1import random
2from pathlib import PurePath, Path
3from typing import List, Callable, Dict, Optional
4
5from torchvision import datasets, transforms
6
7import torch
8from labml import lab
9from labml import monit
10from labml.configs import BaseConfigs
11from labml.configs import aggregate, option
12from labml.utils.download import download_file
13from torch.utils.data import DataLoader
14from torch.utils.data import IterableDataset, Dataset

17def _mnist_dataset(is_train, transform):
18    return datasets.MNIST(str(lab.get_data_path()),
19                          train=is_train,
20                          download=True,
21                          transform=transform)

24class MNISTConfigs(BaseConfigs):

44    dataset_name: str = 'MNIST'
45    dataset_transforms: transforms.Compose
46    train_dataset: datasets.MNIST
47    valid_dataset: datasets.MNIST
48
49    train_loader: DataLoader
50    valid_loader: DataLoader
51
52    train_batch_size: int = 64
53    valid_batch_size: int = 1024
54
55    train_loader_shuffle: bool = True
56    valid_loader_shuffle: bool = False

59@option(MNISTConfigs.dataset_transforms)
60def mnist_transforms():
61    return transforms.Compose([
62        transforms.ToTensor(),
63        transforms.Normalize((0.1307,), (0.3081,))
64    ])
65
66
67@option(MNISTConfigs.train_dataset)
68def mnist_train_dataset(c: MNISTConfigs):
69    return _mnist_dataset(True, c.dataset_transforms)
70
71
72@option(MNISTConfigs.valid_dataset)
73def mnist_valid_dataset(c: MNISTConfigs):
74    return _mnist_dataset(False, c.dataset_transforms)
75
76
77@option(MNISTConfigs.train_loader)
78def mnist_train_loader(c: MNISTConfigs):
79    return DataLoader(c.train_dataset,
80                      batch_size=c.train_batch_size,
81                      shuffle=c.train_loader_shuffle)
82
83
84@option(MNISTConfigs.valid_loader)
85def mnist_valid_loader(c: MNISTConfigs):
86    return DataLoader(c.valid_dataset,
87                      batch_size=c.valid_batch_size,
88                      shuffle=c.valid_loader_shuffle)
89
90
91aggregate(MNISTConfigs.dataset_name, 'MNIST',
92          (MNISTConfigs.dataset_transforms, 'mnist_transforms'),
93          (MNISTConfigs.train_dataset, 'mnist_train_dataset'),
94          (MNISTConfigs.valid_dataset, 'mnist_valid_dataset'),
95          (MNISTConfigs.train_loader, 'mnist_train_loader'),
96          (MNISTConfigs.valid_loader, 'mnist_valid_loader'))
97
98
99def _cifar_dataset(is_train, transform):
100    return datasets.CIFAR10(str(lab.get_data_path()),
101                            train=is_train,
102                            download=True,
103                            transform=transform)
104
105
106class CIFAR10Configs(BaseConfigs):

125    dataset_name: str = 'CIFAR10'
126    dataset_transforms: transforms.Compose
127    train_dataset: datasets.CIFAR10
128    valid_dataset: datasets.CIFAR10
129
130    train_loader: DataLoader
131    valid_loader: DataLoader
132
133    train_batch_size: int = 64
134    valid_batch_size: int = 1024
135
136    train_loader_shuffle: bool = True
137    valid_loader_shuffle: bool = False

