BayesianGPLVMMiniBatch and infer_newX

Hello,

When I try to use infer_newX in my data I get this error:

------------------------------------------
ValueErrorTraceback (most recent call last)
<ipython-input-12-999e9683e141> in <module>
     38     for i in range(Y_test.shape[0]):
     39         Y_new = Y_test[i][np.newaxis,:]
---> 40         data_embedded, mi = m.infer_newX(Y_new, optimize=True)
     41         fitted_data = m.predict(data_embedded)
     42         Y_predicted = fitted_data[0] * Y_std + Y_mean

~/anaconda3/envs/gpyenv/lib/python3.6/site-packages/GPy/core/gp.py in infer_newX(self, Y_new, optimize)
    676         """
    677         from ..inference.latent_function_inference.inferenceX import infer_newX
--> 678         return infer_newX(self, Y_new, optimize=optimize)
    679 
    680     def log_predictive_density(self, x_test, y_test, Y_metadata=None):

~/anaconda3/envs/gpyenv/lib/python3.6/site-packages/GPy/inference/latent_function_inference/inferenceX.py in infer_newX(model, Y_new, optimize, init)
     20     :rtype: (GPy.core.parameterization.variational.VariationalPosterior, GPy.core.Model)
     21     """
---> 22     infr_m = InferenceX(model, Y_new, init=init)
     23 
     24     if optimize:

~/anaconda3/envs/gpyenv/lib/python3.6/site-packages/paramz/parameterized.py in __call__(self, *args, **kw)
     51         #import ipdb;ipdb.set_trace()
     52         initialize = kw.pop('initialize', True)
---> 53         self = super(ParametersChangedMeta, self).__call__(*args, **kw)
     54         #logger.debug("finished init")
     55         self._in_init_ = False

~/anaconda3/envs/gpyenv/lib/python3.6/site-packages/GPy/inference/latent_function_inference/inferenceX.py in __init__(self, model, Y, name, init)
     81         self.Y = Y
     82         self.X = self._init_X(model, Y, init=init)
---> 83         self.compute_dL()
     84 
     85         self.link_parameter(self.X)

~/anaconda3/envs/gpyenv/lib/python3.6/site-packages/GPy/inference/latent_function_inference/inferenceX.py in compute_dL(self)
    127                 self.dL_dpsi2 = beta/2.*(self.posterior.woodbury_inv[:,:,self.valid_dim] - tdot(wv)[:, :, None]).sum(-1)
    128             else:
--> 129                 self.dL_dpsi2 = beta/2.*(output_dim*self.posterior.woodbury_inv - tdot(wv))
    130             self.dL_dpsi1 = beta*np.dot(self.Y[:,self.valid_dim], wv.T)
    131             self.dL_dpsi0 = - beta/2.* np.ones(self.Y.shape[0])

ValueError: operands could not be broadcast together with shapes (31,31,817) (31,31) 

My training set has 817 columns and my trained model:

m = GPy.models.bayesian_gplvm_minibatch.BayesianGPLVMMiniBatch(training_input, 30, 
                                                                                 num_inducing=31,
                                                                                 missing_data=True, 
                                                                                 init = "PCA",
                                                                                 stochastic = False,
                                                                                 initialize = True)

My new data is a (1, 817) numpy array.

Regards,
Lerko

Start with GPy/inference/latent_function_inference/inferenceX.py, especially compute_dL at the reported broadcasting failure, and trace the call from GPy/core/gp.py infer_newX. Reproduce it with BayesianGPLVMMiniBatch, 31 inducing points, 817 features, and a (1, 817) Y_new array. Done means infer_newX handles this input without the shape error and has a regression test.