README.md

metadata

task_categories:
  - image-classification
size_categories:
  - 100K<n<1M

CINIC10 dataset with interface of CIFAR10.

It is faster than the common CINIC10 due to the fact that all images are loaded into RAM while initing dataset instance.

You should save cinic10.py from this repo in local directory. And then import the CINIC10 class from it:

import torchvision
import torch
from torchvision transforms
from cinic10 import CINIC10

data_mean = [0.47889522, 0.47227842, 0.43047404]
data_std = [0.24205776, 0.23828046, 0.25874835]

transform_train = transforms.Compose([
    transforms.RandomCrop(32, padding=4),
    transforms.Resize(32),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    transforms.Normalize(data_mean, data_std),
])

transform_test = transforms.Compose([
    transforms.Resize(32),
    transforms.ToTensor(),
    transforms.Normalize(data_mean, data_std),
])

batch_size = 64
num_workers = 4

trainset = CINIC10(root='./data', train=True, download=True, transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True, num_workers=num_workers)

testset = CINIC10(root='./data', train=False, download=True, transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=batch_size, shuffle=False, num_workers=num_workers)