62ce35a | Graham Neubig | 24 November 2017, 20:05:32 UTC | Fix compile error on gcc 4.9 + CUDA 7.5 | 24 November 2017, 20:05:32 UTC |
1eed174 | Graham Neubig | 24 November 2017, 17:49:30 UTC | Fixed stateful trainers to work with multi-device (#1096) | 24 November 2017, 17:49:30 UTC |
88aa149 | Graham Neubig | 24 November 2017, 16:06:13 UTC | Make possible to set CUDA compute architecture (#1092) | 24 November 2017, 16:06:13 UTC |
07f736b | Graham Neubig | 24 November 2017, 04:08:36 UTC | Added profiling capability | 24 November 2017, 04:08:36 UTC |
d256c73 | Arne Köhn | 24 November 2017, 03:22:17 UTC | MKL: Link against mkl_rt instead of all libraries directly (#1081) Use the "single dynamic library" linking model of MKL, described here: https://software.intel.com/en-us/articles/a-new-linking-model-single-dynamic-library-mkl_rt-since-intel-mkl-103/ | 24 November 2017, 03:22:17 UTC |
4940412 | msperber | 23 November 2017, 14:53:01 UTC | profile fwd/bwd runtimes and memory (#1088) * profile fwd/bwd runtimes and memory * double counted node memory to account for both fx and dEdfx * print nodes sorted by memory when profiling flag >= 2 * percentages for time and memory profiles * print 100% mark * aggregating memory consumption over nodes * fix previous commit | 23 November 2017, 14:53:01 UTC |
afa1934 | shuheik | 22 November 2017, 14:12:15 UTC | remove superfluous line (#1085) | 22 November 2017, 14:12:15 UTC |
cc9cc21 | Chris Dyer | 21 November 2017, 14:56:32 UTC | fix doc string | 21 November 2017, 14:56:32 UTC |
11625dd | Daniel Hershcovich | 18 November 2017, 18:57:30 UTC | After deployment, test pip installation from PyPI works (#1076) * Add Travis build stage to test pip install after deployment * Run all Travis jobs on tag | 18 November 2017, 18:57:30 UTC |
a29790b | shuheik | 18 November 2017, 18:56:55 UTC | Update Scala bindings (#1075) * update Scala bindings * fix * Expression.hingeDim and UnsignedVectorVector | 18 November 2017, 18:56:55 UTC |
9d5815f | Vu Cong Duy Hoang | 18 November 2017, 18:53:48 UTC | bug in select_cols (#1073) * fixed a bug of dim in select_cols * corrected typo in concatenate's doc * added impl. for TensorTools::access_element() | 18 November 2017, 18:53:48 UTC |
f9550b7 | Graham Neubig | 14 November 2017, 16:37:57 UTC | Made max pooling 2d be able to take in 2d tensors (#1065) * Made max pooling 2d be able to take in 2d tensors * Fixed copying of dimension | 14 November 2017, 16:37:57 UTC |
2f497b9 | Matthew Richardson | 08 November 2017, 20:58:38 UTC | Fixing doc on needing CUDA/cuDNN in path for Windows (was missing a blank line). | 08 November 2017, 20:58:38 UTC |
960e558 | Matthew Richardson | 08 November 2017, 20:07:33 UTC | Added note about CUDA/cuDNN paths when running in WIndows. | 08 November 2017, 20:07:33 UTC |
0653b1d | Daniel Hershcovich | 08 November 2017, 11:25:00 UTC | Fixes to macOS Python install (#1055) * Fix #1054: copy libdynet.dylib only in Python install * No need for DYLD_LIBRARY_PATH * Fix #1056: install libdynet.dylib to site-packages | 08 November 2017, 11:25:00 UTC |
2924bff | Daniel Hershcovich | 08 November 2017, 05:45:57 UTC | Fix install name in macOS wheels with install_name_tool (#1048) * Remove unneeded delocate * Use @loader_path on macOS to fix loading wheels * Fix README to allow properly generating RST from it * Remove table of contents when converting README to rst for PyPI PyPI does not support internal document links (starting with #). https://stackoverflow.com/a/16594755/223267 | 08 November 2017, 05:45:57 UTC |
