|
|
Deformetrica implements three estimation methods for minimizing the loss functions of the different models: a simple gradient ascent, the L-BFGS algorithm in scipy and a stochastic version of the EM algorithm.
|
|
|
Deformetrica implements three estimation methods for minimizing the loss functions of the different models: a simple **gradient ascent**, the **L-BFGS algorithm** from [SciPy.org](https://docs.scipy.org/doc/scipy/reference/optimize.minimize-lbfgsb.html), and a stochastic version of the EM algorithm (still unstable).
|
|
|
|
|
|
In general, we advise the use of the L-BFGS algorithm, which converges faster. There are particular situations where this method is not advised:
|
|
|
- When estimating a template mesh with borders. In this case, deformetrica will perform a convolution operation on the gradient of this shape to smooth its appearance. In a lot of cases, it will cause the L-BFGS method to fail because it is given an altered version of the true gradient.
|
|
|
- When the number of variables to estimate is really large (e.g. to estimate a template image or an atlas with a lot of observations), L-BFGS is greedy in memory. The **memory-length** optimization parameter can be decreased to circumvent this. But in extreme cases, the gradient ascent will use less memory.
|
|
|
In general, we advise the use of the `ScipyLBFGS` estimator, which converges faster. The `GradientAscent` estimator might prove more robust in some situations.
|
|
|
|
|
|
The stochastic version of the EM algorithm can only be used with the Bayesian and longitudinal atlas. Documentation of this algorithm is on our roadmap. |
|
|
\ No newline at end of file |
|
|
The `McmcSaem` estimator can only be used with the bayesian and longitudinal atlas. This algorithm is still under development, its full support with the bayesian atlas is scheduled for Deformetrica 4.1.0. |
|
|
\ No newline at end of file |