Uj2HdZddlmZddlZddlZddlZddlZddlmZm Z ddl m Z e dvrCejjdee ej#j$dddlmZmZmZmZmZdd lmZdd lmZmZd Zd Zd Z dZ!e"dk(re!yy)zURebuild averages, per-design scores, details and explanations from Summary JSON only.) annotationsN)Counter defaultdict)Path)N)ROOTdumpread stage_dirmodel_size_key)METRICS)weak_model_tables weak_resultc tDcic]I}||Dcgc]}|dj||d|!c}x}rt|t|z ndKc}}Scc}wcc}}w)Nmetrics)rgetsumlen)recordskeyrvaluess @/workspace/moltex-experiments/poc_experiments_v2/Summary/main.pyaveragesrs  U\:CU\PQ`abk`l`p`pqt`uaB!I,s:KU\:C0Cv0CVs6{*JN N  :C s A!A A A!A!ct|}t|dz }t|dz jdDcgc] }t|}}t t t t }}|D]>}||d|dfj |||d|d|dfj |@g}t|jD]t\\}} } | D cgc] } | ddk(s | } } |j || t| t| t| ttd | Dd | d d d vt|jD cgc]_\\}} } } || | t| D cgc] } | ddk(s | c} | D cgc]"} dDcic]}|| j|c}$c}} da}} } } } }}ddddd|jdijddddddddttd|D|D cgc])} | dd vr d!Dcic]}|| j|c}+c}} d" }|d#jd$ijd%r8d&|d'<d(|d)<d*|d+d,<d-|d.<|d#d$jd/r d0|d'<d1|d.<t t }|D]E}t|s|jd2r!||d|dfj t|Gg}t|jD]\\}} } || t| d3d4}|d#jd$ijd5rt|d#d$||d6<d7D]D}| D cgc]} | j|| |}} |rt|t|z nd||<F|j |t!|d8z |t!|d9z |t!|d:z |t!|d;z ||d#|dd<|dd=t|td>|D|||gd?d@ }d dAlm}|||jdB|dC<||dD<|d#jd$ijdEg|dE<|dEdFdG|dH<|d#jd$ijdIr|d'xxdJz cc<|d.xxdKz cc<|d#jd$ijdLr |dMxxdNz cc<t!|d;z |t!|dOz |g}|D]j}|dP}|d|dt'|dQg}|dRDcgc]}||dSn||dTc}z }|j dUdVj)dW|DzdXzl|d#}t*j,j/t0dYz dZz |d[z |d\z d]z |}t*j,j/t0d^z d]z |d_|d[d`|d\z}dV}|rdat3|z}db|d#d[dc|d#d\dd|d<de|d=df|dgdh|dCdidj|dCdkdl|dCdmdnt5j6|dot5j6|dp|dqdVj)|drt5j6t'|ds} |d]z j9| dtu|Scc}wcc} wcc} wcc}wcc}} wcc}} } } } }wcc}wcc}} wcc} wcc}w)vNdata_provenance.jsonresultsz*.jsonmodel_idprotocol design_id score_statusscoredc3&K|] }|d ywstatusN.0rs r zrebuild..%s1Mu!!H+uzDunweighted mean across scored dataset records; null metrics excludedr evaluation)rr macro_average_scoresscored_datasetstotal_result_records status_counts aggregationr,) dataset_idfold_idr&r"rdetail)rr r!scoresdatasetsz{All records sharing a design-file SHA256 are held out together. Fit and preprocessing statistics use training records only.z^Fit each design's reviewed records and evaluate those same records. This is training fit only.zTFit all reviewed records and evaluate those same records. This is training fit only.zAUntrained actual-input forward passes. No accuracy is calculated.)leave_one_design_out per_design train_all diagnosticdata selectionzDCAD + Excel + PLY; CSV required for NG and optional for explicit OK.zhExternally reviewed common-CAD node risk; raw CSV pixel coordinates are not automatically mapped to CAD.zMAE/RMSE/Brier use reviewed probability targets; AP/AUC and confusion scores threshold reviewed targets. Null means unavailable/undefined.zTMacro means are per dataset, not per node. Unscored records never enter the average.znGP posterior standard deviations are available when fitted. Other models have no calibrated uncertainty claim.zdPhysics GNN deformation/type heads are not supervised by this node-risk contract