File size: 6,283 Bytes
b386992
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
HFDatasetDataModule
===================

`HFDatasetDataModule` is a PyTorch Lightning `LightningDataModule` that wraps Hugging Face (HF) datasets, enabling seamless integration with NeMo Framework. It allows users to load and preprocess datasets from the Hugging Face hub or custom datasets, supporting multiple data splits.


.. warning::
    This class requires the following packages:
    - `lightning.pytorch`
    - `torch`
    - `datasets`
    - `nemo`

    Ensure they are installed in your environment before using this module.

Class Definition
----------------

.. code-block:: python

    class HFDatasetDataModule(pl.LightningDataModule):
        ...

Initialization
--------------

.. autofunction:: HFDatasetDataModule.__init__

**Parameters:**

- **path_or_dataset** (`str` or `datasets.Dataset` or `datasets.DatasetDict`):
  - Path to the dataset on the Hugging Face Hub or a pre-loaded Hugging Face `Dataset` or `DatasetDict` object.

- **split** (`str` or `list`, optional):
  - Specifies which split(s) of the dataset to load. Can be a single split name (e.g., `"train"`) or a list of split names (e.g., `["train", "validation"]`).

- **collate_fn** (`callable`, optional):
  - A function to collate samples into batches. If `None`, a default collate function is used.

- **num_workers** (`int`, default=2):
  - Number of subprocesses to use for data loading.

- **pin_memory** (`bool`, default=True):
  - If `True`, the data loader will copy Tensors into CUDA pinned memory before returning them.

- **persistent_workers** (`bool`, default=True):
  - If `True`, the data loader will not shutdown the worker processes after a dataset has been consumed once.

- **seq_length** (`int`, default=1024):
  - Sequence length for the data sampler.

- **micro_batch_size** (`int`, default=2):
  - Batch size per micro-batch.

- **global_batch_size** (`int`, default=2):
  - Total batch size across all GPUs.

- **pad_token_id** (`int`, default=0):
  - Token ID used for padding sequences.

- **train_aliases** (`list` of `str`, default=["train", "training"]):
  - Synonyms for the training split.

- **test_aliases** (`list` of `str`, default=["test", "testing"]):
  - Synonyms for the testing split.

- **val_aliases** (`list` of `str`, default=["val", "validation", "valid", "eval"]):
  - Synonyms for the validation split.

- **kwargs**:
  - Additional keyword arguments passed to `datasets.load_dataset`.

Attributes
----------

- **dataset_splits** (`dict`):
  - Dictionary containing the dataset splits (`'train'`, `'val'`, `'test'`).

- **_collate_fn** (`callable`):
  - Function used to collate batches.

- **num_workers** (`int`):
  - Number of workers for data loading.

- **pin_memory** (`bool`):
  - Pin memory flag for data loading.

- **persistent_workers** (`bool`):
  - Persistent workers flag for data loading.

- **seq_length** (`int`):
  - Sequence length for data sampling.

- **micro_batch_size** (`int`):
  - Micro batch size.

- **global_batch_size** (`int`):
  - Global batch size.

- **pad_token_id** (`int`):
  - Padding token ID.

Methods
-------

.. autoclass:: HFDatasetDataModule
    :members:
    :exclude-members: __init__, __repr__, __str__

Detailed Method Descriptions
~~~~~~~~~~~~~~~~~~~~~~~~~~

train_dataloader()
~~~~~~~~~~~~~~~~~

Returns the training `DataLoader`.

**Returns:**

- **DataLoader**:
  - Data loader for the training split.

val_dataloader()
~~~~~~~~~~~~~~~

Returns the validation `DataLoader`.

**Returns:**

- **DataLoader**:
  - Data loader for the validation split.

test_dataloader()
~~~~~~~~~~~~~~~~

Returns the testing `DataLoader`.

**Returns:**

- **DataLoader**:
  - Data loader for the testing split.

map(function=None, split_names=None, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Applies a function to the specified splits of the dataset.

**Parameters:**

- **function** (`callable`, optional):
  - Function to apply to each example in the dataset.

- **split_names** (`str` or `list`, optional):
  - Specific splits to apply the function to. If `None`, applies to all splits.

- **kwargs**:
  - Additional keyword arguments passed to `Dataset.map`.

**Returns:**

- **None**

Properties
----------

.. py:attribute:: train

    Returns the training split of the dataset.

.. py:attribute:: val

    Returns the validation split of the dataset.

.. py:attribute:: test

    Returns the testing split of the dataset.

Static Methods
--------------

from_dict(dataset_dict, split, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Creates an `HFDatasetDataModule` from a dictionary.

**Parameters:**

- **dataset_dict** (`dict`):
  - Dictionary representing the dataset.

- **split** (`str` or `list`):
  - Split(s) to load.

- **kwargs**:
  - Additional keyword arguments passed to the constructor.

**Returns:**

- **HFDatasetDataModule**:
  - An instance of `HFDatasetDataModule`.

collate_fn(batch, pad_token_id=0)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Default collate function to pad and batch sequences.

**Parameters:**

- **batch** (`list` of `dict`):
  - Batch of samples to collate.

- **pad_token_id** (`int`, default=0):
  - Token ID used for padding.

**Returns:**

- **dict**:
  - Batched and padded tensors.

Usage Examples
--------------

**Loading a Single Split:**

.. code-block:: python

    from nemo_lightning import HFDatasetDataModule

    data_module = HFDatasetDataModule(
        path_or_dataset="rajpurkar/squad",
        split="train",
        micro_batch_size=8,
        global_batch_size=64,
        pad_token_id=0
    )

    train_loader = data_module.train_dataloader()

**Loading Multiple Splits:**

.. code-block:: python

    from nemo_lightning import HFDatasetDataModule

    data_module = HFDatasetDataModule(
        path_or_dataset="rajpurkar/squad",
        split=["train", "validation"],
        micro_batch_size=8,
        global_batch_size=64,
        pad_token_id=0
    )

    train_loader = data_module.train_dataloader()
    val_loader = data_module.val_dataloader()

Mapping a Function Over the Dataset
-----------------------------------

.. code-block:: python

    def preprocess(example):
        example['input_ids'] = tokenizer.encode(example['text'])
        return example

    data_module.map(preprocess, split_names=["train", "val"])