[feature request] Implement torch.stack() / _cat() for also for torch.HalfTensor

[timothygebhard]

Issue Description

It seems that stacking is currently (PyTorch 0.4) is not supported for tensors of type HalfTensor; see code example.

I don't know if this is intentional for some reason (hence I didn't label this issue as a bug), but I couldn't find any issue or documentation pointing to that. Since some methods, like torch.utils.data.DataLoader, depend on torch.stack() (in this case implicitly through the default_collate() method), it would be great if either stacking for HalfTensors could get implemented, or the documentation could be updated to contain some kind of hint.

Code Example

Python 3.6.5 (default, Mar 30 2018, 06:41:49)
Type 'copyright', 'credits' or 'license' for more information
IPython 6.3.1 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import torch

In [2]: torch.__version__
Out[2]: '0.4.0'

In [3]: tensor_1_float = torch.tensor([1, 2, 3], dtype=torch.float)

In [4]: tensor_2_float = torch.tensor([4, 5, 6], dtype=torch.float)

In [5]: torch.stack([tensor_1_float, tensor_2_float])
Out[5]:
tensor([[ 1.,  2.,  3.],
        [ 4.,  5.,  6.]])

In [6]: tensor_1_half = torch.tensor([1, 2, 3], dtype=torch.half)

In [7]: tensor_2_half = torch.tensor([4, 5, 6], dtype=torch.half)

In [8]: torch.stack([tensor_1_half, tensor_2_half])
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-8-10f49aa85ea5> in <module>()
----> 1 torch.stack([tensor_1_half, tensor_2_half])

RuntimeError: _cat is not implemented for type torch.HalfTensor

On PyTorch 0.3.1, one gets basically the same error when one tries to do the same thing:

TypeError: Type torch.HalfTensor doesn't implement stateless method cat

[soumith]

we dont support any operation other than copy and serialization on torch.HalfTensor (the CPU version). The GPU HalfTensor supports all operations

[apaszke]

@soumith while we don’t want to have any math operations for half on CPU I feel like the argument given by @timothygebhard for having cat is very reasonable. Cat can be implemented efficiently because it doesn’t care how the data looks like, it’s only copying raw bytes around.

[ezyang]

If you do this, it would probably be a good idea to port THTensor_(cat) to ATen while you're at it.

[adam-dziedzic]

I got a very similar error: RuntimeError: _cat_out is not implemented for type torch.HalfTensor. To my mind, we could add support for this method for data preparation/loading on CPU so that a single tensor type (namely torch.half) is used on CPU and GPU.

[towr]

I ran into this issue with a DataLoader that works on a numpy.fp16 dataset.
Writing a custom collate_fn for the dataloader solves the immediate problem, but it would be nice if it works out of the box.

[TimZaman]

I'm confused, why did we close this issue? There was an MR for this that looked like it worked, but was subsequently closed, together with this issue/

[soumith]

as replied on the PR, it's merged into master, it's available in the nightlies, and will be part of 1.1