6416283 | Chris Dyer | 07 November 2017, 10:37:14 UTC | better autobatching style Former-commit-id: fc5ffc248ee3714f1904d954e784e3f1fad6a76f | 07 November 2017, 10:37:14 UTC |
9275cc5 | Chris Dyer | 06 November 2017, 09:53:45 UTC | add comment describing how to run, minor bug fix in example Former-commit-id: c57f9da5778b6921a08f544f70d550d3977c9b73 | 06 November 2017, 09:53:45 UTC |
15699df | Daniel Hershcovich | 05 November 2017, 21:54:45 UTC | Deploy macOS wheels (#1043) * Fix Miniconda path on AppVeyor * Run brew update before brew install on Travis CI * Add libdynet.dylib to wheel on macOS * Deploy wheels on macOS too * Build sdist only on Linux On macOS it tries to include miniconda too and we don't need it. It's enough the upload the sdist from one job anyway. Former-commit-id: 4a2ed3e2060ae8344079e6441c913916afbcf25a | 05 November 2017, 21:54:45 UTC |
10b4ff4 | Chris Dyer | 05 November 2017, 20:45:26 UTC | MNIST example in Python, also example of conv2d Former-commit-id: 23cb8def4fefc84945bc7c9c0a58dda2c4a4a2e7 | 05 November 2017, 20:45:26 UTC |
4872889 | Graham Neubig | 05 November 2017, 09:49:54 UTC | Added CuDNN to Python documentation (#1041) Former-commit-id: e136b186f02fb84974407de5d0c20e51fb644247 | 05 November 2017, 09:49:54 UTC |
fc819a3 | Graham Neubig | 05 November 2017, 01:48:09 UTC | Updated boundary check (#1040) Former-commit-id: 73d744f59158eb4fa88db96978149bfa63e8879c | 05 November 2017, 01:48:09 UTC |
517174c | Jonathan Kummerfeld | 03 November 2017, 14:40:18 UTC | Adding the 'SILU / Swish' activation function (#1030) * Adding the 'swish' activation function Proposed by Ramachandran, Zoph, and Le: https://arxiv.org/pdf/1710.05941.pdf swish(x) = x . sigmoid(beta * x) * Adding python support and a test * Switching to SILU and adding additional citations * Bugfixes * Updated test * Minor edits to make merge clearer to git Former-commit-id: 252a53fb871160ca496613bb88bf7c5149d19d11 | 03 November 2017, 14:40:18 UTC |
5fcea4d | Daniel Hershcovich | 03 November 2017, 14:38:08 UTC | Leave out intermediate top directory from Eigen zip in Python setup (#1033) * Remove unused files in python/ * Omit intermediate top directory from Eigen zip When extracting Eigen from zip, avoid the "eigen-eigen-..." top directory, and just put everything under eigen/. Otherwise running setup.py again will skip downloading eigen (because it already exists) but fail in building DyNet (because its contents are not what we expect but that intermediate top directory). * Do not fail if pandoc cannot be installed Former-commit-id: 6e4ab5b54c56c28f85130c03e41d7adfe60a67fc | 03 November 2017, 14:38:08 UTC |
417f396 | msperber | 30 October 2017, 11:34:29 UTC | Softmax across rows (#1028) * softmax: use scratch memory instead of aux memory * implement softmax_rows * merge softmax and softmax_rows Former-commit-id: ef22817c3d9108b44a91e16ecee322c9363280fe | 30 October 2017, 11:34:29 UTC |
b7068f3 | Hong Wu | 27 October 2017, 00:46:12 UTC | fix small memory issue, at least 1MB for each type. (#1027) Former-commit-id: 844c6900b6da20871f29135f8c29b7a8bb3c6c9e | 27 October 2017, 00:46:12 UTC |
44a152c | Graham Neubig | 26 October 2017, 18:03:59 UTC | Boundary check in PickNegLogSoftmax (#1026) * Boundary check in PickNegLogSoftmax * Bug fix Former-commit-id: 3753675d419feb44b3194a9eaf9bb812b22e0ecf | 26 October 2017, 18:03:59 UTC |
