Function evaluate_performance

engine.py:85–179

Args: use_gt (bool, optional): Whether to use GT as detection results to verify the upper bound of person search performance. Defaults to False. use_cache (bool, optional): Whether to use the cached features. Defaults to False. use_cbgm (b

(
    model, gallery_loader, query_loader, device, use_gt=False, use_cache=False, use_cbgm=False, gallery_size=100)

@torch.no_grad()
def evaluate_performance(
    model, gallery_loader, query_loader, device, use_gt=False, use_cache=False, use_cbgm=False, gallery_size=100):
    """
    Args:
        use_gt (bool, optional): Whether to use GT as detection results to verify the upper
                                bound of person search performance. Defaults to False.
        use_cache (bool, optional): Whether to use the cached features. Defaults to False.
        use_cbgm (bool, optional): Whether to use Context Bipartite Graph Matching algorithm.
                                Defaults to False.
    """
    model.eval()
    if use_cache:
        eval_cache = torch.load("data/eval_cache/eval_cache.pth")
        gallery_dets = eval_cache["gallery_dets"]
        gallery_feats = eval_cache["gallery_feats"]
        query_dets = eval_cache["query_dets"]
        query_feats = eval_cache["query_feats"]
        query_box_feats = eval_cache["query_box_feats"]
    else:
        gallery_dets, gallery_feats = [], []
        for images, targets in tqdm(gallery_loader, ncols=0):
            images, targets = to_device(images, targets, device)
            if not use_gt:
                outputs = model(images)
            else:
                boxes = targets[0]["boxes"]
                n_boxes = boxes.size(0)
                embeddings = model(images, targets)
                outputs = [
                    {
                        "boxes": boxes,
                        "embeddings": torch.cat(embeddings),
                        "labels": torch.ones(n_boxes).to(device),
                        "scores": torch.ones(n_boxes).to(device),
                    }
                ]

            for output in outputs:
                box_w_scores = torch.cat([output["boxes"], output["scores"].unsqueeze(1)], dim=1)
                gallery_dets.append(box_w_scores.cpu().numpy())
                gallery_feats.append(output["embeddings"].cpu().numpy())

        # regarding query image as gallery to detect all people
        # i.e. query person + surrounding people (context information)
        query_dets, query_feats = [], []
        for images, targets in tqdm(query_loader, ncols=0):
            images, targets = to_device(images, targets, device)
            # targets will be modified in the model, so deepcopy it
            outputs = model(images, deepcopy(targets), query_img_as_gallery=True)

            # consistency check
            gt_box = targets[0]["boxes"].squeeze()

            assert (
                gt_box - outputs[0]["boxes"][0]
            ).sum() <= 0.001, "GT box must be the first one in the detected boxes of query image"

            for output in outputs: