1035 commits

This branch

This branch

All branches

Author	SHA1	Message	Date
Alan Saul	57001851c4	Trying to debug kernel parameters learning (fails even when noise fixed) ... may be some instablility, seems like it can get it if it starts close	2013-07-09 20:05:03 +01:00
Alan Saul	ee980227ac	Fixed 2*variance plotting instead of 2*std plotting, tidied up	2013-07-09 11:35:42 +01:00
Alan Saul	2a366619b3	Changed incorrect naming	2013-07-08 16:09:20 +01:00
Alan Saul	4e5cefb4b5	Reparameratised in terms of sigma2	2013-07-08 15:48:53 +01:00
Alan Saul	ab6a3a571e	Playing trying to find what makes it want to go so low	2013-07-02 11:14:48 +01:00
Alan Saul	a7169ab1ab	Fixed bug where B wasn't refering to current f location	2013-07-01 15:21:47 +01:00
Alan Saul	26b3855af5	Everything seems to be gradchecking again	2013-07-01 10:06:20 +01:00
Alan Saul	c90b1f0c99	Added minimizer for finding f, doesn't help	2013-06-27 15:04:57 +01:00
Alan Saul	617d73ca32	Now checkgrads a lot more of the time, but still fails in optimisation, seems also odd that when parameter is fixed kernel parameters go to infinity	2013-06-26 15:44:26 +01:00
Alan Saul	064efd5535	Added another optimisation which doesn't use gradients. Seems like F is almost always found, but Y can be off, suggesting that Wi__Ki_W is wrong, maybe W?	2013-06-25 18:20:00 +01:00
Alan Saul	e80fad197c	trying to fix optimisation problem, fixed a few bugs but still fails at very low noise	2013-06-24 15:39:38 +01:00
Alan Saul	d4bfd99c21	Starting to fiddle with mode finding code	2013-06-20 14:30:25 +01:00
Alan Saul	e900509a7c	Fixed a sign wrong, now gradchecks weirdly only above certain points	2013-06-19 16:13:11 +01:00
Alan Saul	de689fa8e9	Now gradchecks everytime but student_t fit is bad, noise is underestimated by a long way	2013-06-19 12:00:00 +01:00
Alan Saul	ac461e1b2a	Checkgrads with explicit and implicit components half the time	2013-06-18 17:55:58 +01:00
Alan Saul	f3b8dfb222	about to input new derivations for Z's...	2013-06-03 14:51:09 +01:00
Alan Saul	6c29750795	Took out all the asserts and using pure broadcasting method of diagonal now	2013-05-31 16:59:54 +01:00
Alan Saul	e842f6e687	Made it use the fact that W is diagonal and put assertions in to ensure that the results are the same	2013-05-31 16:45:22 +01:00
Alan Saul	f9857e08c0	Broken it by getting rid of squeeze, but now working on making it faster using proper vector multiplciation for diagonals	2013-05-31 11:55:32 +01:00
Alan Saul	20227fb2ac	Made more numerically stable in a hope that it will work and I will find a bug...	2013-05-30 16:17:37 +01:00
Alan Saul	23ed2a2d15	Lots of name changing and went through all likelihood gradients again	2013-05-29 17:33:06 +01:00
Alan Saul	117c377d13	Ripped out all things Laplace parameter estimation, starting again with new tactic	2013-05-29 14:02:03 +01:00
Alan Saul	d63d370641	About to rip out old chain rule method of learning gradients	2013-05-29 13:46:55 +01:00
Alan Saul	146d7e2458	Trying to fix dL_dytil gradient	2013-05-17 17:42:00 +01:00
Alan Saul	48d693791e	changed name	2013-05-16 12:22:37 +01:00
Alan Saul	e5d7ee9728	FIXED DYTIL_DFHAT	2013-05-16 12:06:09 +01:00
Alan Saul	21ae81de29	Workong on doing explicit gradients	2013-05-16 12:00:15 +01:00
Alan Saul	569311b510	Gradients almost there for dytil_dfhat, diagonal terms are right	2013-05-15 19:25:55 +01:00
Alan Saul	cad2c271eb	merged with devel	2013-05-14 16:26:57 +01:00
Alan Saul	787a038401	Still getting closer to grads for likelihood	2013-05-14 16:23:18 +01:00
Ricardo	47a7df9756	Merge branch 'devel' of github.com:SheffieldML/GPy into devel	2013-05-14 15:38:08 +01:00
Ricardo	3626154c02	remove profile	2013-05-14 15:38:00 +01:00
James Hensman	4bbf31eb1d	removed RA's profile deco	2013-05-14 11:26:16 +01:00
James Hensman	350fa39685	generalised backsub_both_sides	2013-05-14 11:20:44 +01:00
Ricardo	225c35d55d	Merge branch 'devel' of github.com:SheffieldML/GPy into devel	2013-05-13 19:41:10 +01:00
Ricardo	a1036fb2a5	DSYR is being used now	2013-05-13 19:40:51 +01:00
Ricardo	cb4c363c11	Gradients are working now	2013-05-13 19:38:11 +01:00
Alan Saul	5472c5c6ba	Almost have likelihood gradients working but kernels still way off	2013-05-13 18:36:02 +01:00
James Hensman	c859687d43	Merge branch 'devel' of github.com:SheffieldML/GPy into devel	2013-05-13 17:23:54 +01:00
James Hensman	542759353b	minor changes to the symmetrify function	2013-05-13 17:23:49 +01:00
Nicolo Fusi	ccde6c421f	Merge branch 'devel' of github.com:SheffieldML/GPy into devel	2013-05-13 17:21:45 +01:00
Nicolo Fusi	80c25c9aaf	moved linalg function to GPy.linalg	2013-05-13 17:04:59 +01:00
Ricardo	3cc9547cab	Merge branch 'devel' of github.com:SheffieldML/GPy into devel	2013-05-13 15:09:43 +01:00
Ricardo	b84f70b205	gradietns check :)	2013-05-13 15:09:20 +01:00
James Hensman	9c701f6e5b	added DSYR for ricardo	2013-05-13 12:54:55 +01:00
Ricardo	fda9cae3c3	Merge branch 'devel' of github.com:SheffieldML/GPy into devel	2013-05-13 11:11:29 +01:00
Ricardo	68f493b86c	New functions for EP-matching moments	2013-05-13 11:10:27 +01:00
Ricardo	c521c243e1	minor change	2013-05-13 11:03:19 +01:00
James Hensman	c1f1e44b09	using weaved symmetrify in pdinv now	2013-05-13 10:36:15 +01:00
James Hensman	529d7534ca	fixed the bug in stick	2013-05-13 10:26:54 +01:00