How to output aleatoric and epistemic uncertainties associated to the class
BrenoAV opened this issue · 7 comments
Greetings,
I was trying to print out the uncertainties just the way it is shown in Figure 3 from the paper:
Where should I tweak the code, so that I can output those 4 uncertainty values for each image?
In /layers/functions/detection_gmm.py there's one output variable with should contain uncertainties, but I wasn't able to understand the output. Is there anything I'm missing/misunderstanding?
Hello @BrenoAV
As you can see this line, output contains scores, box coordinates, aleatoric & epistemic uncertainties of each task.
You can print out each uncertainty value using detections variable, like scores and box coordinates.
To output normalized values, you need to know the mean and variance of each uncertainty distribution for the dataset.
For doing this, you can use this function.
It makes sense now for me, thank you very much @jwchoi384
hello please if you have a concrete example how to recover them for each image?
@jwchoi384 @BrenoAV
What I do understand, please correct me if I'm wrong.
You can take these values from here layers/functions/detection_gmm.py:
For example, in the demo.py file... When you test some image you can use the while loop to obtain the uncertainty for each bbox, something like this:
max_loc_al_uc = torch.max(detections[0][i][j][5:9]) # 4 values (four coordinates values)
max_loc_ep_uc = torch.max(detections[0][i][j][9:13]) # 4 values (four coordinates values)
cl_al_uc = torch.max(detections[0][i][j][13:14]) # 1 value
cl_ep_uc = torch.max(detections[0][i][j][14:15]) # 1 valueThank you @BrenoAV very much for your answer indeed the same thing that I understood but I have a small concern concerning the four values for the random and epistimological localisation why it sends 4 values instead of a single value as the case of the classification. and for the function can we use it if we want to select a set of images to label ?
For the localization, the 4 values are the uncertainty about the center, height, and weight {x, y, w, h}. On pg. 4 of the paper in section 3.3 Scoring Function: "[...] For localization,
And in the function you cited, you want to take the labeled_set because they have high score values, labeling them and continue the active learning loop.