dataloader

DataLoader is a pytorch class that wraps around a dataset to create a iterator that makes it easy to shuffle and create batches for training.

from torch.utils.data import DataLoader

coll = range(15)
dl = DataLoader(coll, batch_size=5, shuffle=True)
list(dl), next(iter(dl))
 

([tensor([ 8, 12,  4,  1,  7]),
  tensor([ 0,  3,  5, 13,  9]),
  tensor([11,  6, 10, 14,  2])],
 tensor([ 1,  5,  9,  2, 12]))

Two kinds of datasets can be supplied to a pytorch DataLoader, a iterable-style dataset or a map-style dataset.

A iterable dataset can iterable over a data collection. The easiest way to test it is to use iter() over the dataset. For example, a range object is a iterable dataset:

type(coll), iter(coll)

(range, <range_iterator at 0x7fc7f8111950>)

A map-style dataset represents a map from (possibly non-integral) indices/keys to data samples (should have key-value pairs). It also implements the __getitem__() and __len__() methods.

import string
 
 
class AlphabetDataset:
    def __init__(self, letters):
        self.letters = list(letters)
 
    def __len__(self):
        return len(self.letters)
 
    def __getitem__(self, idx):
        return self.letters[idx]
 
 
dataset = AlphabetDataset(string.ascii_lowercase)
dataloader = DataLoader(dataset, batch_size=6, shuffle=True)
next(iter(dataloader))
 

['o', 'v', 'k', 's', 't', 'p']

or more simply, use a dictionary

dt = dict(enumerate(string.ascii_lowercase))
dt.__getitem__(2), dt.__len__()

('c', 26)

dicLoader = DataLoader(dt, batch_size=6, shuffle=True)
next(iter(dicLoader))

['w', 's', 't', 'm', 'c', 'd']