Fixed tensor type issue in affinetransform and warp

[chsasank]

Refer to #187.

Previously the following code would've given an error:

require 'torch'
require 'init'
torch.setdefaulttensortype('torch.FloatTensor')

local im = image.lena()
local matrix = torch.Tensor({{1,0.5},{0,1}}):double()
local im2 = image.affinetransform(im,matrix,'bilinear')

local fld = torch.zeros(2, 512, 512):double()
local im2 = image.warp(im,fld,'bilinear')

This seems to be because src.image.warp expects field to be of default tensor type. This issue is fixed by coercing passed matrix or field to default tensor type.

[soumith]

Thanks!

[TimZaman]

If the default tensor type is double, but the image is a float, this PR will make it fail.
/Users/tzaman/torch/install/bin/luajit: /Users/tzaman/torch/install/share/lua/5.1/image/init.lua:966: bad argument #3 to 'warp' (torch.FloatTensor expected, got torch.DoubleTensor)
Reproduce:

require 'torch'
require 'image'
torch.setdefaulttensortype('torch.DoubleTensor')
pr = image.lena():float()
makes_me = torch.zeros(2, 512, 512):double()
sad = image.warp(pr,makes_me,'bilinear')

[chsasank]

Hi @TimZaman ,
I think you are mistaken. This PR didn't modify anything related to wrap. It modified affinetransform. In fact it fixed the exact bug you've faced with warp except for affinetransform.

[TimZaman]

@chsasank , how about line 955?
local field = field:typeAs(torch.Tensor()) -- coerce matrix to default tensor type
Should have been
local field = field:typeAs(src) -- coerce matrix to src tensor type

[chsasank]

That doesn't seem to work if src is not default tensor type.

This seems to be because src.image.warp expects field to be of default tensor type.

[TimZaman]

Do you agree that the above lua snippet a few comments back does not
currently work, and that this PR changes code (one line L955) for warp as
well?

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

That doesn't seem to work if src is not default tensor type.

This seems to be because src.image.warp expects field to be of default
tensor type.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#189 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/AHXSRESpyDgxzwprV_8xd9fKZsyViwPfks5qwjDKgaJpZM4J7wfR
.

[chsasank]

Yeah, that does seem to be the case. The PR seems to be a bad attempt to fix bug #187 . I'm still figuring out where exactly the error is raised. I'd appreciate any help.

One problem with local field = field:typeAs(src) is that src can be a CharTensor. That'll round field to nearest integer and therefore producing unexpected results.

[TimZaman]

It makes sense to introduce some code testing, also with the affineTransform. The warp function's argument-3 (field) has to be the same as the image's.
I will suggest a PR to fix this. If you want to use a byte or char-tensor, you can manually char:float():warp():char() it back to its original state. This wasn't possible before this PR either, and has never seemed to raise any issues.
I'll make a new PR and add a minor test too so that we know what we're going in the right direction.

[chsasank]

The warp function's argument-3 (field) has to be the same as the image's.

It'd be nice if you can point where this check is enforced. I have gone through generic/image.c but couldn't find it.