You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
@jamesfishbaugh please comment and let us know what has been done:
Speed up measures for shape regression: So far, the estimation procedure of subject-specific growth
trajectories is iterative, and it can take up to several minutes. Parameter tuning can be difficult and time consuming.
This becomes exacerbated when considering large population studies where each observation is highly
dimensional. Without tremendous resources and specialized hardware (i.e. GPUs), a population-based study may
take several days. Since the estimation involves an iterative solver, we will pursue two avenues for speedup: 1) Improving convergence, in which we will accelerate computations by doing fewer iterations, converging to the solution quicker. Although
we are currently using gradient descent, we plan to use other optimization schemes such as the Fast Iterative Shrinkage-
Thresholding Algorithm (FISTA) 59; and 2) Improving efficiency/speed of core computations, by enabling GPU support via the KeOps library60, which will do faster computations while being memory efficient for larger data.
The text was updated successfully, but these errors were encountered:
@jamesfishbaugh please comment and let us know what has been done:
Speed up measures for shape regression: So far, the estimation procedure of subject-specific growth
trajectories is iterative, and it can take up to several minutes. Parameter tuning can be difficult and time consuming.
This becomes exacerbated when considering large population studies where each observation is highly
dimensional. Without tremendous resources and specialized hardware (i.e. GPUs), a population-based study may
take several days. Since the estimation involves an iterative solver, we will pursue two avenues for speedup: 1) Improving convergence, in which we will accelerate computations by doing fewer iterations, converging to the solution quicker. Although
we are currently using gradient descent, we plan to use other optimization schemes such as the Fast Iterative Shrinkage-
Thresholding Algorithm (FISTA) 59; and 2) Improving efficiency/speed of core computations, by enabling GPU support via the KeOps library60, which will do faster computations while being memory efficient for larger data.
The text was updated successfully, but these errors were encountered: