An output with one or more elements was resized since it had shape [2], which does not match the required output shape [1, 2]. This behavior is deprecated, and in a future PyTorch release outputs will not be resized unless they have zero elements. You can explicitly reuse an out tensor t by resizing it, inplace, to zero elements with t.resize_(0). (Triggered internally at ../aten/src/ATen/native/Resize.cpp:26.) return torch.stack(batch, 0, out=out)