Captum: Relu missing in implementation of LayerGradCam.attribute()

[vipinpillai]

I was looking at the code in grad_cam.py for LayerGradCam.attribute() and noticed that there is no Relu operation applied after computing the summed_grads * layer_eval here: https://github.com/pytorch/captum/blob/master/captum/attr/_core/layer/grad_cam.py#L177

[NarineK]

@vipinpillai, we did it on purpose in order to be more flexible and show negative attribution as well. See doc: https://github.com/pytorch/captum/blob/master/captum/attr/_core/layer/grad_cam.py#L47
You need to apply relu externally if you need to.

[vipinpillai]

Thanks @NarineK. I got confused because of the return variable being named _non_neg_scaled_act_.

I think it might be confusing for anyone expecting a GradCAM visualization from this API without explicitly applying relu. Wouldn't it be better to have an argument specifying whether the caller needs a relu / non-relu version of the attribution?

[NarineK]

yeah, we could add the flag. cc: @vivekmig

[vivekmig]

Hi @vipinpillai, thanks for the feedback, adding the argument here #181.

[vipinpillai]

Thanks @vivekmig

[NarineK]

Closing this since it got fixed by: #181