and are not scored.zSix distinct designs are the current holdout groups. All complete records are NG; OK without CSV is excluded. No pixel-to-CAD calibration or validated deformation target was supplied.c3&K|] }|d ywr%r'r(s rr*zrebuild..8s >g8gr+r&) completeduntrained_diagnostic)rr r!r2r&reason) evaluation_protocolsdata_selectiontargetsr5 averaging uncertainty deformation limitationsstatusesfailuresrunconfigallow_weak_labelszAutomatic scan_projected_weak CAD-node targets from CSV pixels, measured height rasters and proper rigid scan-to-CAD fits. Unannotated covered nodes outside defect buffers are assumed negative.rCzIndependently reviewed accuracy is unavailable for weak supervision. Real optimizer epoch losses and checkpoints are saved; weak-label predictions never enter reviewed score averages.r5zExperimental fit on eligible automatic scan-projected weak labels. Predictions are training-fit diagnostics, not independently reviewed accuracy.rAr9a Raster calibration and registration are inferred and unreviewed. CSV source conflicts and poor-overlap cases are recorded. Datasets with failed projection gates are excluded from weak training. All source records are NG; unannotated regions are not reviewed healthy.rGuser_selected_fit_datasetszUser-selected scan/CAD pairs with recorded mirrored frames and CAD-supported scan regions. Node targets remain automatic weak CSV/raster projections; unannotated covered nodes outside guarded regions are assumed negative.zUser confirmation applies to pair inclusion. CAD-node ground truth remains unreviewed. Rejected CSV rows and their nearby assumed negatives are masked; source choices, reflection, raw residuals and partial scan coverage remain recorded.training_excludedF)rr r6independently_reviewedsize_named_outputs input_size)bceaverage_precisioniouavg_scores.jsonindividual_scores.jsondetailed_scores.jsonexplanations.json dataset_count design_countc3,K|] }|ddk(yw)r"r#Nr'r(s rr*zrebuild..Us)YQXA!N*;x*GQXs)rrUrVrWrXzresults/**/*.json) rJrYrZresult_record_countscored_record_countaverage_scoresindividual_scores explanationsself_contained_files)training_diagnosispreparedrbweak_model_checksexcluded_train_dataset_idszyExcluded from every model fit and model-check cohort; retained for inspection and separately marked held-out predictions.) dataset_idspolicytraining_exclusionsinclude_ok_weak_labelsze Explicit OK production records provide weak healthy CAD-node targets without requiring a defect CSV.zv Production OK status is a weak part-level healthy assumption; it is not independently reviewed CAD-node ground truth.evaluation_reference_preparedrDzg Size comparisons use frozen reference CAD-node positions, targets and masks, shared across every size. summary.jsonr-r.)maermserSrTdiceu—z.5fzrc3NK|]}dtj|dyw)zzN)htmlescape)r)values rr*zrebuild..ls&$YSX%tDKK,>+?u%ESXs#%zPlotsoutputsdaterun_id index.html Visualizerz?run=/z

