Closed tomekd closed 6 years ago
what's the compile flags for this configuration?
nevermind, ive got it cmake .. -DCUDA=off
Build was fixed!
Build 'amunmt_compilation_cpu' is failing!
Last 50 lines of build output:
[...truncated 7.80 KB...]
[ 56%] Building CXX object src/amun/3rd_party/yaml-cpp/CMakeFiles/libyaml-cpp-amun.dir/contrib/graphbuilder.cpp.o
[ 57%] Building CXX object src/amun/3rd_party/yaml-cpp/CMakeFiles/libyaml-cpp-amun.dir/node_data.cpp.o
[ 58%] Building CXX object src/amun/3rd_party/yaml-cpp/CMakeFiles/libyaml-cpp-amun.dir/emitterutils.cpp.o
Scanning dependencies of target extract_lex
[ 60%] Building CXX object src/amun/3rd_party/extract_lex/CMakeFiles/extract_lex.dir/extract-lex-main.cpp.o
[ 61%] Building CXX object src/amun/3rd_party/yaml-cpp/CMakeFiles/libyaml-cpp-amun.dir/parse.cpp.o
[ 62%] Building CXX object src/amun/3rd_party/extract_lex/CMakeFiles/extract_lex.dir/utils.cpp.o
[ 64%] Building CXX object src/amun/3rd_party/yaml-cpp/CMakeFiles/libyaml-cpp-amun.dir/ostream_wrapper.cpp.o
[ 64%] Built target libyaml-cpp-amun
[ 65%] Building CXX object src/amun/3rd_party/extract_lex/CMakeFiles/extract_lex.dir/exception.cpp.o
[ 66%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/history.cpp.o
[ 67%] Building CXX object src/amun/CMakeFiles/cpumode.dir/cpu/dl4mt/gru.cpp.o
[ 69%] Linking CXX executable ../../../../extract_lex
[ 69%] Built target extract_lex
[ 70%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/hypothesis.cpp.o
[ 71%] Building CXX object src/amun/CMakeFiles/cpumode.dir/cpu/dl4mt/model.cpp.o
[ 73%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/loader.cpp.o
[ 74%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/logging.cpp.o
[ 75%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/output_collector.cpp.o
[ 76%] Building CXX object src/amun/CMakeFiles/cpumode.dir/cpu/dl4mt/encoder_decoder.cpp.o
[ 78%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/printer.cpp.o
[ 79%] Building CXX object src/amun/CMakeFiles/cpumode.dir/cpu/nematus/encoder.cpp.o
[ 80%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/processor/bpe.cpp.o
[ 82%] Building CXX object src/amun/CMakeFiles/cpumode.dir/cpu/nematus/model.cpp.o
[ 83%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/scorer.cpp.o
[ 84%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/search.cpp.o
[ 85%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/sentence.cpp.o
[ 87%] Building CXX object src/amun/CMakeFiles/cpumode.dir/cpu/nematus/gru.cpp.o
[ 88%] Building CXX object src/amun/CMakeFiles/cpumode.dir/cpu/nematus/transition.cpp.o
[ 89%] Building CXX object src/amun/CMakeFiles/cpumode.dir/cpu/nematus/encoder_decoder.cpp.o
[ 91%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/sentences.cpp.o
[ 92%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/types.cpp.o
[ 93%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/utils.cpp.o
[ 94%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/vocab.cpp.o
[ 96%] Building CXX object src/amun/CMakeFiles/libcommon.dir/common/translation_task.cpp.o
/var/lib/jenkins/workspace/amunmt_compilation_cpu/src/amun/common/translation_task.cpp:4:33: fatal error: thrust/system_error.h: No such file or directory
compilation terminated.
src/amun/CMakeFiles/libcommon.dir/build.make:542: recipe for target 'src/amun/CMakeFiles/libcommon.dir/common/translation_task.cpp.o' failed
make[2]: *** [src/amun/CMakeFiles/libcommon.dir/common/translation_task.cpp.o] Error 1
make[2]: *** Waiting for unfinished jobs....
CMakeFiles/Makefile2:182: recipe for target 'src/amun/CMakeFiles/libcommon.dir/all' failed
make[1]: *** [src/amun/CMakeFiles/libcommon.dir/all] Error 2
make[1]: *** Waiting for unfinished jobs....
[ 96%] Built target cpumode
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2
Build step 'Execute shell' marked build as failure
Archiving artifacts
GitHub Issue Notifier: Build has started failing again, reopend GitHub issue #109
Changes since last successful build:
[hieuhoang] 860326e7855a7dbb68db8177eaaa656d0e4aadbb - sm_50. Turn off marian compile
[hieuhoang] 5d7d4efd61ffb1c2038bbb10a02294c6ab993ca5 - use uint instead of size_t
[hieuhoang] b821b986201b3140e98c61b46e8c092838ea4211 - start using array
[hieuhoang] 4531e784a134fbdd5324ff0e7c38d53617da1577 - use array for cost
[hieuhoang] 813851b7fa61c1847da40f44a87cbaad92f227e7 - use array for nBest
[hieuhoang] 1882ec136f758414f21b3bd7ad75cc6a21f54f7b - use array for mapping
[hieuhoang] 5a543e2e55c656d03915cffcfd803750a6c170a7 - add Assemble() for Array
[hieuhoang] b390685a61efa50cf615a163803ac5f1b6a07fa8 - use Array in Decoder::Filter()
[hieuhoang] 802d97d3dbf394727ff2c1fe3461be5b979e643a - use Array in Embeddings::Decoder::Filter()
[hieuhoang] 98e3aef319366ab367aca359ffed2a4d474acdf8 - completely delete DeviceVector
[hieuhoang] 93379a0858d0fad600d087593ab46bcfc89f2c90 - completely delete HostVector
[hieuhoang] 99e41abef537259fc77edb6c45b5266887af2b17 - sm_35
[hieuhoang] 6d3b79754dd53cdbe44b7fc02e0ba5cb8e884043 - delete unused constructor
[hieuhoang] 6b79df21d2d3bcf42aac97c0138cf57ac8f18d9b - delete unused functions
[hieuhoang] f2701597efc63f457abe00365392b59a7fc7c391 - delete raw_pointer_cast
[hieuhoang] 05c841e49d6c84d71a8f6ac9683c9de264a79fb6 - start test of cudaMemCpy sm25 and sm50
[hieuhoang] 288082b98949a61ccd23968a653374c14037f06c - start test of cudaMemCpy sm25 and sm50
[hieuhoang] 1631e5c35a4ba103776b91bf51daa6e9687fbc9f - test of cudaMemCpy sm25 and sm50
[hieuhoang] c49b316794db5e78eb4e9ca06bfb7747538620cd - debug
[hieuhoang] 8e8c882df70f68bec8d03e7d1aa04caef81da6bc - debug
[hieuhoang] 81295d42bea245cfda8e15163a3c23817e9732e1 - debug
[hieuhoang] ef4d2c8c232d15d3ed968fdb9c4e3182ec921bfe - debug kernel
[hieuhoang] b52005fe23bec266bfda873ee8b9c3ee9ad140b2 - generate code for both sm_35 and sm_50
[hieuhoang] 94d80c8cfaeaf39992c242ba468cdcf6335f7d49 - more compile flag
[hieuhoang] 89ce664765575d816cb3d885def125fe0ea57ee2 - max compatibility for cuda 9
[hieuhoang] e7ae2bee9098f25a0dcf7891c95026f08de3e930 - turn marian compilation back on
Build was fixed!
Build 'amunmt_compilation_cpu' is failing!
Last 50 lines of build output:
[...truncated 1.83 KB...]
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Forcing compilation without CUDA.
-- Looking for pthread.h
-- Looking for pthread.h - found
-- Looking for pthread_create
-- Looking for pthread_create - not found
-- Looking for pthread_create in pthreads
-- Looking for pthread_create in pthreads - not found
-- Looking for pthread_create in pthread
-- Looking for pthread_create in pthread - found
-- Found Threads: TRUE
-- Boost version: 1.58.0
-- Found the following Boost libraries:
-- system
-- filesystem
-- program_options
-- timer
-- iostreams
-- python
-- thread
-- chrono
-- regex
-- date_time
-- atomic
-- Found PythonLibs: /usr/lib/x86_64-linux-gnu/libpython2.7.so (found suitable version "2.7.12", minimum required is "2.7")
-- Found Python
-- Found ZLIB: /usr/local/lib/libz.so (found version "1.2.8")
-- Found Git: /usr/bin/git (found version "2.7.4")
-- Git version: c07f42a
-- Found SparseHash: /usr/include
-- Configuring done
CMake Error at src/amun/CMakeLists.txt:11 (add_library):
Cannot find source file:
cpu/decoder/best_hyps.cpp
Tried extensions .c .C .c++ .cc .cpp .cxx .m .M .mm .h .hh .h++ .hm .hpp
.hxx .in .txx
-- Generating done
-- Build files have been written to: /var/lib/jenkins/workspace/amunmt_compilation_cpu/build
Build step 'Execute shell' marked build as failure
Archiving artifacts
GitHub Issue Notifier: Build has started failing again, reopend GitHub issue #109
Changes since last successful build:
[hieuhoang] 407c16e120ac7fc7fffdc5c698679ade239870e2 - move to cpp
[hieuhoang] d0ee1f2f1a3a931982d8988f691a90e6555c937a - comment out stats collection
[hieuhoang] a1aae560d5ba233f04f112ed422b02ff19f9a5e8 - reduce output to cerr
[hieuhoang] d851af28946a5a8de4ac1b5ceac518c63454c9cc - timings for cpu code
[hieuhoang] 2773b932481f2e7231ad02980ffeb8fe0619f039 - comment out timings output
Build was fixed!
Build 'amunmt_compilation_cpu' is failing!
Last 50 lines of build output:
[...truncated 157 B...]
Cloning the remote Git repository
Cloning repository https://github.com/marian-nmt/marian
> git init /var/lib/jenkins/workspace/amunmt_compilation_cpu # timeout=10
Fetching upstream changes from https://github.com/marian-nmt/marian
> git --version # timeout=10
> git fetch --tags --progress https://github.com/marian-nmt/marian +refs/heads/*:refs/remotes/origin/*
> git config remote.origin.url https://github.com/marian-nmt/marian # timeout=10
> git config --add remote.origin.fetch +refs/heads/*:refs/remotes/origin/* # timeout=10
> git config remote.origin.url https://github.com/marian-nmt/marian # timeout=10
Fetching upstream changes from https://github.com/marian-nmt/marian
> git fetch --tags --progress https://github.com/marian-nmt/marian +refs/heads/*:refs/remotes/origin/*
> git rev-parse refs/remotes/origin/master^{commit} # timeout=10
> git rev-parse refs/remotes/origin/origin/master^{commit} # timeout=10
Checking out Revision 46b0232339ef13efc06b3458facb2cba831411cc (refs/remotes/origin/master)
Commit message: "Update README to reflect main repo"
> git config core.sparsecheckout # timeout=10
> git checkout -f 46b0232339ef13efc06b3458facb2cba831411cc
> git rev-list bda9b18b6cede63b0476e9c144da3f62f03515b1 # timeout=10
[amunmt_compilation_cpu] $ /bin/sh -xe /tmp/jenkins5171408368335421806.sh
+ mkdir build
+ cd build
+ cmake -DCUDA=OFF ..
-- The CXX compiler identification is GNU 5.4.0
-- The C compiler identification is GNU 5.4.0
-- Check for working CXX compiler: /usr/bin/c++
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Check for working C compiler: /usr/bin/cc
-- Check for working C compiler: /usr/bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compile features
-- Detecting C compile features - done
-- Project name: marian
-- Project version: v1.6.0+46b0232
CMake Error at /usr/share/cmake-3.5/Modules/FindCUDA.cmake:617 (message):
Specify CUDA_TOOLKIT_ROOT_DIR
Call Stack (most recent call first):
CMakeLists.txt:40 (find_package)
-- Configuring incomplete, errors occurred!
See also "/var/lib/jenkins/workspace/amunmt_compilation_cpu/build/CMakeFiles/CMakeOutput.log".
Build step 'Execute shell' marked build as failure
Archiving artifacts
GitHub Issue Notifier: Build has started failing again, reopend GitHub issue #109
Changes since last successful build:
[Roman Grundkiewicz] 22d9477ef8405b5b8a77544b01771671dd9cde4a - Postpone concatenations in GRUNematus; add gruNematusOps
[Roman Grundkiewicz] b75615baeeb05b92c16436f755b7f1874af67bcc - Towards encoder layer normalization
[Roman Grundkiewicz] b1afc5278d46c8d39d73720b37d5678a3792c5d4 - Pass layerNorm_ variable into GRUNematus kernel
[Roman Grundkiewicz] 8dfcf3a5457c93563979b2e2ca6b72635efddfe8 - Add encoder layer normalization
[Roman Grundkiewicz] 2918d95e02be8a3f196e645b531142966c6daddf - Add working Nematus model with layer normalization, but break s2s
[Roman Grundkiewicz] 8233706b5d87b3b4b1cc485033bd01304b1bcd23 - Restore non-Nematus layer normalization
[Roman Grundkiewicz] fc48c3c47b56b80dceb78176619860142bad0cac - Refactorize GRUNematus cell
[Roman Grundkiewicz] d1b82ade77b596a3fe7c5515dc0931689221f513 - Restore Nematus model without layer normalization
[Roman Grundkiewicz] 76e6832b83978d6fbb061dec2b049a73ad37513b - Refactorize GRUNematus unit; use one GRU kernel
[Roman Grundkiewicz] 7e07b0605127bae0c9cace50b8d0a1472de16e13 - Add S2S-Nematus mapping
[Roman Grundkiewicz] dc4b8c0dbe58302b603967fb8a742f9ee374e09e - Make b0 and bx0 constant
[Roman Grundkiewicz] e3935b0c7f47c881e50b72c3a87e8474068613ab - Remove sparse b/bx nodes
[Roman Grundkiewicz] 00145dd65a1ed313793bd7740c886bff5598c547 - Refactorize Nematus model name mapping
[Roman Grundkiewicz] c132bca562698335c33b50475fd71a7439c49698 - Restore threading in translator
[Roman Grundkiewicz] 1f9fef583672400c066f0302ed80fa8136b81bdc - Show commit SHA with the project version
[Ulrich Germann] 8d89b9ca362245c77d0ac8de203dc7b0b7d64196 - Bug fix in corpus.cpp: i < vocabPaths.size() - 1 => i + 1 <
[Ulrich Germann] 1dde94e9c44532a7ad35c68aba48bbab62f6466b - Sanity checks with respect to vocabularies and their dimensions.
