Conversation
Interfacing still needs cleaning up
Interfacing still needs cleaning up
Interfacing still needs cleaning up
Codecov Report
@@ Coverage Diff @@
## master #65 +/- ##
=========================================
+ Coverage 99.78% 99.8% +0.01%
=========================================
Files 16 18 +2
Lines 928 1004 +76
=========================================
+ Hits 926 1002 +76
Misses 2 2
Continue to review full report at Codecov.
|
…ue to multiple wrap operations
| from . import objective | ||
| from . import models | ||
| from . import pareto | ||
| from . import models |
| N = tf.shape(fmean)[0] | ||
| D = tf.shape(fmean)[1] | ||
|
|
||
| fmeanf = tf.reshape(fmean, [N * D, 1]) # N*D x 1 |
There was a problem hiding this comment.
No, has to be N*D x 1 so I can use rowwise_gradients, then I reshape later
| @AutoFlow((float_type, [None, None])) | ||
| def predict_f(self, Xnew): | ||
| """ | ||
| Compute the mean and variance of the latent function(s) at the points |
There was a problem hiding this comment.
update doc string? marginalised around ...
| Compute the mean and variance of the latent function(s) at the points | ||
| Xnew. | ||
| """ | ||
| theta = self._predict_f_AF_storage['free_vars'] |
There was a problem hiding this comment.
No other way, have to wait for GPflow issue
|
I skipped the ModelWrapper class in my review as it is a separate PR #64 |
|
I'd like to comment that this isnt to be merged until GPflow/GPflow#480 is addressed, or we hack some way around it. |
|
@nknudde what's the state on this? Is it more compatible with GPflow 1.0? Or is using MCMC to marginalize hyperparameters fast and accurate enough? |
4 participants
In-house version of #60 . Full credits to @nknudde for implementing this.
I already took care of merging #64 and derived the MGP from it. Also introduced a tf_wraps file for the rowwise_gradients. To be merged after releasing.