2f14435 | Hong Wu | 26 October 2017, 13:13:30 UTC | [dynet-1018]. Avoid allocate CPU memory under GPU mode. (#1024) * Avoid allocate CPU memory under GPU mode * fix dismatch type * modify default mem size Former-commit-id: c6420c2adc88b94180ae12028cc89cd9a845e943 | 26 October 2017, 13:13:30 UTC |
ebf32b0 | Matt Richardson | 26 October 2017, 13:13:19 UTC | Re-enable test-io unit test on Windows after Chunyang's fix #1014. (#1022) Former-commit-id: 4c6bc5801c795054b2e0966bdc4677a26444a461 | 26 October 2017, 13:13:19 UTC |
13e3184 | Alexander Rudnicky | 25 October 2017, 20:58:27 UTC | path issue (#1019) Fixed so that cudnn path resolution works on Windows (removed hardcoded *nix path) Former-commit-id: b5c4e4b1cb7ecabf13233f75223af9d3eb5969c9 | 25 October 2017, 20:58:27 UTC |
55a0ed4 | Chunyang Wen | 25 October 2017, 20:57:42 UTC | FIX: TextFileSaver/Loader under windows (#1014) Former-commit-id: de9bc3681a1d37df15b0d554bf5ceb83d3b15d1e | 25 October 2017, 20:57:42 UTC |
e7de791 | Matt Richardson | 25 October 2017, 20:56:51 UTC | Added C++ unit tests to Appveyor (#1013) * Adding unit testing to appveyor * Fixing change to build directory. Adding boost path * Remove test-io for now because it causes infinite loop in windows. Added compiling example code too. * Python unit testing seems to need to be in build\python directory Former-commit-id: 30f8d509b99e6e534af9c50c84683399ff4df11d | 25 October 2017, 20:56:51 UTC |
82ff6a9 | Chunyang Wen | 25 October 2017, 20:56:11 UTC | FIX: compiler warnings (#1015) * FIX: compiler warnings * FIX: unsigned long and long comparison Former-commit-id: 3ebc4c02c777bf38f1abb528f415dc47244a8655 | 25 October 2017, 20:56:11 UTC |
e0362d8 | Chunyang Wen | 23 October 2017, 19:29:23 UTC | fix parameter set_value (#1006) * fix parameter set_value * replace np.array with np.asarray Former-commit-id: 4c07dbfbbb15e7c8ae7bb2249dc13eeed166b4d8 | 23 October 2017, 19:29:23 UTC |
ab89af2 | Daniel Hershcovich | 23 October 2017, 19:28:31 UTC | Fix creation of wheels on Windows (#1007) * Update to new PyPI upload API * Include DyNet library in wheel created on Windows Former-commit-id: bf09b4841d9155f243dfd7cee12a73094cd805a0 | 23 October 2017, 19:28:31 UTC |
7ff4e74 | shuheik | 23 October 2017, 19:28:04 UTC | Fix scala bindings (#1008) * adapt to f9ec2c3 * adapt to 830e0fb Former-commit-id: 0f476022853d5f5b95beeb0f93b6a4c12002054a | 23 October 2017, 19:28:04 UTC |
face8e7 | Chunyang Wen | 20 October 2017, 14:53:31 UTC | fix inproper use of DYNET_ASSERT (#1005) * #1004 replace DYNET_ASSERT with DYNET_ARG_CHECK properly * Fix format Former-commit-id: 25be4893dad9b85b56ec5fad057af3380f060e06 | 20 October 2017, 14:53:31 UTC |
b72d542 | Paul Michel | 20 October 2017, 09:05:02 UTC | Fix examples (#1003) * Link to example README in doc * Forgot to update requirements * Fix hyperlinks Former-commit-id: 8e775a8d840b392a9f775cdb8f51c309fab09cdd | 20 October 2017, 09:05:02 UTC |
c634a1a | Graham Neubig | 19 October 2017, 14:00:28 UTC | Logsumexp dim (#1001) * Added dimension option to tensortools logsumexp * Initial addition of logsumexp_dim * Added binding and bugfix * Added python bindings * Fixed build on windows? Former-commit-id: 6efe3b76685281f6bf7b44fa7be138da19244f8a | 19 October 2017, 14:00:28 UTC |
0f2e9e1 | Hong Wu | 19 October 2017, 11:14:42 UTC | Mutli-gpu support for Tensor::copy_element. (#999) Former-commit-id: c48031d067d4c3d43d0dfb1ad9dedcf36fc2c17a | 19 October 2017, 11:14:42 UTC |
6ef606f | Daniel Hershcovich | 19 October 2017, 11:14:21 UTC | Build Windows wheels on AppVeyor (#1000) * Build Windows wheel on AppVeyor * Add Python versions to AppVeyor Former-commit-id: 8d54af5fd2d5f08efbc0aedf346e6790c7b1c2f5 | 19 October 2017, 11:14:21 UTC |
