Commit Graph
16161 Commits
Author SHA1 Message Date
Vivek Pal 408f0e8352 fixup! Rename RMSprop as RMSProp 2017-04-06 21:58:44 +05:30
Vivek Pal 73b5c68f0a Rename RMSprop as RMSProp 2017-04-06 21:58:44 +05:30
Vivek Pal 1b9cfeafde Update the comment about the RMSpropUpdate constructor 2017-04-06 21:54:00 +05:30
Vivek Pal b8a51f3348 Correct the smoothing parameter name 2017-04-06 21:54:00 +05:30
Vivek Pal c1dd7308c1 Fix alignment issues 2017-04-06 21:54:00 +05:30
Vivek Pal 612c5d98a8 Fix a typo: RmsProp -> RMSprop 2017-04-06 21:54:00 +05:30
Vivek Pal 2c65ade104 Fix a mistake 2017-04-06 21:54:00 +05:30
Vivek Pal 8116544ede Add implementation of the RMSprop update policy 2017-04-06 21:54:00 +05:30
Marcus Edel 6b097f28d5 It isn't guaranteed that the recurrent network test will converge in the specified number of iterations using random weights (if the test pased in 1 of 5 times that's fine). 2017-04-05 20:22:05 +02:00
Ryan Curtin 0242c23b2c Merge pull request #899 from ivmarkp/SMORMS3
Implementation of SMORMS3 optimizer
2017-04-04 16:21:56 -04:00
Marcus Edel fcf165b77d Merge pull request #971 from amoudgl/fix_cnn_test
Fix VanillaNetworkTest.
2017-04-03 22:13:33 +02:00
Ryan Curtin eca65e46a5 Merge pull request #975 from sidak/master
Correct minor typo in ffn.hpp
2017-04-02 18:42:33 -04:00
sidak 49ca92ac96 Merge branch 'master' of https://github.com/sidak/mlpack 2017-04-02 21:27:51 +05:30
sidak c64c610374 Correct minor typo in ffn.hpp 2017-04-02 21:18:34 +05:30
Vivek Pal d23d7eb55c Remove the comment for an unused parameter 2017-04-02 19:59:12 +05:30
Vivek Pal 16a27fc242 Fix macros 2017-04-02 19:55:11 +05:30
Marcus Edel 4057d144ec Merge pull request #974 from sagarbhathwar/add_merge_correction
#include "add_merge.hpp" instead of #include "add_merge_impl.hpp" in add_merge_impl.hpp.
2017-04-02 11:25:13 +02:00
Sagar B Hathwar 177690bada Include add_merge.hpp instead of itself in add_merge_impl.hpp 2017-04-02 09:30:23 +05:30
Abhinav Moudgil 82933523e0 Improve comments 2017-04-01 03:16:50 +05:30
Vivek Pal 1f0310b52f Try transform() instead of for_each() to fix Travis build failure 2017-03-30 21:04:11 +05:30
Marcus Edel a836378271 Don't initialize model parameters for an empty model (Avoid integer divide by zero erro). 2017-03-30 15:54:46 +02:00
Vivek Pal 93e817b998 fixup! Fix a couple of comments 2017-03-30 19:06:04 +05:30
Vivek Pal 14a0091884 Use the combination of for_each and a lambda expression
Avoids allocating the stepSizeMat completely.
2017-03-30 19:01:37 +05:30
Vivek Pal 81b8d03faf Minor changes addressing review comments 2017-03-30 00:57:15 +05:30
Vivek Pal b057f881ab Remove those parameters from SMORMS3 class that:
are modified by the optimizer object which are
stepSize, maxIterations, tolerance and shuffle
2017-03-30 00:57:15 +05:30
Vivek Pal dacb95ead8 Use constructor to initialise previousStepSize variable 2017-03-30 00:57:15 +05:30
Vivek Pal 98b2f5a639 Add Epsilon function to the SMORMS3Update update policy 2017-03-30 00:57:15 +05:30
Vivek Pal 5fee74edec Fix a couple of comments 2017-03-30 00:57:15 +05:30
Vivek Pal ac8492b0e0 Initialise previousStepSize before its first use
Fixes valgrind error: Conditional jump or move
depends on uninitialised value(s)
2017-03-30 00:57:15 +05:30
Vivek Pal 24f6cec66e Fix incorrect instantiation of SMORMS3 constructor 2017-03-30 00:57:15 +05:30
Vivek Pal 0dd928840e The arguments not necessary to define the type 2017-03-30 00:57:15 +05:30
Vivek Pal 7bbe4359de No need to include smorms3_update.hpp here
It was moved to optimizer/smorms3 directory.
2017-03-30 00:57:15 +05:30
Vivek Pal c4121ff2c6 Rename n_rows and n_cols as rows and cols respectively 2017-03-30 00:57:15 +05:30
Vivek Pal 782c4a93d1 Address review comments with the following changes:
* Rename parameter eps as epsilon everywhere.
* Make paramter stepSize & gradient const in SMORMS3Update::Update.
* Move smorms3_update.hpp to /optimizers/smorms3 directory.
* Reset stepSizeMat only when stepSize != previousStepSize.
* Update tests.
2017-03-30 00:57:15 +05:30
Vivek Pal 4950f54b20 fixup! Let's stick with stepSize instead of lRate 2017-03-30 00:57:15 +05:30
Vivek Pal 3c720d3b31 Let's stick with stepSize instead of lRate
We currently use stepSize in other optimizers as well
e.g. RMSprop etc and also, consistent code looks good.
2017-03-30 00:57:15 +05:30
Vivek Pal a281494636 Implement SMORMS3 optimizer as an update policy of SGD
* SMORMS3Update policy class implements the update steps of SMORMS3 algorithm.
* SMORMS3 class as a wrapper leveraging the SGD class.
2017-03-30 00:57:15 +05:30
Vivek Pal d798b3dfb1 Rename function LRate as LearningRate 2017-03-30 00:57:15 +05:30
Vivek Pal 8ec825c738 Rename eps parameter in the constructor as epsilon 2017-03-30 00:57:15 +05:30
Vivek Pal 03408c5742 Remove leading underscores in the macros 2017-03-30 00:57:15 +05:30
Vivek Pal 558de23e1d Move initialisation of lRateMat outside the for loop 2017-03-30 00:57:15 +05:30
Vivek Pal dac421821e Remove unnecessary newlines 2017-03-30 00:57:15 +05:30
Vivek Pal 9945b04f25 Use (%) Schur product instead of (*) Matrix multiplication 2017-03-30 00:57:15 +05:30
Vivek Pal c812c18d83 Fix a typo 2017-03-30 00:57:15 +05:30
Vivek Pal ddb3128601 Fill matrix lRateMat with lRate not matrix x 2017-03-30 00:57:15 +05:30
Vivek Pal 4b4c2d76c7 Update optimizers and tests CMakeLists 2017-03-30 00:57:15 +05:30
Vivek Pal 946ba643d0 Add tests for SMORMS3 optimizer 2017-03-30 00:57:15 +05:30
Vivek Pal c7872c20c6 Implement the SMORMS3 optimizer 2017-03-30 00:57:15 +05:30
Ryan Curtin bd7f32794f Merge pull request #970 from amoudgl/style_fix
Remove all trailing spaces
2017-03-29 09:11:47 -04:00
Abhinav Moudgil 6cb13c1c92 Fix VanillaNetworkTest 2017-03-29 17:56:29 +05:30