Skip to content

N
NiuTrans.Tensor
Commit 4bbd6a27 by hello
Options

Merge with NiuTrans.NMT

parent 4a3a47f1
正在显示 包含 648 行增加 和 507 行删除
.gitignore
......@@ -5,3 +5,6 @@ vc140.pdb
NiuTrans.Tensor.vcxproj.user
NiuTrans.Tensor.aps
data/
build/
xxx/
bin/
CMakeLists.txt
# cmake minimum version
# CMake minimum version
cmake_minimum_required(VERSION 2.8)
# Project's name
project(NiuTensor)
# The name of the generated executable file
# The name of the dynamic link library
set(NIUTENSOR_EXE "NiuTensor")
set(NIUTENSOR_DLL "${NIUTENSOR_EXE}")
# The prefix of the generated executable file
set(NIUTRANS_EXE "NiuTensor")
set(NIUTRANS_DLL "${NIUTRANS_EXE}")
# Generated file path
set(EXECUTABLE_OUTPUT_PATH ../bin)
set(LIBRARY_OUTPUT_PATH ../lib)
# Use CMAKE_MACOSX_RPATH for macOS
# Use CMAKE_MACOSX_RPATH for MacOS
set(CMAKE_MACOSX_RPATH 1)
# Open floder manage
set_property(GLOBAL PROPERTY USE_FOLDERS ON)
set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
option(USE_CUDA "Use CUDA" OFF)
option(USE_HALF_PRECISION "Use Half Precision in CUDA Codes" OFF)
option(USE_MKL "Use MKL" OFF)
option(USE_OPENBLAS "Use OpenBLAS" OFF)
option(USE_FP16 "Use FP16" OFF)
option(GEN_DLL "Generate Dynamic Link Library" OFF)
# If set USE_CUDA ON, please modify CUDA_ROOT below.
# If set USE_MKL ON, please modify the INTEL_ROOT below.
# If set USE_OPENBLAS ON, please modify the OPENBLAS_ROOT below.
if (USE_CUDA)
if(NOT EXISTS ${CUDA_ROOT})
if(NOT DEFINED CUDA_TOOLKIT_ROOT_DIR)
if(WIN32)
set(CUDA_ROOT "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v9.2")
message(STATUS "HERE cuda")
set(CUDA_TOOLKIT_ROOT_DIR "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.1")
else()
set(CUDA_ROOT "/usr/local/cuda-9.2")
set(CUDA_TOOLKIT_ROOT_DIR "/usr/cuda-9.0")
endif()
endif()
set(CUDA_TOOLKIT_ROOT_DIR ${CUDA_ROOT})
message(STATUS "CUDA_ROOT: ${CUDA_ROOT}")
message(STATUS "CUDA_TOOLKIT_ROOT_DIR: ${CUDA_TOOLKIT_ROOT_DIR}")
endif()
if(USE_MKL)
if(NOT DEFINED INTEL_ROOT)
if(WIN32)
message(STATUS "HERE mkl")
set(INTEL_ROOT "C:/Program Files (x86)/IntelSWTools/compilers_and_libraries_2020.2.254/windows")
else()
set(INTEL_ROOT "/opt/intel/compilers_and_libraries_2020.2.254/linux")
set(INTEL_ROOT "/usr/intel/compilers_and_libraries_2020.2.254/linux")
endif()
endif()
message(STATUS "INTEL_ROOT: ${INTEL_ROOT}")
......@@ -52,9 +49,9 @@ endif()
if(USE_OPENBLAS)
if(NOT DEFINED OPENBLAS_ROOT)
if(WIN32)
set(OPENBLAS_ROOT "C:/Program Files/OpenBLAS")
set(OPENBLAS_ROOT "D:/software/BaiduNetdiskDownload/thirdparty20170624/OpenBLAS")
else()
set(OPENBLAS_ROOT "/opt/OpenBLAS")
set(OPENBLAS_ROOT "/usr/OpenBLAS")
endif()
endif()
message(STATUS "OPENBLAS_ROOT: ${OPENBLAS_ROOT}")
......@@ -93,66 +90,57 @@ endfunction(my_add_executable)
# Set libs and compiler options for CUDA
if(USE_CUDA)
add_definitions(-DUSE_CUDA)
if(USE_HALF_PRECISION)
if(USE_FP16)
add_definitions(-DHALF_PRECISION)
endif()
