.
 
 **Solver pipeline (in order):**
-1. **Analytical solvers** (instant, near-zero cost):
-   identity → constant → color_map → transpose → flip → rotate → shift → tile → upscale → kronecker → nonuniform_scale → mirror_h → mirror_v → quad_mirror → concat → concat_enhanced → diagonal_tile → fixed_crop → spatial_gather → varshape_spatial_gather
+1. **Analytical solvers** (instant, near-zero cost, 66 solvers):
+   identity → constant → color_map → transpose → flip → rotate → shift → tile → upscale → kronecker → nonuniform_scale → mirror_h → mirror_v → quad_mirror → concat → concat_enhanced → diagonal_tile → fixed_crop → spatial_gather → varshape_spatial_gather → gravity_unrolled → edge_detect → mode_fill → [wave1-5 solvers] → marker_extract → **wave5**: row_uniform_fill, majority_replace_5, find_color_with_hole, fill_bbox_per_color, border_detect, symmetry_check_lr
 
 2. **Conv solvers** (learned via least-squares, validated against arc-gen):
    - `conv_fixed` — Slice → Conv → ArgMax → Equal+Cast → Pad
@@ -93,18 +103,16 @@ python -m neurogolf_solver.main --data_dir ARC-AGI/data/training/ --arcgen_dir A
    - `conv_diffshape` — Slice → Conv → Slice(crop) → ArgMax → Equal+Cast → Pad
    - `conv_var_diff` — Conv(30×30) → ArgMax → Equal+Cast → Mul(input_mask)
 
-## v5 Changes from v4
+## v5.1 Wave 5 Solvers
 
-| Change | Impact |
-|--------|--------|
-| **Opset 10 → 17, IR 10 → 8** | Enables Slice(step=-1) for zero-cost transforms |
-| **s_flip: Slice(step=-1)** | 0 MACs (was ~165K with Gather) |
-| **s_rotate k=2: double Slice reverse** | 0 MACs (was ~165K) |
-| **s_rotate k=1,3 (square): Slice+Transpose** | 0 MACs (was ~165K) |
-| **All Pad nodes: tensor-based pads input** | Required for opset 17 compatibility |
-| **All ReduceSum nodes: axes as tensor input** | Required for opset 13+ compatibility |
-| **lstsq crash fix: try/except LinAlgError** | Prevents SVD non-convergence crash (task 313) |
-| **Refactored to 16-file package** | Maintainable, testable, no more monolith |
+| Solver | Task | ARC-GEN | Cost | Technique |
+|--------|------|---------|------|-----------|
+| `row_uniform_fill` | 52 | 262/262 | 37,901 | Check col equality via int-encode + Slice + Equal |
+| `majority_replace_5` | 229 | 262/262 | 38,850 | ArgMax channel counts → majority mask → Where |
+| `find_color_with_hole` | 291 | 261/261 | 96,172 | Per-channel bbox_area - pixel_count → ArgMax deficit |
+| `fill_bbox_per_color` | 132 | 262/262 | 310,001 | CumSum forward + backward > 0 = range fill per color |
+| `border_detect` | 98 | 262/262 | 270,849 | Pad+Slice shifted copies → Min (erosion) → Sub from non_bg |
+| `symmetry_check_lr` | 103 | 215/215 | 37,785 | Gather(reverse cols) → Equal → ReduceMin → index color_select |
 
 ## Scoring
 
@@ -112,8 +120,8 @@ python -m neurogolf_solver.main --data_dir ARC-AGI/data/training/ --arcgen_dir A
 Score per task = max(1.0, 25.0 - ln(MACs + memory_bytes + params))
 ```
 
-- Analytical solvers (Slice/Transpose/Gather) → near-zero cost → ~20-25 pts
-- Conv solvers → cost proportional to kernel size → ~7-14 pts
+- Analytical solvers (Slice/Transpose/Gather) → near-zero cost → ~14-16 pts
+- Conv solvers → cost proportional to kernel size → ~8-11 pts
 - Unsolved → 1.0 pt minimum
 
 ## Competition Rules
@@ -127,6 +135,14 @@ Score per task = max(1.0, 25.0 - ln(MACs + memory_bytes + params))
 | Excluded tasks | {21, 55, 80, 184, 202, 366} |
 | Validation | Models checked against train + test + arc-gen (ALL splits) |
 
+## v5.0 → v5.1 Changes
+
+| Change | Impact |
+|--------|--------|
+| **+wave5.py** (6 solvers) | +7 tasks solved (52, 98, 103, 132, 229, 291 + conv_fixed now catches 305) |
+| **No existing code modified** | Only import + 6 entries appended to ANALYTICAL_SOLVERS |
+| **LB 1071 → 1164** | +93 points (+8.7%) |
+
 ## Key Docs
 
 - **SKILL.md** — Competition rules, architecture, methodology, checklist