pytorch batch size

系统管理员 2022-10-19 12:44 164阅读 0赞

当数据集的图片大小不一致时，设置batchsize>1，报如下错误：

epoch 0, processed 0 samples, lr 0.0000001000
    Traceback (most recent call last):
      File "train.py", line 347, in <module>
        main()        
      File "train.py", line 138, in main
        train(train_list, model, criterion, optimizer, epoch,f)
      File "train.py", line 199, in train
        for i,(img, target)in enumerate(train_loader):
      File "/home/wanglin/anaconda2/lib/python2.7/site-packages/torch/utils/data/dataloader.py", line 615, in __next__
        batch = self.collate_fn([self.dataset[i] for i in indices])
      File "/home/wanglin/anaconda2/lib/python2.7/site-packages/torch/utils/data/dataloader.py", line 232, in default_collate
        return [default_collate(samples) for samples in transposed]
      File "/home/wanglin/anaconda2/lib/python2.7/site-packages/torch/utils/data/dataloader.py", line 209, in default_collate
        return torch.stack(batch, 0, out=out)
    RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 401 and 440 in dimension 2 at /opt/conda/conda-bld/pytorch_1544080996887/work/aten/src/TH/generic/THTensorMoreMath.cpp:1333

本来以为是anaconda2/lib/python2.7/site-packages/torch/utils/data/sampler.py中的BatchSampler类的batch.append(idx)出错，最后发现是在读取数据时，程序会调用anaconda2/lib/python2.7/site-packages/torch/utils/data/dataloader.py 的default\_collate(batch)函数，当执行到torch.stack(batch, 0, out=out)会报错。原因是，torch.stack函数（以及torch.cat()函数）要求输入是相同size的。验证如下：

>>> a=torch.rand((2,3))
    >>> b=torch.rand((2,3))
    >>> a
    tensor([[0.1102, 0.0474, 0.6739],
            [0.1565, 0.1700, 0.4528]])
    >>> b
    tensor([[0.3654, 0.1404, 0.6968],
            [0.0738, 0.0108, 0.4125]])
    >>> c=torch.cat((a,b))
    >>> c
    tensor([[0.1102, 0.0474, 0.6739],
            [0.1565, 0.1700, 0.4528],
            [0.3654, 0.1404, 0.6968],
            [0.0738, 0.0108, 0.4125]])
    >>> c.size()
    (4, 3)
    >>> d=torch.stack((a,b))
    >>> d
    tensor([[[0.1102, 0.0474, 0.6739],
             [0.1565, 0.1700, 0.4528]],
    
            [[0.3654, 0.1404, 0.6968],
             [0.0738, 0.0108, 0.4125]]])
    >>> d.size()
    (2, 2, 3)
    >>> e=torch.rand((2,3))
    >>> f=torch.stack((d,e))
    Traceback (most recent call last):
      File "<stdin>", line 1, in <module>
    RuntimeError: invalid argument 0: Tensors must have same number of dimensions: got 4 and 3 at /opt/conda/conda-bld/pytorch_1544080996887/work/aten/src/TH/generic/THTensorMoreMath.cpp:1324
    >>> f=torch.cat((d,e))
    Traceback (most recent call last):
      File "<stdin>", line 1, in <module>
    RuntimeError: invalid argument 0: Tensors must have same number of dimensions: got 3 and 2 at /opt/conda/conda-bld/pytorch_1544080996887/work/aten/src/TH/generic/THTensorMoreMath.cpp:1324