c5d1684 | Graham Neubig | 19 October 2017, 01:27:46 UTC | Sum dim (#997) * Some modification of sum_dim to match moment_dim * Update of python doc * Fixed default batch value Former-commit-id: 830e0fb70d5cf826025e12340979f037751d0de4 | 19 October 2017, 01:27:46 UTC |
b3f26b0 | Paul Michel | 18 October 2017, 21:35:52 UTC | Remove useless external dependency (#998) Former-commit-id: a8fa71d2749804ca81b55e4600014ec1b1091ffa | 18 October 2017, 21:35:52 UTC |
768b99f | Hong Wu | 18 October 2017, 12:23:05 UTC | Add devices/xor-multidevice in example (#993) Former-commit-id: a5f230572688d36246b5937b41f42ebaadc207a4 | 18 October 2017, 12:23:05 UTC |
abd3676 | Hong Wu | 18 October 2017, 12:22:51 UTC | Fix several places missing support for multi-gpu. (#994) Former-commit-id: 4910c210bc9a3823cd981ad52fd30cdd56df7c09 | 18 October 2017, 12:22:51 UTC |
bddfedd | Graham Neubig | 17 October 2017, 16:52:12 UTC | Full multi-GPU support (#967) * A start at full multi-GPU support * Made sure appropriate device used before forward * add across device memcpy for copy_elements function * fix several bugs for matrixmultiply and Tensor::zero * fix several multi-gpu bugs * typo fix * clean up * Fixed multi-device example * Address comments in pr Former-commit-id: 9ab97efb3daa5c2dd8ca01bc51780dda6b0c265a | 17 October 2017, 16:52:12 UTC |
321bfbe | Daniel Hershcovich | 17 October 2017, 16:31:47 UTC | Change model parameters to be stored as shared pointers (#986) * Fix #930 and #954: model params are now shared_ptr * Add test for parameter access after deleting model * Keep reference to parent model in Python * Test deleting parent model * Fix model save/load test to handle shared pointers properly * Remove assertions, just verify no segfault occurs * Add back std:: prefix Former-commit-id: f9ec2c37a96c2b5036684e197ca0a66f2177c182 | 17 October 2017, 16:31:47 UTC |
56049b5 | Graham Neubig | 17 October 2017, 11:22:47 UTC | Updated Eigen version Former-commit-id: f865da5d1a21e747fafb3ac034407e4f444f3da8 | 17 October 2017, 11:22:47 UTC |
a9045bb | Daniel Hershcovich | 15 October 2017, 13:35:51 UTC | Run full Travis CI test suite only on cron jobs (#985) * Run full Travis CI test suite only on cron jobs * Only build the master branch, avoid double builds Former-commit-id: cabbb552df3798eca85dfe55f94320dd6519162b | 15 October 2017, 13:35:51 UTC |
0b104c4 | Daniel Hershcovich | 15 October 2017, 13:35:34 UTC | Run Python unit tests on AppVeyor (#984) * Create conda env on AppVeyor * Run Python unit tests on AppVeyor * PyPI deploy on AppVeyor on tags * Update AppVeyor badge link * Fix AppVeyor build directory Former-commit-id: dd29db7942297981816e763d8c33103c65732204 | 15 October 2017, 13:35:34 UTC |
19b2538 | Graham Neubig | 13 October 2017, 17:30:20 UTC | An attempt at more numerically stable softmax (#979) * An attempt at more numerically stable softmax * Made softmax faster on CPU (4x?) * Made log_softmax faster * Fixed bugs Former-commit-id: c1b3bae4d679726498fc2e6bd6fa96eeb7b76dd7 | 13 October 2017, 17:30:20 UTC |
d1c672d | Paul Michel | 13 October 2017, 16:57:42 UTC | fix Former-commit-id: 065dbaf3066480326083488440f3ceb9cbae1303 | 13 October 2017, 16:57:42 UTC |
614dfdd | Paul Michel | 13 October 2017, 16:57:03 UTC | Remove stale symlinks Former-commit-id: 50725955693741a38ff1db3313dea84f9d6f9ce6 | 13 October 2017, 16:57:03 UTC |
