RuntimeError: invalid argument 2: size 'xx' is invalid

1

Hello,
I encounter an issue with RuntimeError but could not solve it. My code structure looks like

1 for data in data_loader:
2     # data has shape (batch, num, channel, height, width)
3     b, n, c, h, w = data.size()
4     data = data.view(b*n, c, h, w)
5     out = model(data) # out has shape (b*n, feat_dim)

I set batch size to 3, and drop_last=False for data_loader. num is the number of frames, which is 10 for example, so (num, channel, height, width) is a sequence of frames in the same video. Each time the initial data has shape (3, 10, 3, 224, 224) and I reshape it into (30, 3, 224, 224) by line 4 so that it can be processed by a 2d CNN. However, the last batch has the number 2, so data becomes (2, 10, 3, 224, 224). When reshaping the last batch by line 4, the systems returns an error

RuntimeError: invalid argument 2: size '[20 x 3 x 224 x 224]' is invalid for input with 2949120 elements at /pytorch/torch/lib/TH/THStorage.c:41

Any idea about what is wrong and how to fix this issue?

Thanks!

2

I tested this with sample input of size (4, 5, 3, 224, 224) and it does not error out. The error you’re getting is a little bizarre; it implies that the number of elements in the tensor is less than the new view shape.

What exactly is data.size()?

3

The exact data.size() is (3, 15, 3, 256, 128), and the last batch is (2, 15, 3, 256, 128). The specific code is

arr = []
for data in data_loader:
    # data has shape (batch, num, channel, height, width)
    data = Variable(data.cuda())
    b, n, c, h, w = data.size()
    data = data.view(b*n, c, h, w)
    out = model(data) # out has shape (b*n, feat_dim)
    out = out.data.cpu()
    out = out.view(b, n, -1)
    out = out.mean(1) # so shape is (b, feat_dim)
    arr.append(out)
arr = torch.cat(arr, 0)

The origin is here.

4

This issue has been solved. It was a typo mistake.:grinning:

1 Like
5

Where was the typo? I got a similar error and I can’t find nothing

6

The typo was in my code, not here. It was like

b, n, c, h, w = data1.size()
data1 = data1.view(b*n, c, h, w)
.
.
.
data2 = data2.view(b*n, c, h, w) # forgot to get size of data2 but used size of data1

lol

1 Like
7

your post helped me cause I had a typo mistake and got the same result too, lol