Draft
Conversation
Adds walk-history to proto_extend_point and uses the halfway walk point as break_pt when dot(dir1, dir1_prev) < -0.5, preventing spurious zig-zag stubs from near-front kinks. Raises dl_vtx_cut default to 2.5 cm (was 2.0 cm in TaggerCheckNeutrino default_configuration) and removes the redundant dl_vtx_cut parameter from clus.jsonnet so C++ is the single source of truth. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…e_shower_1
In examine_shower_1, when building a candidate shower starting from a weakly-
directed muon track (sg) at main_vertex, complete_structure_with_start_segment
flood-fills the graph and pulls in any same-cluster neighbor — including a
Michel electron shower (Shower_C) attached at sg's far-end junction. The
inner associated-showers loop then finds Shower_C via the angle/distance test
(angle ≈ 0° and distance = 0 because they share a vertex), satisfies the
energy/length commit condition, and flips sg's pdg from 13 → 11, converting
the stopping muon into an electron.
Fix: in the low-energy else-branch of the associated-showers loop, skip any
conn_type=1 shower whose start_vertex is already inside shower1's vertex set.
Such a shower is a downstream decay product (e.g. a Michel electron starting
at the muon's stopping vertex), not an external EM shower that should cause the
start segment to be reclassified. High-energy (>80 MeV) downstream showers
are handled by the existing if-branch above and are unaffected.
Also carries forward two earlier fixes committed to examine_shower_1:
- set_mass(electron) when flipping start_segment pdg to 11 (mass consistency)
- dedup loop to remove stale single-segment muon showers that share a
start_segment with the newly created shower1 (prevents duplicate bee ids)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
In fill_bee_pf_tree, indirect-connection showers (conn_type 2/3) get a pseudo mother node to bridge the gap to the main vertex. The prototype (fill_psuedo_reco_tree, fill_ssmsp_psuedo overloads 1 & 3) labels this pseudo mother as gamma (22) for EM showers and neutron (2112) for all others — the physical reading being that an isolated proton activity visible only weakly must have been produced by an unseen neutral (neutron). The toolkit's SSM tagger (fill_ssmsp_pseudo_1/_3) already applied this rule correctly. The Bee PF tree helper (append_gamma_shower) did not: it always wrote "gamma" regardless of the underlying shower's particle type, silently mislabeling isolated proton activities. The pf_pdg_to_name helper already had case 2112 → "n" but it was never reached. Fix: rename append_gamma_shower → append_pseudo_shower and add the PDG test (e/γ → 22, otherwise → 2112) matching the SSM and prototype logic. Pi0 children are unaffected (they are always EM, condition evaluates to 22). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…n break_segments segment_search_kink's flag_switch direction-swap trigger had no minimum-window guard: when a kink sat ~2-3 cm from a track endpoint, only 3-4 post-kink fit points existed and were trivially collinear (path≈chord to within 0.03), causing flag_switch to fire on degeneracy rather than real geometry. This dispatched proto_extend_point backward into the track body, producing spurious ~2-3 cm tail segments at track endpoints. Fix 1 (PRSegmentFunctions.cxx, segment_search_kink): require the full 9-point post-kink window AND an absolute chord > 3 cm before trusting the straightness ratio. Both branches (flag_switch and flag_search else-if) are guarded symmetrically. Fix 2 (NeutrinoPatternBase.cxx, break_segments geometry check): for degree-1 end vertices, extend the replace_segment_and_vertex condition from (min_dis < 1.5 cm && angle > 120°) to (min_dis < 2.0 cm && kink_angle_at_break > 30°). kink_angle_at_break measures the approach-vs- departure angle in the steiner path at break_wcp, which is a more physically meaningful discriminant than the overall-segment angle used previously. This absorbs the residual ~1.7 cm tail that survives after Fix 1 when flag_continue is false and the walker snaps to the kink itself. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…nalysis Replace the previous 13-component (5 live, 8 commented-out) composite score with a clean 7-component formula tuned on 36 annotated events: score = s_dl + s_snap + s_fwd_z + s_clen + s_isol + s_main + s_fv Key changes vs. the previous scheme: - Re-enable s_clen (cluster length bonus, +2.0 max at 60 cm) and boost s_main (+2.0) and s_isol (-2.0) so geometric signals can compete in the uncertain-DL regime (where s_dl ≈ 5 for all candidates). - Cap s_fwd_z at ±0.25 (was -(z-z_min)/200cm, reaching -3 on wide events) so it can no longer swamp the main-cluster bonus. - Replace the hard 2*dl_vtx_cut snap-distance gate with a soft s_snap penalty (-min(2, snap/5cm)), fixing 3 false rejections on long main-cluster candidates at 5-8 cm snap distance. - Delete all dead commented-out scoring code (segs, ltrk, mult, flg_in, conf, ptbk) per "delete unused code" policy. - Add DL score mean/std/regime log line for future diagnosis. - Bump dl_vtx_min_accept_score default 0.0 → 4.0: empirically separates correct uncertain-regime picks (8-12) from true-failure events (3-5). Net effect on the 36-event log: +6 correct outcomes (3 wrong picks flipped, 3 false rejections accepted), 0 regressions. Also includes Phase 1 changes (top-K DL payload from Python, dl_vtx_rerank toggle, dl_vtx_score_scale knob). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The degree-1 terminus branch of use_replace in break_segments (introduced in 7725a9c to absorb ~1.7 cm fold-back tail artifacts) was also absorbing genuine Michel electrons. Diagnostics on two events revealed the key discriminant: Event 6852 (fold-back tail): tv1-vs-tv2 angle = 27.8°, use_replace=1 ✓ Event 6528 (12.94 MeV Michel): tv1-vs-tv2 angle = 79.2°, use_replace=1 ✗ tv1 = end_v - start_v (whole segment direction) tv2 = end_v - break_wcp (local stub direction) A fold-back artifact at a track endpoint has the stub pointing in roughly the same direction as the main track (small angle), because the fitter has simply extended a few points past the true end. A real physical secondary (Michel electron, delta ray) diverges from the parent axis at a significant angle. kink_angle_at_break cannot discriminate (both cases ≈50–90°). Add `angle < 45°` to the terminus condition: - fold-back tail (27.8°): still absorbed ✓ - Michel electron (79.2°): falls through to break_segment_into_two ✓ Also removes the temporary std::cout diagnostic prints and #include<iostream> added to PRSegmentFunctions.cxx during investigation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…a output Previously m_fitted_charge_2d was cleared at the start of every fill_fitted_charge_2d() call, so T_proj_data only captured the last cluster-filtered do_multi_tracking() pass — all other beam-flash clusters were silently dropped. Add a per-cluster snapshot cache m_cluster_fitted_charge_2d that stores the result of each fill_fitted_charge_2d() call keyed by m_cluster_filter. Overwriting the same key on each re-fit gives "latest fit wins per cluster", handling the case where a cluster is re-fit during examine_structure / merge_vertices. A new assemble_fitted_charge_2d() method merges all snapshots into m_fitted_charge_2d at the end; TaggerCheckNeutrino::visit calls it just before grouping.set_track_fitting() so write_proj_data sees all clusters' cells with correct cluster_id and pred_charge. Also consolidate kPlaneChOffset usage in UbooneMagnifyTrackingVisitor and apply it consistently to pu/pv/pw fields in T_rec_charge. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
New IEnsembleVisitor that opens the existing tracking ROOT file in UPDATE mode and adds T_tagger (all TaggerInfo fields + nu_x/y/z) and T_kine (all KineInfo fields), matching prototype output tree names. Also adds tagger_output() factory to clus.jsonnet and a tagger validation plan document. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
New comparison binary that reads T_tagger and T_kine from a prototype ROOT file and a toolkit ROOT file, compares all scalar (Float_t/Int_t) and vector (vector<float>/vector<int>) branches event by event, and groups results by originating tagger function (cosmic, gap, mip, ssm, shw_sp, pio_family, stem, lem, br, tro, hol_lol, vis, numu, nue, kine, etc.). Writes per-category histograms and a T_summary tree to report.root; prints a compact per-category table and exits non-zero if any branch differs. Replaces the earlier separate build-verify and single-event manual inspection steps. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Section 5.4 now documents the actual C++ comparison app (wire-cell-uboone-tagger-compare) replacing the earlier Python script placeholder. Implementation sequence table updated with DONE/TODO status for each step. Key file reference table extended with the new visitor header/impl and comparison app. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ration Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Reclassify all SPDLOG_LOGGER_DEBUG/->debug() calls in clus sub-loggers (clus.Cluster, clus.TrackFitting, clus.NeutrinoPattern, and other clus component loggers) to TRACE level. MultiAlgBlobClustering high-level flow logs remain at DEBUG. With -L clus:debug (default), only high-level flow is shown. To re-enable detail per subsystem, add e.g. -L clus.NeutrinoPattern:trace or -L clus:trace. Requires building with --with-spdlog-active-level=trace (new wscript default). