Transform module in Anomalib

[yxteo2]

I want to use center crop for fastflow, but after adding it the performance drop quite a lot.
I not sure is is the way I add transform got issue, can someone help me to check?

def _build_eval_transform(h: int, w: int) -> dict:
    s = 0.875
    pre_h = int(math.ceil(h / s))
    pre_w = int(math.ceil(w / s))
    return {
        "class_path": "torchvision.transforms.v2.Compose",
        "init_args": {"transforms": [
            {"class_path": "torchvision.transforms.v2.Resize",
             "init_args": {"size": [pre_h, pre_w], "antialias": True}},
            {"class_path": "torchvision.transforms.v2.CenterCrop",
             "init_args": {"size": [int(h), int(w)]}},
            {"class_path": "torchvision.transforms.v2.Normalize",
             "init_args": {"mean": [0.485, 0.456, 0.406],
                           "std":  [0.229, 0.224, 0.225]}}
        ]}
    }

def _patch_yaml(cfg_path: str, args) -> dict:
    yaml = YAML()
    with open(cfg_path, "r", encoding="utf-8") as fp:
        cfg = yaml.load(fp)

    trainer_cfg = cfg.setdefault("trainer", {})
    data_args = cfg["data"]["init_args"]
    model_args = cfg["model"].setdefault("init_args", {})

    # ----------------- Device & accelerator -----------------
    if args.device and args.device != "cpu" and torch.cuda.is_available():
        trainer_cfg["accelerator"] = "gpu"
        try:
            torch.cuda.set_device(int(str(args.device).split(",")[0]))
        except Exception:
            pass
        print(f"[INFO] Using GPU cuda:{args.device}")
    else:
        trainer_cfg["accelerator"] = "cpu"
        trainer_cfg["devices"] = 1
        args.device = None
        print("[INFO] Using CPU")
    
    trainer_cfg.setdefault("benchmark", True)          # cudnn autotune
    trainer_cfg.setdefault("precision", "16-mixed")    # AMP (fast on GPU)

    # ----------------- Trainer settings -----------------
    if args.max_epochs:
        trainer_cfg["max_epochs"] = args.max_epochs

    # Early stopping patience (if config uses callbacks.init_args.patience)
    if args.patience is not None:
        callbacks = trainer_cfg.get("callbacks")
        if isinstance(callbacks, dict) and "init_args" in callbacks:
            callbacks["init_args"]["patience"] = args.patience

    if args.output_ckpt_folder and args.output_ckpt_folder.lower() != "none":
        cfg["default_root_dir"] = args.output_ckpt_folder

    # ----------------- Data settings -----------------
    data_args["image_size"] = list(args.image_size)
    if args.num_workers is not None:
        data_args["num_workers"] = int(args.num_workers)
    else:
        # reasonable default (you can tune)
        data_args.setdefault("num_workers", max(2, (os.cpu_count() or 8) // 2))

    if args.num_workers is not None:
        data_args["num_workers"] = int(args.num_workers)

    # ----------------- Transforms (train / eval) -----------------
    h, w = list(args.image_size)

    train_transform = _build_train_transform(h, w, args)
    eval_transform = _build_eval_transform(h, w)

    # Always overwrite here (since this entrypoint is explicitly in charge)
    data_args["train_transform"] = train_transform
    data_args["eval_transform"] = eval_transform
    data_args["train_batch_size"] = int(args.batch_size)
    data_args["eval_batch_size"] = int(args.batch_size)

    scores_path = Path(cfg['default_root_dir']) / "val_scores_for_anomaly.txt"
    if scores_path.exists():
        scores_path.unlink(missing_ok=True)  # py>=3.8

    # ----------------- Write back -----------------
    with open(cfg_path, "w", encoding="utf-8") as fp:
        yaml.dump(cfg, fp)

    return cfg