def generate(self, orig_image, input_image, target_class):
"""
Generates and returns multiple saliency maps, based on guided backpropagation.
Args:
orig_image: Original resized image.
input_image: Preprocessed input image.
target_class: Expected category.
Returns:
Colored and grayscale gradients for the guided backpropagation.
Positive and negative saliency maps.
Grayscale gradients multiplied with the image itself.
"""
guided_grads = self.generate_gradients(input_image, target_class)
color_guided_grads = normalize_gradient(guided_grads)
grayscale_guided_grads = normalize_gradient(convert_to_grayscale(guided_grads))
pos_sal, neg_sal = get_positive_negative_saliency(guided_grads)
pos_sal_grads = normalize_gradient(pos_sal)
neg_sal_grads = normalize_gradient(neg_sal)
grad_times_image = guided_grads[0] * input_image.detach().numpy()[0]
grad_times_image = convert_to_grayscale(grad_times_image)
grad_times_image = normalize_gradient(grad_times_image)
return {
"guided_grads_colored": color_guided_grads,
"guided_grads_grayscale": grayscale_guided_grads,
"guided_grads_grayscale_grad_times_image": grad_times_image,
"saliency_maps_positive": pos_sal_grads,
"saliency_maps_negative": neg_sal_grads,
}
def generate(self, orig_image, input_image, target_class):
"""
Generates and returns multiple saliency maps, based on vanilla backpropagation.
Args:
orig_image: Original resized image.
input_image: Preprocessed input image.
target_class: Expected category.
Returns:
Colored and grayscale gradients for the vanilla backpropagation.
Grayscale gradients multiplied with the image itself.
"""
vanilla_grads = self.generate_gradients(input_image, target_class)
color_vanilla_bp = normalize_gradient(vanilla_grads)
grayscale_vanilla_bp = convert_to_grayscale(vanilla_grads)
grayscale_vanilla_bp = normalize_gradient(grayscale_vanilla_bp)
grad_times_image = vanilla_grads[0] * input_image.detach().numpy()[0]
grad_times_image = convert_to_grayscale(grad_times_image)
grad_times_image = normalize_gradient(grad_times_image)
return {
"vanilla_colored_backpropagation": color_vanilla_bp,
"vanilla_grayscale_backpropagation": grayscale_vanilla_bp,
"vanilla_grayscale_grad_times_image": grad_times_image,
}
def generate(self, orig_image, input_image, target_class):
"""
Generates heatmaps using the integrated gradient method.
Args:
orig_image: The original image.
input_image: Preprocessed input image.
target_class: Expected category.
Returns:
The heatmaps.
"""
integrated_grads = self.generate_integrated_gradients(
input_image, target_class, 5)
grayscale_integrated_grads = normalize_gradient(
convert_to_grayscale(integrated_grads))
grad_times_image = integrated_grads[0] * input_image.detach().numpy(
)[0]
grad_times_image = convert_to_grayscale(grad_times_image)
grad_times_image = normalize_gradient(grad_times_image)
return {
"integrated_gradients": grayscale_integrated_grads,
"integrated_gradients_times_image": grad_times_image,
}
def generate(self, orig_image, input_image, target_class):
"""
Guided gradcam is just pointwise multiplication of the cam mask and
the guided backprop mask.
Args:
orig_image: Original resized image.
input_image: Preprocessed input image.
target_class: Expected category.
Returns:
Colored and grayscale gradients for the guided Grad-CAM.
"""
cam = self.gradcam.generate_cam(input_image, target_class)
guided_grads = self.gbp.generate_gradients(input_image, target_class)
cam_gb = np.multiply(cam, guided_grads)
guided_gradcam = normalize_gradient(cam_gb)
grayscale_cam_gb = convert_to_grayscale(cam_gb)
guided_gradcam_grayscale = normalize_gradient(grayscale_cam_gb)
return {
"guided_gradcam": guided_gradcam,
"guided_gradcam_grayscale": guided_gradcam_grayscale,
}
def generate(self, orig_image, input_image, target_class):
"""
Generates and returns multiple sensitivy heatmaps, based on SmoothGrad technique.
Args:
orig_image: Original resized image.
input_image: Preprocessed input image.
target_class: Expected category.
Returns:
Colored and grayscale gradients for the SmoothGrad backpropagation.
"""
param_n = 5
param_sigma_multiplier = 4
smooth_grad = self.generate_smooth_grad(
input_image, target_class, param_n, param_sigma_multiplier
)
color_smooth_grad = normalize_gradient(smooth_grad)
grayscale_smooth_grad = convert_to_grayscale(smooth_grad)
grayscale_smooth_grad = normalize_gradient(grayscale_smooth_grad)
return {
"smooth_grad_colored": color_smooth_grad,
"smooth_grad_grayscale": grayscale_smooth_grad,
}