akko/diffvg
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Resolve crash in 'backward' when a background image with only 3 channels is passed to 'forward'

Browse Source 
The existing code adds the fourth channel to the background image directly inside 'forward'. However, this breaks back propagation because Torch's autograd framework records the shapes of all inputs to the 'forward' function and expects shapes passed to 'backward' to match. By adding a channel to the background image inside 'forward' and passing this to 'backward', there is an extra channel that autograd does not expect, and it crashes.

The resolution is to instead raise an exception with a useful error message for the end user that they need to add a channel of all ones to the background image.
This commit is contained in:
Dan Nissenbaum
2022-11-05 15:07:47 -04:00
committed by Akko
parent f5e2ad9c43
commit a5e26c4395
1 changed files with 1 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
pydiffvg/render_pytorch.py
View File

@@ -384,9 +384,7 @@ class RenderFunction(torch.autograd.Function):
if background_image is not None:
background_image = background_image.to(pydiffvg.get_device())
if background_image.shape[2] == 3:
background_image = torch.cat((\
background_image, torch.ones(background_image.shape[0], background_image.shape[1], 1,
device = background_image.device)), dim = 2)
raise NotImplementedError('Background image must have 4 channels, not 3. Add a fourth channel with all ones via torch.ones().')
background_image = background_image.contiguous()
assert(background_image.shape[0] == rendered_image.shape[0])
assert(background_image.shape[1] == rendered_image.shape[1])