def evaluate_test_p2m()

in shapenet/evaluation/eval.py [0:0]

• 49 lines of code
• 11 McCabe index (conditional complexity)

def evaluate_test_p2m(model, data_loader):
    """
    This function evaluates the model on the dataset defined by data_loader.
    The metrics reported are described in Table 1 of our paper, following previous
    reported approaches (like Pixel2Mesh - p2m), where meshes are
    rescaled by a factor of 0.57. See the paper for more details.
    """
    assert comm.is_main_process()
    device = torch.device("cuda:0")
    # evaluation
    class_names = {
        "02828884": "bench",
        "03001627": "chair",
        "03636649": "lamp",
        "03691459": "speaker",
        "04090263": "firearm",
        "04379243": "table",
        "04530566": "watercraft",
        "02691156": "plane",
        "02933112": "cabinet",
        "02958343": "car",
        "03211117": "monitor",
        "04256520": "couch",
        "04401088": "cellphone",
    }

    num_instances = {i: 0 for i in class_names}
    chamfer = {i: 0 for i in class_names}
    normal = {i: 0 for i in class_names}
    f1_1e_4 = {i: 0 for i in class_names}
    f1_2e_4 = {i: 0 for i in class_names}

    num_batch_evaluated = 0
    for batch in data_loader:
        batch = data_loader.postprocess(batch, device)
        imgs, meshes_gt, _, _, _, id_strs = batch
        sids = [id_str.split("-")[0] for id_str in id_strs]
        for sid in sids:
            num_instances[sid] += 1

        with inference_context(model):
            voxel_scores, meshes_pred = model(imgs)
            # NOTE that for the F1 thresholds we take the square root of 1e-4 & 2e-4
            # as `compare_meshes` returns the euclidean distance (L2) of two pointclouds.
            # In Pixel2Mesh, the squared L2 (L2^2) is computed instead.
            # i.e. (L2^2 < τ) <=> (L2 < sqrt(τ))
            cur_metrics = compare_meshes(
                meshes_pred[-1], meshes_gt, scale=0.57, thresholds=[0.01, 0.014142], reduce=False
            )
            cur_metrics["verts_per_mesh"] = meshes_pred[-1].num_verts_per_mesh().cpu()
            cur_metrics["faces_per_mesh"] = meshes_pred[-1].num_faces_per_mesh().cpu()

            for i, sid in enumerate(sids):
                chamfer[sid] += cur_metrics["Chamfer-L2"][i].item()
                normal[sid] += cur_metrics["AbsNormalConsistency"][i].item()
                f1_1e_4[sid] += cur_metrics["F1@%f" % 0.01][i].item()
                f1_2e_4[sid] += cur_metrics["F1@%f" % 0.014142][i].item()

            num_batch_evaluated += 1
            logger.info("Evaluated %d / %d batches" % (num_batch_evaluated, len(data_loader)))

    vis_utils.print_instances_class_histogram_p2m(
        num_instances,
        class_names,
        {"chamfer": chamfer, "normal": normal, "f1_1e_4": f1_1e_4, "f1_2e_4": f1_2e_4},
    )