bdf4a42 | Paul Michel | 13 October 2017, 16:44:43 UTC | Fix path to tutorials in docs Former-commit-id: d50993150570bbe1766c117f2252e93dbe656630 | 13 October 2017, 16:44:43 UTC |
5365dd5 | msperber | 13 October 2017, 14:59:17 UTC | Multidim2 (#977) * fix problem with broadcasting batch dimension * overwrite_n for moment operations Former-commit-id: 0a9193a2928a7d0bc9dd6866c0d60dd3191d53fc | 13 October 2017, 14:59:17 UTC |
8c76a31 | Paul Michel | 13 October 2017, 13:31:13 UTC | Fix shape of lookup aprameters in python (#975) Former-commit-id: 1ca25fe9649397b8f3532abac56d4827f1dbb96e | 13 October 2017, 13:31:13 UTC |
085254c | Chunyang Wen | 13 October 2017, 13:30:14 UTC | add doc and reimplement Dim::truncate (#970) * remote use GPU:0 by default when use --dynet-gpu * fix error * Comment DeviceMempool enum * #945 1. add doc for Dim::truncate 2. reimplement Dim::truncate 3. add tests for dim Former-commit-id: 0c6c6de9faa65254263fbbf28760fb828a852111 | 13 October 2017, 13:30:14 UTC |
28d580f | Stefan Schweter | 13 October 2017, 13:26:27 UTC | Compile multiprocessing examples only when Boost is enabled (#978) * doc: mention -DENABLE_BOOST=ON option to compile multiprocessing c++ examples * cmake: compile multiprocessing c++ examples only when boost is enabled Former-commit-id: 29536c4204e37c85b2e9f77ad64d8f87df244130 | 13 October 2017, 13:26:27 UTC |
bf5729e | Chris Dyer | 10 October 2017, 13:47:52 UTC | comments, C++ style Former-commit-id: a69c5158ccfd7887ee3174d2e99924a445b26fea | 10 October 2017, 13:47:52 UTC |
0c35fe4 | Chris Dyer | 09 October 2017, 23:13:02 UTC | git rid of needless memcpy Former-commit-id: e0f39178a1df1319ad97aafd9430ac8d81fb1e78 | 09 October 2017, 23:13:02 UTC |
e145ce8 | Chris Dyer | 09 October 2017, 22:40:04 UTC | clean up exec code Former-commit-id: 707e8d01c8d3be62b2e2b9e4443e6b65e2d7e36c | 09 October 2017, 22:40:04 UTC |
27705e5 | Chris Dyer | 09 October 2017, 11:26:34 UTC | improve readability of complex code Former-commit-id: d06dae2d211e954797e2a361ed4d17490342a154 | 09 October 2017, 11:26:34 UTC |
d3b5013 | Chris Dyer | 09 October 2017, 10:57:58 UTC | cleanup C++ for readability Former-commit-id: e666b45a5b1ddc62ccb6bdaaccba6d286308684f | 09 October 2017, 10:57:58 UTC |
78cbe81 | Chris Dyer | 08 October 2017, 21:31:11 UTC | c++ style Former-commit-id: 3cb47b3cd85b4388ab4b1e808318483a8031a394 | 08 October 2017, 21:31:11 UTC |
db74246 | Chris Dyer | 08 October 2017, 21:25:25 UTC | better C++ style Former-commit-id: 4681c1795020e2e569632448d2bd18cc62a037ff | 08 October 2017, 21:25:25 UTC |
08d3bb5 | Chris Dyer | 08 October 2017, 19:40:19 UTC | fix cuda compile error i introduced earlier Former-commit-id: b992b4032bc47fb0bb8a49d689d7b25f677346a3 | 08 October 2017, 19:40:19 UTC |
ce532af | Chris Dyer | 08 October 2017, 13:52:01 UTC | C++ cleanup (don't pass vector<> by value) Former-commit-id: 07381369dd1277a2b951e137f8695cedf1da546c | 08 October 2017, 13:52:01 UTC |
a9a5666 | Chris Dyer | 08 October 2017, 13:28:19 UTC | small C++ fix Former-commit-id: 778c481761565148501c28070e3f2d883fda19e4 | 08 October 2017, 13:28:19 UTC |
b5c4893 | Chris Dyer | 08 October 2017, 13:17:08 UTC | small C++ fix Former-commit-id: 6fb48caf4ff695c9c95022f54a86efae074465d8 | 08 October 2017, 13:17:08 UTC |
44bfa11 | Chris Dyer | 07 October 2017, 10:05:45 UTC | fix uninitialized variable when built without CUDA Former-commit-id: 78b61357890dc56fce0ce2ca38bbe382a104aab2 | 07 October 2017, 10:05:45 UTC |
b4e8170 | Graham Neubig | 06 October 2017, 17:31:48 UTC | Fix eigen version (#966) * Changed Eigen version to work with pip on clang 9 * Updated setup * Update * Update Eigen version to compile Former-commit-id: c9cc4c7c3d261dda2065b320582038a4600af72f | 06 October 2017, 17:31:48 UTC |
