Large values for saliency of uninitialized Inception_v3

[AsimJalwana]

Hi,

There is something strange that happens when one used the torchvision inception_v3 model (pretrained=False), the resultant saliency map has very large values (order of 10^8).

Saliency Algorithm: Guided Back Propagation (from torchray.attribution.guided_backprop import guided_backprop)

Such randomization of weights helps to perform sanity checks.

Regards,
Asim