find_package(CUDA REQUIRED)
if(GPU_ARCH STREQUAL K) # Kepler cards (CUDA 5 until CUDA 10)
set(ARCH_FLAGS -arch=compute_30 -code=compute_30,sm_30,sm_35,sm_37)
elseif(GPU_ARCH STREQUAL M) # Maxwell cards (CUDA 6 until CUDA 11)
set(ARCH_FLAGS -arch=compute_50 -code=compute_50,sm_50,sm_52,sm_53)
elseif(GPU_ARCH STREQUAL P) # Pascal (CUDA 8 and later)
set(ARCH_FLAGS -arch=compute_60 -code=compute_60,sm_60,sm_61,sm_62)
elseif(GPU_ARCH STREQUAL V) # Volta (CUDA 9 and later)
set(ARCH_FLAGS -arch=compute_70 -code=compute_70,sm_70,sm_72)
elseif(GPU_ARCH STREQUAL T) # Turing (CUDA 10 and later)
set(ARCH_FLAGS -arch=compute_75 -code=sm_75)
elseif(GPU_ARCH STREQUAL A) # Ampere (CUDA 11 and later)
set(ARCH_FLAGS -arch=compute_80 -code=sm_80)
endif()
if(USE_HALF_PRECISION)
if(NOT DEFINED GPU_ARCH)
set(ARCH_FLAGS -arch=sm_60
-gencode=arch=compute_60,code=sm_60
-gencode=arch=compute_61,code=sm_61
-gencode=arch=compute_62,code=sm_62
-gencode=arch=compute_70,code=sm_70
-gencode=arch=compute_72,code=sm_72
-gencode=arch=compute_70,code=compute_70
)
elseif(${GPU_ARCH} STREQUAL K OR ${GPU_ARCH} STREQUAL M)
message(FATAL_ERROR "your GPU cannot use the function half precision")
endif()
endif()
find_package(CUDA ${CUDA_VERSION} REQUIRED)
if(WIN32)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4819")
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-maxrregcount=0 -m64 -Wno-deprecated-gpu-targets -use_fast_math")
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} ${ARCH_FLAGS})
SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4819")
set(CMAKE_CUDA_FLAGS ${CMAKE_CUDA_FLAGS} "-maxrregcount=0 -m64 -use_fast_math -DUSE_CUDA")
set(CMAKE_CUDA_FLAGS ${CMAKE_CUDA_FLAGS} -arch=sm_30
-gencode=arch=compute_30,code=sm_30
-gencode=arch=compute_50,code=sm_50
-gencode=arch=compute_52,code=sm_52
-gencode=arch=compute_60,code=sm_60
-gencode=arch=compute_61,code=sm_61
-gencode=arch=compute_62,code=sm_62
-gencode=arch=compute_70,code=sm_70
-gencode=arch=compute_70,code=compute_70
)
set(CMAKE_POLICY_DEFAULT_CMP0028 NEW)
link_directories("${CUDA_ROOT}/lib/x64")
include_directories("${CUDA_ROOT}/include")
set(CUDA_LIB_DIR "${CUDA_ROOT}/lib/x64/")
link_directories("${CUDA_TOOLKIT_ROOT_DIR}/lib/x64")
include_directories("${CUDA_TOOLKIT_ROOT_DIR}/include")
set(CUDA_LIB_DIR "${CUDA_TOOLKIT_ROOT_DIR}/lib/x64/")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}cublas.lib")
if(CUDA_VERSION_MAJOR EQUAL 11)
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}cublasLt.lib")
endif()
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}npps.lib")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}nppc.lib")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}cudadevrt.lib")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}curand.lib")
else()
set(CMAKE_CXX_FLAGS "-fPIC -msse4.2 -w -march=native -Wno-enum-compare -Wno-sign-compare -Wno-format -Wno-dev -O3 -DNDEBUG -rdynamic")
