Getting pred. probabilities

[georgefairs]

Hi there,

Is there an obvious way to to extract the prediction probabilities instead of just no. correct labels vs no. ground truth labels? I have checked the print_acc.py and eval_prob_adaptive.py but I can't see anything obvious. Any help would be super appreciated.

Thanks.

[njuaplusplus]

I am also interested in how to get the prediction probability (ie, prediction confidence). From the current eval_prob_adaptive function, it seems impossible to get the probability.

I tried to compute the softmax over the data, the probability is close to 1/n for each label among n classes in my cases.