import torch import os from torch.utils.data import DataLoader from .data import create_skeleton from .data import create_skeleton, custom_collate_for_mmgt, get_DatasetClass_for_evaluation os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" def prepare_eval_dataset(config, split, shuffle=False, augmentation=0, da_mirroring=0, da_rotations=0, drop_last=False, num_workers=None, batch_size=None ,dataset=None, stats_mode="probabilistic", **kwargs): data_loader_name = f"data_loader_{split}" config[data_loader_name]["shuffle"] = shuffle config[data_loader_name]["da_mirroring"] = da_mirroring config[data_loader_name]["da_rotations"] = da_rotations config[data_loader_name]["augmentation"] = augmentation config[data_loader_name]["drop_last"] = drop_last config[data_loader_name]["if_permute_keypoints"] = False if "probabilistic" in stats_mode.lower(): config[data_loader_name]["if_load_mmgt"] = True else: config[data_loader_name]["if_load_mmgt"] = False if batch_size is not None: config["batch_size"] = batch_size if num_workers is not None: config["num_workers"] = 0 skeleton = create_skeleton(**config) if dataset == None: split = data_loader_name.split("_")[-1] if 'eval' not in data_loader_name else "_".join(data_loader_name.split("_")[-2:]) DatasetClass = get_DatasetClass_for_evaluation(config['dataset_name']) dataset = DatasetClass(split=split, skeleton=skeleton, **config) data_loader = DataLoader(dataset, pin_memory= True, shuffle=config[data_loader_name]["shuffle"], batch_size=config["batch_size"], num_workers=config["num_workers"], drop_last=config[data_loader_name]["drop_last"],collate_fn=custom_collate_for_mmgt) assert len(data_loader) >0, "Dataset is too small and last batch was dropped" return data_loader, dataset, skeleton