I currently have this idiom:
auto ious = at::zeros({batches, num_anchor, height, width, max_gt}, xy.options());
Is it the correct way to basically build a tensor that has same options as xy
but different shape?
Another example working in Pytorch 1.4:
at::Tensor mask =
at::empty({dets_num * col_blocks}, dets.options().dtype(at::kLong));
One reason ask this is because in newest Pytorch I get this error:
ImportError: /opt/python/lib/python3.7/site-packages/mtorch/smtl_cpu.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZN2at6tensorEN3c108ArrayRefIiEERKNS0_13TensorOptionsE