Task 319 ONNX Guide: full algorithm for 267/267, memory estimate ~100-200KB

medal-solvers/TASK319_ONNX_GUIDE.md

@@ -0,0 +1,64 @@
+# Task 319 — ONNX Build Guide
+
+## Rule (267/267 VERIFIED)
+
+**CRITICAL**: Object binary = `(inp == color)` in bbox, NOT `(inp != bg)`!
+
+### Algorithm (2-pass, template = largest then second-largest):
+
+```
+1. bg = channel with most pixels (argmax of per-channel sums)
+2. Sort non-bg channels by pixel count → [ch1_largest, ch2, ch3]
+3. For template_channel in [ch1_largest, ch2]:
+   a. Find largest uniform block (bs_r, bs_c) in template's binary
+   b. Downsample template by block → pattern (ph × pw)
+   c. For each other channel (candidate):
+      - Cross-correlate pattern with candidate binary at all positions
+      - If max correlation == sum(pattern) → MATCH FOUND
+      - Output = candidate channel's binary (one-hot encoded)
+   d. If no match with uniform block, try row/col grouping:
+      - Group consecutive identical rows, group consecutive identical cols
+      - Take first row/col of each group → downsampled pattern
+      - Cross-correlate with candidates
+4. Output = matched candidate in one-hot format
+```
+
+### Statistics:
+- Largest template + block method: 248/267
+- + row/col grouping: 250/267  
+- + second-largest template: 267/267 (100%!)
+
+### ONNX Implementation Notes:
+
+**Input**: [1, 10, 30, 30] — channels ARE the per-color masks
+
+**Key Operations Needed**:
+1. Per-channel pixel sum → ReduceSum [10] → find bg (ArgMax) and rank others
+2. For template channel: try block sizes (2,2), (2,3), (3,2), (3,3), (2,4), (4,2), (4,3), (3,4), (4,4), (5,5)
+3. Block uniformity check: reshape to [h/bs_r, bs_r, w/bs_c, bs_c], check min==max per block
+4. Cross-correlation: Conv2D with downsampled pattern as kernel (or MatMul approach)
+5. Match detection: max(correlation) == sum(pattern)
+6. Output construction: select matched channel
+
+**Memory Estimate**:
+- Correlation maps: ~[30,30] per block_size × ~10 sizes = 36KB
+- Template/candidate masks: [30,30] × 3 = 11KB
+- Auxiliaries: ~50KB
+- **Total: ~100-200KB → score 12-13 pts → gain +4-5**
+
+**Block Size Enumeration** (covers 248/267):
+- Valid (bs_r, bs_c) where bs divides object dims (max 10×10)
+- Try: (2,2), (2,1), (1,2), (3,3), (3,1), (1,3), (2,3), (3,2), (4,2), (2,4), (5,5), (4,4), (5,2), (2,5)
+
+**Row/Col Grouping in ONNX** (covers remaining 2 from 1st template):
+- Row boundary: ReduceSum(abs(mask[r] - mask[r-1])) > 0
+- Take rows AT boundaries → downsampled rows
+- Similarly for cols
+
+**Second Template Pass** (covers final 17):
+- Same algorithm but on second-largest channel
+- If first template found a match, skip this pass
+
+### Files
+- `task319_solver_267.py` — Complete Python solver (267/267)
+- `build_task255_onnx.py` — Reference for ONNX building patterns