pytorch数据抽样 随机取样

WeightedRandomSampler

不放回只能用WeightedRandomSampler，其他都是放回取样

只需将权重都设为1(or相同值)

from torch.utils.data import WeightedRandomSampler,DataLoader

a = torch.Tensor([1, 2, 3, 4, 5])
dataloader = DataLoader(dataset=a,
                    batch_size=4,  # 组织成batch大小
                    sampler=WeightedRandomSampler(torch.ones(a.size()), 20))
for i in dataloader:
     print(i)

结果

tensor([1., 5., 1., 2.])
tensor([5., 1., 5., 2.])
tensor([1., 4., 2., 2.])
tensor([4., 5., 4., 3.])
tensor([5., 2., 2., 1.])

参数

Arguments:
        weights (sequence)   : a sequence of weights, not necessary summing up to one
        num_samples (int): number of samples to draw
        replacement (bool): if ``True``, samples are drawn with replacement.
            If not, they are drawn without replacement, which means that when a
            sample index is drawn for a row, it cannot be drawn again for that row.

RandomSampler编组输出

#emb_pos[5,1600]
for num in [1, 2, 3, 4]:
    print('num', num)
    emb_pos_ = emb_pos.repeat(num,1)
    loader_pos = DataLoader(dataset=emb_pos_,
                            batch_size=num,  # 组织成batch大小
                            sampler=RandomSampler(emb_pos_))
    for i in loader_pos:
        print(i.size())

结果

num 1
torch.Size([1, 1600])
torch.Size([1, 1600])
torch.Size([1, 1600])
torch.Size([1, 1600])
torch.Size([1, 1600])
num 2
torch.Size([2, 1600])
torch.Size([2, 1600])
torch.Size([2, 1600])
torch.Size([2, 1600])
torch.Size([2, 1600])
num 3
torch.Size([3, 1600])
torch.Size([3, 1600])
torch.Size([3, 1600])
torch.Size([3, 1600])
torch.Size([3, 1600])
num 4
torch.Size([4, 1600])
torch.Size([4, 1600])
torch.Size([4, 1600])
torch.Size([4, 1600])
torch.Size([4, 1600])