[Roman Grundkiewicz] b3d87151cc1812ac848c8dbc133babd243d8cbe1 - Print port number
[Roman Grundkiewicz] 53c80e0d37c3e37b8c9bdf821b890c285517e746 - Update amun.yml for deep Nematus models
[Roman Grundkiewicz] 9c307606152fc28a7d017c39d8621433e6485034 - Add script injecting s2s options to nematus .npz models
[Roman Grundkiewicz] faf784f9f5e15bf6440e5d8bfa6f3cf9a93bd062 - Fix --relative-paths option
[Marcin Junczys-Dowmunt] 4898020b3aadde0956fa833ff9fadded27b469f0 - sum in debug
[Marcin Junczys-Dowmunt] f3f95d4a4e85168fbf44feeda24459f51235501c - fix rows op
[Marcin Junczys-Dowmunt] ec1d7fd2ad8e2e17fdc06ced3d4074dc6909cac5 - fixed error with too small fakeInput for deep transition cells
[Roman Grundkiewicz] 4e4e508cbe2220610061341065889ed6849ebacf - fix empty input for deep transition network cells
[Roman Grundkiewicz] efd46910d2911aec44adfc4a1c8d182c1e861dd8 - Change max vocab size to 100,000
[Roman Grundkiewicz] 040caa0f6e10bc2a4d65ebff68bff2e502f35fa7 - Remove unwanted newline
[Marcin Junczys-Dowmunt] bdaa4e4bd85f6c0a65eb14b1ed022d4771ad3329 - batched sgemm
[Marcin Junczys-Dowmunt] 5cbf89b8cb2f7179b2cac14b87c2f2ef401e5e10 - towards a transformer-style encoder
[Marcin Junczys-Dowmunt] e28281bd746a01436bd1a60a92a4deb64907855d - fixed positonal embedding
[Marcin Junczys-Dowmunt] 2599864de1befdbfd2caa852b4f1590889f5b324 - better masking and multiply embeddings
[Marcin Junczys-Dowmunt] f2beb4f083ac426641b969ad977fc99e096daa85 - catch NaNs
[Marcin Junczys-Dowmunt] 66cdae04bc17a08c5f9d6f4d978ad6c73f751034 - broadcasting for softmax mask
[Marcin Junczys-Dowmunt] 5f74749ab0435ef3c9d2238eb860cc1a97ce8b8e - correct mask for fake batch
[Marcin Junczys-Dowmunt] 11e3f44ff87fcfa5d34f6f4d506b9db21db5f174 - clean-up transformer encoder
[Marcin Junczys-Dowmunt] d6d9c85b584054dbcb041cbf4d11c04ad03419a4 - fix batch width
[Marcin Junczys-Dowmunt] c59d45a15d5a2e2984f79b9b173d495b9e05373b - comments and add back multi-gpu translation
[Marcin Junczys-Dowmunt] b252d2df804bf6f5cdc9a823db716448feb4375f - rearrange normalization
[Marcin Junczys-Dowmunt] 90ee3b1e130734796ad314bb742d440d2163edff - re-enable FFN
[Marcin Junczys-Dowmunt] 41039ba9aa9d9f642390342bd7300acec91e12c4 - larger hash for params and consts
[Marcin Junczys-Dowmunt] fa2e827873e7a79c187e9afc9c048d8a690363bf - add dropout for embeddings, change relu/LN
[Marcin Junczys-Dowmunt] 5f8cbe7f006b000aa14101d7b05439ecfdd90181 - source word dropout
[Marcin Junczys-Dowmunt] 00dcf68c5b918332b8d7f5fd9f5a521455bafcc0 - transformer decoder, first attempt
[Marcin Junczys-Dowmunt] 4efe8c2c4118f631014f405a7d1fc2c0502644e4 - reset optimizer stats with lr decay
[Marcin Junczys-Dowmunt] ed0c5fc79f6914b0022fb697e9a658e0cfc89fab - google transforer-style warmup lr-schedule
[Marcin Junczys-Dowmunt] ab3a793f70a001e6b027b9a66758f7e99f83a90f - modify transformer schedule
[Marcin Junczys-Dowmunt] 23a9e8c66c7e656c8c8a7acb897f9300eabec267 - register all optimizers in scheduler
[Marcin Junczys-Dowmunt] 8356fe2a0debd291b791af91bf2b571af8070296 - remove debug message
[Marcin Junczys-Dowmunt] 494c95533f438c52017b0b8da16ff27f44444c84 - fix missing scheduler registration for multi-gpu training
[Roman Grundkiewicz] 50d5e43836d7a71ba11123d665de5a8c635f378b - fix optimizer resets
[Marcin Junczys-Dowmunt] e42c1fbc0652846536f73bf0791eb43664e71031 - better hashing
[Marcin Junczys-Dowmunt] 444309fa9e7614e43dbbebdfa51df0dea5b70321 - hash collision resolution
[Marcin Junczys-Dowmunt] b2d93dfeacfd83a886e98f191473fccc449b4b97 - collision resolution
[Marcin Junczys-Dowmunt] adf8f955f9fa158da634409d7e84333ced97aecd - GRU-style gates for transformer
[Marcin Junczys-Dowmunt] b08687e5689d9218dc6b3dcac36570271154f093 - combine self-att and att in gate
[Marcin Junczys-Dowmunt] 9c5443f867c928c4ded04de6150c5ec8ccf2dd5a - go back to original transformer implementation
[Marcin Junczys-Dowmunt] 9f79c919dc886408d7180e207f60a2a3aa9af3d6 - add biases
[Roman Grundkiewicz] 621a7ddc0319bc728c6173e2ebe02cae5d2ea8f7 - Set proper eps in Nematus LN; fix GRUNematus
[Marcin Junczys-Dowmunt] aed9f48660bdf504a2b0db6ad6261f17e679c591 - more complrex transformer
[Roman Grundkiewicz] 683b1dd68fc5e94194298cc4fd6e315f27bd25c2 - Set proper eps for Nematus LN; fix LN in GRUNematus
[Marcin Junczys-Dowmunt] 2746829d883c397567c59abe72bd23c0054c847b - separate transformer models
[Marcin Junczys-Dowmunt] ba95dd5e707ae2ac6f52f95a0c435336e1cd7ea0 - decoder masking
[Marcin Junczys-Dowmunt] 57a826ed5bf2efbd5354f54cc75300869a776617 - comments on masking
[Marcin Junczys-Dowmunt] 2de0c09299ace80c2a6b999d305446741d4ea961 - add -inf mask to transformer_gru
[Marcin Junczys-Dowmunt] 8ec7f4647b356bf47b483eecaa78d0f7ec7d5261 - add lr to scheduler
[Marcin Junczys-Dowmunt] 8ab6928ff39305bd8751f07b6b250ffd1a244679 - more changes for transformer model
[Marcin Junczys-Dowmunt] 03dc42a26e891e6d990135e972b929b07ac912c6 - highway connections in transformer
[Roman Grundkiewicz] 4dacccb691ec4cbf55afc7b4e9bbff930666df0f - add inference
[Marcin Junczys-Dowmunt] 0b7e110473775be986e73e65dc0ab0c41b6b567d - add more LN
[Marcin Junczys-Dowmunt] da41acdb74ecc104005d1b03c49d6a013653c503 - special kernel for highway connections
[Roman Grundkiewicz] 108316603150d76dcfb6e634e066bef3bf9b9156 - Set a fixed seed in Iris example
[Roman Grundkiewicz] 9abd56ad81276d63f1e21c0828867006f88b47e4 - Use absolute path in Iris example
[Marcin Junczys-Dowmunt] 2abdd4b0895274e763a59302bc637b9ed2929918 - add perplexity
[Roman Grundkiewicz] 0965ab76b2b0584b98ba94afe86777ae1e1c392f - Add --optimizer-params option
[Roman Grundkiewicz] de6f03ee5072ce5c31979b634fb734a84f4714f8 - Remove keywords from optimizers
[Roman Grundkiewicz] c3b432a7d5493d4d065dff621dcff9e2c808a889 - Refactorize optimizers
[Marcin Junczys-Dowmunt] 67f87fc871ae2803cfaea3a441f4b2756a8fa9b7 - cleaned up validator and costs
[Marcin Junczys-Dowmunt] 9e4bf68fb473d331c0666ee4c712a7b59642275b - fix lr warmup
[Marcin Junczys-Dowmunt] 43afe059b71865b64d4a6e09ed900c1d13bfee1c - switch on proper highway connection
[Marcin Junczys-Dowmunt] bf16f10b4d2026f96cb57fa9792db322cf956b2e - add back LN
[Roman Grundkiewicz] 884289c58a09399cee0df96cd03d59c6bb53c571 - Fix Jenkins job name
[Roman Grundkiewicz] f35338a3de2a3470bbc385d720472b455890cf40 - fix transpose
[Marcin Junczys-Dowmunt] d7f40b10fea042283bde25c3f9a77066cac1120c - google style learning schedule
[Roman Grundkiewicz] de300f5c576e4a74377d9b70d067a997902057f5 - Fix optimizer params
[Marcin Junczys-Dowmunt] de101b797da581a2b120825d5fbf324765abf9bc - repeated warmup, fixed gradients of gemm with transposed matrices
[Marcin Junczys-Dowmunt] 2b3fd1810f05d40b76703bb4799b0490a8dac9b8 - change format to scientific notation
[Marcin Junczys-Dowmunt] 1b07cea5938f21169f72ef7d4d85de7ae998604d - add missing decay multiplier
[Marcin Junczys-Dowmunt] 296efd6845a405d3ab6c20256d07af5a9aac4f2d - fix google warmup
[Marcin Junczys-Dowmunt] ebd53cf54c49499f25428d9db8ded55f0c7238ff - separated multiplies for google-style decay
[Roman Grundkiewicz] 99b189ff6beef54df13dc5c4134a1edea6a04a74 - Add 'translation' action as validation metric
[Roman Grundkiewicz] 9f0708da4c4ceb4b515e2ea29ace7663b5a77a4e - Check for options in BeamSearch
[Roman Grundkiewicz] 5cbd77c38a506a4d9399d3b2b4d5ed78940f73fe - Add draft translation as a validation metric
[Roman Grundkiewicz] 31782faed2cbefbe536d57ba02ef833a834a79d6 - Refactorize Corpus class
[Roman Grundkiewicz] 3dc3e3eb4c508a1e2d28f8105a9450cedf26750f - Fix translation in validator
[Tomasz Dwojak] 7e1ce9aa26ad49dea0907d4c6ab5a7c553ced607 - Add Model Base
[Tomasz Dwojak] 213e2e7f021114505538cfe2f0bb07663a762049 - Replace EncoderDecoder with ModelBase
[Tomasz Dwojak] 0ca70a8ae386ff9a215bb6d56429a5766e590e06 - Fix mnist examples
[Tomasz Dwojak] 24ea42dcb63f32947d7b49c7676739b62cb79520 - Add mnist models to model factory
[Roman Grundkiewicz] 0e0a0db26613bbbc69a530a89e9c5790ccff66c0 - Use builder_ in translation action
[Roman Grundkiewicz] 90f454ef14b04491f884696ae9d3fc0cadb172e2 - Use inference=true during validation
[Marcin Junczys-Dowmunt] cb54a2573368aa4e56f4bf677a64681c4a16cbac - inference passing
[Roman Grundkiewicz] 279ea0593c086b9b2edf9a01044877017c7633a3 - Fix an issue with no default validation metric
[Roman Grundkiewicz] e14f8a9a0b6401c6814ce317dd56cf5050ff3766 - Remove S2S validator
[Roman Grundkiewicz] 74493149bddc8c9400f54c79485c9948bc1192cb - Add translation options for validation
[Roman Grundkiewicz] e3748c6802d48e2987c609b945fb272adbcd066a - Use a post-processing script after validating translation
[Roman Grundkiewicz] 3fe8866f9239734d96a38ba0c44a25e9980bb98f - Refactorize validators
[Roman Grundkiewicz] 7d8ae22ae7f12e6d1089cc32716b5c30405469a4 - Update .clang-format
[Marcin Junczys-Dowmunt] a4523f4cb3695fb2f1eb8c439258787fe2b82432 - towards beam search with transformer model
[Roman Grundkiewicz] aea3cc90943575abff1c7743dc884b1f860011e6 - Use temporary files during validating translation by default
[Marcin Junczys-Dowmunt] a93951eecd5b335cc0519d7aa96cb9475392ef16 - scale up context to match batch size for translation
[Roman Grundkiewicz] a18fb01340f60b31c9c59124e33030f7bb010475 - Move model_base.h from common/ to models/
[Marcin Junczys-Dowmunt] b920f18221ed3efcce814982c85043edee15b8e4 - move masking
[Roman Grundkiewicz] 6b11f617870284103dc4ece967570e6643be98c4 - Fix MNIST examples
[Roman Grundkiewicz] b0630a1c806b9ff55683d83b814466fdd9dcb4f9 - Disable obsolete marian_test
[Marcin Junczys-Dowmunt] 63efddf5811a4e26c02b18c382cd5210cf200f5b - make validator more precise
[Marcin Junczys-Dowmunt] 0b07a2f5ca9d9ce391afa55720ba604e15db80fd - unify validation across different tools
[Marcin Junczys-Dowmunt] 847f7f7ce0d66a4b008f34f34f19484062fd7929 - round max length in batch statistics to next larger multitude of step
[Roman Grundkiewicz] e595c1141b0f81093acedd8f625ae17af7d8439e - Add --valid-max-lenght option
[Roman Grundkiewicz] 9d7f1fe616366e81a9736152950478b8c014d45c - Organize includes in marian.h
[Roman Grundkiewicz] 86021ab6e5ba777c4b54daeee4670bbe48a133a9 - Refactorize validators
[Roman Grundkiewicz] ea9551a610a74b29343ca399b3ec30d4d199df2c - Rename attributes in GraphGroup
[Roman Grundkiewicz] bf2a9304e8bf7e62c66e17dcbeef00c211851a2c - Remove builder_type in GraphGroup
[Marcin Junczys-Dowmunt] 736da62f6a66e8b124ba63beda424fbd468ea9b5 - better specification of summary for rescorer
[Roman Grundkiewicz] 4793bff5b4155226d09d39a6202a107e66310bbd - Refactorize Scheduler
[Marcin Junczys-Dowmunt] 2463072ef28556a96f0022918d628fe853d6a9f8 - multi-threaded model loading for translator, faster startup
[Roman Grundkiewicz] eda9f6abed829e59ac6e7ea420abd4c881204efa - Remove forceBatchSize() method
[Roman Grundkiewicz] b0e4aadd0dc0805df300fdbe9529fe629b8c38d1 - Remove template from Scheduler
[Roman Grundkiewicz] e658e23d02f4196f03d034d5ca76f0b3fc9bcdc4 - Remove template from GraphGroup
[Marcin Junczys-Dowmunt] a237a59746d796417241df47c2e34441e31273f6 - working (I think) sync sgd
[Marcin Junczys-Dowmunt] 6c3eea9dd853d662b3e433cc6c9ec13ef3d021ea - fix wrong gradient summation in sync sgd
[Roman Grundkiewicz] 9e1de8de834ef9b5d606778d8ec4f684f8708f1e - Refactorize validators
[Marcin Junczys-Dowmunt] cf0ab4c0bb9aeddd10f28628c16bd30e6beb55ff - init graphs in sync sgd
[Marcin Junczys-Dowmunt] 92779e4d6542f699a47be806f292e546fc530817 - clean up allocators
[Roman Grundkiewicz] a7fc8e9239cf6a8d516549ed4d7ec939e7b3858b - clean-up
[Marcin Junczys-Dowmunt] 9ed84cd37416bf383b285f5cdbd776e8f2ca3076 - set scheduler for sync sgd
[Roman Grundkiewicz] 0fb3dd2b07a1056b120946589508dca926f43b71 - adjust tensort dims
[Marcin Junczys-Dowmunt] 68d9095af7b6a6dceb0f364b7dcaf01bfddd236e - guard agains empty batch
[Roman Grundkiewicz] 2be38a58da915e085d0db0e5a5dd8ebd34dbb8ed - Restore MNIST LeNet example
[Roman Grundkiewicz] 10fd58b40ba8ffe638091293072d8a1efbc5b9bd - Refactorize model factories
[Roman Grundkiewicz] 40095af4f8466157cbf7f673e69dd4446dd6fe12 - fix dims agains
[Roman Grundkiewicz] 4736ba8f02fa67b04b8321bcb031e99467458430 - Clean headers in graph groups
[Marcin Junczys-Dowmunt] 4497cd82b12a3193c9eae22bbbbc5e8a8098e769 - add running average to sync sgd
[Marcin Junczys-Dowmunt] 31585bb3d939fc3d625d0945ea4b8cd2ff043728 - refactored graph groups and resulting division into heads and cpp files
[Marcin Junczys-Dowmunt] 5620eea38cbccb1ba277ba24772e815b72a585f4 - restart warmup after loading
[Marcin Junczys-Dowmunt] 1494011bdc31597f5f74f27e4797c06a59b8b31e - option to repeat lr warmup at startup
[Marcin Junczys-Dowmunt] e9ba86dba286d8c0da8201bb6055a64a8d922ed9 - attempt to add back hard attention models
[Marcin Junczys-Dowmunt] 24596049108408374bb9009942de0dc1149897f4 - make hard-att work again
[Marcin Junczys-Dowmunt] 62d833cdcbac776121fcb76ec3971e956fc47e52 - solves issue #106
[Marcin Junczys-Dowmunt] 1f77cf0a804b8bdf8d75362be15ba27a76bc419d - unify binary names
[Marcin Junczys-Dowmunt] 90a437671868d3f5cef7d24e34fa6321e17c451c - check if model is present
[github] 4e967086fdc3203bd927ea7b6408e88f632c03b1 - Update README.md
[github] d6269d57f145201763cb4ec2b3409047ab399d5d - Update README.md
[github] a2acf00d9321aaa2ebf470d6f6fb9844ef4f0e5d - Update README.md
[github] 673b3200e0808e4197b13e17c03b8264ffc84403 - Update README.md
[Marcin Junczys-Dowmunt] f542888bc6e913c9f4af5bfefdf34f2581191ce9 - negative cross-entropy cost in scorer summary
[Marcin Junczys-Dowmunt] 19011e368967a7423d05a8ddbf390039673f2ef3 - add cyclic warmup and start learning rate, unify warmup options
[Marcin Junczys-Dowmunt] 059d2f3a17a7f04131c9342229cdc3547c61eb74 - fixed model reloading for complex models
[Roman Grundkiewicz] 798454edc6c7b634461eb855e8405eec975ea433 - Update project version
[Roman Grundkiewicz] 955fb5188edd05be76c058b6d25b3b1d918bd593 - Version v0.9.0
[Marcin Junczys-Dowmunt] 99500867ebcfe20c4741d2f1d40c0e5f011ea06f - add back overwritten options
[Marcin Junczys-Dowmunt] 0457d394a60a53621eb4571c61a6d663a65906b2 - remove old options
[Roman Grundkiewicz] 5543e4add1c9a32c9d926f52c4a71cd8101cd9b0 - Set high logging levels in server command
[Roman Grundkiewicz] 03c7370346edfd82f499b7dce0b75086796fd6c4 - Fix --models option
[Marcin Junczys-Dowmunt] 28d86206f4facbf301cc412c537468d03459d90c - add dropout-src and dropout-trg to transformer
[Roman Grundkiewicz] c56c66cfb910ad493c28e22e632a51da379a0868 - Fix createLoggers with no arguments
[Marcin Junczys-Dowmunt] 58b67cb2f53102cb219dce4218d443391bacb69d - concatenation of attention for multi-encoders
[Roman Grundkiewicz] f9e17f6aedbe3438ce19fec79c25d0e940e29d5b - Fix compilation error in logger test
[Marcin Junczys-Dowmunt] fe626962ea23c226a62e0f47b82f2bf2ca29f65d - Changelog
[Roman Grundkiewicz] 0f04d53d8a287c10b7f7a765700f20a3843fe79f - Add a badge for regression tests
[Roman Grundkiewicz] 2a0eab21707ff4eb4ad8b40dcae854eca21d63c3 - Autoformat easy code
[Roman Grundkiewicz] 936900b121785b439f2c9c2e0b3bf85cb0b9a92d - Autoformat rest of the code
[Marcin Junczys-Dowmunt] 2c2bf41a3288c63934dd54680c0f8de5c6c4173e - typos in changelog
[Roman Grundkiewicz] eae797ff0327f3662b36c2b5020848747b8b69d5 - Add draft contributing
[Roman Grundkiewicz] b3f0c041c6b89a4c39c9acd97d4f0c26cd82cd28 - Update contributing file
[Roman Grundkiewicz] 48e94717eb72f67ada2d017c61b1a2993ffa54af - Update changelog
[Roman Grundkiewicz] 3da1d1d48bc0fb55765f1adfc481d3e9e5a45fb1 - Remove trailling single comments and autoformat
[Roman Grundkiewicz] 71d1f91755f84dfd82f6582d677fcb8467de4e6d - Version v1.0.0
[Marcin Junczys-Dowmunt] edb1465edf828efd70a0214dac2ff4fe83f3e4bf - normalization weight for translation
[Marcin Junczys-Dowmunt] 92893bb7734533b9d84bddfe2af3e9ae04bb996d - update changelog
[Marcin Junczys-Dowmunt] 6d065f47f2703876fea02c5293c2ea08dc6959bd - get rid of dependency on normalize in printer
[Marcin Junczys-Dowmunt] 2b15c83173876c278585927a3b9a532c50bbe4f3 - fixed vocabulary issue in #116 and added fatures in #111
[Roman Grundkiewicz] 10cf2e4ba639a0728b32f3a81adb72777dab341f - Fix corpus after merging
[Marcin Junczys-Dowmunt] f4540c63c65c4df2442e130209d412f98f08a3d9 - Fixes #117
[Marcin Junczys-Dowmunt] 20e25056ca494e3b3269fa6ee76984be61a31d81 - Fixes #119
[Marcin Junczys-Dowmunt] 9453d043d2feeb1fa6cc743dd84fa04ca011daff - ugly workaround for #119, needs more work
[Marcin Junczys-Dowmunt] 51dc848c05be8810353702c2f7d5333c09f2c79e - #119: check index
[Marcin Junczys-Dowmunt] c7df913af7eb6bd3c03b97b21c2cbef3effc714f - replace bindex for index when no broadcast
[Marcin Junczys-Dowmunt] b356b47f4ea1d8f3189b594385b2882e08ed9dfe - clean data copying
[Marcin Junczys-Dowmunt] 113eab7d7955a7a1ab0f1dc6dd14371715bb4439 - use allocator for temp memory
[Marcin Junczys-Dowmunt] 4a5e3878e62a1661a37b16c7dba09b14391e5397 - uses memory allocator for temp memory in kernels
[Roman Grundkiewicz] 53d5d77e2af44dfc897127db792b86f96753fad1 - Re-autoformat after merge
[Roman Grundkiewicz] fce8a3b6ffbdc75e603a0a65f56c5f4fec7cbac9 - Exclude tests and experimental stuff from autoformat
[Marcin Junczys-Dowmunt] 983399c5b4dfa0d7120cbfc8fa2e3c632e65851b - workaround for slow concatenation operator, should be fixed later
[Roman Grundkiewicz] f017e94d894d8bcc1ba0477205cb0fe58814d9eb - Add new LOG macros
[Roman Grundkiewicz] e462eb384955d6b658a620886bb86edf576f2147 - Switch to a new LOG macro
[Roman Grundkiewicz] 10e2f9d40aaa89334983ba0342b854ff918d7d6f - Remove sparse logger sinks
[Roman Grundkiewicz] e5f8bfb3f283ce8f36816118df18c3aa261ca6e1 - Rename sink 'info' => 'general'
[Roman Grundkiewicz] 4fe858e781c51d1e650c6d49cefdd8f2cd33b96d - Add missing log level 'warn'
[Roman Grundkiewicz] f2ea473f67f220a315e1565db23635950cbd02e5 - Document logging macros
[Roman Grundkiewicz] 5a19eec4ac9da58492e2eb8893a8b44e978fbf60 - Fix missing validation logs in translation validator
[Marcin Junczys-Dowmunt] a7891fe98a9cc23c3861accb13f4da5e16f8bebb - swish
[Marcin Junczys-Dowmunt] 642241008419b4756ee80aa6614f6d5debc1a231 - stable logistic function in kernels
[Marcin Junczys-Dowmunt] f69870a7902e79e02e3e7753db6f351ffde4a85e - proper debug display for variable lengthshape
[Marcin Junczys-Dowmunt] 281a3344602deb8cb67a5fd254d9727008e77dab - dynamic gpu shape
[Marcin Junczys-Dowmunt] 0ab3a2e06372bd5616d94b4a2e9ec6abe5b8c49d - shape for gemm
[Marcin Junczys-Dowmunt] 543cca72092fa55a36b842b33588e363037a3ced - constant length GPU shape
[Marcin Junczys-Dowmunt] 674b6908781da58929b5d77fe71062e41386db58 - working mnist for constant GPU shape
[Marcin Junczys-Dowmunt] c5efa0bbf3345b9489ec774d0be0befcd08ac42c - add mnist_ffnn.cpp
[Marcin Junczys-Dowmunt] 4f04d35d8dc83e50cdcf563b32afa91354d7d948 - more changes towards dynamic shape
[Marcin Junczys-Dowmunt] 5097801c2f3a2bb54264548c5d154b0295261198 - more changes towards dynamic shapes
[Marcin Junczys-Dowmunt] 499b812a24576d4c3197eb03e0bbfe49d2041287 - a selection of unit sets for forward operations
[Marcin Junczys-Dowmunt] b7aebb559ff931ac275857ff9d54a9371f9e5e14 - tests for reshape and transpose
[Marcin Junczys-Dowmunt] fdd9516e982dad030b2fac36acab4184e27e532b - fixed bug in transpose for ndarray
[Marcin Junczys-Dowmunt] 5ffff46805c9a94dc603a00ff85180ee1ba5a94b - fixed transpose
[Marcin Junczys-Dowmunt] a6b65423feb2ecf547202f41a5384109035c9768 - tests for reductions
[Marcin Junczys-Dowmunt] 9362c80f7476bf41004bdc977e0a865f17bb913b - correct reductions
[Marcin Junczys-Dowmunt] 1da446a7e702b1e3d1380fe1fff780daadd18e83 - more tests and clean-up
[Roman Grundkiewicz] fc2180251eb1b954adf85455804a5905615d6cf7 - Add ABORT macros
[Roman Grundkiewicz] 3487e830573d8c26deadc3230763f3b332e86938 - Replace UTIL_THROWNs with ABORTs
[Roman Grundkiewicz] 7f9cfa454891cd61e4a8a81fa9df8e4f370aea52 - Update clang-format
[Roman Grundkiewicz] c7e99fc4a503c29a8709269611cefdb9f86ff7c7 - Autoformat
[Marcin Junczys-Dowmunt] 46433253735e79e03613fcbd28e64ff393f72451 - tests for rnn
[Marcin Junczys-Dowmunt] e72d742c45623cdc2f986aeb999740ce4b48798c - disable buggy test for now
[Marcin Junczys-Dowmunt] fe4a804d6692fa3bacd636e8b753f53911c59fbe - add abort
[Marcin Junczys-Dowmunt] b3765f61bdbc30cb5f2a74ac7f882b8a0b9055ba - replace TimeStepNode with more general StepNode
[Roman Grundkiewicz] a9279f1c91f2cc5cf54d036f52a9b9bc14698408 - Add function name, file name and line number in ABORT
[Roman Grundkiewicz] 3d226d9bd526694ffecb39c6ba15cd2748ccf7ab - Add a comment for SwishNodeOp
[Roman Grundkiewicz] 505353d76dc95fa3df703b4ff908fa489eb0e85b - Add comments to ReLU i Swish
[Roman Grundkiewicz] 2038b3cb2dbfa81876c71de6eb4bd953b8ef7587 - Add LeakyReLU
[Roman Grundkiewicz] a6d4f2a9ab1ed4b7f559715e9a5b70082f5efd32 - Fix Leaky ReLU
[Roman Grundkiewicz] 857289d1bc0afa14d1338f7b6a29cb614b03bcac - Add PReLU
[Roman Grundkiewicz] 34e13035203987c856d6ba14f18d855f5f28cb19 - Use PReLU in LeakyReLU
[Roman Grundkiewicz] 9e3ff281639a02756353d4fcc0479a1916e9a839 - Replace UTIL_THROWNs with ABORTs
[Marcin Junczys-Dowmunt] b2c9df7ca9c287552de9f359bfb239ce847d96e4 - define GPU_SHAPE_DIMS
[Marcin Junczys-Dowmunt] 3e40ed80cbcb280fc0597ded1a86d7bfbfc0018a - add test for rnn-specific attention
[Marcin Junczys-Dowmunt] c7192bb4e57c26c37ed4408d9e051ede680740a4 - updated attention test
[Roman Grundkiewicz] 6e40f327e9aaec371fbc02fbfba842cb4ea590bb - Remove translation from valid logs
[Marcin Junczys-Dowmunt] d9a0fd299cb617d79b350a2b4346bbc6fcddae7f - adapt s2s to new shape
[Roman Grundkiewicz] 34ddfe4f5e96c4f5fc38eabaeb9f324cf01dd311 - Add geometric printing strategy for validating translation
[Marcin Junczys-Dowmunt] 680ac5e0ca1a8d24f3e03e6ef545b8f2d7411c05 - attempt at translation with new shape
[Marcin Junczys-Dowmunt] c2a5e93a17f5fb62f5e3475adc133da9750562f1 - adapt model loading and saving to shape
[Marcin Junczys-Dowmunt] 347140724040d4bdc7151dfce5e12dd133e612e7 - error message
[Marcin Junczys-Dowmunt] ff171a810f80a7a4b803029373a5fe394e1f6ed8 - exception on rows
[Marcin Junczys-Dowmunt] d716dac8e6c6c1d68cbc9bbad39ef1f8c9946c80 - add beam dim to empty embedding
[Marcin Junczys-Dowmunt] c18f76069da1c1854f8bb4b571b64fb1cde1787f - working beam search with new shape
[Marcin Junczys-Dowmunt] fd9953df60c2fa1e3b638a752df6158ca733460d - array operators
[Marcin Junczys-Dowmunt] 5a73d915d01fad25ae0094bcc24d3ffdb275850a - adapted nematus and hardatt
[Marcin Junczys-Dowmunt] 7b97cbf3097d5a83a4d8e2fcdb0ca2aa779ebd55 - remove dim
[Marcin Junczys-Dowmunt] c176e1326bc496dcd8f285d366ddbcd8d790477a - move dim check for decoder mask
[Marcin Junczys-Dowmunt] 0daa611030edb23343436c48fc442e41f091383a - simplify shape
[Marcin Junczys-Dowmunt] 11a2b0695aec6ca1edfc175b3845c5a133e4e9f8 - fix transposition
[Marcin Junczys-Dowmunt] 4f14c9175af1db7ab625ad3df89db8687424d0b0 - unroll loop for variable GPU shape length
[Marcin Junczys-Dowmunt] dba29e9c023234aefd8f11a9dedd19b1d9140c8f - new namespace gpu
[Marcin Junczys-Dowmunt] c7d236762a98dcabb8cd0a5f923e0259563134aa - more gpu-specific TMP
[Marcin Junczys-Dowmunt] 274d2cf69225c1c4fb1bc0f8595982f32ec2f27d - a bit more clean up
[Marcin Junczys-Dowmunt] c88f988e5f17db4cc4d247c7dc56552b80413a7c - add missing tensor for broadcasting
[Marcin Junczys-Dowmunt] 0d20bb57c19290eb01c0e1d4ad539d94571b0f80 - clean up permutation
[Marcin Junczys-Dowmunt] cd6fca847dd9aa0b5236306beaaa8f778e2d8726 - commented out offending code make it compile, needs to be fixed when
[Roman Grundkiewicz] c31cb1e450e03d042221aa02b4e86fd01494600f - Fix displaying best translations
[Roman Grundkiewicz] a3b72c2ffea19f0cfe5c9a9077ece4204029f6b0 - Add TCMalloc
[Roman Grundkiewicz] fc6bd2034f762b57f4f779c75595d18d0bf22b4c - Fix repository name
[Marcin Junczys-Dowmunt] ca64c429e4aa4dcd49b68bab6a7d744fe06b44c2 - new functional templates
[Marcin Junczys-Dowmunt] 9d06d786ba568e74f54be98f5a23d9b53d87ed7b - farewell thrust
[Marcin Junczys-Dowmunt] 39859481fe0c63c85929d02f6ee14856b433917f - ternary operator
[Marcin Junczys-Dowmunt] c8c7f2f62a3bbcc6aa468da315279fd1283a3c66 - remove lib
[Marcin Junczys-Dowmunt] 00b8c65c482e10381c7e74d64c8c2c88e9d95c71 - cleaned up functional templates
[Marcin Junczys-Dowmunt] b2999035041745b87949094401e5945f45bcfd72 - validation with multi-gpu
[Marcin Junczys-Dowmunt] 1ae536568b78bd187cb2109654307ee2e2cb76fd - fix segfault
[Marcin Junczys-Dowmunt] be81d84024c0a4e77076a3348d8c7bd19419f7a1 - multi-gpu validation for async graph
[Marcin Junczys-Dowmunt] 31742956f7e3413828bb7368a6ae43ffc6078f9a - set continueValidation to true
[Marcin Junczys-Dowmunt] 8c19d2e0068f4132a53ecd737657f21c2312fa89 - multi-gpu CE validaton
[Marcin Junczys-Dowmunt] 704c739921915004e0fe37caa87c2113e3c5fcdb - scope around thread pool
[Marcin Junczys-Dowmunt] 9ed217cdbebcb0900c172d40fae25acb74518483 - adjust capture list
[Marcin Junczys-Dowmunt] e47546e3b0054253c8f094a76955dc7881a8f086 - multi-gpu rescorer
[Marcin Junczys-Dowmunt] 50a6513095a1fc07e6337417ed8dc19cb2d45a74 - init graos in threads
[Marcin Junczys-Dowmunt] 12a03731713b192b98ca21419849930c8a253276 - clean up waiting and invert moving-average
[Marcin Junczys-Dowmunt] 83185a629b9863409857180ddccb21c23ad4b1f0 - fix type
[Marcin Junczys-Dowmunt] 7f4b9d907fda758db95aadde472256d684fa3e7e - use external lock
[Marcin Junczys-Dowmunt] 13d1461ec428264069ced4c86feeccbc3069c9fc - wait based on bound
[Marcin Junczys-Dowmunt] 58cfa6ef308bdb8fb6e4e419263aa781b26d2fb5 - fix missing continuation criterion
[Marcin Junczys-Dowmunt] 8a15b5b6e0fc8c9478bc7f14dcd4c936217db514 - comments
[Roman Grundkiewicz] d6a644142baef49f28ebff9b659fc0940165bf47 - update changelog
[github] ab898352eda9d4e00648f02287451aac3791ad11 - Update CHANGELOG.md
[github] 73db1b9284911e8fa7b2cd591ca8ac3f1bd748d4 - Update CHANGELOG.md
[Roman Grundkiewicz] 7fdf1bc5ccba43f5bef481b49e297886799de102 - Rename --trans-output to --valid-translation-output
[Roman Grundkiewicz] de2b4016045eb3b954534d0bf255cbd2e3f951fd - Fix GeometricPrinting
[Roman Grundkiewicz] 40b8435d1d9f2bce8202232ed5f4e191e6c80d69 - Add default printing strategy
[Roman Grundkiewicz] a3b22eed5747d096e97d759adaf2c026547ef4d4 - Add --quiet-translation option
[Roman Grundkiewicz] 90cbc1f187bddf4659b4cd71938d01c7afa30f61 - Create config for marian decoder when model is saved
[Roman Grundkiewicz] 57b5bdb31fd84d3b6db30533e86687f41f864e52 - Refactorize save() method
[Roman Grundkiewicz] 654ad9f2a2da8b29606bff8d06eea60b300059b0 - Remove unnecessary Args
[Roman Grundkiewicz] 6b53373df8181a5816bfd0bb9c0ce88699d4c9e6 - Change description for --best-deep
[Roman Grundkiewicz] 82bc57d28a8afae53028568210334807e116fdeb - Use clang-format off/on
[Roman Grundkiewicz] bbb96e2ad7dd79d3ed0017b1d5b54d4838b351ae - Rename script name
[Roman Grundkiewicz] 83f3b3561b2fb50503d3729f36f6b6dd3e48d3bb - Move optimizer-delay option in help
[Roman Grundkiewicz] f52556464d0ff4c8bfde5ff72c6af9db40b63cda - Fix --quiet-translation
[Roman Grundkiewicz] 3e409336d8f17f7c923d47565ed5582997016e19 - Add --quiet-translation in decoder
[Roman Grundkiewicz] 35077fccc67c2a95307165af81aeb2d56c87f7fa - Add basic training example
[Roman Grundkiewicz] 7e53c94d54f4a2cca92b9156c62be51bcf97ae3f - Add dev and test sets
[Roman Grundkiewicz] ad136332655d3c148201f576518cfc22faf45c7a - Update training example
[Roman Grundkiewicz] e7971bf35f46089001937f82d61f13f1e256767c - Add transformer example
[Roman Grundkiewicz] b26cdfb74cc4e2eebd52144c337b87cc4bc01eee - Remove exit command
[Roman Grundkiewicz] 6c5f234655910fcdc2ff88ee26658645c5a1e3ce - Fix cloning tools
[Roman Grundkiewicz] 45f2addb21d07628f0209a3642b632b2f659e9c5 - Fix train set in transformer example
[Roman Grundkiewicz] 30a3a88eefc44cda9fe7604baeef21188c54d4c6 - Add marian-vocab
[Roman Grundkiewicz] b6bbbcba003c416b7767151c005039d3c7ad0145 - Create loggers in marian-vocab
[Roman Grundkiewicz] cb873572a7f5c8b2a108f6c0813be28311cdc203 - Use stdin/stdout in marian-vocab
[Roman Grundkiewicz] 018e9054d266776be50b09e870a20e6d1fbabcb9 - Add --max-size to marian-vocab
[Roman Grundkiewicz] ec4efecc1de2519db0d622addca2e771bd50700d - Use common vocab
[Roman Grundkiewicz] d99e7451a0caf01b163a0a233a5e7c4250145222 - Update README in transformer example
[Roman Grundkiewicz] 07b1a01c9297f85f73c19a60a8988cb8cf31755e - Fix misspellings
[Roman Grundkiewicz] 59e0f2c14db40ca16685d2c25fb199d5a54c9d0a - Suppress stderr in validation scripts
[Tomasz Dwojak] c7bb6458d4ce2b61cdfa8e71fce51768da1b5703 - Add non-const data() access
[Tomasz Dwojak] 54bc3edc9dfae028153cc0e13a6685254c26f0d9 - Separate cudnn code from nodes
[Tomasz Dwojak] 0811e40151281ed5da4b3826d0deef150fd930cc - Add toString method to shape()
[Tomasz Dwojak] 5e66a96990e958f8c5a73332da7a945f24fd850e - Split cudnn wrappers to .h and .cu files
[Tomasz Dwojak] 4c94dc14bfcc5abd65b8be829768539425fa7801 - Add missed headers to shape.h
[Tomasz Dwojak] e8138aaf2bf062820420cbcd131b99a08fdce4a0 - Fix lenet model
[Tomasz Dwojak] d2acc55925b81c42026d164836449f6508c280b9 - Move pooling code to cudnn files
[Tomasz Dwojak] e554e2aa45d9e049aff1accfdd8e5d2d3de62d3c - Remove ifdef CUDA outside cudnn files
[Marcin Junczys-Dowmunt] f517729affc6a9abaefcb5a57b7902f9cf87d199 - make compilation for g++ <= 4.9 work
[Marcin Junczys-Dowmunt] 78387b6ae2493bc526ebbffb6411e8769928325e - fix warnings in g++ <= 4.9
[Marcin Junczys-Dowmunt] 78aa8b51eda8ffdae9d1be984a99100150559cb6 - set default mini-batch size for validation to 32 as discussed in issue
[Tomasz Dwojak] 04d3d59f1210b81704d3e81a28e0f96dc2107e9d - Debugging memory leak
[Roman Grundkiewicz] 843d1d3597662438bd59e2a684859418c2bf5c08 - Remove examples