b947387 | Graham Neubig | 06 October 2017, 15:29:30 UTC | Fix device string encoding on Python 3 (#964) * Fix device string encoding on Python 3 * Added missing conversion Former-commit-id: 4e9611cd824a011993cec1ad7f267edec6ab840d | 06 October 2017, 15:29:30 UTC |
b28ac1f | Graham Neubig | 06 October 2017, 15:18:52 UTC | Runs OSX compile first on travis (#965) Former-commit-id: bb91bdaba5172a626eb6dd5870006a2e2fa9bdae | 06 October 2017, 15:18:52 UTC |
b2305d7 | Graham Neubig | 06 October 2017, 13:35:15 UTC | Added check to matrix multiply Former-commit-id: b94f60334f6265ff9e13540a76c8827659320492 | 06 October 2017, 13:35:15 UTC |
56fe3b2 | Graham Neubig | 06 October 2017, 13:26:21 UTC | Fix doc for pairwise rank loss Former-commit-id: 7f4305b8df46343e89d351f90cffab8531de8156 | 06 October 2017, 13:26:21 UTC |
d816039 | Paul Michel | 05 October 2017, 18:56:40 UTC | Makes it possible to get a list of all parameters in python (#960) * Makes it possible to get a list of all parameters in python * Fix tests Former-commit-id: 227393d763e2d38f36dda405034d2ba6fdb8dedc | 05 October 2017, 18:56:40 UTC |
6760666 | Pengcheng YIN | 04 October 2017, 10:05:45 UTC | exception handling for model related API (#957) Former-commit-id: 719413aa7a198fa4f091951af2688e195b94f53c | 04 October 2017, 10:05:45 UTC |
7acfc28 | Chris Dyer | 03 October 2017, 11:20:53 UTC | Merge branch 'master' of https://github.com/clab/dynet Former-commit-id: 9fbe4f48c76e272a19733157b0a99b9a76c9c37c | 03 October 2017, 11:20:53 UTC |
3f617de | Chris Dyer | 03 October 2017, 11:20:40 UTC | make upgrades easier Former-commit-id: 7fca795a2abf5f01e0381ea59eec42aee951bd72 | 03 October 2017, 11:20:40 UTC |
c81e44f | Daniel Hershcovich | 02 October 2017, 12:35:31 UTC | Native Python on Linux on Travis CI instead of Conda (#950) * Add "u" suffix if Python is compiled with UCS-4 Otherwise we get "undefined symbol: PyUnicodeUCS2_DecodeUTF8" when trying to install a wheel built using this environment if it was first built with cp27-cp27m and then built with cp27-cp27mu, since they were using the same build directory. * Fix #941: native Python on Travis CI, not Conda Former-commit-id: 52c57c805ca2e6c66130ca1658b3177810f2b77b | 02 October 2017, 12:35:31 UTC |
547682c | Hong Wu | 01 October 2017, 13:50:52 UTC | hotfix for https://github.com/clab/dynet/pull/927 (#953) Former-commit-id: f6be87560a24892d2f0a4260883822b2be26fb06 | 01 October 2017, 13:50:52 UTC |
c79848e | Brad Skaggs | 30 September 2017, 18:20:21 UTC | Use a release version of Eigen for building Python package. (#608) * Remove need for hg for building Python package by using an Eigen release. * Fixed some bugs Former-commit-id: d160a7b3f665efb94cb4115c8069e3d2b262204e | 30 September 2017, 18:20:21 UTC |
fdb903a | Hong Wu | 30 September 2017, 18:18:12 UTC | Multi-gpu support for basic add_parameters function. (#947) Former-commit-id: c43f52b7db55d689f26bf6e2995f4ac5f1c9bbd7 | 30 September 2017, 18:18:12 UTC |
0c3ff7b | Graham Neubig | 30 September 2017, 18:17:50 UTC | Restructure examples (#927) * Restructure examples * Fixed compiling C++ examples Former-commit-id: 5049e5995f169fe1798139e1ca4dc98a7c0c4317 | 30 September 2017, 18:17:50 UTC |
