| """Eval conv deep head from checkpoint.""" |
| import os, sys, time, torch |
| sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) |
| from train_conv_deep import DeepConvHead, cofiber_decompose, make_locations |
|
|
| DEVICE = "cuda" |
| COCO_ROOT = os.environ["ARENA_COCO_ROOT"] |
| VAL_CACHE = os.environ["ARENA_VAL_CACHE"] |
|
|
| head = DeepConvHead(hidden=256, n_blocks=10).to(DEVICE) |
| ckpt = torch.load(os.path.join(os.path.dirname(__file__), "heads", "cofiber_threshold", "conv_deep", "conv_deep_256h_10b_step12000.pth"), map_location=DEVICE, weights_only=False) |
| head.load_state_dict(ckpt["head"]) |
| head.eval() |
| print(f"Loaded step {ckpt['step']}, {sum(p.numel() for p in head.parameters()):,} params") |
|
|
| val = torch.load(VAL_CACHE, map_location="cpu", weights_only=False) |
| from pycocotools.coco import COCO |
| from pycocotools.cocoeval import COCOeval |
| coco_gt = COCO(os.path.join(COCO_ROOT, "annotations", "instances_val2017.json")) |
| cat_ids = sorted(coco_gt.getCatIds()) |
| idx_to_cat = {i: c for i, c in enumerate(cat_ids)} |
| strides = [16, 32, 64]; H = 640 // 16 |
| all_locs = torch.cat(make_locations([(H,H),(H//2,H//2),(H//4,H//4)], strides, torch.device(DEVICE))) |
|
|
| all_results = [] |
| t0 = time.time() |
| with torch.no_grad(): |
| for idx in range(len(val)): |
| item = val[idx] |
| spatial = item["spatial"].unsqueeze(0).float().to(DEVICE) |
| img_id = int(item["img_id"]); scale = item["scale"] |
| cls_l, reg_l, ctr_l = head(spatial) |
| cls_s = torch.cat([c.permute(0,2,3,1).reshape(-1, 80) for c in cls_l]).sigmoid() |
| reg_s = torch.cat([r.permute(0,2,3,1).reshape(-1, 4) for r in reg_l]) |
| ctr_s = torch.cat([c.permute(0,2,3,1).reshape(-1) for c in ctr_l]).sigmoid() |
| scores = cls_s * ctr_s.unsqueeze(1) |
| max_s, max_c = scores.max(1) |
| topk = min(100, max_s.shape[0]) |
| top_s, top_i = max_s.topk(topk) |
| tc=max_c[top_i]; tr=reg_s[top_i]; tl=all_locs[top_i] |
| x1=(tl[:,0]-tr[:,0])/scale; y1=(tl[:,1]-tr[:,1])/scale |
| x2=(tl[:,0]+tr[:,2])/scale; y2=(tl[:,1]+tr[:,3])/scale |
| w=(x2-x1).clamp(min=0); h=(y2-y1).clamp(min=0) |
| for i in range(topk): |
| s = top_s[i].item() |
| if s < 0.01: continue |
| all_results.append({"image_id": img_id, "category_id": idx_to_cat[tc[i].item()], |
| "bbox": [x1[i].item(), y1[i].item(), w[i].item(), h[i].item()], |
| "score": s}) |
| if (idx+1) % 1000 == 0: |
| print(f" {idx+1}/{len(val)} ({time.time()-t0:.0f}s)", flush=True) |
|
|
| print(f"\n{len(all_results)} detections") |
| coco_dt = coco_gt.loadRes(all_results) |
| ev = COCOeval(coco_gt, coco_dt, "bbox") |
| ev.params.imgIds = sorted(coco_gt.getImgIds())[:len(val)] |
| ev.evaluate(); ev.accumulate(); ev.summarize() |
| print(f"\nConv Deep 256h 10b (step 12000): mAP={ev.stats[0]:.4f} mAP50={ev.stats[1]:.4f} mAP75={ev.stats[2]:.4f}") |
|
|