[Tomasz Dwojak] e60b214340d7bdecf30a95a26d143522e07531ff - Fix compilation without CUDNN
[Tomasz Dwojak] abe60aee9f9a9e03617f2f6b46810272b27af56f - Fix mnist lenet
[Tomasz Dwojak] 89202635337603032fc7af03b570c82b54b9bf0e - cmalg-format on cudnn files
[Roman Grundkiewicz] 12023d3f50c97693e14c7ab826231191e3b16508 - Clean MNIST example
[github] 76c8bb541e59edf88f127f3b0c28ed0c8f25a84c - chmod +x client
[Tomasz Dwojak] 812865b58b169cec8c4ee07acf494d8a17409482 - Add non-const data() access
[Tomasz Dwojak] db05f5dd990ea7bc748f0ed6d266cbe6a6a0908f - Separate cudnn code from nodes
[Tomasz Dwojak] 5d16efefbfb4c07951100968969ce52a63074532 - Add toString method to shape()
[Tomasz Dwojak] ef372504be743b5d112026a1eb6e8ab81ee1e367 - Split cudnn wrappers to .h and .cu files
[Tomasz Dwojak] 29283e0caf0bf72aee362fb29d07adc39ffebdd3 - Add missed headers to shape.h
[Tomasz Dwojak] 72761ed08425cc1d20a9f522c619052bef6f5dd2 - Fix lenet model
[Tomasz Dwojak] 9b24f9c6ac23d50ca698daa7ca2f13ecc1e96ead - Move pooling code to cudnn files
[Tomasz Dwojak] e918657b87fe9e59d56f54a39a599ad29b137f44 - Remove ifdef CUDA outside cudnn files
[Tomasz Dwojak] 5d2ee1517237ca727da683820f3ce065e4adecfa - Debugging memory leak
[Tomasz Dwojak] fb34dbcb1bc053727757d4ca96992420829e8584 - Fix compilation without CUDNN
[Tomasz Dwojak] 0add974d484db152d1570a381e85ddf1690e9666 - Fix mnist lenet
[Tomasz Dwojak] 2fdbfdff5b9b9615c273b223bc8cdc96914ccb01 - cmalg-format on cudnn files
[Marcin Junczys-Dowmunt] 923e03f38cc7129177d1b445680d4a22c052e19f - closes #133
[Roman Grundkiewicz] 50c17e0c708431ccd26ed8db6e290b3d8f0c6d7b - Update CONTRIBUTING
[Roman Grundkiewicz] ad47d7cada566fb4c22cf52c4a5aaf8c02d4f483 - Abort if no vocabs provided
[Roman Grundkiewicz] e978d41f7829bd45696a6642fd3ce97f36632434 - Throw exceptions in validateOptions()
[Marcin Junczys-Dowmunt] 3ef3b421a39e264c222eb98da26cf0cb02faeec5 - towards batched decoding
[Marcin Junczys-Dowmunt] 48c72831b0725e95f58db1bef223f7e300bf9876 - add CUBLAS_TENSOR_OP_MATH for cuda 9.0
[Tomasz Dwojak] 28586b6ad1d08b36008d63fd8e1962888f99b859 - Move building embs in s2s enc to a separate method
[Tomasz Dwojak] 8d33f14785650efd7f1cd620e9f6c50c8602b111 - Set cudnn version at least 7
[Tomasz Dwojak] b60ccd0bee215cf5930e34819892d5e7ac4c28df - Fix indents in s2s.h
[Tomasz Dwojak] 339a504c4fdee04caf24913f2a89c05263f57ceb - move highway impl to expr ops
[Tomasz Dwojak] fd97c84e0680df952e53fe2c5d65e89748b0c638 - Add ConvPooling
[Tomasz Dwojak] 6e244033ed2da970066256db271a07d526d66a0f - Add function converting to and form cudnn data format
[Tomasz Dwojak] ce44cd9e287804860538dc80af0f87ae3ff8cfec - Simplify convolution layers
[Tomasz Dwojak] 0effc8d28d25fbf80e3cea34d6f5b4e42490a7f2 - Refactor convolution and and poolingWithMasking
[Tomasz Dwojak] 8eaa532d8e50507893267d46c14c7c4df6f45651 - Add charS2S encoder
[Tomasz Dwojak] f43829275bba04c952e263332ca2ad0bb10f27e8 - Add char-s2s model
[Tomasz Dwojak] 9b935f7c8b737bead5e3aebda9f0e86fbe8c11ca - encoder s2s: Use input shape instead of from options
[Tomasz Dwojak] 4011aebb5ab498fe8b7c182991f6ec2085ee58ab - Fix reshaping in convolution
[Tomasz Dwojak] 15947c6061c009679b0a00ca873d768c1159cd6f - Fix char model
[Marcin Junczys-Dowmunt] 5fe987df0613ffd5c57ce03c503f90e3239044bc - working batched decoding for beam-size 1
[Marcin Junczys-Dowmunt] 359bfeebcb39a2643d7e2825bbd9844031e8e1b9 - working batched translation for rnn models
[Marcin Junczys-Dowmunt] fba6778c38de40f876217dc79bfcbbcea986655b - fixed segfault
[Marcin Junczys-Dowmunt] 51ff08a15350a91b4faee0d26d53889e0735a530 - second attempt at fixing segfault
[Marcin Junczys-Dowmunt] 0e39a5f212c750a4abd7c220d37c58cb183f6d3c - restrict length by source batch width
[Marcin Junczys-Dowmunt] 57770ba0f988e53914ddf281ea7048abde0a5d1e - clean up code
[ubuntu] 57a7790edf472f445787ad77574121dfd82b872d - reduce beamsize when possible
[ubuntu] e4d6cd95ccdf907c9a9dfac56888928c401fa6ec - fix beam search with reduction
[afaji321] f67c36088740d687e83db40d1bdecefb56f8d02b - reimplement gradient dropping
[afaji321] abb6c48e9266c683efe511eff132da1652759655 - reimplement gradient dropping
[Marcin Junczys-Dowmunt] 51c8d4499b67cacef81c1edc762e1ca1f2087df0 - working batched translation
[Marcin Junczys-Dowmunt] 7432024c7de7c2b928b1654d62afb7b9834ed934 - maxi-batch-sort for translator
[Marcin Junczys-Dowmunt] f6792e3f49c5b941267bbbc2083363c376293132 - make tests a bit more tolerant
[Marcin Junczys-Dowmunt] e378a764ce25374412f91d5013772f86a9ada5da - update changelog
[Roman Grundkiewicz] 19b9a9d49294dd285f9d7756acfe3a2d1c8e1bd5 - Add missing marian-vocab do CHANGELOG
[Roman Grundkiewicz] a6abebac2046ee91a4677e953ae5cedd169b21c9 - Create a separate command for each unit test
[Roman Grundkiewicz] 6f77efa0565ea0f6b7076f0931c35654cc4153f0 - Version v1.1.0
[github] c50994d8f3643ce7bea2171445a533b8ac0efd6c - Update README.md
[afaji321] 7da15d4832db7c3c2d1a04850441ba9cd7a8c5af - add warmup
[Roman Grundkiewicz] 9023667939b0fdd645f971cdeb0ab4e764b07057 - Fix #139
[Tomasz Dwojak] 79cdf1fd560286fdb40e25a06e71eed7912826a9 - Support char-s2s model options in command line
[Marcin Junczys-Dowmunt] 39e15a6d3eb3ba86a8f9d9aeb96c2fa039581d25 - remove tcmaloc_minimal
[hieuhoang] a1b13e8cf07ff88a8387b58387786038bf952ae7 - gpu arch consistent with amun
[hieuhoang] fe7ca63a66e3ab79423ef94024f14ddd019659a4 - compilable with cuda 8
[Marcin Junczys-Dowmunt] 66655e76a3d1574f4f7a6fe8984979180d6bf74e - remove redundant exceptions
[Marcin Junczys-Dowmunt] 1089be80f96e479a51cbe2983839021f6d27c6d6 - disable tensor cores due to cublas bug
[Marcin Junczys-Dowmunt] c512b9eb20a9b011d61af119fde36fc0bf7d7f3a - option max-length-crop
[Marcin Junczys-Dowmunt] fed955ae472e5394b06f65a280ad2254c2cca915 - fix saving of best translation model
[ubuntu] e7f2c57dbea6a7fdb9cceaac5e7b8f741b83e716 - fix bugs for lm
[Marcin Junczys-Dowmunt] b985786190b88ceba65e714dbad4396eae2e04ce - ignore model config
[ubuntu] d327ca2afd98d5fc461136d4bb2537ed6b54f8ed - reenable tensor cores
[afaji321] 214c9fc3ef7fce3908fa0acfad4972303f98c9d3 - fix bug when dropping while moving average is activated
[ubuntu] 3b5c3aeca7187973feebf7d993680dfbf6657948 - ignore model config
[Marcin Junczys-Dowmunt] 7905040f87cd1c2e5b8073b40d88db0e3d560e14 - switch off tensor cores again
[Tomasz Dwojak] 0dfb8abc49dc71d16119f54bb1d8bdd9c7c907fa - Turn off conv_test
[Tomasz Dwojak] 7101efe2b1fcb233ee74b1688ee437bd2ebbd08e - Add ifdef around char-model options
[Marcin Junczys-Dowmunt] 2ab491beab2b57db98912d8df6237a282480e5e4 - add check for vocabs
[Marcin Junczys-Dowmunt] 004aeab930c67b813193720e2ae387ba3f628ad8 - fix sanity checking and a number of error messages
[Marcin Junczys-Dowmunt] 0a626593b11ea5b1cf6501f89a30ecc776db0a13 - add missing ifdef
[Marcin Junczys-Dowmunt] 25000e31ae2308441048c321286ba42e2288ae78 - v1.1.1
[Marcin Junczys-Dowmunt] f395e139aa4c90f4c87682baf67e8b9cd382a9dc - fix a number of issues with ensembling and batched-decoding
[Marcin Junczys-Dowmunt] e696962be0cf3373b0f1fd478a783e9a0edcfa6a - update to v1.1.2
[Marcin Junczys-Dowmunt] 6b843c585786f5132a177ea7b48ec8ff47660c72 - fix bug with tied-embeddings
[Roman Grundkiewicz] acdd9078a7608b1d8757a157e57e0b9cdff6629d - Autoformat files related to gradient dropping
[Marcin Junczys-Dowmunt] ad8d3adddad7ce4b440a1d9b982dd134cf77ac25 - update to v1.1.3
[Marcin Junczys-Dowmunt] 3deb9015a8729f4453f0f06b27385e764cce1965 - add gradient dropping to changelog
[Tomasz Dwojak] 4a44b53e7c824f0f31c4f8a0ececd09850d9d553 - Fix char-s2s kernels
[Tomasz Dwojak] dc819b14ed5ffeceec8c2c9082d28d96af6249b9 - Fix compilation after renamed chars2s file
[Marcin Junczys-Dowmunt] 5c4bcf9a9ca76d547ed56f9248ff057a37b269f5 - fix error with protected data member
[afaji321] 4dbf986566b7afda31befc0e20a34099831bef6b - Add sparse momentum correction and masking
[Roman Grundkiewicz] d9063529eaef092279002c1c22d5b5da40320384 - Add COMPILE_SERVER option to CMake
[Roman Grundkiewicz] 379dd9071e49d5f937fd9e18f03b57e725196057 - Update Simple-WebSocket-Server
[Roman Grundkiewicz] 2c26cb1a4185c95bca03d5c92292e1aa6970a5f3 - Update CHANGELOG
[Roman Grundkiewicz] 9b1273f70272252aeb72b9081959e05fc2607256 - Abort if no vocabs provided for scoring
[Roman Grundkiewicz] 879938f578d7a33c76373dc6b9422760ae7f1a8f - Abort if model does not exist
[Roman Grundkiewicz] 89b23ce8abd259ace055c35b0da47428bdcbc37a - Fix decoding error with python2
[Roman Grundkiewicz] 2b369a54f9a127a18e3025a2ffa57d11153c5fc3 - Fix misspelling in version patch
[Roman Grundkiewicz] 43fbaa6c109c51a77832d6f09e0543794882b518 - Add scripts printing special:mode.yml from model.npz
[Roman Grundkiewicz] e6970cb5d5711997f689709396e1ef3ae9c74ac4 - Add version number to model.npz
[Roman Grundkiewicz] 641d6a9059170459dced3d1c7221c6d5715979f2 - Update CHANGELOG
[Roman Grundkiewicz] ba0ea7491fab383992013a8379592657eedfe1ce - Add printing value for any key from model.npz
[Roman Grundkiewicz] 4a02655975e5f4eaca462bf44a14f704747e2c43 - Add weights initialization from pre-trained model
[Roman Grundkiewicz] 04b8985ac80e0ba81e1c130b5c5d84fb0cc9cec7 - Update CHANGELOG
[Marcin Junczys-Dowmunt] 97ddb30948d94d25367a6e7d07b67498a635079e - disable server if no openssl found
[Marcin Junczys-Dowmunt] 0d61c14ee255859ca663798d09d3af482c9b4203 - right-to-left training, decoding, scoring
[Marcin Junczys-Dowmunt] e7e357c7165a25ce4e44231a167b6321acae619c - Updated to v1.2.0
[Marcin Junczys-Dowmunt] 98de8e65812925d6f8fe35755c32e9c99fbc78c1 - translation should use valid-mini-batch for mini-batch size
[Marcin Junczys-Dowmunt] 69b37bae7e2881e6371e8addc168d7cf61796682 - fix gradient averaging in sync-sgd
[Marcin Junczys-Dowmunt] 958222ba9cc3dd6273d9f544dafb2b1ff771877d - fix delay between saving and validation for async SGD
[Marcin Junczys-Dowmunt] b054b4c565b1b0f1b89ac07a4ea2ac37ca801db1 - update changelog
[Marcin Junczys-Dowmunt] 20bcb30c3e3b525efd499cbe3677d1e4bb95eea4 - change versioni to 1.2.1
[Marcin Junczys-Dowmunt] b5b10a5766295256dbfbc2f42b9dd4991257211f - asynchronous batch pre-loading
[Marcin Junczys-Dowmunt] cbbc56b0e12ee654ce0d94d03aa12cfb397c731c - remove debug messages
[Marcin Junczys-Dowmunt] 4653a62221c5f8660e287ea054e823e416b226e1 - clean up threading
[Marcin Junczys-Dowmunt] 6d5accce578547043115145608c5581ff09b8c20 - add comments
[tomneckermann] 208d42e11dc45ada2ec89cfaeb880bb57711a062 - Add multi-node graph group header and source files
[tomneckermann] 0a1cdbc9d502a1d8d6722db73e72fe1c55c16205 - Add logic and compile config for multi-node
[tomneckermann] 168a570dcfccc45d4a12fc60952a90518cc3c6ca - Add command line options for multi-node
[tomneckermann] fb1663a1c012097d8e7587921827d845b1308cf1 - Add new-line at EOF in various places
[tomneckermann] a7f28376d0b9672f55c022c50b5403077008db33 - Fix multi-node-overlap command option so it can be disabled
[Marcin Junczys-Dowmunt] 64e9ed37bdd2a654098a15d0522fd57aea8cfaa5 - fix sorting
[Marcin Junczys-Dowmunt] e20d578f7df4db71eb8476a38e66bf84740ff14c - use cublas transpose for tied transposed embeddings
[Marcin Junczys-Dowmunt] bc3d6c9c703ba82e666fb69bee105fe6a9669684 - fix for --mini-batch=1
[github] 41c12eac738e51bcc8bff847d9da94cd65f776fd - Update CHANGELOG.md
[Marcin Junczys-Dowmunt] bfe6839880275180de2a46e032ef837e7df65024 - sqlite3-based corpus source
[Marcin Junczys-Dowmunt] 65d9cce2ef943532fb6209a177e7beb38fe3c64c - fast sqlite corpus container, optional
[Marcin Junczys-Dowmunt] 6e223c2fc32841abc66246951a066ee677a6cc15 - cleaner inheritence
[Marcin Junczys-Dowmunt] 13e3d5fc81df5ad7a4ab2214886617acbce13a9d - update to v1.3.0
[Marcin Junczys-Dowmunt] 3c46f28541a7451b05d5571029992607fdb8038e - missing next() in MNIST
[Roman Grundkiewicz] 57aeea26e816b283f4109174d3f7c32e30288ae4 - Autoformat multi-node code
[Roman Grundkiewicz] 6c314e7ea353117ce573d96e4a9d70dd39e22048 - Use std::vector
[github] b41213a744bbc4be2520a77054000e1078a3ca85 - Missing include for
[Roman Grundkiewicz] 30e59e9f2ef86cce476ae7bf5f9179ec8f377f9b - Option --devices as a vector of strings
[Roman Grundkiewicz] 74017efd1cb6591f7fad7b8380ce6a29b6655c47 - Add Split() returning a vector
[Roman Grundkiewicz] 1b829b9724be2653b03bbf4944c41e9bbdafe826 - Modify --devices format in multi-node setting
[Roman Grundkiewicz] 0acca3f8813a9cdb773acb0e13c149f6068c964c - Add HimL acknowledgements
[Roman Grundkiewicz] 3d94f546cecca4f8144ceb74db7762e305f5de4f - Abort if no scheduler for exponential smoothing
[Roman Grundkiewicz] 869e93436f4b1ab45442ba58ca3d9e21602caa44 - Fix #155 - validate final models
[Marcin Junczys-Dowmunt] 8c2c596732b34cb304c9f346bdd9fcb0c616fcc5 - Fixes issue #169
[Marcin Junczys-Dowmunt] 8b00026868ce0e3cb6107aa4a866c59447aa4270 - update to version 1.3.1
[Roman Grundkiewicz] d7b1f21eac0d1c291e1d397de1d9b0bf5073675c - Add word-level weights
[github] 69ac5482c892ae4c3d3537681c07a0cdcbeedc86 - tansitional -> transitional
[Marcin Junczys-Dowmunt] fa824ed8a6f9eb849d6e9fd314a95e6f92b187df - persistent and reusable sqlite storage, issue #172
[Roman Grundkiewicz] 083b86acd4ae7cd8af173fe96e461887c89839b5 - Fix implicit value for --sqlite
[Roman Grundkiewicz] 06741a6fe95a8b2458c10851c4eabda806b10972 - Remove Corpus::setWordAlignment()
[Roman Grundkiewicz] 8ee34741f7a1e700944ddcf54021603296930354 - Rename --sentence-weights to --data-weighting
[Roman Grundkiewicz] 8130d3653ff07f765d4be4a3ae668806f9f6788f - Add sentence-level weighting
[Roman Grundkiewicz] a67124bfdad684ff32315a3d62ba2b64baac9621 - Autoformat CorpusSQLite
[Roman Grundkiewicz] 051c38fa789a1de206cdfae792894d169de5127e - Document SubBatch
[Roman Grundkiewicz] 85d64f172cd23343c72d17baf4983c94ff3baaa6 - Move CorpusBase to separate file
[Roman Grundkiewicz] 1241c075543b8de4cc2ea0773163f1d920203207 - Fix scorer after adding --data-weighting
[Roman Grundkiewicz] 4b150b79019f4ea5042cc7eaa7af54bea26e0003 - Move parts of Corpus and CorpusSQLite to the parent class
[Roman Grundkiewicz] 68cfde726005d893fd494634228a19bd88305b29 - Document SentenceTuple
[Roman Grundkiewicz] 174c4d345f6f48a948fb380743bcf54f8e1b7670 - Iterate weight file in corpus
[Roman Grundkiewicz] 036c89c2b022186182ea9ea407bcc9fa80cc676c - Move word alignment to separate file
[Marcin Junczys-Dowmunt] 232fba88333b741a2394dff46c7de4ebb5b09ef0 - fix lm-transformer
[Roman Grundkiewicz] ec8cf90ba94ca5454ffa05105c68e37eaaf36004 - Iterate word alignment file in corpus
[Roman Grundkiewicz] 03bdd0ddf410f5027993897029a038d813643396 - Restore guided alignment and data weighting in SQLite corpus
[Roman Grundkiewicz] 66473d9b99039010484040e197d03f5dca625fe3 - Refactorize corpus classes
[Roman Grundkiewicz] 35f5d565a4a9f19a19e0f1f6a7c4fb64f86e7f49 - Fix guided alignment and data weighting in SQLite corpus
