Why is the probability of output always around 0.5？

[kaiyangzhou]

Why is the probability of output always around 0.5？

[pandababyer]

Hello mate, have you solve the problem?

[linjingxu]

@pandababyer Not yet

[pandababyer]

@linjingxu I think the probs is the predicted importance scores for each frame, and the author want to select the key-shot which is the key segment. so in the function generate_summary, the author convert the key-shot select problem to a knapsack problem and find the key segment in 'picks'. I think that may explain why the score is around 0.5 rather than approach 0 or 1. Looking forward to your reply :)

[Pager07]

@pandababyer I have experimented with a couple of models that do the regression with sigmoid to output the frame importance score. Because the frame score is so sparse (at least this is true for summe dataset) the model learns to predict all 0, hence, the loss is pretty low. Very little gradient flow backwards. These regressive model tend to not have that much learning, according to my experience.