140@CIFAR10Configs.calc(CIFAR10Configs.dataset_transforms)
141def cifar10_transforms():
142    return transforms.Compose([
143        transforms.ToTensor(),
144        transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
145    ])
146
147
148@CIFAR10Configs.calc(CIFAR10Configs.train_dataset)
149def cifar10_train_dataset(c: CIFAR10Configs):
150    return _cifar_dataset(True, c.dataset_transforms)
151
152
153@CIFAR10Configs.calc(CIFAR10Configs.valid_dataset)
154def cifar10_valid_dataset(c: CIFAR10Configs):
155    return _cifar_dataset(False, c.dataset_transforms)
156
157
158@CIFAR10Configs.calc(CIFAR10Configs.train_loader)
159def cifar10_train_loader(c: CIFAR10Configs):
160    return DataLoader(c.train_dataset,
161                      batch_size=c.train_batch_size,
162                      shuffle=c.train_loader_shuffle)
163
164
165@CIFAR10Configs.calc(CIFAR10Configs.valid_loader)
166def cifar10_valid_loader(c: CIFAR10Configs):
167    return DataLoader(c.valid_dataset,
168                      batch_size=c.valid_batch_size,
169                      shuffle=c.valid_loader_shuffle)
170
171
172CIFAR10Configs.aggregate(CIFAR10Configs.dataset_name, 'CIFAR10',
173                         (CIFAR10Configs.dataset_transforms, 'cifar10_transforms'),
174                         (CIFAR10Configs.train_dataset, 'cifar10_train_dataset'),
175                         (CIFAR10Configs.valid_dataset, 'cifar10_valid_dataset'),
176                         (CIFAR10Configs.train_loader, 'cifar10_train_loader'),
177                         (CIFAR10Configs.valid_loader, 'cifar10_valid_loader'))
178
179
180class TextDataset:
181    itos: List[str]
182    stoi: Dict[str, int]
183    n_tokens: int
184    train: str
185    valid: str
186    standard_tokens: List[str] = []
187
188    @staticmethod
189    def load(path: PurePath):
190        with open(str(path), 'r') as f:
191            return f.read()
192
193    def __init__(self, path: PurePath, tokenizer: Callable, train: str, valid: str, test: str, *,
194                 n_tokens: Optional[int] = None,
195                 stoi: Optional[Dict[str, int]] = None,
196                 itos: Optional[List[str]] = None):
197        self.test = test
198        self.valid = valid
199        self.train = train
200        self.tokenizer = tokenizer
201        self.path = path
202
203        if n_tokens or stoi or itos:
204            assert stoi and itos and n_tokens
205            self.n_tokens = n_tokens
206            self.stoi = stoi
207            self.itos = itos
208        else:
209            self.n_tokens = len(self.standard_tokens)
210            self.stoi = {t: i for i, t in enumerate(self.standard_tokens)}
211
212            with monit.section("Tokenize"):
213                tokens = self.tokenizer(self.train) + self.tokenizer(self.valid)
214                tokens = sorted(list(set(tokens)))
215
216            for t in monit.iterate("Build vocabulary", tokens):
217                self.stoi[t] = self.n_tokens
218                self.n_tokens += 1
219
220            self.itos = [''] * self.n_tokens
221            for t, n in self.stoi.items():
222                self.itos[n] = t
223
224    def text_to_i(self, text: str) -> torch.Tensor:
225        tokens = self.tokenizer(text)
226        return torch.tensor([self.stoi[s] for s in tokens if s in self.stoi], dtype=torch.long)
227
228    def __repr__(self):
229        return f'{len(self.train) / 1_000_000 :,.2f}M, {len(self.valid) / 1_000_000 :,.2f}M - {str(self.path)}'
230
231
232class SequentialDataLoader(IterableDataset):
233    def __init__(self, *, text: str, dataset: TextDataset,
234                 batch_size: int, seq_len: int):
235        self.seq_len = seq_len
236        data = dataset.text_to_i(text)
237        n_batch = data.shape[0] // batch_size
238        data = data.narrow(0, 0, n_batch * batch_size)
239        data = data.view(batch_size, -1).t().contiguous()
240        self.data = data
241
242    def __len__(self):
243        return self.data.shape[0] // self.seq_len
244
245    def __iter__(self):
246        self.idx = 0
247        return self
248
249    def __next__(self):
250        if self.idx >= self.data.shape[0] - 1:
251            raise StopIteration()
252
253        seq_len = min(self.seq_len, self.data.shape[0] - 1 - self.idx)
254        i = self.idx + seq_len
255        data = self.data[self.idx: i]
256        target = self.data[self.idx + 1: i + 1]
257        self.idx = i
258        return data, target
259
260    def __getitem__(self, idx):
261        seq_len = min(self.seq_len, self.data.shape[0] - 1 - idx)
262        i = idx + seq_len
263        data = self.data[idx: i]
264        target = self.data[idx + 1: i + 1]
265        return data, target
266
267
268class SequentialUnBatchedDataset(Dataset):
269    def __init__(self, *, text: str, dataset: TextDataset,
270                 seq_len: int,
271                 is_random_offset: bool = True):
272        self.is_random_offset = is_random_offset
273        self.seq_len = seq_len
274        self.data = dataset.text_to_i(text)
275
276    def __len__(self):
277        return (self.data.shape[0] - 1) // self.seq_len
278
279    def __getitem__(self, idx):
280        start = idx * self.seq_len
281        assert start + self.seq_len + 1 <= self.data.shape[0]
282        if self.is_random_offset:
283            start += random.randint(0, min(self.seq_len - 1, self.data.shape[0] - (start + self.seq_len + 1)))
284
285        end = start + self.seq_len
286        data = self.data[start: end]
287        target = self.data[start + 1: end + 1]
288        return data, target
289
290
291class TextFileDataset(TextDataset):
292    standard_tokens = []
293
294    def __init__(self, path: PurePath, tokenizer: Callable, *,
295                 url: Optional[str] = None,
296                 filter_subset: Optional[int] = None):
297        path = Path(path)
298        if not path.exists():
299            if not url:
300                raise FileNotFoundError(str(path))
301            else:
302                download_file(url, path)
303
304        with monit.section("Load data"):
305            text = self.load(path)
306            if filter_subset:
307                text = text[:filter_subset]
308            split = int(len(text) * .9)
309            train = text[:split]
310            valid = text[split:]
311
312        super().__init__(path, tokenizer, train, valid, '')
313
314
315def _test_tiny_shakespeare():
316    from labml import lab
317    _ = TextFileDataset(lab.get_data_path() / 'tiny_shakespeare.txt', lambda x: list(x),
318                        url='https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt')
319
320
321if __name__ == '__main__':
322    _test_tiny_shakespeare()