1

有一个更好的方法吗？如何在不创建新张量对象的情况下用零填充张量？我需要输入始终相同batchsize，所以我想填充小于batchsize零的输入。就像序列长度较短时在 NLP 中填充零一样，但这是批处理的填充。

目前，我创建了一个新张量，但正因为如此，我的 GPU 将出现内存不足。我不想将批处理大小减少一半来处理此操作。

import torch
from torch import nn

class MyModel(nn.Module):
    def __init__(self, batchsize=16):
        super().__init__()
        self.batchsize = batchsize
    
    def forward(self, x):
        b, d = x.shape
        
        print(x.shape) # torch.Size([7, 32])

        if b != self.batchsize: # 2. I need batches to be of size 16, if batch isn't 16, I want to pad the rest to zero
            new_x = torch.zeros(self.batchsize,d) # 3. so I create a new tensor, but this is bad as it increase the GPU memory required greatly
            new_x[0:b,:] = x
            x = new_x
            b = self.batchsize
        
        print(x.shape) # torch.Size([16, 32])

        return x

model = MyModel()
x = torch.randn((7, 32)) # 1. shape's batch is 7, because this is last batch, and I dont want to "drop_last"
y = model(x)
print(y.shape)
4

1 回答 1

0

您可以像这样填充额外的元素：

import torch.nn.functional as F

n = self.batchsize - b

new_x = F.pad(x, (0,0,n,0)) # pad the start of 2d tensors
new_x = F.pad(x, (0,0,0,n)) # pad the end of 2d tensors
new_x = F.pad(x, (0,0,0,0,0,n)) # pad the end of 3d tensors
于 2021-03-29T08:32:56.453 回答