d9f110a | msperber | 30 September 2017, 18:17:09 UTC | Multidim operations & speed-improvements (#776) * some initial code * implemented multidim moments * broadcasting for cmult * cmult: use templates to get rid of code duplication * implemented cadd with broadcasting support * broadcasting for cdiv * speedup and todos * weight noise for RNNs * first stab at lstm node, forward pass * added comment * LSTM forward: memory layout of h/c vector + restructuring/fixes * bugfix + prepared unit test * summarized 4 matrix multiplication into one operation; forward computes without error * forward tests passing * lstm node forward pass working with minibatches * lstm 3-part node: forward working & tested * backward pass for vanilla_lstm_h and vanilla_lstm_c * lstm-gates: parts of backward pass * some more missing parts * removed old code, cleaned up tests * vanilla_lstm_c: backward passing test * some testing / fixing * some comments * CPU matrix multiply * re-added LSTM node * lstm gates fwd passing test * tests * lstm_h backward passing gradient checks * lstm gates backward: replaced contraction by matrix multiply * lstm gates bwd implemented but not yet passing tests * fixed math in lstm_h backward * fixed the same math error in lstm_gates bwd * improved tests; all tests passing * unified CPUMatrixMultiply / CUDAMatrixMultiply -> MatrixMultiply * python interface for LSTM node * changed use of slice for potential speed improvement * speed-up: replace .sum() by manual summation over batches * speed-up by replacing shuffle with reshape * speed by replacing shuffle with transpose * marked places that need speed-up * speed up for outer product * Some fixes for GPU * Comment out extra scratch memory * add CompactVanillaLSTMBuilder * added weight norm to vanilla_lstm_gates * updated doc & python interface * add weight noise to CompactVanillaLSTMBuilder * added missing free() of scratch allocator * initial code to integrate dropout into lstm nodes * fix missing CwiseSum * integrated dropout into LSTM node * removed unused variables * removed unused variables * speed up StdDimension by precomputing mean in temporary memory * speed up cdiv backward pass by precomputing square in scratch memory * improved speed of weight_norm by precomputing stuff before broadcasts in temporary memory * deprecated mean_elems, moment_elems, std_elems, moment_batches, mean_batches, std_batches * fixed scope warning * remove unused code * fixed fast LSTM * more flexible broadcasting for CwiseSum * more flexible broadcasting for CwiseMultiply and CwiseQuotient * undo fast-lstm workaround * deprecated mean_elems, std_elems, moment_elems, mean_batches, std_batches, moment_batches * small bugfix * debugging: Revert "undo fast-lstm workaround" This reverts commit c1c4dae57e8adbfb70476946821ebfed67f3b704 [formerly 09d3050fb2a1632e3a56128517bf7494006bc9ca]. * removed nodes-arith-scalar * remove cadd (operator+ now has same functionality) * avoid broadcasting for cmult if dimensions are equal * same for cmult forward * same for cmult backward * same for cdiv * debug * debug * debug * debug * debug * add missing files * Added escaping for backward as well * Fixed several bugs in backward * De-deprecated some functions Former-commit-id: 11963baead216509cc89dd43002a5fb6db5b1cfa | 30 September 2017, 18:17:09 UTC |
b9cae34 | Hong Wu | 29 September 2017, 14:45:08 UTC | Dynet-906. Fix hardcoded build folder name. (#910) * dynet-906. Fix hardcoded build folder name. * use PYTHON_DIR * fix CI script * fix path * fix osx path * make path format consistent Former-commit-id: 6e1e9f496973c7f77d85aea44481e9a50c654249 | 29 September 2017, 14:45:08 UTC |
