Features at the bottleneck

[ArielleZhang]

Hi,

I noticed that in the code implementaion, there is a (232,) feature for each trajectory at each timestep. When scoring and classifying the difficulty of the trajectory we only have one score per trajectory, then in the contrastive loss step, how are anchors selected? Is it the average of all the features at each timestep of the same trajectory?

Thanks,

Arielle

[os1a]

Hi Arielle,

Sorry for the late response. Yes, the features for each trajectory has a dimension of 232. When applying the contrastive learning we multiply the features with its transpose (matrix multiplication) in line 24. The result will be a matrix of size (num_trajectories, num_trajectories), which are the anchors. Then we select the maximum one.

Best, Osama