Fix setting torch's tensor default to float (was double)

[TimZaman]

Probably fixes #1134. Was due to an (unnesecary imo) breaking change in torch/image#189 .

[TimZaman]

cc @gheinrich . I'm now mostly checking what Travis thinks about this PR. It seems we're not using double tensors anywhere ever though.

[chsasank]

Hi,

The error was

digits.webapp: ERROR: Train Torch Model: ...e/travis/torch/install/share/lua/5.1/threads/threads.lua:183: [thread 3 callback] /home/travis/torch/install/share/lua/5.1/image/init.lua:966: bad argument #3 to 'warp' (torch.FloatTensor expected, got torch.DoubleTensor)

Very similar error(torch/image#187) was fixed by torch/image#189

You may take cues from this PR and fix wrap in torch/image

[TimZaman]

Hmm it is similar but no, torch/image#189 introduced it.

On Tuesday, 4 October 2016, Sasank Chilamkurthy [email protected]
wrote:

Hi,

The error was

digits.webapp: ERROR: Train Torch Model: ...e/travis/torch/install/share/lua/5.1/threads/threads.lua:183: [thread 3 callback] /home/travis/torch/install/share/lua/5.1/image/init.lua:966: bad argument #3 to 'warp' (torch.FloatTensor expected, got torch.DoubleTensor)

Very similar error(torch/image#187
torch/image#187) was fixed by torch/image#189
torch/image#189

You may take cues from this PR and fix wrap in torch/image

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#1139 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/AHXSRC58xLfdKJXBQzEv__h2mOKOuiRxks5qwg_DgaJpZM4KNX8x
.

[TimZaman]

Okay, tracked it down. The bug was introduced by torch/image#189 indeed. It tried to solve another bug but introduced another, and that affected us. The fix is a one-liner in torch, fixing it while supporting the previous and current torch/image versions in torch is more painful. Before and after the warp() statement we could switch between
torch.setdefaulttensortype('torch.FloatTensor')
and then back to
torch.setdefaulttensortype('torch.DoubleTensor')
afterwards. But I think we should be able to set the torch.FloatTensor as default globally in any case.

edit
actually we cannot do the above, because we have multiple data loaders there'd be race conditions. That means I need to convert the image to double and then back to float (until the torch PR is merged)

[lukeyeager]

Since the bug was fixed in torch/image#196, let's push an update to torch/distro instead of putting in a temporary fix.

@TimZaman agreed?

[TimZaman]

Aye captain