The current interface precludes the application of the GPR module to cases where the training samples are sequences, trees, graphs etc. In fact, GPR should not care about whether its input are fixed-length feature vectors or not --- that should be handled by the kernel.

Expected usage scenario:

k = StructuredDataKernel()
gp = GaussianProcessRegressor( k )
X = [ object1, objects2, ... ]
Y = [ object3, objects4, ... ]
y = [ scalar_values... ]
gp.fit( X, y ) 
gp.predict( Y )

Error information (sklearn/utils/validation.py", line 448, in check_array):

ValueError: setting an array element with a sequence.

Expect outcome:
The learning workflow should proceed just fine as long as k returns valid inner product matrices between X and itself or Y.