关于Dataloader中的collate_fn参数

以MNIST为例

from torchvision import datasets
mnist = datasets.MNIST(root='./data/', train=True, download=True)
print(mnist[0])

结果

(<PIL.Image.Image image mode=L size=28x28 at 0x196E3F1D898>, 5)

MINIST数据集的dataset是由一张图片和一个label组成的元组

dataloader = torch.utils.data.DataLoader(dataset=mnist, batch_size=2, shuffle=True,collate_fn=lambda x:x)
for each in dataloader:
    print(each)
    break

结果

[(<PIL.Image.Image image mode=L size=28x28 at 0x2CB3B105630>, 0), (<PIL.Image.Image image mode=L size=28x28 at 0x2CB3B105668>, 2)]

collate_fn为lamda x:x时表示对传入进来的数据不做处理

下面自定义collate_fn看看什么效果

def collate(data):
    img = []
    label = []
    for each in data:
        img.append(each[0])
        label.append(each[1])
    return img,label
dataloader = torch.utils.data.DataLoader(dataset=mnist, batch_size=2, shuffle=True,collate_fn=collate)
for each in dataloader:
    print(each)
    break

结果

([<PIL.Image.Image image mode=L size=28x28 at 0x241433A36D8>, <PIL.Image.Image image mode=L size=28x28 at 0x241433A3710>], [9, 3])

说明：若不设置collate_fn参数则会使用默认处理函数
但必须保证传进来的数据都是tensor格式否则会报错