Differences with torchvision.datasets.CIFAR10

torchvision.datasets.CIFAR10

class torchvision.datasets.CIFAR10(
    root: str,
    train: bool = True,
    transform: Optional[Callable] = None,
    target_transform: Optional[Callable] = None,
    download: bool = False)

For more information, see torchvision.datasets.CIFAR10.

mindspore.dataset.Cifar10Dataset

class mindspore.dataset.Cifar10Dataset(
    dataset_dir,
    usage=None,
    num_samples=None,
    num_parallel_workers=None,
    shuffle=None,
    sampler=None,
    num_shards=None,
    shard_id=None,
    cache=None)

For more information, see mindspore.dataset.Cifar10Dataset.

Differences

PyTorch: Read the CIFAR-10 dataset(only support CIFAR-10 python version). API integrates the transformation operations for image and label.

MindSpore: Read the CIFAR-10 dataset(only support CIFAR-10 binary version). Downloading dataset from web is not supported. Transforms for image and label depends on mindshare.dataset.map operation.

Categories	Subcategories	PyTorch	MindSpore	Difference
Parameter	Parameter1	root	dataset_dir	-
	Parameter2	train	-	Specify whether it is the training set，supported by usage in MindSpore
	Parameter3	transform	-	Supported by mindspore.dataset.map operation
	Parameter4	target_transform	-	Supported by mindspore.dataset.map operation
	Parameter5	download	-	Not supported by MindSpore
	Parameter6	-	usage	Specify a subset of the dataset
	Parameter7	-	num_samples	The number of images to be included in the dataset.
	Parameter8	-	num_parallel_workers	Number of worker threads to read the data
	Parameter9	-	shuffle	Whether to perform shuffle on the dataset
	Parameter10	-	sampler	Object used to choose samples from the dataset
	Parameter11	-	num_shards	Number of shards that the dataset will be divided into
	Parameter12	-	shard_id	The shard ID used in distributed training
	Parameter13	-	cache	Specify single-node data caching service

Code Example

# PyTorch
import torchvision.transforms as T
import torchvision.datasets as datasets
from torch.utils.data import DataLoader

root = "/path/to/dataset_directory/"
dataset = datasets.CIFAR10(root, train=True, transform=T.RandomCrop((28, 28)), download=True)
dataloader = DataLoader(dataset)

# MindSpore
import mindspore.dataset as ds
import mindspore.dataset.vision as vision

# Download the dataset files, unzip into the following structure
#  .
#  └── /path/to/dataset_directory/
#       ├── data_batch_1.bin
#       ├── data_batch_2.bin
#       ├── data_batch_3.bin
#       ├── data_batch_4.bin
#       ├── data_batch_5.bin
#       ├── test_batch.bin
#       ├── readme.html
#       └── batches.meta.txt
root = "/path/to/dataset_directory/"
ms_dataloader = ds.Cifar10Dataset(root, usage='train')
ms_dataloader = ms_dataloader.map(vision.RandomCrop((28, 28)), ["image"])