pytorch module 정리

torchvision.datasets (cococaptions)

yuuuun 2021. 4. 20. 15:59
반응형

pytorch.org/vision/stable/datasets.html

 

torchvision.datasets — Torchvision master documentation

torchvision.datasets All datasets are subclasses of torch.utils.data.Dataset i.e, they have __getitem__ and __len__ methods implemented. Hence, they can all be passed to a torch.utils.data.DataLoader which can load multiple samples in parallel using torch.

pytorch.org

img_data = torchvision.datasets.ImageNet('[imagenet directory]')
data_loader = torch.utils.data.DataLoader(img_data, batch_size=4, shuffle=True, num_workers=args.nThreads)

CocoCaptions

from torchvision.dataset import CocoCaptions

CLASS torchvision.datasets.CocoCaptions(root: str, annFile: str, 
					transform: Union[Callable, NoneType] = None, 
					target_transform: Union[Callable, NoneType] = None, 
					transforms: Union[Callable, NoneType] = None)

root: image가 다운 받아져있는 directory

annFile: json annotation file 저장되어 있는 directory

transform: 어떤 return을 반환해줄지 정의

target_transform: 어떤 data를 선택할지 정의

CLIP 코드 中 (check code)

import torchvision.transforms as transforms
import torchvision.datasets as dset
improt random

def rand_choice(x):
    return random.choice(x)

tfms = transforms.Compose([transforms.Resize((128, 128)), 
                                transforms.ToTensor(),
                                transforms.Normalize(mean=(0.485, 0.456, 0.406), 
                                                        std=(0.229, 0.224, 0.225))])

target_tfm = rand_choice
cap = dset.CocoCaptions(root = './datasets/train2014/',
                                annFile = './datasets/annotations/captions_train2014.json',
                                transform=tfms,
                                target_transform=target_tfm,)

tfms: return할 데이터의 형태

target_tfm: 기존에는 lambda로 설정되어 있었으나 오류가 나서 수정

root내의 파일 형태는 jpg 파일

반응형