[Roman Grundkiewicz] 300ed96cf4187dc7a36a9fda2859577101d792cf - Clean options and comments
[Roman Grundkiewicz] dca77499c6c09a07df1cea9d79d0b3f188309dd3 - Update CHANGELOG
[Roman Grundkiewicz] f7345ffa8d9d1ec8f75cae7f98027353fc4b31f0 - Use custom random function in SQLite
[Roman Grundkiewicz] 03d9fad5e1cee923a82be25a5fbdb168f71fcc32 - Use std::default_random_engine instead of std::srand
[Roman Grundkiewicz] e1739bbb7633e8740f536207e88eec06e688abd8 - Update CHANGELOG
[Roman Grundkiewicz] d2957c6695d64f60af4ca3aef8817db98822305a - Fix data weighting in validation
[Roman Grundkiewicz] a5aeb1dffa71ae899187684858e96d25e8b50384 - Document CorpusBatch
[Roman Grundkiewicz] 45f1d9623140270a5604be17837328c0a87fb2ce - Fix guided alignment and data weighting in sync-SGD
[Roman Grundkiewicz] f9003d9f2ad2fe879edffdf8c97551de2d090182 - Fix data weighting in sync-SGD & disable guided alignment
[Marcin Junczys-Dowmunt] ed3499353dcf2b5d454689c47aa71f60fff41fe8 - make weights work for sync-sgd, a couple of simplifications,
[Marcin Junczys-Dowmunt] b18cdfc2a124c1e154580ac8c0b9d8053a0fd91d - use common item-wise comparator
[Marcin Junczys-Dowmunt] 8f3208c0a124f0b19bb4b1af500b25f308163509 - make example compile
[Roman Grundkiewicz] 8b2b0fe17f7b614b61c07e56d5c6e3736b311abf - Fix comments
[Roman Grundkiewicz] 129323f3ce765c3553b12456585fac377307260f - Fix --max-length-crop in data weighting
[Roman Grundkiewicz] 0a355a6aa7817d1edd7dbb10f0d906c81c3681b0 - Remove needless attribute
[Roman Grundkiewicz] 5f0870d766b42a489603aaf1f2eea8fb21b86e22 - Add save/load methods in training state
[Roman Grundkiewicz] e3d49ed00dfd0e76a35576e550155256392325bb - Move display statistics to training state
[Marcin Junczys-Dowmunt] b8c5bcf1143696c8c39ad5cb12dd5157019f04ea - fix merge error
[fseide] b28eeb62e805d2447e64dbbcc2e67596604921d8 - aligned with the Dynamite versions
[fseide] 34fc3b0088d763f6c85affe12f9bddf2db153f47 - CNTK compat: changed some #includes; moved Cost() to header; support for
[Roman Grundkiewicz] 6f639d5dd622455cd4c981a3435e7a3f74d2191a - Update CHANGELOG
[Roman Grundkiewicz] f3b8d9139abf300a8603cb68bec69194a5a66636 - Save Adam params in single-GPU training
[fseide] 49510a4110076912d5aa2b2a9cc028472b54617c - towards CNTK compat
[fseide] 486876b70f4608fbf028554c9ef2ed261ab68296 - undid a few CNTK-specializations in original Marian sources, instead
[fseide] 1817351bb921cdc9a1fb8c46392a1f426c2f7bd2 - moved guidedAlignmentCost() to header; lookup() now takes a graph
[fseide] 291cc6c96b44095abab51588b7561e8183997c22 - cleaned up commented-out #includes; deleted the two functions that were
[fseide] 5f3353c5a5ad05629e7fb6dcfec2a7fafe3bcc63 - moved attention to models; deleted two CPPs
[fseide] 1ce352fb58fb39976624d9884ab78adbabd22f9b - added file missing in last commit
[fseide] 034f52a91ced8f2c004da516992f9f53121eed58 - removed the CNTK-specific embedding code
[fseide] 8282937cdfdc501190a7d39262f4f191d9e8601a - SubBatch::indices() renamed to data()
[fseide] e9e966777c2f686ae68afdd1a2e5b49cc4607a04 - SubBatch::indices() renamed to data() in rest of code
[fseide] d565d522129328aa6cae3974d49843ffdcd3d9aa - removed the last CNTK-specific #ifdef
[fseide] 117db2cd93351e64d8928f4303fd117662493ff5 - fixed some options for enums
[fseide] 97626c3c922942ff12604d426bab3e265b5d5c44 - few more headers; Config::YamlNode; options(int) casts
[Marcin Junczys-Dowmunt] 191bf7f6c9f9d0f92f332a8a7cda88114981f161 - steps towards separating GPU and CPU code
[Marcin Junczys-Dowmunt] a332c04d15dd72282f532991938dd2b12150cb36 - dispatching
[Marcin Junczys-Dowmunt] e09a9fae09204b5f32a478fd83be79d4c3cb971d - more cleanup
[Marcin Junczys-Dowmunt] 215177471a5ba1f3dff0e17801bae653d9005072 - add --cpu flag
[Marcin Junczys-Dowmunt] eeb1afbeae4c1908aeabc20d86183cec199d721d - fix missing cudaSetDevice in NthElement
[Roman Grundkiewicz] f41e564e6857a37bf061bee44896452054ca3d21 - Abort if multi-node enabled but MPI not found
[Roman Grundkiewicz] 1a13077392bb2deb4eaedeabf4b7cb70af989ffd - Save and load the last best validation score, eta and factor
[Roman Grundkiewicz] e15a0e50118ffa838ba06b7c913cb144ca0df248 - Load optimizer params only if --no-reload is not set
[Roman Grundkiewicz] 9ee5b64e4e416598eaa0982e97105c4dc9f0b469 - Refactorize and document TrainingState
[Roman Grundkiewicz] fac29e5fc631b94f69f8d16e9df9f81a0a471b71 - Generalize save/load methods in OptimizerBase
[Roman Grundkiewicz] 51be6d0d41e19746ef71b2cce3c1bbce3a5fca43 - Support multi-GPU in save/load methods in Adam
[Roman Grundkiewicz] 0d59c07aca0e4c0eeb900cde003dea445863fcec - Load and save Adam parameters in async- and sync-SGD
[Marcin Junczys-Dowmunt] 20e4ee3fe2d3f77ea54db3a2d1ef4af0bb827d7b - add deviceid to multinode
[Marcin Junczys-Dowmunt] 16913f3fb9c7c63a2631b5ada245858a81ea3f6d - update tests to use new deviceId
[Marcin Junczys-Dowmunt] 3e929e7054a6e06a40bf887673de3a8bd3210b55 - missing adaptation in cudnn wrappers
[Marcin Junczys-Dowmunt] 372240e23c0c02bb2d9aeee099a6613bf23e5831 - fix wrong include
[Marcin Junczys-Dowmunt] 744c8f46d09f0b5b8fe74ab68ba8974dd3e513af - add missing includes in tests
[Marcin Junczys-Dowmunt] 483426798930bc17309fecfb401b28ec8917529c - clean up allocators and devices
[Marcin Junczys-Dowmunt] b4fdf6134b47220955a41a4b60806c53829eb1c7 - remove tensor_cpu.h
[Marcin Junczys-Dowmunt] ae600b3ef7cefb82c569444b1477873e4c2229f3 - more clean up in backend
[Marcin Junczys-Dowmunt] dd296e77f76143033fc1589c7dce6d12196bbfdd - whitespace
[Roman Grundkiewicz] 8c898c4b0ac3cfb6b3e959eac3a71503c271b221 - Fix loading Adam parameters
[Roman Grundkiewicz] e5337f9fcaf1ffe6d26eac5a03de3d1f1fc81f94 - Fix Adam optimizer after merge with master
[Roman Grundkiewicz] 3cefe96c434b3bef23eea05ff20dc98906982e1b - Fix restoring validation
[Roman Grundkiewicz] 492e7dda4f72c8f8479fdeb6d2b7dd4d40d45a9b - Rename Adam params in .npz file; warn if Adam params not loaded
[afaji321] a0e74ad35ebbbecaa0d30007edd0e8d84a4681dd - remove debug
[Roman Grundkiewicz] 403927c287d8a513d74716acd649657040a23087 - Handle the case when the size of parameters is not dividable by the
[Marcin Junczys-Dowmunt] 327cfc1cc3fbe3ab92927c800aa30aef1c41d517 - pass through backend
[Marcin Junczys-Dowmunt] 6931f6c3c8ad5440e151a29787e89bab69ef9745 - make tensor device independent
[Marcin Junczys-Dowmunt] 0c959c1a7f0c0a95f2fb43b3ddc577fb43f22ebf - clean up tensors
[fseide] cf95c04ea65f9b6b207635d0f4c833537fb3d81d - commented the shape axes in transformer.h
[fseide] 449651811603925365551bcee4c53d4d45cd6295 - all-layers attention experiment
[Marcin Junczys-Dowmunt] 5e1b95e0bb0a6b86b75515c1069bf55596b6fe6c - finshed backend separation
[Marcin Junczys-Dowmunt] 79c385b894ed41c6ef699fb310ce1f59250bfc91 - reformatting
[Marcin Junczys-Dowmunt] 6e421f7a741dca8d7181f87acd398da60bb77f7d - add cpu tests for graph
[fseide] 952912d9b971b7a3c43708d6c162fb1ec2025f1d - added a comment
[fseide] 4f06ca2acdb9c308f9876165ee8c65cf7814c73c - test features crossLayerAttention and noQKProjection
[fseide] 13420687e9767ac4f39a992542d00ff6cc59769c - disabled crossLayerAttention and noQKProjection
[Roman Grundkiewicz] 01925e75551ecfec9fdae80c350b4cbaae5b4c60 - Fix allocated memory size
[Roman Grundkiewicz] 212ac9700e80109df2511f477e4e04f1bd5a57f5 - Fix loading scheduler statistics
[afaji321] 63ed7ca94118444154c767fea36aca31b4830280 - optimize memory usage on gradient dropping
[Roman Grundkiewicz] d4317805a2054dc86ea1b52fbc0841bb3bcb1f77 - Add build-* to .gitignore
[Roman Grundkiewicz] 0fc0ff234adc0f5dfff92463dd9f299d9e21f8dd - Fix after automatic merge with master
[Roman Grundkiewicz] 7c6110a712669e3f0a49204ac946c79c8a232fb8 - Fix loading Adam params; move load/save methods to .cu file
[Roman Grundkiewicz] e1a43361be83eeeeb06011a82fff14bdebfa6654 - Refactorize optimizers
[Roman Grundkiewicz] f1b426bb0b021ce72a52077275b1c4380d2a6d48 - Add saving Adagrad params
[Roman Grundkiewicz] 4c7df27d6bbad00a248ef87b2015b098e67b7d26 - Update CHANGELOG
[Marcin Junczys-Dowmunt] d9d66f416b3284516f44e5b8e7b1fc7a0623e33f - prototype cpu version
[Marcin Junczys-Dowmunt] f224db8257c1e0662510af76efa2252c54bd5f76 - add FinMKL
[Marcin Junczys-Dowmunt] 8d3f70599f3ff3944a158948ea9e147744909048 - batched sgemm
[Marcin Junczys-Dowmunt] ed2f6ec969bab2191569249a21031ac72639b9ac - sort rescorer batches by trg>src
[Marcin Junczys-Dowmunt] 1d6130e4c4c40464debaef6941791bb9d18b5f31 - add copyright notice
[Roman Grundkiewicz] 91c70e35addae0e49cf15dcbb2a588c4e5f5e930 - Restore corpus state
[Roman Grundkiewicz] b76182130db1795f98aba2daaa783c4df8160a6d - Add marian::data::RNGEngine class
[Roman Grundkiewicz] 193bc83e8826c0946188220bf7f55a0f859a23d3 - Save scheduler statistics into model.npz.progress.yml
[Roman Grundkiewicz] 03d56b47c17149122f27a1e27890c698068ef57e - Fix displayed costs for a very first batch in restored training
[Roman Grundkiewicz] aee256ee6f4d06269592e8fcb1d442e238e2aed6 - Make corpus batch generator a training observer
[Marcin Junczys-Dowmunt] 73c38cf82c83153210befb302b82764495477d91 - clean up cpu version
[Marcin Junczys-Dowmunt] 37399aeb9ff9ebd1b08cfc8550ad6ef170ddac7b - make scorer use cpu-threads
[Marcin Junczys-Dowmunt] c9103bc2af59764aba3ef9c4a030fc34cfa98ab3 - remove omp threads for now
[Marcin Junczys-Dowmunt] e4c651e20a7b10d3a8f32ebc733d19f935795549 - remove more omp
[fseide] 92cb1a4fcb46d067d4354b4e8602535631a16087 - special:model.yml now written to a different file for Philly HDFS
[fseide] f0205dc4c4eba3cc6167d08adf1e7b8e2e0b7f79 - minor fix after last merge
[Marcin Junczys-Dowmunt] 5bd46ef1e181f29dd090d2b5135c8fe428b451dc - add number of missing operators and cpu-based training
[Marcin Junczys-Dowmunt] 6dd57c38f0e8cb896453d0455ed5ac5f9e936dd0 - add exception safety to thread pool, handle different devices better
[Marcin Junczys-Dowmunt] ccbbbb95d3ed57c1517df0220841f857f35e5c7f - fix FindMKL
[Marcin Junczys-Dowmunt] 845063b3429f9304b7d09a7c43037308cc4d06a4 - re-enable mpi
[Marcin Junczys-Dowmunt] dbba0f220dc16d6c6104f67010e9ce3b9f2a204b - add cudnn back
[Marcin Junczys-Dowmunt] b18fab0823585c864a513baaf5d5c0a1e0ad8e4b - update changelog
[Marcin Junczys-Dowmunt] cead1b4feb753066f052e7a3727a13d1261fc64e - Restrict to OpenBLAS
[Marcin Junczys-Dowmunt] 736445379f22b8e135ceb6da04f671d3226e3e6a - skip tests for CPU if no BLAS
[Roman Grundkiewicz] a3a787c074b5735cf757ff969001a0a5942d0c66 - Update CHANGELOG
[Roman Grundkiewicz] 013e119677dccc13de8619e526ec4b8ea86b4906 - Fix restoring lower-is-better validators
[Roman Grundkiewicz] 3dc1a13415f678a8cc9e87ce94898a1b2a774214 - Do not validate twice after training is restored
[Marcin Junczys-Dowmunt] 162dafe8c1e9c2648e77f8a931b80d0ad7149253 - move try-catch block further inward in threadpool
[Marcin Junczys-Dowmunt] 76a1c3be8e8e8da9b2e0f0d707502de05acaa110 - catch exceptions in sub-threads
[Marcin Junczys-Dowmunt] b2410113362c6485726361f958daea995e8f5a74 - make search for MPI optional
[Roman Grundkiewicz] 3a88c763f2825c16819303bb2f97b9e8c13e1bf8 - Autoformat afaji's files
[Roman Grundkiewicz] 35b9438b7a1058dc1914b53afeb51db591c4fa77 - Fix segmentation fault before the last model->save() in async training
[Marcin Junczys-Dowmunt] 36e8077a7d1c1bfb537d1bac0d451a2312bba4ea - fix issue #183 by using boost regex, version prior to g++-4.9 fail
[Marcin Junczys-Dowmunt] d97d7bf3fd2a59fa3b9bc1d0c7489ba42d7d3691 - fix vocab loading for lms during decoding
[Marcin Junczys-Dowmunt] 3753573d3dea8e432bb10162353f2b3b01f02d65 - put back missing cudaStreamSynchronize(0)
[Marcin Junczys-Dowmunt] 71f911940c91d2fd5c337ddf4d8e88108d5ed822 - reduce memory usage due to lambda captures
[rihards.krislauks] 76b6cb4a75a390e041ef4c0c34b030c9256c53d5 - Add a bias in the multiplicative wrapper's xWm calculation
[Marcin Junczys-Dowmunt] fd208d841b0ec4fdbfdc81df601e86986528a051 - simplified interface, get slowly rid of keywords
[Marcin Junczys-Dowmunt] 14ce7ea7b93d1ca4ffc1c64f22c7397d0bec1f8f - updated tests
[Roman Grundkiewicz] 059a45e661fc486d113120e5e601e33acfbe827f - Add CMake option: COMPILE_CPU
[afaji321] 3db8cc1ba9bee8b32441ca71097e8249ef8cea10 - sum gradients in sync sgd if cost-type is ce-sum
[nheart] f4e0ad95ce9ce54124f33d433cf12c28d580577c - Add the ability to get statistics about target words
[afaji321] b5490b587bf71072bfd2db5e5f5804dcafe73ad9 - do not average the cost if ce-sum is used
[Marcin Junczys-Dowmunt] 716ca2dd7ca697c1a3575987a7a438c3f34cc40f - move gpu::Tensor to functional::Tensor
[Marcin Junczys-Dowmunt] edd10947a92f9dd89b0c41acccc8c6ccd73d6539 - separated gpu and cpu code some more
[Marcin Junczys-Dowmunt] 42293372caf2a052dc8d6693b4855413d8de5a14 - move to cpp
[Marcin Junczys-Dowmunt] 8ec6eef9d57c886513a581ce3c6d3502f5e019b7 - more coda separation
[fseide] ce335560fca37a865d5060e595d924c186eee5c6 - Vocab::load() now supports plain-text files as well
[fseide] edf9d250a73d941ee05fd6a420d2df05d5ab36cd - new option --word-penalty, similar to --normalize but additive
[fseide] 2d36e9435513f5a679caa6010f022a33575bdac7 - my strange experiment with alignments (--transformer-heads-top,
[fseide] 6be99bbb67f6cae4459815e3d392427d9d4f216b - added debug code to dump simply everything, show the op, and dump the
[fseide] 415b8b44092fa4bc70fc44c86807056e7c122249 - (added commented-out debug code)
[Marcin Junczys-Dowmunt] 4f72feabe069d3cf91a9f4fdf7d2c2c1afc0623c - disable gradient-dropping for CPU-only version
[Marcin Junczys-Dowmunt] d053a107c2815deee22bf3184fcddefca1467631 - place aborts in gradient dropping
[Marcin Junczys-Dowmunt] 83e9e8927ae7b1fc8d0597db439d2d852a993921 - add missing flag to CUDA_NVCC_FLAGS
[Marcin Junczys-Dowmunt] b8646ffe30f7578129f1746a75b126e6697ad67d - make pooling_test compile
[Marcin Junczys-Dowmunt] 9373561f2464f4953abcd105dfea265e3cf2d255 - cleaned line endings
[Marcin Junczys-Dowmunt] 1d3496613d583464c12b97dc448db04756f55019 - remove old code
[fseide] 65c5eb17d71649ace05e4dfcd9cdf223a7904706 - cherry-picked threadpool fix from master
[fseide] 61fd6ad11fad70563c93ca95a41877a831c91cf2 - added a total-word counter, for direct comparability of progress with
[fseide] 4be2bbe55514ee4c8123266bc29331ca599ebae3 - made transformer-heads-top an optional option
[Marcin Junczys-Dowmunt] 94650d9ea85a5a58404a9933cb15c60b2ce218dc - remove debugging artifact
[fseide] 28035f97f487f8dc5475ee76fdd6b783e1f66830 - fixed disabling of my experimental code path if not specified
[fseide] de4e26b23cec522277e0f538f244afb731be79a9 - bug fix: word-penalty option should have no single-character alternative
[github] 417ada1049e3e2fae96f6fa050cfac51f51fc589 - Updates graph_group_multinode.h with cuda_runtime
[Marcin Junczys-Dowmunt] 989335c243c24c57de6a59b0bb12e947bcafa6ca - sort vocab alphabetically if the frequency is the same, solving issue