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Keep DEBUG only for configure-time roll-call, event milestones (loading tensor set, Produce pctrees, performance summary, EOS), and the rare diagnostic skip. Demote to TRACE: the dump() ensemble printouts (fired before/after every visitor, ~60 lines/event), per-segment fitted-point counts, and bee-points fill summaries. With -L clus:debug the MABC log footprint drops from ~72 to ~22 lines/event. Full detail returns with -L clus:trace. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Pass "!V:Silent" to all TMVA::Reader constructors in UbooneNueBDTScorer and UbooneNumuBDTScorer to suppress the verbose booking/loading messages that fire at startup for every BDT reader. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- TaggerCheckNeutrino: add per-step DEBUG-level wall-clock timing in visit() gated by m_perf, covering preload, PR, vertex determination, taggers, kine fill, and a TOTAL line (~12 lines/event when enabled) - NeutrinoPatternBase: fix init_first_segment always passing &cluster as main_cluster causing false flag/pointer mismatch WARN for non-main clusters; pass nullptr for non-main clusters instead; demote the nullptr fallback message from WARN to TRACE - BlobSampler: demote wire-index out-of-range from WARN to TRACE since the boundary clamping is already correct; consolidate into one TRACE line Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Extends the existing `struct Perf` in MultiAlgBlobClustering with std::chrono so each perf checkpoint emits a DEBUG-level line showing the step name, its wall-clock duration, and cumulative time since the start of operator(). Gated by `m_perf` (existing config knob). Also minor whitespace cleanup in SteinerGrapher.cxx and ClusteringRecoveringBundle.cxx (trailing spaces, missing newline). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Rename three fields to match prototype output tree names: shw_sp_br3_7_shower_main_length → shw_sp_br3_7_main_length br3_7_shower_main_length → br3_7_main_length numu_cc_3_acc_track_length → numu_cc_3_track_length Also remove match_isFC from the output visitor (not written by prototype). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
numu_tagger3.weights.xml was trained with "numu_cc_3_acc_track_length" so the numu3 Reader::AddVariable must keep that label, even though the TaggerInfo field and xgboost reader now use the prototype name "numu_cc_3_track_length". Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When -v is given and a branch differs, print the prototype and toolkit values from the first differing event alongside the existing stats. Scalars show inline; vectors print on separate indented lines. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add -n<NeutrinoTaggerInfo.h> option: parses the header at runtime to
build a branch→default map (float/int fields read {init}, vectors → []).
Verbose output now prints a fixed-width table with aligned columns for
n_diff/n_cmp, max|diff|, proto, toolkit, and default. Vectors get a
two-row proto/toolkit dump below the stats line.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
detect_type now returns SCALAR_D for Double_t leaves. setup_branch_stat allows mixed Double_t/Float_t pairs (canonicalised as SCALAR_D) by reading each side into its native buffer and promoting to double in compare_one. Verbose output prints SCALAR_D values the same way as SCALAR_F. Fixes [SKIP unsupported type] for nu_x/y/z after prototype changed their type to Double_t. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
nu_x/y/z are written under alias names in the output visitor but live as kine_nu_x/y/z_corr in NeutrinoTaggerInfo.h. Add a small static alias table so the default-value lookup follows the indirection. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The header is installed at <prefix>/include/WireCellClus/NeutrinoTaggerInfo.h alongside the binary at <prefix>/bin/. Use readlink(/proc/self/exe) to find the prefix at runtime so defaults are shown without needing -n. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…se mode
Two additions to the -v output:
1. Fingerprint cluster report (before the per-branch table): groups scalar
branches that share identical (proto, toolkit) first-diff values and prints
any cluster of >= 3 branches. This surfaces the common pattern where one
upstream quantity (e.g. main-shower energy, main-shower length, main vertex)
drives dozens of apparent diffs. For the nue_5384_132_6604 event the
energy cluster alone has 45 members, collapsing the apparent 16-category
diff into ~5 root causes.
2. 'note' column in the per-branch table: sentinel_note() flags rows where one
side is a known sentinel and the other is not:
|v| >= 1e7 -> [SENT proto] or [SENT tool] (no-segment-found style)
int == -1 -> [UNINIT proto] or [UNINIT tool]
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Make this new one, as #444 seems not working for the tracking purpose