Model checks against weak labels

Training fit and held-out-design agreement are separate; these are not independently reviewed accuracy.

uMoltex v2 summary

Moltex v2 — z / z

u datasets · u designs · r]z0 scored records

Training status: training_statusu · trained_prediction_recordsu trained prediction records · folds_with_epoch_lossesz) folds with epoch losses

PNG results, losses & training diagnosis · CAD / scan / label / model-input visualizer

— = no valid score. Diagnostic and training-fit outputs are separate from unseen-design evaluation.

z

Average scores

u 
ModelProtocolScored datasetsMAERMSEAPIoUDice

Individual and detailed scores

Individual scores · Detailed predictions/targets · Data and code provenance

Explanations

z
utf-8encoding)rr sortedrglobrlistappenditemsrrdictrrrr rr Plots.mainrbstrjoinospathrelpathr rrprq write_text)! directory provenancerrgroupeddesignsrecordavgmodelr grouprr#didk individualr` weak_groups weak_checksentrymetricrsummaryrbrowsrcellsrrJ plots_link viewer_link weak_sectioncontents! rrebuildrs YIi"889J&,i).C-J-J8-T&UV&UdtDz&UGV"4(+d*;WG #VJ%789@@H #VJ%7 9LMNUUV\] C$*7==?$; 5"DUa&78&C!UD 8U]^dUe'*6{CPUJ%)'1Mu1M*M%N#i"'(<"8 : ;%<8>gmmo7NPP8O3,uhe %(%%&Y%Q1^;LPX;Xq%&YZJO PJOEF6~ 6~AEE!H6~ JO PQ8OP !^vkY -[ %..488GMN}_kH}Q >g >>?%c%8$? 5"c%jlqr e  2 . 2 23G H"0E1B81Le"TE, 9F).LA!%%-2KaiFL7=CKF 34E&M: 5!%@ & &,- -z:+ +W5( (,7 '*V:L_:])&1.AZ]^eZf&))YQX)Y&Y!$:Wc(j kG .$6w z@Z$[G !#.G  ,6u,=,A,A(B,O,S,STprt,uG ():AB^:_N+OL&'%Xr*../GHY$K K]#(` `#%Xr*../NO[!&O O!( (,7^ #W- D/0 #VJ%7VDU=V9WX WJKWJs73</% S7IJWJK K FRWW$YSX$YYY\ccd  %.C)!;c&k!ICPXM!Y\h!hjstJ''//$"5 "DiPUZ[^_e[fZgghilmuivhwSxxKLMO`alOmn 5>&)*#genX.F-GH O]7>+B*CKs}[ [$[ [/8 [[ [/ [)[$$7[)<[/[>'[9[>6\ \ \ [/$[))[/9[>c .td|}|||tt|ddz tt|ddz d}t|dz ||dD]?}t|}|dz |d z |d z |d z |d z |d dz }t||At |} t dz |dz |dz } | j dd|dz jd} dD]d} | jtjjt | z |dztjjt | z | dz} f| dz j| dt| dz | dD]} t| | z t|| z | |fS)NSummary manifest_dirzcode_provenance.jsonzdependencies.json)rJr;rccode dependenciesr metric_filesrrr r3r!r2z.jsonreportsrurvT)parentsexist_okrwr}r~)rsrxryz results.htmlrk)rVrWr) r r rr rr mkdir read_textreplacerrrr)rJmanifestrcindexrrsourcerrr report_dirrstagenames rcollectrs)S)IhHtC$78;QQR"&tC,?'@CV'V"WYJ + +Z8'fI%z(::VJ=OORXYbRcc%&+1,+?*@(FG T6 ( i G !CK/#h-?JTD1<'22G2DG(//"''//$, "JS"P"$''//$, "Kc"QS). ,,Ww,Gn $g.Z Z$ Y%5 67[ I ctjt}|jdtd|j }t |j}td|dd|jy)N) descriptionz --summary-dirT)typerequiredzRebuilt r\z! records from Summary JSON only: ) argparseArgumentParser__doc__ add_argumentr parse_argsr summary_dirprint)parserargsrs rmainrsj  $ $ 9F dTB    Dd&&'G HW2344UVZVfVfUg hir__main__)#r __future__rrrprsys collectionsrrpathlibr __package__rinsertr__file__resolvercommonr r r r r Postprocess.mainrSummary.model_checksrrrrrr__name__r'rrrs[" ,*HHOOAs4>113;;A>?@>>$? hV4j z46r