fee92d5 | Graham Neubig | 29 September 2017, 13:59:25 UTC | Check cwise multiply batch dims (#944) Former-commit-id: 8a5d091af12389b473f64de61d79718418164b5e | 29 September 2017, 13:59:25 UTC |
3e02fcf | Daniel Hershcovich | 29 September 2017, 13:31:22 UTC | AppVeyor badge (#932) Former-commit-id: 24a67b45d8fef7c84052ff6d6c9d86e023862194 | 29 September 2017, 13:31:22 UTC |
b3058ef | Pengcheng YIN | 29 September 2017, 13:30:58 UTC | FIx typo in the doc string of pick_batch_elems (#943) Former-commit-id: 076cefed0d275950168d61a2a106eda52280d5f1 | 29 September 2017, 13:30:58 UTC |
0b82b7b | WZY | 29 September 2017, 06:46:39 UTC | Fix typo in tutorial-1-xor.ipynb (#946) Former-commit-id: e47e9b44809d65a2cf07db343805b2cc258bea5f | 29 September 2017, 06:46:39 UTC |
9fb2342 | Paul Michel | 27 September 2017, 21:57:54 UTC | Exception passing in RNN (#939) Former-commit-id: d14858a2dba1316ab1e282271120b4dc42da8cca | 27 September 2017, 21:57:54 UTC |
ca830aa | Chunyang Wen | 27 September 2017, 15:03:51 UTC | add doc for DeviceMempool (#936) (#937) * remote use GPU:0 by default when use --dynet-gpu * fix error * Comment DeviceMempool enum Former-commit-id: 300fb8eebeb0b679b6a7556c91cc63df25aadaed | 27 September 2017, 15:03:51 UTC |
738e1aa | Edgar Chen | 27 September 2017, 13:07:38 UTC | Add const_parameter to python docs (#935) Former-commit-id: 7770887e937d93b34272c68f61230c46cae6e36a | 27 September 2017, 13:07:38 UTC |
c2a4342 | Chunyang Wen | 27 September 2017, 07:18:34 UTC | replace --dynet-gpu with --dynet-gpus 1 instead of --dynet-devices GPU:0 (#934) Former-commit-id: d783e2ed5cb9612c12503a98a745597232431828 | 27 September 2017, 07:18:34 UTC |
b548c04 | Daniel Hershcovich | 25 September 2017, 18:51:46 UTC | Compile with CUDA on Travis CI (#911) * Disable warnings, as log is too big for Travis CI * Test on Travis CI with CUDA Build DyNet with -DBACKEND=cuda. Exclude on osx. On Travis CI, force tests to run on CPU (no GPU) * Disable tests when compiling with CUDA on Travis for now Former-commit-id: ae79d2e7c033e0976f930594adf5ffd17c2ac340 | 25 September 2017, 18:51:46 UTC |
acd98ec | Daniel Hershcovich | 25 September 2017, 18:50:57 UTC | Continuous integration for Windows using AppVeyor (#914) * Fix "conversion from 'double' to 'float' requires a narrowing conversion" * Fix #913: Continuous integration for Windows using AppVeyor Former-commit-id: 03aaa6c39f5f29a735e263abda3003eb694013fc | 25 September 2017, 18:50:57 UTC |
a614a19 | Daniel Hershcovich | 22 September 2017, 17:59:18 UTC | Allow adding model parameters after update (#893) * Failing test for #889 * #889: allocate more shadow parameters as needed Fix indentation in test Fix missing semicolons Fix return type * Test both SGD and Adam, and 100 parameters to be sure * Fix #889: segmentation fault due to l2 norm scratch size mismatch * Fix potential segmentation fault in project_weights too * Less repetition by passing reference to vector Former-commit-id: a8b496ac99badfd2c32ed01469728654883942ca | 22 September 2017, 17:59:18 UTC |
129d39e | xunzhang | 22 September 2017, 14:47:07 UTC | fix indent Former-commit-id: 77ebe199df946ab44ab53a1ee0adbf84c4b8ba76 | 22 September 2017, 14:48:45 UTC |
9f7eda9 | Graham Neubig | 22 September 2017, 14:47:07 UTC | Merge branch 'master' of github.com:clab/dynet Former-commit-id: 87a6daeb9b5318cf14377b53b3bf0f8b6439496f | 22 September 2017, 14:47:07 UTC |
1d5fefd | Graham Neubig | 22 September 2017, 14:46:41 UTC | Hotfix for Python3 Former-commit-id: 5d44a8d948e425bd1504760a3c091e48c33a2067 | 22 September 2017, 14:46:41 UTC |