Open bhattg opened 4 years ago
Good catch. I wonder if this would fix the issue described in Figure 4b.
Hey, do you mean this in the caption of Figure 4 - "One trajectory (in the center) strongly deviates from typical trajectories seen during training, and the model struggles to predict the correct transition." ??
Yes, exactly.
I also wander why not apply transition_model to negative state
According to the paper, the negative component of the contrastive loss is the difference between the negative states (randomly sampled from embedding at timestamp t, (z{t}~)) and the ground truth state (z{t+1}).
However, as per the line 113 of modules.py, given no trans, you are effectively taking the difference between randomly sampled from embedding at timestamp t (z{t}~) and z{t} (rather than z_{t+1}).
` def contrastive_loss(self, obs, action, next_obs):
` Thus, I feel instead of the state as the first argument of the energy function, next_state should have been the argument. Please let me know if I am misconstruing at any point.
Thanks.