Hello!
Thank you so much for the great work you've done, it is really impressive. I'm trying to get to know it in detail, modify some parts of the model to possibly improve it and am going through the code at the moment.
There's a part in the sampling of positions along the conditional Gaussian probability path in the Euclidean flow matcher (R3FM), which I don't get. Precisely, here:
|
# This seems like it should be right but its not |
|
x_t = x_t - x_t.mean(-2, keepdim=True) |
So what do you mean by your comment that "This seems like it should be right but its not" and why do you subtract that mean from the positions? I assume you center them to have zero mean, but I can't really find that step mentioned anywhere in your paper on Minibatch OT. Is it done for stability of the training?
Hello!
Thank you so much for the great work you've done, it is really impressive. I'm trying to get to know it in detail, modify some parts of the model to possibly improve it and am going through the code at the moment.
There's a part in the sampling of positions along the conditional Gaussian probability path in the Euclidean flow matcher (R3FM), which I don't get. Precisely, here:
FoldFlow/foldflow/models/r3_fm.py
Lines 115 to 116 in 9d2c260
So what do you mean by your comment that "This seems like it should be right but its not" and why do you subtract that mean from the positions? I assume you center them to have zero mean, but I can't really find that step mentioned anywhere in your paper on Minibatch OT. Is it done for stability of the training?