[Marcin Junczys-Dowmunt] 944efc9650d260a88555530e7e2126adf8892be3 - fixing infinite loop causing issue #203
[Marcin Junczys-Dowmunt] 2a4bfb29bd8c952d13ac62241710fc44bf03e5f8 - add dropout and proper activation
[Marcin Junczys-Dowmunt] 7f025279a44cd09e000a0b96c18e65f69471ad35 - fix dropout
[fseide] 4c593ea7f33b2eea7d65b95004df290ff02b98f5 - merged Marcin's Transformer updates (dropout)
[fseide] 0a5dada11bdd84e4a60cbe56b35d41fe6517b47d - added a workaround for an earlier bug in my strange experiment
[Marcin Junczys-Dowmunt] 3ce0bceee51c932299992f67b0d069622366afb8 - update dropout operator, make dropout variational
[Marcin Junczys-Dowmunt] 2479866c9b7a914049cc078938103a1954ea9bb0 - update comment
[Marcin Junczys-Dowmunt] 89b7bb08176f11609e651abd2dc3b1f66ee9ca38 - add n-best-list input/output format to marian-scorer
[Marcin Junczys-Dowmunt] ade07fec025aaef22c2c2818224c8e21e69d6368 - fix dropout
[fseide] a172f87dee104942617e0ed7b61e7f429c5981d9 - bug workaround: initial LR should be 0 when using lr-warmup
[Marcin Junczys-Dowmunt] 94475d3a545e9878ff588022bb3f58babd9fba46 - change dropout back to full mask
[fseide] d1556aff1af1dc5e5f8736b4933bbf909b6986a9 - bug fix: Vocab::load() should not just overwrite reserved word ids
[Marcin Junczys-Dowmunt] ba5e11f1ae340c69cf990a375c76b94c809515f1 - init learning rate in constructor
[fseide] d62fb91d30f9cc72d98411dbc4ebc4216444ff1c - added pieces towards explicit sent-end prob and offset embedding, both
[Marcin Junczys-Dowmunt] fd1821ea909959a3b467da942f1204082470eec4 - add option to splitting, add init for trainingObserver
[Marcin Junczys-Dowmunt] c275afee1d5cc069479fafe81bc3463c2c945a3f - set initial learning rate based on scheduler
[afaji321] bc734248ac5fc2e4a75a1597cdbb7e1895a0c4fd - separate gpu and cpu code for gradient dropping
[afaji321] 76a7f2abdbcf3844f422d8491154c40e2cd3df40 - fix spacing
[afaji321] 548d71497c7df7867a7fb09018b6132aebb35cb5 - fix spacing
[Roman Grundkiewicz] 5f2eedc6e505eecf5bdef474be3e4f7066702fa7 - Restoring corpus state as a default behavior
[fseide] 78be6d12551395ff985215872f77d31273e9dbc0 - added minor stuff for debugging
[fseide] 4b396fc4923716e8d85aa486af4b940cb0740649 - new optional parameter transformer-offset-embedding-range
[Roman Grundkiewicz] 6d0c75cf48bab913e2c9c52f1c4c6cd0d656005d - Autoformat files
[Roman Grundkiewicz] 43a22f2210902a8ea9a091b00071a1701e2fe3ae - Clean includes
[Roman Grundkiewicz] 08cea72872098d8149470c29c72ede6ed8d403e3 - Autoformat missing files
[Marcin Junczys-Dowmunt] 31e2e19a6258c096e48ddf5f6b89a5e0b443d09c - use smoothed graph for final validation if available
[Marcin Junczys-Dowmunt] f5ef175c568f2ea81de6ff3eb231a12f78537a7c - update changelog
[fseide] a587b4c410f74a6a333e4785149ab338cce15989 - coded offset embeddings
[fseide] 615558ca3a52347ebb85506dfe211763d96eb555 - minor update of offset embedding
[Roman Grundkiewicz] 5843da93691016a8bdceb4a4a96c6aa0fa6ca820 - Store all validators in .progress.yml
[fseide] 6ea70c064278a344a7852c5242493d269ec8bc3e - bug fix in offset embeddings
[Marcin Junczys-Dowmunt] aa51ccef144179dfb3c030924e6feb3620b6e51f - update version to 1.4.0
[fseide] 82f91df3714ee6bfdf7aef48b5e9dce7a3d01164 - simplified the offset embedding code to avoid some reshapes
[fseide] b45dd06fa5dda51208c47e913a315788bf6ddb73 - added a fake operator delay()
[fseide] ecddeff6b10afc256ae6e0540a2bae02882f606f - commented a bug
[fseide] 870ada4bace4de5ea9faf27097eed36612d0fd4b - added a correction term to make offset embeddings work in decoding;
[fseide] ad0e1f31b7a8dd08589e9ebf251ebb80e2ef229e - (some tidying-up)
[fseide] c6eeed6404d483229899365bb48fbb72278f6f37 - (minor change to diagnostic messages)
[fseide] ee58dacd662df0b753c120dee2f7af1158ff8644 - minor changes to sent-end prob experiment
[fseide] 51ed00baefc52b5d388b1ecc04e6a81be4a1f5a4 - post-merge fixes (minor)
[fseide] 694a18da3666b6ae7d86ca800579070a20f01d37 - removed my personal research stuff again from merged version, leaving
[fseide] cd01415167ee7ed27f12f5a57ada6e56e3142ff8 - added a BUGBUG comment
[fseide] 3a8a10a2e5fc4865b7633c7187127cc319ea1f17 - added build command to README.md
[fseide] 5ca724bc590b2a1d1047444fe4f33ea2ef8778dc - moved npz_save_all() to cnpy.h
[fseide] 2713dfb62107a7c7d3da3f005c7fa56ab83c8993 - refactored for saving model weights and runtime parameters in one go to
[fseide] 807f2170a780445bf2fac9591ad5c6bc7c2fe753 - saving runtime parameters to .npz file again, in a HDFS-compliant manner
[fseide] 0c68d0f3f2e9cdfe1490f8b95d091807aa2e7309 - bug fix: Vocab::load() should detect the .json extension as an
[fseide] 4fd2dcf8f2b18fe139e7c593fde204979a5a5862 - bug fix: after merging with latest master, Scheduler::update() should
[fseide] bb622d052203ec679bd7d9239ed96ad6836eaf19 - -c option now accepts multiple YML files, with the later overriding the
[fseide] a728253daf2b25bb0717244dfb140231bee696d2 - new option --disp-label-counts per PR feedback from Marcin
[fseide] 89b1913be9abaf5bcda9efd8bf00954e34fb1f82 - new CMake target marian.zip for Philly execution
[fseide] a659955dbb80bcd06b40bd441fd25e44ec15d492 - Scheduler::update() now reverts back to previous normalization (by
[fseide] c1edd478996d569f249a8f3f861547adedb8829c - bug fix: with disp-label-counts, we should always display cost per label
[fseide] c4076833475f3179cf93275220d4b9375dd37351 - bug fix: NpzItem must remember its type tag
[fseide] 2559632b77b72ef08c417d97bfc62cf64a5c789b - bug fix: word-penalty parameter had a dangling alternative "-n"; bug
[fseide] 4b4cad63969faf8050b128308e3c6e6fad0c4cf1 - minor clean-up: removed an #if 1 in scheduler.h; avoiding a malloc() in
[fseide] 6f7b200c3d703e28443bd3f7911a511b3e92f908 - (removed some left-over code)
[fseide] bc95c059bacadf31ced8c3681c765775f67d498c - bug fix: AddYamlToNpzItems() should continue to use varName, not the
[fseide] c3d6deaa13efdb764a3196f82f13ba3736ba85a7 - minor clean-up of ExpressionGraph::save(npzItems)
[fseide] 461942e0978665479e30bd3b03f42e252434bdf7 - AddYamlToNpz() should not double-count the terminating zero of the
[Marcin Junczys-Dowmunt] e3f890aa2e8581105004ed2c9c8bcb6a616dc157 - add delay to sync sgd
[Roman Grundkiewicz] f314826eedf79a1a01aa5532e05fc90508b44a88 - Acknowledge IARPA
[Roman Grundkiewicz] 681113b5c8378c1f4c8af0ca36c4109fdc5c5f6b - Fix spacing in README
[Marcin Junczys-Dowmunt] a4d9144a3221b25b54186def3b54e6a1894a333d - refactor encoder decoder a bit
[Marcin Junczys-Dowmunt] 3dd4bb009d07b41beb62690aaa2dfa6682688afe - change keywords
[afaji321] b8d0a9bb6fd8569fdca83d7b4b2fe79d92287ae1 - fix scheduler update on delayed gradient
[Marcin Junczys-Dowmunt] 3f91e7dafe5f22b9328a0b7a1018c66e27b40780 - using passed graph instead of field, fixing issue #216
[Marcin Junczys-Dowmunt] 210ab16a7a79784d3f3df60dc0a7403a33971a77 - remove cost and softmax from model and encoder-decoder definitions
[Marcin Junczys-Dowmunt] 9548cbdcc60febf85f4bfe4cbf857ae01604b138 - towards short listed softmax and CE
[catarina.cruz.csilva] ade15ca0a20ecd8999125f6ac0b0d14b47661334 - Fix CMakeLists.txt
[Marcin Junczys-Dowmunt] 0e68aeec9948b82e5f89b2187dffddd337cd88f6 - shortlits
[Marcin Junczys-Dowmunt] 92794423086d321aac66b9f2b790b4d5cef17d11 - approx exp
[Marcin Junczys-Dowmunt] 9928625001414589b90ca3954a3d1ecbf3a33d86 - add gemm from sharp models, dull hardware
[catarina.cruz.csilva] 25f853fd01eb0f4c3350cc806f3dd19cbc6d281c - Fix ENDIF clause condition
[Marcin Junczys-Dowmunt] 9bbca0ac38235273a9599276efae1bcabc65fb81 - built-in quantization
[Marcin Junczys-Dowmunt] 46aa1082b4c8d32d219b43686a91d089e9b25e1a - make test
[Marcin Junczys-Dowmunt] b88d97876e9cb3a8779427c5aaefda7cd891b9c1 - adding bias and specialized transpose hacks
[Marcin Junczys-Dowmunt] 47ef6cc2b397e95ac945fcae32c3c86f8a1c6a6e - add transpose
[Marcin Junczys-Dowmunt] b0622430caab902efd9681ac9582d0f274168764 - fix bias add bug, comment out optmizations for int gemm
[Marcin Junczys-Dowmunt] afc3bde59710c2d5b8f4b0b4b731752dc2cf9a41 - add int16 operators, attempt at memoization
[Marcin Junczys-Dowmunt] f92db836a71f306d686d98eb2c1e8d4c87dbff25 - fixed batching
[Marcin Junczys-Dowmunt] bd06e1919ea908662ce6fe12d468be69c4f8c6f4 - rename parameter
[Marcin Junczys-Dowmunt] 482e33009bfa46adcc171e768194b1d4c309b9b7 - add a few more options to transformer model
[Marcin Junczys-Dowmunt] 2cc3cf16acf21293446afea85c62a8803acede6b - change layer counting
[Marcin Junczys-Dowmunt] 37018f8a6fb42516c48c7a81ab8775bf3cf874b0 - remove debug message
[Marcin Junczys-Dowmunt] 78a99473749ee038778f8b9ec37b16d0a62b86b7 - working memoization
[fseide] 75ee6ab2dc152d106eac650028731ac0250e39d3 - added marian-invoke-local script and libpython2.7 into marian.zip, for
[Marcin Junczys-Dowmunt] f8ef6f5a5f86706a1cda7cbb2ea17abfbcae2374 - transformer mods
[Marcin Junczys-Dowmunt] 07e51050756a7e87c3c60c69554cef6c609b3872 - update test
[Marcin Junczys-Dowmunt] 370183122da7b625df89cb6770598a1a1e9648e0 - autotuning for cpu gemm
[Marcin Junczys-Dowmunt] 11c83bb3d348c54549bcec64511b4dfaf5ffe4c1 - initialize new validators if resume and newly specified in options
[Marcin Junczys-Dowmunt] f0b18f240fc567e03420b92b3068efe139836b99 - add CMake build profiles
[Marcin Junczys-Dowmunt] d4c4e3201b43b49fd13d09bde1dca922c2684b82 - add option to skip linear projection after multi-head attention
[Marcin Junczys-Dowmunt] 1a9efd70b15f5e752c1e047532525164b598e36e - add back delete option
[Marcin Junczys-Dowmunt] d5f82ec999fce2fe483e6b3d01c6bf95fb26b200 - add optimization to scorer
[Marcin Junczys-Dowmunt] 54a533d45e5bcd6b08b0f9d880ad2043ef782fbf - disable type checking for tensor::set
[Marcin Junczys-Dowmunt] d49a711474fa1118b2bb90dea3384d2ea910d2cf - Prefer static linking wherever possible
[Marcin Junczys-Dowmunt] 3f1667726e3205cfcde1e624219aaf7b7a15c4d6 - get rid of obsolete python depdencies
[Marcin Junczys-Dowmunt] 2e77d5ecebcf48dd4114878cc9c136a7d09c9823 - use transpose10 only for matrices with blocksize div 16
[Marcin Junczys-Dowmunt] 5a9bb9792b48b5b3a680e4d309d95f13bf3f9f85 - continue linking against shared zlib
[Marcin Junczys-Dowmunt] 5b868db6dbe62387328ee1e936743325f5d37260 - make static tests optional
[Marcin Junczys-Dowmunt] 992494426c7c4fec0a6aa74bbb882af17f54c4ad - handle regex problems with gcc-4.8
[Marcin Junczys-Dowmunt] d59bd754bb9017a4b2258eb120a850c5a1c81318 - adapt save function for philly
[Marcin Junczys-Dowmunt] 5ba6c776bb39ad5b3463e2122014bc8bfbc5379e - hide sgemm if no blas
[Marcin Junczys-Dowmunt] 31efd4246883567eae69d5d96f3423f9cb55b8e9 - only add target if statically built, also create tar
[Marcin Junczys-Dowmunt] 6c7e4d21d6034297cc357edfab4a267364bd989a - avoid circular references
[hieuhoang] 00242739cb06f817ab329178f26c0c247aca9b6d - missing include
[Roman Grundkiewicz] 3fd204102fb70a40d90206cf5572dc4dd09141e5 - Speed up SQLite corpus restoration
[Roman Grundkiewicz] ac0bf443fa5ea6ab15bd494279b273be839fe073 - Fix restoring multiple validators
[Marcin Junczys-Dowmunt] 0fcfce9f4773c42e74fcfa8ed86ea04c0246bc0e - fix behaviour for final model saving accross different graph-groups
[Marcin Junczys-Dowmunt] a99d7386c88fd1c521225deb3c35fdf9b9965773 - make tests compile with static libs
[Roman Grundkiewicz] 8c34648365bec6c3c516c8d250ec1c8733484e90 - Remove std::random_shuffle
[Roman Grundkiewicz] 9101c00d31a3013cf06e3ccd6a700eb0aa9c0322 - Add option -f for printing whole arrays
[Roman Grundkiewicz] 93495c9fe96706c94d265a3fdbe6fab2ed50c3c9 - Fix segfault after restarting finished training
[maximiliana.behnke] 06d1aef10d5436cd68d47c3c7f6cb3207334e766 - Code refactorisation - Cost function from generic.h into separate
[maximiliana.behnke] 68cdaaf1dc35b4dbb5c3a458b81be159780d7f7d - Remove local inference check and costType
[maximiliana.behnke] be850f66787747b2e7f155b8e70fed4ed07d4a6b - Remove old headers
[Roman Grundkiewicz] 077379ee6a38ed4750b63ea315414f4466c4eeff - Fix saving the final model after training is restarted
[Marcin Junczys-Dowmunt] 79969dc7e446d6d6c66488787ff1cc23869112b1 - update according to PR comments
[Marcin Junczys-Dowmunt] 80b0a3ac2d9977cb83c3728da6364845191b6e4f - address comments from PR review add two new models types
[fseide] 5510c369a9813a4301bf7b510b9b0ad53168cc1e - cmake now builds .tgz file as well
[fseide] e83a1b0690c3d6327313c41a0c5c9dfd1654c513 - new option --interpolate-env-vars
[fseide] ecce13b943fc99d10c8d0f85d8a1660cfd2e1bd4 - bug fix: cmake dependency for zip/tgz should be marian_train, not marian
[fseide] 44d77d98e5d7f43f48c39324cf71aeed8d8dfa6f - validation log now contains Epoch number (needed for Philly)
[Marcin Junczys-Dowmunt] 15f780ec91ff9daf626dfb95481fba73e1c6783f - duplicate library linking to solve issue with static libs
[github] 31f4b43b870833002c555bfc82380f8659f5e07f - This shouldn't be executable
[Marcin Junczys-Dowmunt] fa79736a92348e0784bfb18786f3b12df6358d38 - add more layer types for transformer
[Marcin Junczys-Dowmunt] b28a157b35f69f67e8f4d94b632980bf8b189656 - fix segfault due to missing decoderState
[Marcin Junczys-Dowmunt] 68f166e03b1bd4753940361a4e4a105ed3d07379 - fix another segfault
[Marcin Junczys-Dowmunt] 92fa782d3ba3fd6917632597825c08a9326e027d - add missing rnn cell
[Marcin Junczys-Dowmunt] 86bcd6aa7118564f1267f7f45a1bd63949dcc722 - add prefix
[Marcin Junczys-Dowmunt] 90d32cb63df780cc156bb01d3a17cdff76788f9d - invoke rnn construction
[Marcin Junczys-Dowmunt] fef9498d9631e46aca79a06fd5a06b09f213040a - grow dims
[Marcin Junczys-Dowmunt] 0cfe3181fb4f6b8bd046a9d7a1a7a29fadb311ae - proper gate
[Marcin Junczys-Dowmunt] d94c8a872224254a23774df47595d87d5fad56a6 - fix gate
[Roman Grundkiewicz] f428bac71c6b07dde937c6604a8cb2673047533c - Fix parsing devices in multi-node training
[Marcin Junczys-Dowmunt] 6597f6c90cd806b1784eef3592b037ddc4ec7b93 - fixed decoding with AAN
[Marcin Junczys-Dowmunt] 93c513bcb9ce449286a9e17656f12f4ebfe04015 - change to simple highway gate
[Marcin Junczys-Dowmunt] d25102051723ab11e27788db2d7a9c443c3bf7d7 - add back FFN and gate
[nheart] 19ec55e57f8657f1bf18f0cf5aa27b1c7cdc4088 - Fix wrong argument
[nheart] ab893208dc6f47231871655ab24432df7504ca5a - Ensure just one device is saving
[alham] f697d6de500609cd38e9e1b2f99fb42be6e28eef - blocks other node when validating
[afaji321] faae5c594a4fde84680bedefd1df1082d4d63525 - add barrier
[Marcin Junczys-Dowmunt] 64e2cccb50b99006ce8c583592410a35fc913692 - add options for AAN customization
[nheart] ff219e307c3ef9e8b071ac3a57bbefe86309d239 - Revert "Fix wrong argument". Apparently that is intended
[afaji321] dadf975d482c623fd67f622f6f7248639b7bef80 - change gradient push tag to avoid race condition
[nheart] 1973b473c824610a4ff335ff827e3b016220fa03 - Syncrhonize before training for the first time.
