add an option to get unnormalized gradient maps

[eyaler]

in saliency functions we have

utils.normalize(grads)[0]

i want an option to get the raw gradients for thresholding or comparing between filters

[keisen]

Hi, @eyaler . Thank you for your request.

To create the option you said is so easy, but I'm wondering if I should do it. Something doesn't feel right with that the saliency function return the raw gradients, because it is the Gradient, not the Saliency. I believe we should use vis.optimizer.Optimizer when we want to get raw gradients.

opt = Optimizer(input_tensor, losses, wrt_tensor=wrt_tensor, norm_grads=False)
grads = opt.minimize(seed_input=seed_input, max_iter=1, grad_modifier=grad_modifier, verbose=False)[1]

I'd like to get your thoughts on this.

[keisen]

But, if we think the gradient is just used as the saliency, it should do it. I'm little bit confusing :confused:

[eyaler]

also there is a potential problem with this line: normalized = (array - arr_min) / (arr_max - arr_min + K.epsilon()) since array values can be in the magnitude of epsilon... maybe we should use max(arr_max - arr_min, K.epsilon())

[eyaler]

and the thing is that normalizing the gradients will always highlight something. but for finding negative evidence the magnitude is important

[eyaler]

also: grads = grads / (np.max(grads) + K.epsilon()) in grad cam