Acceleration for calculating Stability Loss and Ground Loss

[yufu-liu]

Hi, sorry for bothering you again.

Because Stability Loss and Ground Loss use 20000 vertices to do a series of calculations, it makes the training time much longer than the cases without using them, especially if hardware resources are limited.

Could you give me some advices to accelerate the calculation of the losses?
(Ex. vectorization of a function, parallelism techniques..etc.)

[sha2nkt]

Hi,

You could train a new regressor with fewer vertices. We chose 20k arbitrarily so that is something to experiment with. The biggest advantage of the regressor is its ability to sample points uniformly. How many points you sample depends entirely on the application.

For advice on vectorization or parallelism, I would have to look into the code a bit. If you find some easy improvements, feel free to start a pull request. I will appreciate your contribution.

Best,
Shashank

[yufu-liu]

Hi, thanks for your answering!

Decreasing the number of the vertices might work if its CoM is still correct enough. Could you explain more about training a new regressor (HDfier)?
I have checked the code, and think that the calculations of volume could be vectorized, like calculating 10 parts of volume parallelly. However, the number of vertices in each part are different, so I cannot directly apply APIs (Ex. np.vectorize or tf.vectorized_map ).

Best regards,
YuFu