| """Side-by-side webcam: same frame & detection, LEFT = INT8 (ONNX Runtime static), RIGHT = INT4 (simulated weight-only). |
| usage: python live_ab.py [tag=v2] q=quit s=snapshot n=toggle norm""" |
| import sys, time, collections, copy, numpy as np, cv2, torch, torch.nn.functional as F, onnxruntime as ort |
| from insightface.app import FaceAnalysis |
| from insightface.utils.face_align import norm_crop |
| from model import TinyFaceNet |
| TAG=sys.argv[1] if len(sys.argv)>1 else "v2"; THR=0.55; VOTE=7; NORM=True |
| ck=torch.load(f"dataset/model_{TAG}.pt",weights_only=False); names=list(map(str,ck["names"])); C=ck["centroids"].numpy() |
| |
| s8=ort.InferenceSession(f"dataset/{TAG}_int8.onnx",providers=["CPUExecutionProvider"]); in8=s8.get_inputs()[0].name |
| |
| m=TinyFaceNet(n_cls=len(names)); m.load_state_dict(ck["model"]); m.eval(); torch.set_num_threads(4) |
| for n_,p in m.named_parameters(): |
| if p.dim()<2 or n_.startswith(("distill","cls")): continue |
| w=p.data; wf=w.reshape(w.shape[0],-1); sc=wf.abs().amax(1,keepdim=True).clamp_min(1e-8)/7; p.data=((wf/sc).round().clamp(-7,7)*sc).reshape(w.shape) |
| def emb8(x): e=s8.run(None,{in8:x})[0][0]; return e/np.linalg.norm(e) |
| def emb4(x): |
| with torch.no_grad(): return F.normalize(m(torch.from_numpy(x)),dim=1)[0].numpy() |
| app=FaceAnalysis(name="buffalo_l",allowed_modules=["detection"],providers=["CPUExecutionProvider"]); app.prepare(ctx_id=-1,det_size=(320,320)) |
| clahe=cv2.createCLAHE(clipLimit=2.0,tileGridSize=(8,8)) |
| def normalize(img): |
| mean=img.mean() |
| if mean<110: |
| gamma=np.log(110/255)/np.log(max(mean,1)/255); lut=(np.linspace(0,1,256)**gamma*255).astype(np.uint8); img=cv2.LUT(img,lut) |
| lab=cv2.cvtColor(img,cv2.COLOR_BGR2LAB); lab[...,0]=clahe.apply(lab[...,0]); return cv2.cvtColor(lab,cv2.COLOR_LAB2BGR),mean |
| def iou(a,b): |
| ix=max(0,min(a[2],b[2])-max(a[0],b[0])); iy=max(0,min(a[3],b[3])-max(a[1],b[1])); i=ix*iy |
| return i/((a[2]-a[0])*(a[3]-a[1])+(b[2]-b[0])*(b[3]-b[1])-i+1e-6) |
| tracks={"INT8":[],"INT4":[]} |
| def annotate(img,bb,sim,side,fr): |
| j=int(sim.argmax()); s=float(sim[j]); inst=names[j] if s>=THR else "unknown"; T=tracks[side] |
| tr=max(T,key=lambda t:iou(t[0],bb),default=None) |
| if tr is None or iou(tr[0],bb)<0.3: tr=[bb,collections.deque(maxlen=VOTE),collections.deque(maxlen=VOTE),fr]; T.append(tr) |
| tr[0]=bb; tr[1].append(inst); tr[2].append(s); tr[3]=fr |
| lab,cnt=collections.Counter(tr[1]).most_common(1)[0]; conf=np.mean(tr[2]); col=(0,255,0) if lab!="unknown" else (0,0,255) |
| x1,y1,x2,y2=bb; cv2.rectangle(img,(x1,y1),(x2,y2),col,2) |
| cv2.putText(img,f"{lab} {conf:.2f} ({cnt}/{len(tr[1])})",(x1,y1-8),0,0.65,col,2) |
| cv2.putText(img," ".join(f"{names[k]}:{sim[k]:.2f}" for k in np.argsort(-sim)[:3]),(x1,y2+18),0,0.42,(255,255,0),1) |
| cap=cv2.VideoCapture(0,cv2.CAP_DSHOW); cap.set(3,640); cap.set(4,480) |
| if not cap.isOpened(): sys.exit("webcam open nahi hua") |
| n=0; t0=time.time(); fr=0; t8=t4=0. |
| while True: |
| ok,raw=cap.read() |
| if not ok: continue |
| img,mean=normalize(raw) if NORM else (raw,raw.mean()); L=img.copy(); R=img.copy() |
| for f in app.get(img): |
| crop=norm_crop(img,f.kps,112); x=((crop.transpose(2,0,1)[None].astype(np.float32)/255.)-0.5)/0.5; bb=tuple(map(int,f.bbox)) |
| a=time.perf_counter(); e8=emb8(x); b=time.perf_counter(); e4=emb4(x); c=time.perf_counter(); t8=0.9*t8+0.1*(b-a)*1000; t4=0.9*t4+0.1*(c-b)*1000 |
| annotate(L,bb,C@e8,"INT8",fr); annotate(R,bb,C@e4,"INT4",fr) |
| for k in tracks: tracks[k]=[t for t in tracks[k] if fr-t[3]<15] |
| cv2.putText(L,f"INT8 ORT {t8:.1f}ms/face",(8,20),0,0.6,(255,255,255),2); cv2.putText(R,f"INT4 sim {t4:.1f}ms/face",(8,20),0,0.6,(255,255,255),2) |
| out=np.concatenate([L,np.full((img.shape[0],4,3),255,np.uint8),R],1) |
| fr+=1; cv2.putText(out,f"{TAG} {fr/(time.time()-t0):.1f}fps norm={'ON' if NORM else 'off'} light={mean:.0f} q=quit s=save n=norm",(8,out.shape[0]-10),0,0.5,(255,255,255),1) |
| cv2.imshow("INT8 | INT4",out); k=cv2.waitKey(1)&0xFF |
| if k==ord("q"): break |
| if k==ord("n"): NORM=not NORM; tracks={"INT8":[],"INT4":[]} |
| if k==ord("s"): n+=1; cv2.imwrite(f"analysis/ab_{n}.jpg",out); print("saved",n,flush=True) |
| cap.release(); cv2.destroyAllWindows() |
|
|
