Number of labels: 1000, Dataset size: 200000 images Final probability for 1000 labels is in the range of 0.3 to 0.34. I was expecting large variation in probabilities. Can someone tell me what I am doing wrong. I am following this tutorial
$\begingroup$
$\endgroup$
4
-
$\begingroup$ Could you elaborate on what you mean by score here? $\endgroup$– Nischal HpCommented Jan 3, 2018 at 6:09
-
$\begingroup$ By score I mean probability $\endgroup$– RavikrnCommented Jan 3, 2018 at 8:11
-
$\begingroup$ Probability of what? The highest probability for the corresponding class? The probability for one specific class? $\endgroup$– Jan van der VegtCommented Jan 3, 2018 at 8:44
-
$\begingroup$ Probability of each class $\endgroup$– RavikrnCommented Jan 3, 2018 at 11:59
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
In my experience, the example code for a low number of classes (<200) works well. When moving to more classes the imbalance data makes the network converge to the same numbers. You have imbalance data because now each output is a binary classifier by its own, this doesn't happen with softmax. The way to mitigate the problem is to use weighted_cross_entropy_with_logits and set pos_weight to a positive number > 1 (10 works). But I still don't get very good results.