set(CUDA_NVCC_FLAGS "-Xcompiler -fPIC -maxrregcount=0 --disable-warnings -use_fast_math -DUSE_CUDA -Wno-deprecated-gpu-targets -std=c++11")
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} ${ARCH_FLAGS})
link_directories("${CUDA_ROOT}/lib64")
include_directories("${CUDA_ROOT}/include")
set(CUDA_LIB_DIR "${CUDA_ROOT}/lib64/")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}libcublas_static.a")
if(CUDA_VERSION_MAJOR EQUAL 11)
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}libcublasLt_static.a")
if(USE_FP16)
set(CUDA_NVCC_FLAGS "-Xcompiler -fPIC -maxrregcount=0 -use_fast_math -DUSE_CUDA -DHALF_PRECISION -Wno-deprecated-gpu-targets -std=c++11 ")
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -arch=sm_60
-gencode=arch=compute_60,code=sm_60
-gencode=arch=compute_61,code=sm_61
-gencode=arch=compute_62,code=sm_62
-gencode=arch=compute_70,code=sm_70
-gencode=arch=compute_70,code=compute_70
)
else()
set(CUDA_NVCC_FLAGS "-Xcompiler -fPIC -maxrregcount=0 -use_fast_math -DUSE_CUDA -Wno-deprecated-gpu-targets -std=c++11 ")
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -arch=sm_60
-gencode=arch=compute_60,code=sm_60
-gencode=arch=compute_61,code=sm_61
-gencode=arch=compute_62,code=sm_62
-gencode=arch=compute_70,code=sm_70
-gencode=arch=compute_70,code=compute_70
)
endif()
link_directories(${CUDA_TOOLKIT_ROOT_DIR}/lib64)
include_directories(${CUDA_TOOLKIT_ROOT_DIR}/include)
set(CUDA_LIB_DIR "${CUDA_TOOLKIT_ROOT_DIR}/lib64/")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}libcublas_static.a")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}libculibos.a")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}libnpps_static.a")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}libnppc_static.a")
......@@ -164,12 +152,12 @@ endif()
# Set libs and compiler options for MKL
if(USE_MKL)
add_definitions(-DUSE_BLAS -DMKL)
add_definitions(-DMKL)
set(COMPILER_DIR "${INTEL_ROOT}/compiler")
set(MKL_DIR "${INTEL_ROOT}/mkl")
set(CPU_ARCH intel64)
if(WIN32)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DNDEBUG")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DNDEBUG -DMKL")
link_directories(${MKL_DIR}/lib/intel64/)
link_directories(${COMPILER_DIR}/lib/intel64)
include_directories(${MKL_DIR}/include)
......@@ -181,9 +169,9 @@ if(USE_MKL)
set(MKL_LIB_PATH ${MKL_LIB_PATH} "${COMPILER_LIB_DIR}libiomp5md.lib")
else()
if(USE_CUDA)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-reorder")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-reorder -DMKL")
else()
set(CMAKE_CXX_FLAGS "-std=c++11 -msse4.2 -w -march=native -Wno-enum-compare -Wno-sign-compare -Wno-reorder -Wno-format -O3 -flto -DNDEBUG -rdynamic -fkeep-inline-functions -fPIC")
set(CMAKE_CXX_FLAGS "-std=c++11 -msse4.2 -w -march=native -Wno-enum-compare -Wno-sign-compare -Wno-reorder -Wno-format -O3 -flto -DNDEBUG -rdynamic -fkeep-inline-functions -fPIC -DMKL")
