ExplainableML
Evaluation code
@udion @sgk98 Thank you. Could you provide the evaluation code, please?
Is the following code for calculating uncertainty correct?
def mean(tensor):
log_tensor = torch.log(tensor)
mean_log = torch.mean(log_tensor)
geometric_mean = torch.exp(mean_log)
return geometric_mean.item()
imdict=get_features_uncer_ProbVLM(CLIP_Net,ProbVLM_Net,cub_test_loader)
uncertainty=imdict['i_u']
uncertainty=[mean(i) for i in uncertainty]
The mean() function calculates the geometric mean.
@asdffasdfaf Have you solved the problem? In the function of get_features_uncer_ProbVLM, I do not find the specific definition for the function multi_fwpass_BayesCap
It is a function that computes the variance (or uncertainty) with the alpha, beta predictions from the network as mentioned in Eq. 9 in paper.
The aleatoric uncertainty can be computed by the formula of variance mentioned here: https://en.wikipedia.org/wiki/Generalized_normal_distribution
I think we forgot to update that function, both of us have graduated now, so can't update the code, but the above function should be straight forward to write.
