siclib/datasets/base_dataset.py

"""Base class for dataset.

See mnist.py for an example of dataset.
"""

import collections
import logging
from abc import ABCMeta, abstractmethod

import omegaconf
import torch
from omegaconf import OmegaConf
from torch.utils.data import DataLoader, Sampler, get_worker_info
from torch.utils.data._utils.collate import default_collate_err_msg_format, np_str_obj_array_pattern

from siclib.utils.tensor import string_classes
from siclib.utils.tools import set_num_threads, set_seed

logger = logging.getLogger(__name__)

# mypy: ignore-errors


class LoopSampler(Sampler):
    """Infinite sampler that loops over a given number of elements."""

    def __init__(self, loop_size: int, total_size: int = None):
        """Initialize the sampler.

        Args:
            loop_size (int): Number of elements to loop over.
            total_size (int, optional): Total number of elements. Defaults to None.
        """
        self.loop_size = loop_size
        self.total_size = total_size - (total_size % loop_size)

    def __iter__(self):
        """Return an iterator over the elements."""
        return (i % self.loop_size for i in range(self.total_size))

    def __len__(self):
        """Return the number of elements."""
        return self.total_size


def worker_init_fn(i):
    """Initialize the workers with a different seed."""
    info = get_worker_info()
    if hasattr(info.dataset, "conf"):
        conf = info.dataset.conf
        set_seed(info.id + conf.seed)
        set_num_threads(conf.num_threads)
    else:
        set_num_threads(1)


def collate(batch):
    """Difference with PyTorch default_collate: it can stack of other objects."""
    if not isinstance(batch, list):  # no batching
        return batch
    elem = batch[0]
    elem_type = type(elem)
    if isinstance(elem, torch.Tensor):
        # out = None
        if torch.utils.data.get_worker_info() is not None:
            # If we're in a background process, concatenate directly into a
            # shared memory tensor to avoid an extra copy
            numel = sum([x.numel() for x in batch])
            try:
                _ = elem.untyped_storage()._new_shared(numel)
            except AttributeError:
                _ = elem.storage()._new_shared(numel)
        return torch.stack(batch, dim=0)
    elif (
        elem_type.__module__ == "numpy"
        and elem_type.__name__ != "str_"
        and elem_type.__name__ != "string_"
    ):
        if elem_type.__name__ in ["ndarray", "memmap"]:
            # array of string classes and object
            if np_str_obj_array_pattern.search(elem.dtype.str) is not None:
                raise TypeError(default_collate_err_msg_format.format(elem.dtype))
            return collate([torch.as_tensor(b) for b in batch])
        elif elem.shape == ():  # scalars
            return torch.as_tensor(batch)
    elif isinstance(elem, float):
        return torch.tensor(batch, dtype=torch.float64)
    elif isinstance(elem, int):
        return torch.tensor(batch)
    elif isinstance(elem, string_classes):
        return batch
    elif isinstance(elem, collections.abc.Mapping):
        return {key: collate([d[key] for d in batch]) for key in elem}
    elif isinstance(elem, tuple) and hasattr(elem, "_fields"):  # namedtuple
        return elem_type(*(collate(samples) for samples in zip(*batch)))
    elif isinstance(elem, collections.abc.Sequence):
        # check to make sure that the elements in batch have consistent size
        it = iter(batch)
        elem_size = len(next(it))
        if any(len(elem) != elem_size for elem in it):
            raise RuntimeError("each element in list of batch should be of equal size")
        transposed = zip(*batch)
        return [collate(samples) for samples in transposed]
    elif elem is None:
        return elem
    else:
        # try to stack anyway in case the object implements stacking.
        return torch.stack(batch, 0)


class BaseDataset(metaclass=ABCMeta):
    """Base class for dataset.

    What the dataset model is expect to declare:
    default_conf: dictionary of the default configuration of the dataset.
    It overwrites base_default_conf in BaseModel, and it is overwritten by
    the user-provided configuration passed to __init__.
    Configurations can be nested.

    _init(self, conf): initialization method, where conf is the final
    configuration object (also accessible with `self.conf`). Accessing
    unknown configuration entries will raise an error.

    get_dataset(self, split): method that returns an instance of
    torch.utils.data.Dataset corresponding to the requested split string,
    which can be `'train'`, `'val'`, or `'test'`.
    """

    base_default_conf = {
        "name": "???",
        "num_workers": "???",
        "train_batch_size": "???",
        "val_batch_size": "???",
        "test_batch_size": "???",
        "shuffle_training": True,
        "batch_size": 1,
        "num_threads": 1,
        "seed": 0,
        "prefetch_factor": 2,
    }
    default_conf = {}

    def __init__(self, conf):
        """Perform some logic and call the _init method of the child model."""
        default_conf = OmegaConf.merge(
            OmegaConf.create(self.base_default_conf),
            OmegaConf.create(self.default_conf),
        )
        OmegaConf.set_struct(default_conf, True)
        if isinstance(conf, dict):
            conf = OmegaConf.create(conf)
        self.conf = OmegaConf.merge(default_conf, conf)
        OmegaConf.set_readonly(self.conf, True)
        logger.info(f"Creating dataset {self.__class__.__name__}")
        self._init(self.conf)

    @abstractmethod
    def _init(self, conf):
        """To be implemented by the child class."""
        raise NotImplementedError

    @abstractmethod
    def get_dataset(self, split):
        """To be implemented by the child class."""
        raise NotImplementedError

    def get_data_loader(self, split, shuffle=None, pinned=False, distributed=False):
        """Return a data loader for a given split."""
        assert split in ["train", "val", "test"]
        dataset = self.get_dataset(split)
        try:
            batch_size = self.conf[f"{split}_batch_size"]
        except omegaconf.MissingMandatoryValue:
            batch_size = self.conf.batch_size
        num_workers = self.conf.get("num_workers", batch_size)
        if distributed:
            shuffle = False
            sampler = torch.utils.data.distributed.DistributedSampler(dataset)
        else:
            sampler = None
            if shuffle is None:
                shuffle = split == "train" and self.conf.shuffle_training
        return DataLoader(
            dataset,
            batch_size=batch_size,
            shuffle=shuffle,
            sampler=sampler,
            pin_memory=pinned,
            collate_fn=collate,
            num_workers=num_workers,
            worker_init_fn=worker_init_fn,
            prefetch_factor=self.conf.prefetch_factor,
        )

    def get_overfit_loader(self, split: str):
        """Return an overfit data loader.

        The training set is composed of a single duplicated batch, while
        the validation and test sets contain a single copy of this same batch.
        This is useful to debug a model and make sure that losses and metrics
        correlate well.
        """
        assert split in {"train", "val", "test"}
        dataset = self.get_dataset("train")
        sampler = LoopSampler(
            self.conf.batch_size,
            len(dataset) if split == "train" else self.conf.batch_size,
        )
        num_workers = self.conf.get("num_workers", self.conf.batch_size)
        return DataLoader(
            dataset,
            batch_size=self.conf.batch_size,
            pin_memory=True,
            num_workers=num_workers,
            sampler=sampler,
            worker_init_fn=worker_init_fn,
            collate_fn=collate,
        )