@@ -173,7 +173,8 @@ def load(self, file_prefix):
173173 self .num_training_samples = params ['num_training_samples' ]
174174
175175
176- class StudentsTProcessRegressor (GaussianProcessRegressor ):
176+ class StudentsTProcessRegressor (BayesianModel ,
177+ GaussianProcessRegressorMixin ):
177178 """ StudentsT Process Regression built using PyMC3.
178179
179180 Fit a StudentsT process model and estimate model parameters using
@@ -204,8 +205,15 @@ class StudentsTProcessRegressor(GaussianProcessRegressor):
204205 Rasmussen and Williams (2006). Gaussian Processes for Machine Learning.
205206 """
206207
207- def __init__ (self , prior_mean = 0.0 ):
208- super (StudentsTProcessRegressor , self ).__init__ (prior_mean = prior_mean )
208+ def __init__ (self , prior_mean = None , kernel = None ):
209+ self .ppc = None
210+ self .gp = None
211+ self .num_training_samples = None
212+ self .num_pred = None
213+ self .prior_mean = prior_mean
214+ self .kernel = kernel
215+
216+ super (StudentsTProcessRegressor , self ).__init__ ()
209217
210218 def create_model (self ):
211219 """ Creates and returns the PyMC3 model.
@@ -241,13 +249,17 @@ def create_model(self):
241249 degrees_of_freedom = pm .Gamma ('degrees_of_freedom' , alpha = 2 ,
242250 beta = 0.1 , shape = 1 )
243251
244- # cov_function = signal_variance**2 * pm.gp.cov.ExpQuad(
245- # 1, length_scale)
246- cov_function = signal_variance ** 2 * pm .gp .cov .Matern52 (
247- 1 , length_scale )
252+ if self .kernel is None :
253+ cov_function = signal_variance ** 2 * RBF (
254+ input_dim = self .num_pred ,
255+ ls = length_scale )
256+ else :
257+ cov_function = self .kernel
248258
249- # mean_function = pm.gp.mean.Zero()
250- mean_function = pm .gp .mean .Constant (self .prior_mean )
259+ if self .prior_mean is None :
260+ mean_function = pm .gp .mean .Zero ()
261+ else :
262+ mean_function = self .prior_mean
251263
252264 self .gp = pm .gp .Latent (mean_func = mean_function ,
253265 cov_func = cov_function )
@@ -277,7 +289,8 @@ def load(self, file_prefix):
277289 self .num_training_samples = params ['num_training_samples' ]
278290
279291
280- class SparseGaussianProcessRegressor (GaussianProcessRegressor ):
292+ class SparseGaussianProcessRegressor (BayesianModel ,
293+ GaussianProcessRegressorMixin ):
281294 """ Sparse Gaussian Process Regression built using PyMC3.
282295
283296 Fit a Sparse Gaussian process model and estimate model parameters using
@@ -308,9 +321,15 @@ class SparseGaussianProcessRegressor(GaussianProcessRegressor):
308321 Rasmussen and Williams (2006). Gaussian Processes for Machine Learning.
309322 """
310323
311- def __init__ (self , prior_mean = 0.0 ):
312- super (SparseGaussianProcessRegressor , self ).__init__ (
313- prior_mean = prior_mean )
324+ def __init__ (self , prior_mean = None , kernel = None ):
325+ self .ppc = None
326+ self .gp = None
327+ self .num_training_samples = None
328+ self .num_pred = None
329+ self .prior_mean = prior_mean
330+ self .kernel = kernel
331+
332+ super (SparseGaussianProcessRegressor , self ).__init__ ()
314333
315334 def create_model (self ):
316335 """ Creates and returns the PyMC3 model.
@@ -344,13 +363,17 @@ def create_model(self):
344363 noise_variance = pm .HalfCauchy ('noise_variance' , beta = 5 ,
345364 shape = 1 )
346365
347- # cov_function = signal_variance**2 * pm.gp.cov.ExpQuad(
348- # 1, length_scale)
349- cov_function = signal_variance ** 2 * pm .gp .cov .Matern52 (
350- 1 , length_scale )
366+ if self .kernel is None :
367+ cov_function = signal_variance ** 2 * RBF (
368+ input_dim = self .num_pred ,
369+ ls = length_scale )
370+ else :
371+ cov_function = self .kernel
351372
352- # mean_function = pm.gp.mean.Zero()
353- mean_function = pm .gp .mean .Constant (self .prior_mean )
373+ if self .prior_mean is None :
374+ mean_function = pm .gp .mean .Zero ()
375+ else :
376+ mean_function = self .prior_mean
354377
355378 self .gp = pm .gp .MarginalSparse (mean_func = mean_function ,
356379 cov_func = cov_function ,
0 commit comments