#!/usr/bin/env python3 # -*- coding: utf-8 -*- """v5 : ablation texte (v4) vs texte+vision (v5). GBM + NN robuste + blend, 5-fold CV.""" import json, re, warnings import numpy as np, pandas as pd import torch, torch.nn as nn from sklearn.ensemble import HistGradientBoostingRegressor from sklearn.model_selection import KFold, cross_val_predict from sklearn.metrics import r2_score warnings.filterwarnings("ignore") torch.manual_seed(42); np.random.seed(42) CUR_YEAR=2026 dev="mps" if torch.backends.mps.is_available() else "cpu"; print("device:",dev) # ---- GPT specs ---- specs={} for l in open("data/processed/specs.jsonl",encoding="utf-8"): r=json.loads(l) if r.get("specs"): specs[r["uid"]]=r["specs"] # ---- vision ---- facade={} for l in open("data/processed/facade_specs.jsonl",encoding="utf-8"): r=json.loads(l) if r.get("facade"): facade[r["uid"]]=r["facade"] print("specs GPT:",len(specs),"| façades:",len(facade)) COND={"a_renover":0,"original":1,"bon":2,"renove":3,"refait_neuf":4,"neuf":5,"inconnu":2} BOOL=["turnkey","recently_renovated","garage","pool","waterfront","finished_basement","fireplace","central_ac","new_construction","income_property","renovated_kitchen","renovated_bathroom","large_lot","quiet_area"] def gfeat(uid): s=specs.get(uid);o={"g_"+b:0 for b in BOOL};o.update({"g_cond":2,"g_hardwood":0,"g_quartz_granite":0,"g_view_water":0}) if not s: return o for b in BOOL: o["g_"+b]=int(bool(s.get(b))) o["g_cond"]=COND.get(s.get("condition","inconnu"),2) fl=s.get("flooring") or [];o["g_hardwood"]=int("bois_franc" in fl) o["g_quartz_granite"]=int(s.get("countertop") in ("quartz","granit","marbre"));o["g_view_water"]=int(s.get("view")=="eau");return o RENO={"a_renover":0,"date":1,"standard":2,"renove":3,"haut_de_gamme":4,"inconnu":2} LAND={"aucun":0,"minimal":1,"entretenu":2,"amenage_soigne":3} SETS=["urbain_dense","banlieue","rural","bord_eau","boise"] MATS=["brique","pierre","vinyle","bois","aluminium","stucco","crepi","beton"] def vfeat(uid): f=facade.get(uid) o={"v_cond":5,"v_curb":5,"v_photoq":5,"v_reno":2,"v_land":1,"v_luxury":0,"v_garage_vis":0,"v_is_ext":0} for s in SETS: o["v_set_"+s]=0 for m in MATS: o["v_mat_"+m]=0 if not f: return o o["v_cond"]=f.get("condition_score") or 5;o["v_curb"]=f.get("curb_appeal") or 5;o["v_photoq"]=f.get("photo_quality") or 5 o["v_reno"]=RENO.get(f.get("renovation_level","inconnu"),2);o["v_land"]=LAND.get(f.get("landscaping","minimal"),1) o["v_luxury"]=int(bool(f.get("curb_luxury")));o["v_garage_vis"]=int(bool(f.get("has_garage_visible")));o["v_is_ext"]=int(bool(f.get("is_exterior"))) st=f.get("setting") if st in SETS: o["v_set_"+st]=1 for m in (f.get("exterior_materials") or []): if m in MATS: o["v_mat_"+m]=1 return o # ---- parsing structurel ---- def pa(s): if not s or not isinstance(s,str): return np.nan