0e3c512 | karenyyng | 27 September 2015, 23:10:58 UTC | fixed version info | 27 September 2015, 23:10:58 UTC |
e2e2775 | Michael D. Schneider | 16 August 2015, 06:19:50 UTC | Clean up kernels C++ classes to avoid ambiguities in subclass method and member resolutions | 16 August 2015, 06:19:50 UTC |
81f62bd | Michael D. Schneider | 16 August 2015, 06:18:58 UTC | Use bounds checking in accessing the Metric class parameter vector | 16 August 2015, 06:18:58 UTC |
3979483 | karenyyng | 10 June 2015, 09:54:01 UTC | C++ level implementation of l_sq as 1 / beta should be fixed | 10 June 2015, 09:54:01 UTC |
5789efb | karenyyng | 09 May 2015, 04:51:51 UTC | found version setup.py that generates correct debugging symbols | 09 May 2015, 04:51:51 UTC |
0e82b9f | karenyyng | 09 May 2015, 04:47:47 UTC | adding debugging symbols | 09 May 2015, 04:47:47 UTC |
fdcfe79 | karenyyng | 08 May 2015, 05:50:21 UTC | deleted const keyword for parent class gradient methods | 08 May 2015, 05:50:21 UTC |
8513960 | karenyyng | 08 May 2015, 05:32:27 UTC | first draft of gradient function inheritance | 08 May 2015, 05:32:27 UTC |
b512bc1 | karenyyng | 01 May 2015, 23:07:22 UTC | commented out print statements | 01 May 2015, 23:07:22 UTC |
d1ccf80 | karenyyng | 01 May 2015, 18:16:09 UTC | version that passed all tests | 01 May 2015, 18:16:09 UTC |
e428cf2 | karenyyng | 27 April 2015, 22:05:28 UTC | gives approximately correct results - due to numerical precision loss one must use approximate comparison for tests to pass | 27 April 2015, 22:05:28 UTC |
70bf0d6 | karenyyng | 27 April 2015, 20:47:44 UTC | changed value function signature by removing const - inheritance works properly now | 27 April 2015, 20:47:44 UTC |
8662218 | karenyyng | 25 April 2015, 01:28:24 UTC | confirmed metric bug is due to improper inheritance of the value method (ExpSquaredKernel value method not overridden | 25 April 2015, 01:28:24 UTC |
1796309 | karenyyng | 24 April 2015, 01:30:06 UTC | bug fix removed const from function defs | 24 April 2015, 01:30:06 UTC |
5c07019 | karenyyng | 24 April 2015, 01:05:49 UTC | compiles again after implementing modified value function | 24 April 2015, 01:05:49 UTC |
d7e5671 | karenyyng | 24 April 2015, 00:20:49 UTC | bug fixes and adding to implementation of an alternative product operator | 24 April 2015, 00:20:49 UTC |
aa2c964 | karenyyng | 23 April 2015, 23:54:29 UTC | almost finished adding methods to .h file | 23 April 2015, 23:54:29 UTC |
e239a3a | karenyyng | 23 April 2015, 23:12:34 UTC | almost finished adding methods to .h file | 23 April 2015, 23:12:34 UTC |
150a904 | karenyyng | 23 April 2015, 07:04:18 UTC | too tired to continue | 23 April 2015, 07:04:18 UTC |
9191d0a | karenyyng | 23 April 2015, 04:51:58 UTC | adding more methods and compiles | 23 April 2015, 04:51:58 UTC |
b5c9609 | karenyyng | 23 April 2015, 04:50:44 UTC | adding more methods | 23 April 2015, 04:50:44 UTC |
d4706b8 | karenyyng | 23 April 2015, 03:59:47 UTC | adding more methods | 23 April 2015, 03:59:47 UTC |
5acd475 | karenyyng | 23 April 2015, 03:29:31 UTC | implemented more of the C++ version of the kernel | 23 April 2015, 03:29:31 UTC |
7fcd336 | karenyyng | 22 April 2015, 23:55:56 UTC | package compiles after adding new methods | 22 April 2015, 23:55:56 UTC |
f9f5110 | karenyyng | 22 April 2015, 23:43:34 UTC | trying different cpp implementation of the kernel after taking derivatives | 22 April 2015, 23:43:34 UTC |
d74434b | karenyyng | 22 April 2015, 19:33:49 UTC | debugging | 22 April 2015, 19:33:49 UTC |
13e64d7 | karenyyng | 17 April 2015, 06:57:55 UTC | slight changes | 17 April 2015, 06:57:55 UTC |
