[REQUEST]

[surajpaib]

Problem Faced:
Backprop visualize doesn't consider models with grayscale input.

Description:
Refer to this line:

flashtorch/flashtorch/saliency/backprop.py

Line 222 in ba5e3db

module.in_channels == 3:

In this line we add a check for 3 channels which results in skipping grayscale inputs to the conv layer

My current workaround:
I have a copy of the backprop class with the in_channels set to 1 for my grayscale image use case.

Possible Fix:
Would it make sense to remove the in_channels == 3 check since the first conv would be where the input image is supplied and should be good to register the hook at or am I missing something important here?

[MisaOgura]

Hi @surajpaib, thanks for the request!

Accommodating for models that handle greyscale input images definitely makes sense in terms of improving the versatility of the tool.

I have opened a PR for this (#31) and will release a new version after the merge.

[MisaOgura]

Hi @surajpaib,

The latest version 0.1.3 implements this improvement.

It would be great if you could use this version and see if it provides the functionality you require.

Many thanks!

[surajpaib]

@MisaOgura Perfect! Will test it over the weekend and report my findings, if any! Thanks