eclique / rise Goto Github PK

Since D-RISE is archived, I cannot post an issue on this repo, but I actually think it could happen to have the same issue with RISE, I apologize if it does not suit.
So, I have a bad model (with a very small recall, around 0.1): when the mask is applied, I get very few labels and thus the saliency map remains at 0 often. I guess I could decrease the probability of the masked superpixels but is it a good idea to decrease it too much ? I still have pretty bad results with 0.3. I also decreased the detect threshold of the model (I have no predicted labels for the masks over 0.2, and I start to have a correct percentage of null saliency maps at 0.08... but is it a good idea ?).
Anyways, thanks a lot for your work!

I believe both RISE and D-RISE (cvpr'21) are surprising works. When will you be releasing your D-RISE codes? What's your codebase for detectors? Thanks!

In def generate_masks(N, s, p1):

Couldn't find what does N,s,p1 stands for in the code. I assume N stands for the number of filters generated? Since variable names don't imply their purpose, Commenting on its purpose or what it stands for would be great.

Is any code available for evaluation using of tensorflow.

Hi ! My question is about the computation of the AuC in the Causal Metrics script.
I wonder why the computation is :
(arr.sum() - arr[0] / 2 - arr[-1] / 2) / (arr.shape[0] - 1)
Instead of the regular area under the curve:
arr.sum() / (arr.shape[0] - 1)
To me, the normalized version implemented substracts the mean of the first and last score and I don't understand why that is.
Could you clarify that?
Thank you.

Hi, could you give an explanation for the reason of substracting the halves of the head and tail elements in an array when calculating the AUC?

I tried setting a threshold to distinguish it, but the effect was not very good

Is any possible to see "random-masking/run02/explanations/exp_{:05}-{:05}.npy " while evaluating a batch of explanations?

@eclique thanks for open sourcing the code just having a query

1. can we use this method to perform operation on object detection architecture like yolo , retina net and semantic segmentation mdoels like mask rcnn yolact
   2.If so what is the modification to be made int he code

Hi there,

I am trying to implement RISE for greyscale images. However, in easystart.ipynb with def load_img (see below), the input size is converted from greyscale into RGB images.

def load_img(path):
img = image.load_img(path, target_size=model.input_size)

So, do you know if RISE is only usable with RGB images?

Best regards,
Daan

Hello,
The masks that have been generated are stored somewhere? Is it possible to to get access to the path where the masks stored? Also, what is 's' in the generate_mask function?

It is not clear to me what exactly the values on the colorbar indicate. It seems obvious that the higher values indicate more salient regions and vice versa. In the heatmap below:

It is not clear how the value of 0.0016 is to be interpreted for example

Hi,

I noticed that your code of "evaluation.py" does not include the normalization of images in the single_run or evaluate.
Shouldn't the evaluation of the insertion and deletion game includes the normalization of images?

Thank you in advance,
Hanwei