fix saliency map bugs

[BillyCheung10botics]

This PR fixed the following two bugs:

• tf2 version incomplatibility
  First, the codes constructing saliency models were incompatible with the tensorflow version 2.
  

  donkeycar/donkeycar/management/makemovie.py

  Lines 186 to 196 in 20bf83c

  	modifier_fn = get('guided')
  	sal_model_mod = modifier_fn(sal_model)
  	losses = [
  	(ActivationMaximization(sal_model_mod.layers[layer_idx], None), -1)
  	]
  	self.opt = Optimizer(sal_model_mod.input, losses, norm_grads=False)
  	return True
  	def compute_visualisation_mask(self, img):
  	grad_modifier = 'absolute'
  	grads = self.opt.minimize(seed_input=img, max_iter=1, grad_modifier=grad_modifier, verbose=False)[1]

  
  We should use the GradientTape() of tensorflow to calculate the gradient.

• uint8 conversion bug
  Second, the conversion of RGB values from float (0.0, 1.0) to int (0, 255) was wrong,
  

  donkeycar/donkeycar/management/makemovie.py

  Lines 239 to 240 in 20bf83c

  	image = image * 255
  	image = image.astype('uint8')

  
  which creates artifacts as the bottom left figure shown. The red/yellow/blue parts are artifacts with RGB values over 255. Since some float values of the image array can be great than 1.0. I use cv2.normalize to properly convert the RGB values from float to uint. The correct saliency map is represented as the yello-purple clould as the right figure shown.
  
  (left): video with artifacts (right): video with correct saliency mask

[DocGarbanzo]

This looks good, but I'm wondering if the Gaussian Blur is the source of having the salient mask showing much less localised impact than in the original algorithm. Could you make this switchable, so the user could choose between the 'old' version of overlay vs the new one?
It seems you have done this completely without using the keras-viz package, that is great. From what I recall that was using a single step of backprop to utilise a single gradient calculation, but it seems this can be done much better with using the keras gradient (or GradientTape in tf 2.0) directly between the input and output of the salient-adjusted model. Can you confirm?

[DocGarbanzo]

Great change, many thanks.