endif(USE_CUDA)
link_directories(${MKL_DIR}/lib/intel64/)
link_directories(${COMPILER_DIR}/lib/intel64)
......@@ -199,9 +187,10 @@ endif()
# Set libs and compiler options for OpenBLAS
if(USE_OPENBLAS)
add_definitions(-DUSE_BLAS -DOPENBLAS)
add_definitions(-DUSE_BLAS -DMKL)
set(OPENBLAS_INCLUDE_DIR "${OPENBLAS_ROOT}/include")
set(OPENBLAS_LIB_DIR "${OPENBLAS_ROOT}/lib")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DUSE_BLAS")
if(WIN32)
link_directories(${OPENBLAS_LIB_DIR})
include_directories(${OPENBLAS_INCLUDE_DIR})
......@@ -222,15 +211,15 @@ set(OPENBLAS_LIB ${OPENBLAS_LIB_PATH})
# Generate dynamic link library about project
if(USE_CUDA)
if(GEN_DLL)
cuda_add_library(${NIUTENSOR_DLL} SHARED ${CPP_FILES} ${H_FILES} ${CU_FILES} ${CUH_FILES})
cuda_add_library(${NIUTRANS_DLL} SHARED ${CPP_FILES} ${H_FILES} ${CU_FILES} ${CUH_FILES})
else()
my_add_executable(${NIUTENSOR_EXE} ${CPP_FILES} ${H_FILES} ${CU_FILES} ${CUH_FILES})
my_add_executable(${NIUTRANS_EXE} ${CPP_FILES} ${H_FILES} ${CU_FILES} ${CUH_FILES})
endif()
else()
if(GEN_DLL)
add_library(${NIUTENSOR_DLL} SHARED ${CPP_FILES} ${H_FILES})
add_library(${NIUTRANS_DLL} SHARED ${CPP_FILES} ${H_FILES})
else()
my_add_executable(${NIUTENSOR_EXE} ${CPP_FILES} ${H_FILES})
my_add_executable(${NIUTRANS_EXE} ${CPP_FILES} ${H_FILES})
endif()
endif()
......@@ -254,17 +243,17 @@ if(WIN32)
if(GEN_DLL)
message(STATUS "Generate Dynamic Link Library")
message(STATUS "Name of Dynamic Link Library: " ${NIUTENSOR_DLL})
target_link_libraries(${NIUTENSOR_DLL} ${ALL_LIB})
message(STATUS "Name of Dynamic Link Library: " ${NIUTRANS_DLL})
target_link_libraries(${NIUTRANS_DLL} ${ALL_LIB})
else()
message(STATUS "Generate Makefile For Executable File")
message(STATUS "Name of Executable File :" ${NIUTENSOR_EXE})
target_link_libraries(${NIUTENSOR_EXE} ${ALL_LIB})
message(STATUS "Name of Executable File :" ${NIUTRANS_EXE})
target_link_libraries(${NIUTRANS_EXE} ${ALL_LIB})
endif()
message(STATUS "${MESS}")
else()
add_definitions(-std=c++11)
set(MESS ${MESS} "On Linux or macOS")
set(MESS ${MESS} "On Linux")
if(USE_CUDA)
set(MESS ${MESS} " Use CUDA")
set(ALL_LIB ${ALL_LIB} ${CUDA_LIB})
......@@ -285,12 +274,12 @@ else()
if(GEN_DLL)
message(STATUS "Generate Dynamic Link Library")
message(STATUS "Name of Dynamic Link Library: " ${NIUTENSOR_DLL})
target_link_libraries(${NIUTENSOR_DLL} ${ALL_LIB} ${FLAG})
message(STATUS "Name of Dynamic Link Library: " ${NIUTRANS_DLL})
target_link_libraries(${NIUTRANS_DLL} ${ALL_LIB} ${FLAG})
else()
message(STATUS "Generate Makefile For Executable File")
message(STATUS "Name of Executable File: " ${NIUTENSOR_EXE})
target_link_libraries(${NIUTENSOR_EXE} ${ALL_LIB} ${FLAG})
message(STATUS "Name of Executable File: " ${NIUTRANS_EXE})
target_link_libraries(${NIUTRANS_EXE} ${ALL_LIB} ${FLAG})
endif()
message(STATUS "${MESS}")
endif()
\ No newline at end of file
endif()
