Spaces:
Build error
Build error
File size: 3,499 Bytes
daf0288 |
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 |
from typing import Any
from torch.utils.data import DataLoader, Dataset, Sampler
from functools import partial
import tokenizers as tk
import torch
from torch.utils.data import default_collate
from ..utils.mask_generator import MaskGenerator
from ..utils import (
prepare_html_seq,
prepare_cell_seq,
prepare_bbox_seq,
)
class Collator:
def __init__(
self,
vocab: tk.Tokenizer,
max_seq_len: int,
label_type: str,
) -> None:
self.vocab = vocab
self.vocab.enable_truncation(max_seq_len)
self.label_type = label_type
def __call__(self, batch) -> Any:
return self._collate_batch(batch, self.vocab, self.label_type)
def _collate_batch(
self,
batch: list[dict],
vocab: tk.Tokenizer,
label_type: str,
):
if "cell" in label_type:
image_list = [j for i in batch for j in i[0]]
else:
image_list = [i["image"] for i in batch]
image_list = default_collate(image_list)
if "cell" in label_type:
filename = [(j["filename"], j["bbox_id"]) for i in batch for j in i[1]]
else:
filename = [i["filename"] for i in batch]
label = dict(filename=filename)
if "html" in label_type:
html_list = ["".join(prepare_html_seq(i["html"])) for i in batch]
label["html"] = vocab.encode_batch(html_list)
if "cell" in label_type:
cell_list = [
" ".join(prepare_cell_seq(j["cell"])) for i in batch for j in i[1]
]
label["cell"] = vocab.encode_batch(cell_list)
if "bbox" in label_type:
bbox_list = [" ".join(prepare_bbox_seq(i["bbox"])) for i in batch]
label["bbox"] = vocab.encode_batch(bbox_list)
return image_list, label
def generate_mask_for_batch_samples(
batch, grid_size: int, num_mask_patches: int, min_num_patches: int
):
N = len(batch)
mg = MaskGenerator(
input_size=grid_size,
num_mask_patches=num_mask_patches,
min_num_patches=min_num_patches,
)
mask_list = [mg() for _ in range(N)]
return default_collate(batch), default_collate(mask_list)
def dataloader_vae(
dataset: Dataset, batch_size: int, sampler: Sampler = None, **kwargs
) -> DataLoader:
dataloader = DataLoader(
dataset, batch_size, sampler=sampler, num_workers=8, pin_memory=True
)
return dataloader
def dataloader_beit(
dataset: Dataset,
grid_size: int,
num_mask_patches: int,
min_num_patches: int,
batch_size: int,
sampler: Sampler = None,
**kwargs
):
dataloader = DataLoader(
dataset,
batch_size,
sampler=sampler,
collate_fn=partial(
generate_mask_for_batch_samples,
grid_size=grid_size,
num_mask_patches=num_mask_patches,
min_num_patches=min_num_patches,
),
num_workers=8,
pin_memory=True,
)
return dataloader
def dataloader_html(
dataset: Dataset,
batch_size: int,
vocab: tk.Tokenizer,
max_seq_len: int,
label_type: str,
sampler=None,
) -> DataLoader:
collate_fn = Collator(vocab, max_seq_len, label_type)
dataloader = DataLoader(
dataset,
batch_size=batch_size,
shuffle=False,
num_workers=8,
collate_fn=collate_fn,
pin_memory=True,
sampler=sampler,
)
return dataloader
|