4d45fa4 | karenyyng | 30 March 2015, 16:39:59 UTC | important bug fix in Cython code, same bug as 38159cdfd7dbc6d3d135a5c42cb370ad14fc89cc | 30 March 2015, 16:39:59 UTC |
da840ae | karenyyng | 25 March 2015, 00:13:47 UTC | KappaKappaExpSquaredKernel now can be computed ... | 25 March 2015, 00:13:47 UTC |
69f0888 | karenyyng | 24 March 2015, 22:20:24 UTC | bug fixes | 24 March 2015, 22:20:24 UTC |
8ff481a | karenyyng | 24 March 2015, 21:48:03 UTC | added variable types to _kernels.pyx, to check consistencies of arguments of methods | 24 March 2015, 21:48:03 UTC |
92c58d1 | karenyyng | 23 March 2015, 22:32:46 UTC | cleaning up unnecessary definitions, George now compiles but more testing is needed | 23 March 2015, 22:32:46 UTC |
d719ad1 | karenyyng | 21 March 2015, 16:30:58 UTC | testing Cython code method by method | 21 March 2015, 16:30:58 UTC |
2c2e569 | karenyyng | 21 March 2015, 16:05:10 UTC | commented out python and C++ header to get Cython code working first | 21 March 2015, 16:05:10 UTC |
88dbef7 | karenyyng | 16 March 2015, 15:28:39 UTC | added python code to cython file _kernel.pyx - need testing | 16 March 2015, 15:28:39 UTC |
139d858 | karenyyng | 10 March 2015, 01:33:25 UTC | added more header info, imports without errors | 10 March 2015, 01:33:25 UTC |
2b08729 | karenyyng | 10 March 2015, 00:06:19 UTC | added skeleton of the classes | 10 March 2015, 00:06:19 UTC |
fc5dd1c | Dan F-M | 06 January 2015, 17:38:36 UTC | reverting a few docs changes | 06 January 2015, 17:38:36 UTC |
c90d483 | Dan F-M | 06 January 2015, 17:37:32 UTC | Merge branch 'master' of https://github.com/dfm/george | 06 January 2015, 17:37:32 UTC |
2fe6d9e | Dan Foreman-Mackey | 06 January 2015, 17:37:25 UTC | Merge pull request #26 from mindriot101/correct-model-fitting-example-docs Correct model fitting example docs | 06 January 2015, 17:37:25 UTC |
2baf7ed | Dan F-M | 06 January 2015, 17:36:57 UTC | A few words for the paper | 06 January 2015, 17:36:57 UTC |
40800ce | Simon Walker | 06 January 2015, 09:57:57 UTC | Use correct params array for call to `model` In the final example the `model` function is in the source code, and only takes the Gaussian parameters. | 06 January 2015, 09:57:57 UTC |
95b671b | Simon Walker | 06 January 2015, 09:52:33 UTC | Run the second burn in and save lnp This step was not included in the documentation, and must occur to get the maximum lnp before running the second burn in. | 06 January 2015, 09:52:33 UTC |
29f4818 | Simon Walker | 06 January 2015, 09:39:30 UTC | Set the number of walkers | 06 January 2015, 09:48:38 UTC |
10aeb35 | Simon Walker | 06 January 2015, 09:37:53 UTC | Set the seed in the documentation This ensures the user following the example gets the same results as shown | 06 January 2015, 09:37:53 UTC |
6f2a4a5 | Dan F-M | 18 November 2014, 21:47:58 UTC | optimization method | 18 November 2014, 21:47:58 UTC |
3cb331b | Dan F-M | 18 November 2014, 21:24:58 UTC | Merge branch 'master' of https://github.com/dfm/george | 18 November 2014, 21:24:58 UTC |
81056ea | Dan F-M | 18 November 2014, 21:24:52 UTC | typo in hyperparameter docs | 18 November 2014, 21:24:52 UTC |
978781d | Dan F-M | 18 November 2014, 02:51:21 UTC | some documentation about kernel parameterization | 18 November 2014, 02:51:21 UTC |
bba942c | Dan Foreman-Mackey | 29 October 2014, 01:15:03 UTC | Nicer badges | 29 October 2014, 01:15:03 UTC |
830b32e | Dan F-M | 13 October 2014, 21:41:25 UTC | typo | 13 October 2014, 21:41:25 UTC |
1d3405f | Dan F-M | 13 October 2014, 21:25:00 UTC | fixing #23 | 13 October 2014, 21:25:00 UTC |
9c460b5 | Dan F-M | 13 October 2014, 21:13:38 UTC | include Cython files in source distribution on PyPI | 13 October 2014, 21:13:38 UTC |