source/sample/transformer/Decoder.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -40,6 +40,16 @@ AttDecoder::AttDecoder()
decoderLayerNorm = NULL;
selfAttCache = NULL;
enDeAttCache = NULL;
history = NULL;
preNorm = true;
useHistory = false;
finalNorm = false;
devID = -1;
eSize = -1;
hSize = -1;
nlayer = -1;
vSize = -1;
dropoutP = 0.0F;
}
/* de-constructor */
......@@ -53,8 +63,10 @@ AttDecoder::~AttDecoder()
delete[] fnnLayerNorms;
delete[] enDeAtt;
delete[] enDeAttLayerNorms;
if (preNorm)
if (finalNorm)
delete decoderLayerNorm;
if (useHistory)
delete history;
}
/*
......@@ -70,13 +82,12 @@ void AttDecoder::InitModel(Config& config)
vSize = config.tgtVocabSize;
dropoutP = config.dropout;
preNorm = config.preNorm;
finalNorm = config.finalNorm;
useHistory = config.useHistory;
CheckNTErrors(nlayer >= 1, "We have one encoding layer at least!");
CheckNTErrors(vSize > 1, "set vocabulary size by \"-vsizetgt\"");
/* embedding model */
embedder.InitModel(config, false);
selfAtt = new Attention[nlayer];
fnns = new FNN[nlayer];
selfAttLayerNorms = new LN[nlayer];
......@@ -86,10 +97,15 @@ void AttDecoder::InitModel(Config& config)
selfAttCache = new Cache[nlayer];
enDeAttCache = new Cache[nlayer];
if (preNorm)
if (finalNorm)
decoderLayerNorm = new LN;
if (useHistory)
history = new LayerHistory;
/* initialize the stacked layers */
embedder.InitModel(config, false);
for (int i = 0; i < nlayer; i++) {
selfAtt[i].InitModel(config);
fnns[i].InitModel(config);
......@@ -100,8 +116,10 @@ void AttDecoder::InitModel(Config& config)
selfAttCache[i].enable = true;
enDeAttCache[i].enable = true;
}
if (preNorm)
if (finalNorm)
decoderLayerNorm->InitModel(config);
if (useHistory)
history->InitModel(config);
}
/*
......@@ -117,15 +135,26 @@ make the decoding network
XTensor AttDecoder::Make(XTensor& inputDec, XTensor& outputEnc, XTensor* mask,
XTensor* maskEncDec, int nstep, bool isTraining)
{
/* clear the history */
if (useHistory)
history->ClearHistory();
XTensor x;
x = embedder.Make(inputDec, true, isTraining, nstep);
/* dropout */
if (isTraining && dropoutP > 0)
x = Dropout(x, dropoutP);
x = Dropout(x, dropoutP, /*inplace=*/true);
if (useHistory)
history->Add(x);
for (int i = 0; i < nlayer; i++) {
if (useHistory)
x = history->Pop();
XTensor att;
XTensor ende;
XTensor fnn;
......@@ -146,10 +175,10 @@ XTensor AttDecoder::Make(XTensor& inputDec, XTensor& outputEnc, XTensor* mask,
/* dropout */
if (isTraining && dropoutP > 0)
att = Dropout(att, dropoutP);
att = Dropout(att, dropoutP, /*inplace=*/true);
/* residual connection */
res = Sum(att, x);
res = Sum(att, x, /*inplace=*/true);
/* layer normalization with post-norm for self-attention */
selfAttnAfter = LayerNorm(res, selfAttLayerNorms[i], preNorm, false, true);
......@@ -163,10 +192,10 @@ XTensor AttDecoder::Make(XTensor& inputDec, XTensor& outputEnc, XTensor* mask,
/* dropout */
if (isTraining && dropoutP > 0)