t=s.lower();m=("m2" in t) or ("m²" in t);n=re.sub(r"[^0-9.,\s]","",t).strip().replace(" ","").replace(",","") try: v=float(n) except: return np.nan return (v*10.7639 if m else v) if v>0 else np.nan def cc(c): if not c or not isinstance(c,str): return None c=c.split("(")[0].strip();return re.sub(r"\s+"," ",c).lower().strip() or None def ti(x): try: return int(x) except: return np.nan AMEN={"pool":["swimming pool","piscine","pool"],"waterfront":["waterfront","bord de l'eau","navigable","lakefront","riverfront","au bord de"],"garage":["garage"],"fireplace":["fireplace","foyer","wood stove","poêle"],"central_air":["central air","air conditionné","climatisation","a/c","heat-pump","thermopompe","air exchanger"],"finished_bsmt":["totally finished","finished basement","sous-sol fini","basement"],"near_water":["near water","waterfront","navigable"],"cul_de_sac":["cul-de-sac","cul de sac"],"golf":["golf"],"ski":["ski"],"near_park":["park","parc"],"near_school":["school","école","cegep","university","université","elementary","high school"],"near_highway":["highway","autoroute"],"transit":["public transit","public transportation","transport"],"no_rear_neighbor":["no backyard neighbors","sans voisin arrière"]} def sd(s): s=s.lower() if "ft" in s: f=re.search(r"([\d.]+)\s*ft",s);i=re.search(r"([\d.]+)\s*in",s) return (float(f.group(1)) if f else 0)+(float(i.group(1))/12 if i else 0) m=re.search(r"([\d.,]+)\s*m",s) if m: try: return float(m.group(1).replace(",","."))*3.28084 except: return None return None def ra(d): if not d or "x" not in d: return None a,_,b=d.partition("x");la,lb=sd(a),sd(b);return la*lb if (la and lb and 10 else np.nan;o["n_levels_r"]=len(lv);return o data=json.load(open("data/processed/maisons_avec_prix.json",encoding="utf-8"));rows=[] for d in data: p=d.get("property") or {};ad=d.get("address") or {};ft=d.get("features") or [] blob=(" ".join(ft)+" "+(d.get("description") or "")+" "+(d.get("title") or "")).lower() yb=ti(p.get("yearBuilt"));age=(CUR_YEAR-yb) if (not pd.isna(yb) and 1700<=yb<=CUR_YEAR) else np.nan pt=p.get("buildingType") or p.get("type");st=p.get("storeys") or p.get("levels") try: st=float(str(st).replace(",",".")) except: st=np.nan uid=f"{d.get('source')}:{d.get('listingId')}" rec={"uid":uid,"price":d.get("priceNumber"),"source":d.get("source"),"sqft":pa(p.get("squareFeet")),"lot_sqft":pa(p.get("lotSize")),"beds":ti(p.get("bedrooms")),"baths":ti(p.get("bathrooms")),"parking":ti(p.get("parkingSpaces")),"storeys":st,"age":age,"ptype":(pt or "Unknown").strip(),"city":cc(ad.get("city"))} for n,kw in AMEN.items(): rec[n]=int(any(k in blob for k in kw)) rec["n_images"]=len(d.get("images") or []);rec["virtual_tour"]=int(bool(d.get("virtualTourUrl"))) rec.update(pr(d.get("rooms")));rec.update(gfeat(uid));rec.update(vfeat(uid));rows.append(rec) df=pd.DataFrame(rows) df=df[df["price"].notna()];df=df[(df["price"]>=50_000)&(df["price"]<=15_000_000)] df.loc[(df["sqft"]<300)|(df["sqft"]>20_000),"sqft"]=np.nan df.loc[(df["lot_sqft"]<200)|(df["lot_sqft"]>5_000_000),"lot_sqft"]=np.nan df.loc[df["beds"]>15,"beds"]=np.nan;df.loc[df["baths"]>12,"baths"]=np.nan df["ppsf"]=df["price"]/df["sqft"];lo,hi=df["ppsf"].quantile([0.005,0.995]) df=df[~(df["ppsf"].notna()&((df["ppsf"]hi)))] ov=df[df["sqft"].notna()&df["room_area"].notna()&(df["room_area"]>200)];ratio=(ov["sqft"]/ov["room_area"]).median() re_=df["room_area"]*ratio;ok=df["room_area"].notna()&(re_.between(300,20000)) df["sqft_final"]=df["sqft"];fill=df["sqft_final"].isna()&ok;df.loc[fill,"sqft_final"]=re_[fill] df["has_sqft"]=df["sqft_final"].notna().astype(int);df["age2"]=df["age"]**2;df["n_images_c"]=df["n_images"].clip(upper=60) vc=df["city"].value_counts();keep=set(vc[vc>=20].index);df["city_fe"]=df["city"].where(df["city"].isin(keep),"__other__").fillna("__other__") tc=df["ptype"].value_counts();df["ptype_fe"]=df["ptype"].where(df["ptype"].isin(set(tc[tc>=30].index)),"Other") df=df.reset_index(drop=True);print("N=",len(df)) y=np.log(df["price"].values).astype(np.float32) amen=list(AMEN.keys());rv=["n_rooms","n_levels_r","has_primary_bed","has_basement","r_office","r_family","r_workshop","r_winecellar","r_sunroom","r_mezzanine","r_library","r_laundry"] gv=["g_cond","g_hardwood","g_quartz_granite","g_view_water"]+["g_"+b for b in BOOL] vv=["v_cond","v_curb","v_photoq","v_reno","v_land","v_luxury","v_garage_vis","v_is_ext"]+["v_set_"+s for s in SETS]+["v_mat_"+m for m in MATS] base=["sqft_final","has_sqft","lot_sqft","beds","baths","parking","storeys","age","age2","n_images_c","virtual_tour"]+amen+rv NUM_v4=base+gv NUM_v5=base+gv+vv missnum=["sqft_final","lot_sqft","beds","baths","storeys","age","age2"] for c in missnum: df["miss_"+c]=df[c].isna().astype(np.float32) MISS=["miss_"+c for c in missnum] cats=["city_fe","ptype_fe","source"];card={} for c in cats: u=sorted(df[c].astype(str).unique());m={v:i for i,v in enumerate(u)};card[c]=len(u);df[c+"_ix"]=df[c].astype(str).map(m).astype(np.int64) Xcat_ix=df[[c+"_ix" for c in cats]].values.astype(np.int64) kf=KFold(n_splits=5,shuffle=True,random_state=42) def gbm_cv(numcols): X=pd.concat([df[numcols].reset_index(drop=True),pd.get_dummies(df[cats],dtype=float).reset_index(drop=True)],axis=1) m=HistGradientBoostingRegressor(max_iter=600,learning_rate=0.05,max_leaf_nodes=63,l2_regularization=1.0,random_state=42) return cross_val_predict(m,X,y,cv=kf,n_jobs=-1) # ---- NN robuste ---- D_MAIN=256;D_HID=512;NB=3;SEEDS=[42,1,7] class RB(nn.Module): def __init__(s,d,dh,p): super().__init__();s.bn=nn.BatchNorm1d(d);s.l1=nn.Linear(d,dh);s.act=nn.GELU();s.d1=nn.Dropout(p);s.l2=nn.Linear(dh,d);s.d2=nn.Dropout(p) def forward(s,x): z=s.d1(s.act(s.l1(s.bn(x))));return x+s.d2(s.l2(z)) class Net(nn.Module): def __init__(s,nc,nn_): super().__init__();s.embs=nn.ModuleList();tot=0 for c in nc: d=max(2,min(64,round(1.6*c**0.56)));s.embs.append(nn.Embedding(c,d));tot+=d s.ed=nn.Dropout(0.1);s.bn0=nn.BatchNorm1d(nn_);s.proj=nn.Linear(tot+nn_,D_MAIN) s.bl=nn.ModuleList([RB(D_MAIN,D_HID,0.2) for _ in range(NB)]) s.head=nn.Sequential(nn.BatchNorm1d(D_MAIN),nn.GELU(),nn.Linear(D_MAIN,1)) def forward(s,xc,xn): e=[emb(xc[:,i]) for i,emb in enumerate(s.embs)] x=s.proj(torch.cat([s.ed(torch.cat(e,1)),s.bn0(xn)],1)) for b in s.bl: x=b(x) return s.head(x).squeeze(1) def nn_fold(Xnum,tr,te,seed): torch.manual_seed(seed);np.random.seed(seed) med=np.nanmedian(Xnum[tr],0);med=np.where(np.isnan(med),0,med);Xn=Xnum.copy() for j in range(Xn.shape[1]): col=Xn[:,j];col[np.isnan(col)]=med[j];Xn[:,j]=col mu=Xn[tr].mean(0);sg=Xn[tr].std(0)+1e-6;Xn=(Xn-mu)/sg ym=y[tr].mean();ys=y[tr].std()+1e-6;yz=(y-ym)/ys perm=np.random.permutation(len(tr));cut=int(len(tr)*0.9);tri=tr[perm[:cut]];vai=tr[perm[cut:]] T=lambda a,dt: torch.tensor(a,device=dev,dtype=dt) model=Net([card[c] for c in cats],Xn.shape[1]).to(dev) opt=torch.optim.AdamW(model.parameters(),lr=2e-3,weight_decay=1e-4) sch=torch.optim.lr_scheduler.CosineAnnealingLR(opt,T_max=200);lf=nn.SmoothL1Loss(beta=1.0) Xc=T(Xcat_ix,torch.long);Xnt=T(Xn,torch.float32);yt=T(yz,torch.float32) best=1e9;bs=None;pat=0 for ep in range(200): model.train();idx=np.random.permutation(tri) for i in range(0,len(idx),512): b=idx[i:i+512];opt.zero_grad();l=lf(model(Xc[b],Xnt[b]),yt[b]);l.backward() torch.nn.utils.clip_grad_norm_(model.parameters(),5.0);opt.step() sch.step();model.eval() with torch.no_grad(): vl=((model(Xc[vai],Xnt[vai])-yt[vai])**2).mean().item() if vl=20: break model.load_state_dict(bs);model.eval() with torch.no_grad(): return model(Xc[te],Xnt[te]).cpu().numpy()*ys+ym def nn_cv(numcols): Xnum=df[numcols+MISS].astype(np.float32).values;pred=np.zeros_like(y) for tr,te in kf.split(Xnum): ps=np.zeros(len(te)) for s in SEEDS: ps+=nn_fold(Xnum,tr,te,s) pred[te]=ps/len(SEEDS) return pred def stat(p): e=np.abs(np.exp(p-y)-1);return r2_score(y,p),np.median(e)*100,np.mean(e<=.15)*100,np.mean(e<=.25)*100 print("\n>>> GBM ablation") g4=gbm_cv(NUM_v4);g5=gbm_cv(NUM_v5) print("\n>>> NN v5 (peut prendre ~2 min)") n5=nn_cv(NUM_v5) b5=0.5*g5+0.5*n5 print("\n================ RÉSULTATS v5 (5-fold CV OOS) ================") print(f"{'Modèle':<34}{'R²OOS':>8}{'MAEméd':>9}{'±15%':>7}{'±25%':>7}") for nm,p in [("GBM v4 (texte)",g4),("GBM v5 (texte+vision)",g5),("NN v5 (texte+vision)",n5),("BLEND v5 (NN+GBM)",b5)]: r2,m,a,b=stat(p);print(f"{nm:<34}{r2:>8.4f}{m:>8.1f}%{a:>6.0f}%{b:>6.0f}%") print("\nRappel v4 blend: 0.760 | GBM 0.743 | NN 0.741") np.save("data/processed/v5_blend_oos.npy",b5)