3be361f | Dan F-M | 13 October 2014, 17:47:04 UTC | adding documentaion for GP | 13 October 2014, 17:47:04 UTC |
378b72c | Dan F-M | 13 October 2014, 16:27:00 UTC | removing rebuild from HODLRSolver pickle support | 13 October 2014, 16:27:00 UTC |
a93d0f4 | Dan F-M | 13 October 2014, 16:20:15 UTC | documentation for the HODLR solver | 13 October 2014, 16:20:15 UTC |
407d815 | Dan F-M | 12 October 2014, 22:54:38 UTC | some progress on solver docs | 12 October 2014, 22:54:38 UTC |
7a1adfb | Dan F-M | 11 October 2014, 18:46:49 UTC | removing comment about periodic kernel dimension | 11 October 2014, 18:46:49 UTC |
a556175 | Dan F-M | 11 October 2014, 18:39:24 UTC | Merge branch 'custom-kernels' | 11 October 2014, 18:39:24 UTC |
365612a | Dan F-M | 11 October 2014, 18:39:18 UTC | kernel documentation | 11 October 2014, 18:39:18 UTC |
a93d3cd | Dan F-M | 10 October 2014, 21:25:58 UTC | updating alpha when kernel parameters change | 10 October 2014, 21:25:58 UTC |
18c90e1 | Dan F-M | 10 October 2014, 21:09:26 UTC | a few words about Python kernels | 10 October 2014, 21:09:26 UTC |
7cd5b35 | Dan F-M | 10 October 2014, 20:59:52 UTC | adding custom kernels implemented in Python | 10 October 2014, 20:59:52 UTC |
f7b85f8 | Jonah Bernhard | 10 October 2014, 16:39:14 UTC | Update GP test list. | 10 October 2014, 16:39:14 UTC |
80fb447 | Jonah Bernhard | 10 October 2014, 16:22:59 UTC | Add test for GP alpha cache. | 10 October 2014, 16:22:59 UTC |
f4eae9e | Jonah Bernhard | 10 October 2014, 15:50:55 UTC | Compare arrays using native numpy function. | 10 October 2014, 15:50:55 UTC |
0d3ab7d | Jonah Bernhard | 10 October 2014, 15:47:46 UTC | PEP8 fix. | 10 October 2014, 15:47:46 UTC |
ac06bea | Jonah Bernhard | 09 October 2014, 19:57:43 UTC | Accelerate repeated predictions. | 09 October 2014, 19:57:43 UTC |
d1c05e0 | Jonah Bernhard | 09 October 2014, 19:53:40 UTC | Forgot to add new test to __all__. | 09 October 2014, 19:53:40 UTC |
8b8ded6 | Dan Foreman-Mackey | 07 October 2014, 19:09:29 UTC | Add DOI to README | 07 October 2014, 19:09:29 UTC |
b2e4150 | Dan F-M | 07 October 2014, 19:06:09 UTC | bumping version number | 07 October 2014, 19:06:09 UTC |
c41088c | Dan F-M | 07 October 2014, 19:01:58 UTC | HODLR in_place bug | 07 October 2014, 19:01:58 UTC |
514ecab | Dan F-M | 07 October 2014, 18:52:06 UTC | fixing transpose issue in prediction | 07 October 2014, 18:52:06 UTC |
8979de9 | Dan F-M | 07 October 2014, 18:47:09 UTC | Merge branch 'fix_covariance' of https://github.com/jbernhard/george | 07 October 2014, 18:47:09 UTC |
fc46c5c | Jonah Bernhard | 07 October 2014, 16:08:23 UTC | Fix bug in computing the posterior covariance. | 07 October 2014, 16:08:23 UTC |
945e1d0 | Jonah Bernhard | 07 October 2014, 16:07:40 UTC | Add GP prediction / regression test. | 07 October 2014, 16:07:40 UTC |
45d424e | Dan F-M | 05 October 2014, 16:13:12 UTC | save some memory | 05 October 2014, 16:13:12 UTC |
adbc09f | Dan F-M | 23 September 2014, 15:17:53 UTC | a few formal words | 23 September 2014, 15:17:53 UTC |
928132d | Dan F-M | 18 September 2014, 16:58:48 UTC | a few words in the paper | 18 September 2014, 16:58:48 UTC |
aa619df | Dan F-M | 07 September 2014, 17:54:11 UTC | remove extra assert | 07 September 2014, 17:54:11 UTC |
f83d333 | Dan F-M | 07 September 2014, 17:53:55 UTC | slight refactor of grad computation | 07 September 2014, 17:53:55 UTC |
430def1 | Dan F-M | 07 September 2014, 17:37:33 UTC | Merge branch 'speedup-grad' of https://github.com/shoyer/george | 07 September 2014, 17:37:33 UTC |
81eddc3 | Dan F-M | 07 September 2014, 17:37:29 UTC | temp | 07 September 2014, 17:37:29 UTC |
