Torch.cat error

vision
1

This snippet works fine.
aa = torch.FloatTensor(torch.randn(1,1,256))
bb = torch.FloatTensor(torch.randn(1,5332,256))
cc = torch.cat((aa,bb),1)
cc.shape
(1, 5333, 256)
type(cc)
<class ‘torch.Tensor’>
type(aa)
<class ‘torch.Tensor’>
type(bb)
<class 'torch.Tens

However this one throws error :

embeddings = self.embed(captions)
pdb.set_trace()
embeddings = torch.cat((features.unsqueeze(1), embeddings), 1)

Debugging the type and shapes:

(Pdb) embeddings.shape
torch.Size([1, 5332, 256])
(Pdb) features.shape
torch.Size([1, 256])
(Pdb) type(embeddings)
<class ‘torch.Tensor’>
(Pdb) type(features)
<class ‘torch.Tensor’>
(Pdb) embeddings = torch.cat((features.unsqueeze(1), embeddings), 1)
*** RuntimeError: cuda runtime error (59) : device-side assert triggered at /opt/conda/conda-bld/pytorch_1524580938250/work/aten/src/THC/THCCachingHostAllocator.cpp:257
(Pdb)

Can’t understand why? Is it because one of the inputs can be a variable ?

2

Have you solve this problem?
I face the same mistake like you.

3

I would recommend to run the code on the CPU first to get a better error message.
If the code only raises the error using the GPU, run the script via:

CUDA_LAUNCH_BLOCKING=1 python script.py args

and post the stack trace here, please.