[nheart] 0188e1fe680fac88a1d567cc1e8a3e13cf891391 - Initalize the paramShards to have the same parameters as the params
[afaji321] 4599d8828c1cbf01c7ac3e60c5cb1b7cb58f6d96 - synchronize parameter shards
[Marcin Junczys-Dowmunt] adefc92a4eb79a4fa77b0b439ed203d3eabbde0a - optimized element
[Marcin Junczys-Dowmunt] f637b1caab250dedc6823335e47801e86e28a47b - invert recursion, more readable
[Marcin Junczys-Dowmunt] 1b5479cd74f1e84c54df4d7e405dcc3554753d15 - add optimized element-wise cpu function with proper broadcasting
[Marcin Junczys-Dowmunt] 6a7ae80537bda8f1a838cbb5b2ac523cb801640b - small optimizations
[Marcin Junczys-Dowmunt] 7d5ea76d7b671cb3282a964e79e05dd1787d4aca - clip gemm to given value
[Marcin Junczys-Dowmunt] 6571741b735e1fed9f999476361f0d68de6f1118 - clipping gemm
[Marcin Junczys-Dowmunt] d94f28b648ea5c3d70b4a18c73319db9a93453a1 - merge sse avx512
[Marcin Junczys-Dowmunt] 29fa501145d768e0470ca2b361897eef742b9431 - add gradient to clipping function
[Marcin Junczys-Dowmunt] 36e312d2bc39cad255b374c3603f6edd5e939616 - add max-length-factor to training
[afaji321] 780c416f56098ee28d54cd55a96c04b5fe943b08 - add optimizer delay
[nheart] 8e45570417cd086f423c45b950365f6e0406fcb0 - cosmetic fixes
[nheart] ea79da67fd5cf0fc114e43230ad1aa8a93208e65 - Begin implementing local optimizers. Fix device initialization
[nheart] 5484c2f3e2711d9b93327d8e2a7c414c6e2c6950 - Implement sychronous graph group multi-node
[Marcin Junczys-Dowmunt] 1cffa82f2d154f2fa3eb6cfbbb2ada07c2d3f75f - small changes for shared task
[github] a6f55830127eff5bb4ab5cbb06b94872b34eb8c6 - Add int8_t multiplier code (not used yet), refactor intergration
[github] e1bc66fb66a93e0adef0f94ae72a1c655316a438 - Add int8_t Marian wrappers (but not ops yet)
[github] 5a9569d55214e6930a29bf0a7e6b7a6b39f38b4f - Sorry, shouldn't have committed this set to false
[Marcin Junczys-Dowmunt] ca73a7c0bd502dff6c9f07f8c19a60ec0b3b4151 - add batch-size based on words for decoder
[afaji321] a7c89b5dfaac3004b8814c994e3df43d887c183b - bug fixes
[afaji321] 7877795404a9923171d214e615145925144919d7 - remove debugs and add delayed gradient update
[nheart] ae487e0f99be5c212213db1df053a5705796db1c - Removed unimplemented function
[nheart] 2a839bfb69904235307ece41a712b98c8410e0ae - Fix communication size
[nheart] 34ed548cbf574d73f753134e36e5f1066b3c9205 - Cudamemcpy also needs to be smaller
[nheart] 309158d237bac9cd7b16cf8e267f3b8735829a04 - Properly finalize MPI
[github] 1cac15c016e88d1c32d351909f9fd79c15b9df7b - Bugfix int8_t: last rows, -128
[github] dbeeb88b0c93028a114135841002056537f630df - And fix compilation
[nheart] e2651e67ae4d38cf4867a2b6b23497a2a7dbf180 - Don't require thread-multiple for sync multinode. Proper finalization
[afaji321] ce229208848c15bf1c062b5ee94a551f6bf9fa9f - multiple batch stat by the number of devices
[github] 404aded9f46daa0174e0e1ad6823e182062aa8bb - 16-bit with clipping. Untested.
[github] 2e62d8569655eac844e067b99c6195abfe8c51b3 - Revert "16-bit with clipping. Untested." Moved it to another branch,
[Marcin Junczys-Dowmunt] 3b70fe721a5703ec4bb9382df86d06b60a71a5d7 - allocator granularity
[fseide] 43d0db057c9666090230b414e453c5b13bda50de - --config option itself can now use --interpolate-env-vars
[afaji321] f275803b2820b3f1e506c571858bba79a5f336a9 - remove cuda
[fseide] a7e7f52bc9321b0178ab1a89a325da9ab39817db - workaround for Philly, to translate unified Philly paths with cluster to
[afaji321] 6fb921c16b18c9ee849cf4c1b064d09d43e4010b - add exp. smoothing and remove bcast
[afaji321] 5c0307f746774cdb942332cc8cc41866c4047821 - add missing MPI_FOUND
[jonathac] 3e68deff5285ddd22ef3f97c54978e14d60b65e8 - Pile of truly horrible hacks to almost get things building on windows
[fseide] 2f8b093a6d9255ee0d7204c61c03a1fc79f9e136 - npz_save() now saves to tmp file first and renames; Philly progress
[fseide] 2280fd573db65ec25c28733ee463206f8bd45229 - bug fix: Vocab::load() from plain text should handle Windows files that
[fseide] 3ee068a90f24071261806b253f2156a4d48a0374 - EOS_ID and UNK_ID are now determined from the loaded vocabulary file,
[fseide] 450746bca7571c8c394d73144141898154ae974b - updated two more places that instantiated BeamSearch
[jonathac] dec121fc8d358115d532e92e1b4c7d9eb9da92cd - Fix cannot open program_options.lib
[afaji321] f1d4dc95d893d62986cb1395b1eaa839577f1755 - optimize asynchronous gradient dropping
[fseide] 40e4faf63b9e6da48c7140a3692eb1b63ea3224d - back compat fix: vocab reading should accept alternative tokens eos and
[fseide] 8e86ec567941e2492e261e6fe0ff4d604639b63e - should use constants instead of literals for vocab ids
[fseide] d1e0bf474f523ae2d21e66104fcff0adbceb2de4 - updated Nematus back-compat
[Marcin Junczys-Dowmunt] 3e81a742e1f10ab11dc834370ebdd28063a448f9 - remove rnn layer for now
[fseide] fda7c14b8b8f2d093d848b48178c918fd1d3d2ee - added dummy vcxproj
[fseide] 6d10a3532daf8aa52dc92ab4cc5268912babfcd1 - towards a buildable and usable VS project
[fseide] 4d30a86a5110c32fc388bfed4c27cd92cfbb8739 - bug fix: headers should be ClInclude-d, not ClCompile-d
[fseide] d0281dd5926648db642290422da2e13dc716bf65 - temporarily included marian_decoder.cpp to have something to run (later
[fseide] 606557e8056cfb6b35f65e74f45a26e10dfa9570 - MKL now compiles (but produces a garbage result)
[fseide] 33b77b0d2bc82db492b5b956437645f913affbe7 - bug fix: need to define MKL_ILP64, in order to match the linked MKL
[Marcin Junczys-Dowmunt] aec7d5e6e876ec670b8312a88339246b388f233f - fixed merge
[stanczakdominik] e95320a6d1e43dc92475ed10f6a9dcdc65d55a3d - Remove leftover references to Python bindings
[fseide] 9beab3e6bb28a5c8ef90b118731273e52a177640 - temporarily disabled an error dialog box caused by an unaligned free()
[fseide] 9b23b426ab8799729517ccdf231bfbf2eefbc817 - disabled Warn=Error for Release so that we can build it;
[fseide] f570f621ad1e10fb11340d02d4519b546a2d3872 - fixed the build options for MKL
[fseide] 78028e843633809237242b60961ee053cfd5f416 - fixed some hard-commented-out pieces of code that did not use to compile
[Marcin Junczys-Dowmunt] 44c2cc11969fcd3aea742d88c77e9ad268caf914 - address comments by fseide for PR
[Marcin Junczys-Dowmunt] 68d61a662294cb3f26b3935da95a8ce1c404c293 - Use original cmake
[Marcin Junczys-Dowmunt] 7447dcba4fe3a0de4920d740658df18b3cf77b92 - small fixes to make things compile and pass regression tests
[fseide] 7ad6e4f96d29a76f81044ae94a7dd93eee06b5a6 - updated the vcxproj after previous merges
[fseide] e3f32ad8928371f39686eb417670560cc824157c - moved vs project into its own folder
[fseide] 9eee77a663fc2575a4d8cc516afcc20b3a678555 - updated .gitignore
[Roman Grundkiewicz] 7ae48b7e2fd6e0b2a381c68d73ef290fa39b3b5c - Autoformat files
[Roman Grundkiewicz] 63f1845087c97dfe6b4b7b70ee82a0f926d9d995 - Fix MPI and CUDA #ifdef directives
[Roman Grundkiewicz] 4a4723fb3003791e144b9c9b3234aa4c64470696 - Fix CPU only build
[Roman Grundkiewicz] c6ea6f324f96b744aa6b8426c6e1eab3b4943124 - Add Jenkins CPU-only build status
[Jonathan.Clark] 2882b441b7ff615421ed6b52c4357751e2204d10 - Merged PR 3923: replace aligned_alloc with malloc
[Roman Grundkiewicz] cc587480ac3d029f39f4fc7e9fec12b021dab024 - Refactoring loss functions
[Roman Grundkiewicz] 8974fcc0ad6afcc55c0cd88ff873b7dc81a6a973 - Version 1.4.1
[Roman Grundkiewicz] 33a91c35c83b16371209eb9afb70c9e8a75f6dcd - Save and restore exponential smoothing in single-GPU training
[Roman Grundkiewicz] 8a3550ecae286d7b7964a90db14f1dd5ea3ed6a8 - Rename movingAvg to mvAvg
[Roman Grundkiewicz] 3acccf1b04832df75de5ba12b21776e206b5f654 - Add --alignment option
[Roman Grundkiewicz] 4a4b8a33bc8a673effcab16f0df39e3d095d6b6f - Remove commented code
[Roman Grundkiewicz] 3cd64a463e1d1d153964b5288eb29e54c219bb87 - Return alignments for beam 1 and mini-batch 1
[Roman Grundkiewicz] b611bef0f00fed0dba7919d96896253b7692a547 - Autoformat files
[Roman Grundkiewicz] accddbc831f4700f79f5e5597c8d3ed16415af18 - Fix generating alignments for beam sizes > 1
[Roman Grundkiewicz] 515d7bbfaf42b6ef7f578ade5e2418ded719a67d - Fix generating alignments for mini-batches > 1
[Roman Grundkiewicz] 4e8543c74a64df11a95443ccae5e250966c2d3a6 - Use masking for generating alignments
[Marcin Junczys-Dowmunt] 71cb7b5bca0bb2a571474f2a59fa56c286c5f546 - used batched instead of strided gemm
[ubuntu] 54dec27eb6a540df63f657f2e143dcce2c500edf - more efficient concatenation for certain cases
[ubuntu] b942d83e1202eb4904224b6fea1c962cc1ae78a3 - added special function for bias, TODO: make this general
[Marcin Junczys-Dowmunt] 2c6ffbda4b027c8cd2038f6d8866808572825377 - specialized transpose
[Marcin Junczys-Dowmunt] 7901018ca4abfcf89b1c383b492c812d3df069a0 - a bit of clean-up
[Marcin Junczys-Dowmunt] 04c081e92d5cd858134ce7cfd00d8d30d622b6da - remove obsolete code
[Marcin Junczys-Dowmunt] 227b0debcd5da55fe125c200127a5a7f057d678f - do not recompute contexts
[Marcin Junczys-Dowmunt] ab848f81d6de6739d9bf663c733aec5053a4c0a0 - add comment
[Marcin Junczys-Dowmunt] 91c19625e02f4598fb73f0a7b767b3f177487cf9 - update version information
[Roman Grundkiewicz] 591847d374bd1e8bf861fae785429e3363121d18 - Refactorize returning alignments
[Roman Grundkiewicz] 064e4e0f5f269d67c4f60fe9c3aad9fa15cab213 - Fix --alignment in validation
[Roman Grundkiewicz] 8facb11cca0da981fe51f003a3fe9db56545c58a - Update CHANGELOG
[alancucki+github] 4b5a8eedfc03bc832f82f989234e3aca3cd69c82 - Calculate Glorot init scale from trailing dims
[Marcin Junczys-Dowmunt] 38f4b81881f865d2958bf8d30529be94f628a5f9 - make it compile with quicksand
[Marcin Junczys-Dowmunt] 388217698da5895e95980358edbf2a78fc9a8fd2 - change add bias to matrix multi-ly
[Roman Grundkiewicz] 52d72b519ada4425dc577e1536b8bf7ca5b86c40 - Unify variable names
[Roman Grundkiewicz] 8088043e781c0b9ce484d06941ad47a6588d7e0e - Save and restore exp. smoothing in sync training
[Roman Grundkiewicz] a7a59e4fbfb490bf20104adbdc502265d80632b5 - Save and restore exp smoothing in async training
[Roman Grundkiewicz] a67568a5fb5584cd8dd648387d7c99daa42be1c0 - Refactorize sync/async graph groups
[Roman Grundkiewicz] 3964a700b9f1ab7318513cb33a6ab549545233e6 - Extract exponential smoothing to separate class
[Roman Grundkiewicz] 8b5b4386afb9828d9520677e93e2acc5f7d89265 - Use CPU for loading graph with avg params
[fseide] 8adde0787ebc360fd82c455e8ed04e87d42d0b44 - rolled back the accidental change of the vcxproj
[fseide] e87564c6643dc0b1d4344ed69e32bf72e009b12d - bug fix: History::Add() should obey the actual EOS symbol from the given
[fseide] cca9405d2ce807019e5b619be408c9a59c002bd6 - made last commit compile
[fseide] 4eb1cce8557e464ef4c9030ee50094628b6d27d5 - some code simplifications of transformer
[fseide] 79cc3af85a8dc3984235b45dcae149edb438296a - renamed logsum to logaddexp, consistent with numpy
[fseide] fea201254f2f1af29f678a97141565fa86a6d5ea - renamed logit() to sigmoid(), logit() is a different function
[fseide] b0940e0c487bcd39a30aa520d024ffb203c564a1 - renamed max and min operators to maximum and minimum, as in numpy
[fseide] c450ca9fb0e0407af20b61148ef9c67360ba66a9 - further small refactoring in transformer; renamed layer_norm to
[fseide] cd6157bd968896b5e477d26886706e4169d6af83 - some more Logit->Sigmoid replacements for GPU
[Marcin Junczys-Dowmunt] 695fb469dde46c16b6ed560b49c460e978c4dcb9 - changes to make quicksand compile with marian
[Marcin Junczys-Dowmunt] bf4a29b581bd3e2b93ac42d78fee0f06a9f8bcef - remove debug artifac
[fseide] a1c52b05f34715bd7d47509700e9b1bceeb976e9 - further refactored LayerAAN()
[fseide] f712748a5ce9d1dc89bc3afd36c40b977e6372d8 - bug fix: refactored affine() expression should use the correct variable
[fseide] e6aaf2be821e2f56b7e0afc1030c9f19bd6a6e28 - addressed PR feedback
[nheart] d80f587bf46dec3aee60ac7f08a4496eaf151d27 - Remove unused ThreadPool and destructor, causing a crash in mini-batch
[Marcin Junczys-Dowmunt] 5a8edf445bc244c41beaac85e7b6de47a295b4da - fix segfault
[Marcin Junczys-Dowmunt] b4a588669d0f983bc5313ed39e6c23a5c663aa22 - update CHANGELOG.md
[Marcin Junczys-Dowmunt] 20029f534ae219c50c39ef79ac44ab86048b797c - separate batches for sync sgd
[Marcin Junczys-Dowmunt] 50a352587e446c87e8f7c5693f01a2898d785666 - reduce memory usage for syncSGD
[Marcin Junczys-Dowmunt] ff4725063e18c0a06412663ee4556404c624582a - remove threadpool