e14cef0 | Stephan Hoyer | 03 September 2014, 22:11:13 UTC | Remove loop for grad_lnlikelihood calculation The result (using einsum) is cleaner and perhaps slightly faster. | 03 September 2014, 22:11:13 UTC |
9c920c4 | Stephan Hoyer | 03 September 2014, 21:37:57 UTC | Precompute the matrix inverse for grad_lnlikelihood This reduces the number of required operations for grad_lnlikelihood by a factor equal to the number of kernel parameters. e.g., consider this example: import numpy as np import george from george.kernels import Matern32Kernel, ConstantKernel, WhiteKernel x = 10 * np.random.RandomState(12356).rand(2000, 2) yerr = 0.2 * np.ones_like(x[:, 0]) y = np.sin(x[:, 0] + x[:, 1]) + yerr * np.random.randn(len(x)) kernel = ConstantKernel(0.5, ndim=2) * Matern32Kernel(0.5, ndim=2) + WhiteKernel(0.1, ndim=2) gp = george.GP(kernel, solver=george.HODLRSolver) gp.compute(x) %time gp.grad_lnlikelihood(y) Before this change, I get: CPU times: user 8.89 s, sys: 479 ms, total: 9.37 s Wall time: 9.37 s array([-182.28430864, 208.82577854, -463.12445191]) After this change: CPU times: user 3.37 s, sys: 273 ms, total: 3.64 s Wall time: 3.64 s array([-182.27545648, 209.89776176, -479.34696447]) So it's a factor of three speedup, corresponding to the three parameters. There is a similar speedup for basic solver (which for this problem size is actually faster, at least on my laptop, perhaps because scipy's cholesky factorization can use multi-processing). | 03 September 2014, 21:58:48 UTC |
592498f | Stephan Hoyer | 03 September 2014, 20:41:54 UTC | Use np.einsum in grad_lnlikelihood instead of a complex sum/map/dot mix It makes for a 3x constant speedup for that line of code on my machine: In [1]: import numpy as np In [2]: alpha = np.random.randn(5000) In [3]: k = np.random.randn(5000, 5000) In [4]: %timeit sum(map(lambda r: np.dot(alpha, r), alpha[:, None] * k)) 10 loops, best of 3: 148 ms per loop In [5]: %timeit np.einsum('i,j,ji', alpha, alpha, k) 10 loops, best of 3: 43.9 ms per loop More importantly, I think it is also much more readable. | 03 September 2014, 20:41:54 UTC |
0809981 | Dan F-M | 02 September 2014, 18:07:05 UTC | c++ compiler for travis | 02 September 2014, 18:07:05 UTC |
13dae85 | Dan F-M | 02 September 2014, 17:33:52 UTC | making kernels and solvers picklable | 02 September 2014, 17:33:52 UTC |
4cdbdfa | Dan F-M | 30 August 2014, 19:36:43 UTC | silencing some harmless compile warnings | 30 August 2014, 19:36:43 UTC |
297bfca | Dan F-M | 30 August 2014, 19:29:59 UTC | more tol | 30 August 2014, 19:29:59 UTC |
f8a53be | Dan F-M | 30 August 2014, 19:27:43 UTC | more info | 30 August 2014, 19:27:43 UTC |
7508c86 | Dan F-M | 30 August 2014, 19:21:12 UTC | Fixing gradient bug (#13) | 30 August 2014, 19:21:12 UTC |
b48b2a5 | Dan F-M | 30 August 2014, 18:55:41 UTC | fixing kernel dtype bug (#13) | 30 August 2014, 18:56:02 UTC |
d758bbe | Dan F-M | 30 August 2014, 16:09:06 UTC | bug when re-computing product kernels | 30 August 2014, 16:09:06 UTC |
9037511 | Dan F-M | 30 August 2014, 15:39:41 UTC | fixing hodlr apply_inverse behaviour | 30 August 2014, 15:39:41 UTC |
3eb371a | Dan F-M | 30 August 2014, 14:03:01 UTC | allowing 1D kernels | 30 August 2014, 14:03:01 UTC |
5ef0787 | Dan F-M | 30 August 2014, 13:13:25 UTC | ensure contiguous arrays for Cython | 30 August 2014, 13:13:25 UTC |
9304e87 | Dan F-M | 30 August 2014, 12:59:28 UTC | making HODLR solver recompute on kernel changes properly | 30 August 2014, 12:59:28 UTC |
82b7dfe | Dan F-M | 30 August 2014, 03:49:48 UTC | the lnlikelihood should be a scalar | 30 August 2014, 03:49:48 UTC |
faf6ea0 | Dan F-M | 30 August 2014, 03:40:12 UTC | import the HODLR solver | 30 August 2014, 03:40:12 UTC |