-
Notifications
You must be signed in to change notification settings - Fork 5
/
collator.py
59 lines (41 loc) · 1.76 KB
/
collator.py
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
from abc import ABC, abstractmethod
from typing import Dict, List
import torch
from modalities.batch import DatasetBatch
class CollateFnIF(ABC):
"""CollateFnIF class to define a collate function interface."""
@abstractmethod
def __call__(self, batch: List[Dict[str, torch.Tensor]]) -> DatasetBatch:
"""
Process a batch of data.
Args:
batch (List[Dict[str, torch.Tensor]]): A list of dictionaries containing tensors.
Returns:
DatasetBatch: The processed batch of data.
Raises:
NotImplementedError: This abstract method should be implemented in a subclass.
"""
raise NotImplementedError
class GPT2LLMCollateFn(CollateFnIF):
"""GPT2LLMCollateFn class to define a collate function for GPT2 language model."""
def __init__(self, sample_key: str, target_key: str):
"""
Initializes the Collator object.
Args:
sample_key (str): The key for accessing the sample data.
target_key (str): The key for accessing the target data.
"""
self.sample_key = sample_key
self.target_key = target_key
def __call__(self, batch: List[Dict[str, torch.Tensor]]) -> DatasetBatch:
"""
Process a batch of data.
Args:
batch (List[Dict[str, torch.Tensor]]): A list of dictionaries containing tensors.
Returns:
DatasetBatch: A processed batch of data where sample and target sequences are created.
"""
sample_tensor = torch.stack([torch.tensor(d[self.sample_key]) for d in batch])
samples = {self.sample_key: sample_tensor[:, :-1]}
targets = {self.target_key: sample_tensor[:, 1:]}
return DatasetBatch(targets=targets, samples=samples)