[Marcin Junczys-Dowmunt] afabe4e7e9c2e97d7d234d39cd76a37bf5a3af11 - try to test nccl
[Marcin Junczys-Dowmunt] ec8fab721cfe08374344ee29a3543d92751e0752 - some more nccl tests
[Marcin Junczys-Dowmunt] 17f4384a66b9b25f31404be709c5797f4d0f8fd9 - better integration of nccle
[Marcin Junczys-Dowmunt] f68cf82f642bc6ddf7b4f1adccba23adfc73a31f - clean NCCL support
[Marcin Junczys-Dowmunt] fe13bcbf838b263f4a43a25f38a1b5b63fcc8f9a - remove debug info, fix missing mask
[Roman Grundkiewicz] d1d0df98d2b0df859d06fb38ba376a5a717c3bfc - Abort if training nematus model with --dec-cell-high-depth > 1
[Marcin Junczys-Dowmunt] c6350c666f293c64f3745d20d9cb9796eed849c8 - fix transpose operator
[Marcin Junczys-Dowmunt] 94645a31fc93f0a93499027cdad16b7ac33ca42f - fix remaining missing gradient addition
[Marcin Junczys-Dowmunt] 0edf3b3913c04d7a90d2af6797b9b817ac94dca9 - add proper gradient summation to shift operator
[Marcin Junczys-Dowmunt] e7344470d18c182aa8cc19471e755769a0e6e47f - ../CHANGELOG.md
[Marcin Junczys-Dowmunt] b9197c2a5a4714c576f65a825553cfde429a09e1 - adaptations to quicksand
[Marcin Junczys-Dowmunt] 352a437ab49ec00be944e11ed4bba0d52ac49931 - make cols operator non-constant
[Marcin Junczys-Dowmunt] fb1f63e74388ecd606ac9b2215d3c2aa163c82c1 - fix compilation issue
[Marcin Junczys-Dowmunt] 0cdde80c9b4a6e5f4e2ffe030c6d83931cfbc150 - another attempt at fixing gradients for transpose etc
[Marcin Junczys-Dowmunt] c0d526ad76ab01d318a9ce93718389915431145f - proper destructor
[Marcin Junczys-Dowmunt] e55b573b0def0b9c4f4b225a0eecd3c572e9b726 - correct check
[Roman Grundkiewicz] 030c3a178131d9f11a2b003a1fb0e32cd61ccb3f - Fix logging the total number of inserted lines
[Marcin Junczys-Dowmunt] 2b60fc5e04a7f4d64be90352fe5379789c350309 - use NCCL for averaging
[Marcin Junczys-Dowmunt] 2429919ac2e31b807a72fad9cb37dfc3fc69f486 - use non-nccl primitives for now
[Marcin Junczys-Dowmunt] a50bfcf49cbedf208abaf61d8149f7bd394e4e5e - add swapParams to communicator
[Marcin Junczys-Dowmunt] 5fba06c55f03d506c311d87cc577e8cad22a0cd8 - divide createCommunicator across .cpp and .cu
[Marcin Junczys-Dowmunt] 6d4138a4c82f29f08c00fa9194ba77828030c453 - make communicator compile with different defines for NCCL, CUDA
[Marcin Junczys-Dowmunt] 6c765bd616c24ee7312e6784b7e492553dcfaa5a - clean-up and mini-batch-fit bug fix
[Marcin Junczys-Dowmunt] 468678f54549b5351274fecf7ce694347acf58a7 - change option name
[Marcin Junczys-Dowmunt] f85e63d77972b4214550509558a2306f3ddbc989 - Update changelog
[Marcin Junczys-Dowmunt] 365cdc97d013b728449aeba97fab15246c0aaac9 - small changes to fix linux build
[Marcin Junczys-Dowmunt] 58ece991c2fb3275b175bdcbf4c63199633a5928 - re-enable sqlite support
[Roman Grundkiewicz] 9bcb4ee5a30de27f9ac8976f06f1c5ef952f7930 - Fix #262 PoolingWrapper dependency
[Marcin Junczys-Dowmunt] 664b1d01b7578db393e6b95ccbc29cfca18ef405 - add back division by batch count
[Marcin Junczys-Dowmunt] d054de29a70e5950151a04698022a83994b531f5 - option to disable NCCL
[github] 4e6b22f269e85d65dc37203eae1e16544f0f91df - Update README.md
[Roman Grundkiewicz] 1e27a90899011c10f59af3385b75562f1d0187f5 - Standardize attribute name
[Roman Grundkiewicz] d93da1aa004942706bf2ee59ee4ce67e73c10732 - Add exp. smooth. restoration to sync
[Roman Grundkiewicz] 3c54a1a88a9220c068b8bdb1111bf2a9aec884a7 - Fix saving exp. smoothing for singleton
[Marcin Junczys-Dowmunt] 31d634805199f2a2d56f27e8615031d1e8242a3e - add back div lost by merge
[Marcin Junczys-Dowmunt] d2991d58551a9a8f73a0cc5536d2a9e595438185 - make sure that batches in a batch split do not have size zero
[Marcin Junczys-Dowmunt] 2bcc74b52509f56c579741881dd1fc5101aa9e98 - fix batch splitting
[Marcin Junczys-Dowmunt] 63a22690fd4dce326616f46f2eba25318c20b345 - handle empty batches
[Marcin Junczys-Dowmunt] 586eccc534efe592285854da3f22eaad9d0a92f1 - correct div value
[Roman Grundkiewicz] c8a6324eea681bf69ee40e98f2ed4fffd1567854 - Update saving averaged parameters
[Roman Grundkiewicz] d2400fb6c38b4e050a3626c8902cb0a4bc370e50 - Remove virtual function
[Roman Grundkiewicz] c914fe015b0fff9c190e0314eff9b7d46005b940 - Remove unused function
[Roman Grundkiewicz] cbfd1f9fa6d22c0600401a22bcce67214d3c4178 - Fix loading averaged params in sync/async training
[Roman Grundkiewicz] 249f0b2119cb0e2e3992013882aec987396f4d0d - Refactorize AsyncGraphGroup
[Marcin Junczys-Dowmunt] 24e750059fff308d453c928a0cad6f8417b5cf39 - do not produce sub-batch splits of size 0
[Marcin Junczys-Dowmunt] 2c8a91c1b5e0da1c781a5ea9a9b34e73c1127ebd - do not create marian loggers when used in quicksand
[Marcin Junczys-Dowmunt] 0a94e1399fc88216f2f3c979a279c41ed73e7176 - fix batch size after splitting
[Roman Grundkiewicz] 48f1998b74ec917456a81b95a369ada3b4f33cda - Initialize avg params in separate method
[Roman Grundkiewicz] 66316ffbce277616e481bdbe70d8da64f818ba89 - Remove graphAvg_ attribute
[Roman Grundkiewicz] 604e3960081533cb43c53389b9e63d033198fbc9 - Change to Moses-style source-target alignment format
[Roman Grundkiewicz] 54fcaa7a908df411550fca6c3eb4dbda2a60a4ba - Add comments
[Roman Grundkiewicz] bcba5f8b864ac2deaa1c047a3a3604da7eabae9b - Add alignments with thresholds
[Roman Grundkiewicz] 9f9824419ccb6e1feebc6dee9b81be5b06924b59 - Replace method Printer with class OutputPrinter
[Roman Grundkiewicz] e2639deb33c7eb9f21222d7351c79172f8a57332 - Refactorize OutputPrinter
[Roman Grundkiewicz] a35abad143f89aade4985c030dcef683559df89d - Implement alignments for n-best lists
[Roman Grundkiewicz] e1d890131c7d15ecf9f6206aa85f9626c8df9c5e - Update CHANGELOG
[Roman Grundkiewicz] 8b0e2f951b5ce09a622fa7239b2e1e5bd8344fe4 - Fix alignment in validators
[Marcin Junczys-Dowmunt] c93ce4de18a77f6b6623d5b2ca21cf45a6affbdc - merge with caching
[Marcin Junczys-Dowmunt] 5611bc941b23e240ec7dab5eed71f4fc9f27302e - add missing return
[Marcin Junczys-Dowmunt] 68e4e3b9d95cc00bf1febb2e2c6b3c014470e40b - add proper caching
[Marcin Junczys-Dowmunt] 21aa530f7510fd5aa1e84430c30fb707c7571f02 - built-in bleu validator
[Marcin Junczys-Dowmunt] 965db9d8fce16691b6054c39c108d97e43e31af5 - comments
[Marcin Junczys-Dowmunt] 45430353dd854d613739d74245e1f8ba9c2423d7 - explicit types for min
[Marcin Junczys-Dowmunt] 5f2e34caedbf85c643eebd5040037e5169df95a0 - fix compiler warning for unreturned value
[Roman Grundkiewicz] 12be95196312f78db81b3d7f93ca85bf402084d6 - Rename Eclipse build
[Roman Grundkiewicz] 9557b2b82e8bd1de8a5030229425b74824b188e8 - Move VS build
[Roman Grundkiewicz] e3f8f6a9819aadb0e1deae47b0dcae2846d71395 - Install clang-format if not installed
[Roman Grundkiewicz] 00f12c339c9a2c2010d77859fba60626a5f57a25 - Autoformat files, part 1
[Roman Grundkiewicz] 2bfaaf29d2965d7c9b4ef2407dc491a024033b83 - Autoformat files, part 2
[Roman Grundkiewicz] 6cb612d7c632540e20dfa2129ff6f1cfebdc0f50 - Autoformat files, part 3
[Roman Grundkiewicz] f02b98dc4576b4a7d45f72658b702de58a151cd0 - Autoformat files, part 4
[Roman Grundkiewicz] ecce0db01d35f1472c45486b51eec171140edaa6 - Autoformat files, part 5
[Roman Grundkiewicz] 7c45235e4af65e8259e2650261705ee30b86faf3 - Autoformat files, part 6
[Roman Grundkiewicz] c3807da7374176b1253d27c4d334c2af0fa20633 - Autoformat files, last part
[Roman Grundkiewicz] c73a2cf6fbe58dd7b605c63bf6b5c060c380397f - Fix compiler warnings
[Roman Grundkiewicz] 442de7024962f5c88ef4ba6d74de4d3ceb3bfb8c - Make shape_ private
[Roman Grundkiewicz] 0f3c9dbdae6f8248096c31f8fcd6de12409fd9f0 - Add comment on saving loaded model
[Roman Grundkiewicz] d65ec6b9620feca4fe18819f4db3ed47b36f5c9a - Add comment on tracking validation
[Roman Grundkiewicz] 78cc62e1beddc95b96902467f1e9fccf9db590a3 - Move checking model/config file to validateOptions()
[Marcin Junczys-Dowmunt] fab8e011fbcc7c577724aa38fd070edc2195c752 - mapped model loading
[Marcin Junczys-Dowmunt] d63f21e01ab3e7ed7ca28050f7f78ba4ada6bb94 - add a binary mmap file creator
[Marcin Junczys-Dowmunt] ed2199c6d1269e299c4ba1da9a26c50a56b2edbb - comment out automatic mapping
[Marcin Junczys-Dowmunt] 7c1d05c9e4da1af51bdde6e4ce2ce1951ebfdbac - refactored binarizer
[weyo.huang] 3e19457912a5310c3f7d566b668a23f1b9f92c3c - avoid cudaMalloc in CopyRows
[Marcin Junczys-Dowmunt] a2fbe92e16c1d593eb4917c310c1b5b08a992378 - towards cleaner model IO with mapping
[Marcin Junczys-Dowmunt] c7cfdd4017ca010632aacabbccf9223be7f09dc9 - working binary file with conversion
[Marcin Junczys-Dowmunt] 3d0cf1b7afcab8c58b2d7a4e743807e110f0763d - clean up code for model io
[Marcin Junczys-Dowmunt] 28adbe2a92a1641276c271954036f03ade324f9e - more clean-up
[Marcin Junczys-Dowmunt] e448e9f9b8f9286b31d329f02f506ca6bc4ea75a - prepare scorers memory mapping
[Marcin Junczys-Dowmunt] fedab0dc906859806e528c649f61431d7d873ce1 - fix condition
[Marcin Junczys-Dowmunt] d96c03f7cb599f4b1d7ccdfe05e13085242503c9 - find static NCCL
[Marcin Junczys-Dowmunt] 2b70bfbd5c5689c4e023715a09f236778536946a - remove spurious message
[Marcin Junczys-Dowmunt] e7f70b81d117941e34f392173adb8a22dda44a15 - do not use vector copy
[Marcin Junczys-Dowmunt] da43a5288b4d9b1cb924df0481ef6da80899a887 - remove stray p
[Marcin Junczys-Dowmunt] 72325598fdc68a00fe40980511595785a5e1cbd6 - remove cnpy.h from includes where possible
[Marcin Junczys-Dowmunt] 2bff43075c530e58447621eb2598d64a44b1f993 - remove cnpy.h where possible
[Marcin Junczys-Dowmunt] fab48eb3623f4426280069c0fceea64dfe9e4445 - use new io for optimizers
[Marcin Junczys-Dowmunt] ec2825bd3a7f7d079c50120eaaf0f70cffe37041 - use emplace_back
[Marcin Junczys-Dowmunt] c879ee2d787a1d8d1cbbe1488b8d9973081ecb22 - fix data size
[Marcin Junczys-Dowmunt] 08361fd10f1bc5910e4101d1a6efd6033bbad545 - fix dimension
[Marcin Junczys-Dowmunt] 9834e09a88022819262387af714764934c72caf8 - correct integer type
[Roman Grundkiewicz] 1acafbd1b30e9c7169324cb7e6457ecde172ddf7 - Remove ExpressionGraph::forceInit()
[Roman Grundkiewicz] a9f99cc4b94fd828852fca5c0e158dd017b5fd99 - Rename .mvavg.npz to .orig.npz
[Roman Grundkiewicz] 6133e69978bd5cd9a5b9609829951ba9b99ae9e1 - Rename updateMovingAverage
[Marcin Junczys-Dowmunt] fb60415bf604dbecd95d3b58018951ad56caed32 - change class names
[Marcin Junczys-Dowmunt] e5a540b25be5c018a5c51f17a78edf4d67efb775 - rename file to keep VS happy
[Marcin Junczys-Dowmunt] 4fe4b09ebfb9664fa83ae84e26ee82ad19453561 - adapt quicksand module to memory mapping
[Marcin Junczys-Dowmunt] 34aa79d17ee0c119e03762162fb742f24f94ab4b - add rnn layer to transformer
[Marcin Junczys-Dowmunt] 47b2b3b719cbd9069f9209a63265959a7473fb5c - add tied layers option
[Marcin Junczys-Dowmunt] 8fec51fb043fe7a0590cfc46c94976359716ace9 - add forgotten layerNorm_ member
[Marcin Junczys-Dowmunt] be96ca3d9b292713d9e5ec55560e255f12849fa5 - return correct decoder state based on regressive layer
[Marcin Junczys-Dowmunt] cd5a9b92ef41a2148553a3f7586e6cb46649917a - add function approximator
[Marcin Junczys-Dowmunt] cb250c09f148c440f4e8abe67d07eeaa85558560 - approximate sigmoid
[Marcin Junczys-Dowmunt] d6936b365e7c46a62b92c85d21e236a6f9102e9c - add header
[Marcin Junczys-Dowmunt] 5562191a67b31e6de702e3be74105e977f95273a - add correct header
[Marcin Junczys-Dowmunt] ff85854c5e402e5a85ab0211c64ce2b19d80481b - make transformer-tied-layers return default value
[Roman Grundkiewicz] 7ab3f2604677173cbfbc84a5ed3e36b19f4554a2 - Autoformat
[Roman Grundkiewicz] a4ef9b52a668df434aca4be9141f93a74119842f - Uncomment logging
[Roman Grundkiewicz] facd045bf63f9022bde028fb953edc4d9d0f2b76 - Update CHANGELOG
[Roman Grundkiewicz] 832bcba6ac0fa490038ef95f048c35e4cfcc4d0a - Add logging to Amun save()
[Marcin Junczys-Dowmunt] e0f5163afd1c6fafab9b5363d83806f5439f59a5 - set version to 1.6.0
[Marcin Junczys-Dowmunt] 8b799b2d2889132d3607cde2d2e54ed8f81b2546 - move cmake_doze.txt
[Marcin Junczys-Dowmunt] 1a61a9c5154b86228644b3100accd64b60084068 - add submodule examples
[Marcin Junczys-Dowmunt] 5a42fdd3e6776ebbe464174620f37be781a2bff1 - checkout examples automatically
[github] 46b0232339ef13efc06b3458facb2cba831411cc - Update README to reflect main repo
Build 'amunmt_compilation_cpu' is failing!
Last 50 lines of build output:
Changes since last successful build:
[hieuhoang] d3bf3f536468afb915afbf8996ebc36e1b21267c - consistent format for timings
[hieuhoang] 90f7f1cfc47ad43ffa206e7fbc2ed153f654763e - namespace
[hieuhoang] a9cde0f34c2c3ffd18c00d36736ea8aaf8bcdcee - move timings into GPU namespace
View full output