dswah / pygam Goto Github PK

class will read better if all properties that are specific to glms are self.glm_X_ not self.X_glm_

• method to fit multiple times on a logspace grid for lam
• should also know to choose the correct score vis a vis UBRE/GCV
• should be allowed to specify search across each feature's lam if desired

the edges are wrong

want quick way of checking model.

right now its buried in the penalty matrix method.
it should be somewhere else.

this is a thing that people do.

kind of already do it...

X_train = np.array([1,2,3,4])
y_train = np.array([2,3,4,5])
gam = LogisticGAM()
gam.fit(X_train, y_train)

I run the above codes and raised an error"AssertionError: y data is not in domain of link function"
I don't know how to use your API.

this will make it easy to make custom GAMs!

you should be able to penalize each feature differently, duh.

for example, logistic GAM shouldnt be allowed to fit on y in whole real line

would like to be able to see these per feature function

however, would like models to stay in main module

maybe even QR decomposition ?

add these puppies.

needed for generating knots, basis functions, penalty matrices

I want to make predictions of time series. Can this project process time series data ?

• lam variable is float and should be limited in length (ie 5 sig fig)
• wrap text around similar to sklearn

estimated scale should go here otherwise no one will know where to find it.

create a few standard penalty matrices to chose from:

• 2nd order smoothing (TOP)
• circular smoothing (periodic data) (TOP)
• harmonic smoothing (i think this is just p>=1) (TOP)
• monotone smoothing (TOP)
• non-negative impulse response
• varying penalties
• etc

it'll be more clear this way

need to find optimal lambda vector somehow :)

Hello,
I was hoping to try out your implementation of GAM. However I am not sure how I can do that.
Thanks,
Vinod

Using piecewise constant splines, no difference penalty

to document, prioritized:

• each argument
• each user-facing method
• each dev-facing method

see how jacques implemented his versioning in
https://github.com/jwkvam/bowtie/blob/master/bowtie/__init__.py

eliminate corresponding columns and rows from U, D, Vt

lol.
we dont need to compute the whole A (influence, hat) matrix for the edof estimation.
try doing just the diagonal.

pg 177

H is any positive semidefinite
matrix, which may be zero, but may also be used to allow lower bounds to
be imposed on the smoothing parameters, or to regularize an ill-conditioned problem.
For example, if it is required that λ1 ≥ 0.1 and λ2 ≥ 10, then we could set H =
0.1S1 + 10S2.

should not be allowed to fit n basis functions with less than n data-points.

woohoo!

ensure that AIC is computed right:

no constants missing from log likelihood, and deviance is defined correctly.
this is importnat to be able to compare models when data has different scales.

• doing full computation on small problems gives exact answer.
• subsampling approach gives same answer
• text suggests using rotation matrices from SVD = U, D, Vt, doesnt give right answer. Multiplying U. Ut gives identity (unsurprisingly), and trace is always the same...

there are some threads where people specifically request GAMs.

• scikit-learn/scikit-learn#3482
• http://stats.stackexchange.com/questions/199077/generalized-additive-model-python-libraries

for the 2nd order difference penalty to make sense, the knots have to be evenly spaced.
easy.

one should be able to specify multiple penalties per feature, and each penalty should get its own lambda

any distribution might have a known scale, and the API should be abstract enough to allow this.

its almost free.

pg 164 in Wood introduces this concept. elaborated in appendix A.5

maybe someone wants to use their own penalty