File size: 2,135 Bytes
485748e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""PyTorch loading example for Shelly97/multimodal-hate-triplets.

Install: pip install -U datasets pillow torch transformers
The processor is supplied by the caller to match the training model.
"""
import torch
from torch.utils.data import Dataset


def process_image_text(image, text, processor):
    processed = processor(
        text=text, images=image, max_length=77, padding="max_length",
        truncation=True, return_tensors="pt",
    )
    return {key: processed[key].squeeze(0)
            for key in ("pixel_values", "input_ids", "attention_mask")}


def load_and_process(example, processor):
    result = {}
    for key in ("org", "pos", "neg"):
        # Hugging Face Image columns decode directly into PIL images.
        with example[f"{key}_img"].convert("RGB") as image:
            result[key] = process_image_text(image, example[f"{key}_txt"], processor)
    result["label"] = example["org_label"]
    return result


class MultimodalDataset(Dataset):
    def __init__(self, data, processor):
        self.data = data
        self.processor = processor

    def __len__(self):
        return len(self.data)

    def __getitem__(self, idx):
        # Propagate corrupt data/processor errors so training cannot silently lose rows.
        return load_and_process(self.data[idx], self.processor)


def collate_fn(batch):
    if not batch:
        return None
    result = {
        key: {field: torch.stack([item[key][field] for item in batch])
              for field in ("pixel_values", "input_ids", "attention_mask")}
        for key in ("org", "pos", "neg")
    }
    result["label"] = torch.tensor([item["label"] for item in batch], dtype=torch.long)
    return result


# Usage (after publication):
# from datasets import load_dataset
# from transformers import CLIPProcessor
# from torch.utils.data import DataLoader
# processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
# data_pairs = load_dataset("Shelly97/multimodal-hate-triplets", split="train")
# dataset = MultimodalDataset(data_pairs, processor)
# loader = DataLoader(dataset, batch_size=32, shuffle=True, collate_fn=collate_fn)