ende = Dropout(ende, dropoutP);
ende = Dropout(ende, dropoutP, /*inplace=*/true);
/* residual connection */
res = Sum(ende, selfAttnAfter);
res = Sum(ende, selfAttnAfter, /*inplace=*/true);
/* layer normalization with post-norm for encoder-decoder attention */
endeAttnAfter = LayerNorm(res, enDeAttLayerNorms[i], preNorm, false, true);
......@@ -179,94 +208,27 @@ XTensor AttDecoder::Make(XTensor& inputDec, XTensor& outputEnc, XTensor* mask,
/* dropout */
if (isTraining && dropoutP > 0)
fnn = Dropout(fnn, dropoutP);
fnn = Dropout(fnn, dropoutP, /*inplace=*/true);
/* residual connection */
res = Sum(fnn, endeAttnAfter);
res = Sum(fnn, endeAttnAfter, /*inplace=*/true);
/* layer normalization with post-norm for fnn */
x = LayerNorm(res, fnnLayerNorms[i], preNorm, false, true);
}
if (preNorm)
return decoderLayerNorm->Make(x);
return x;
}
/*
make the decoding network
>> inputDec - the input tensor of the decoder
>> outputEnc - the output tensor of the encoder
>> mask - mask that indicates which position is valid
>> maskEncDec - mask for the encoder-decoder attention
>> nstep - the current length of the decoder input
>> isTraining - indicates whether the model is used for training
<< return - the output tensor of the decoder
*/
XTensor AttDecoder::MakeFast(XTensor& inputDec, XTensor& outputEnc, XTensor* mask,
XTensor* maskEncDec, int nstep, bool isTraining)
{
XTensor x;
x = embedder.Make(inputDec, true, isTraining, nstep);
/* dropout */
if (isTraining && dropoutP > 0)
x = Dropout(x, dropoutP);
for (int i = 0; i < nlayer; i++) {
XTensor res;
res = x;
/* layer normalization with pre-norm for self-attn */
x = selfAttLayerNorms[i].Make(x);
/******************/
/* self attention */
x = selfAtt[i].Make(x, x, x, mask, isTraining, &selfAttCache[i], SELF_ATT);
/* dropout */
if (isTraining && dropoutP > 0)
x = Dropout(x, dropoutP);
/* residual connection */
x = Sum(res, x);
res = x;
/* layer normalization with pre-norm for encoder-decoder attention */
x = enDeAttLayerNorms[i].Make(x);
/* encoder-decoder attention */
x = enDeAtt[i].Make(outputEnc, x, outputEnc, maskEncDec,
isTraining, &enDeAttCache[i], EN_DE_ATT);
/* dropout */
if (isTraining && dropoutP > 0)
x = Dropout(x, dropoutP);
/* residual connection */
x = Sum(res, x);
res = x;
/* layer normalization with pre-norm for fnn */
x = fnnLayerNorms[i].Make(x);
/* fnn */
x = fnns[i].Make(x, isTraining);
if (useHistory)
history->Add(x);
}
/* dropout */
if (isTraining && dropoutP > 0)
x = Dropout(x, dropoutP);
if (useHistory)
x = history->Pop();
/* residual connection */
x = Sum(res, x);
}
/* clear the history while not training */
if (useHistory && !isTraining)
history->ClearHistory();
x = decoderLayerNorm->Make(x);
if (finalNorm)
return decoderLayerNorm->Make(x);
return x;
}
......
source/sample/transformer/Decoder.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -74,6 +74,9 @@ public:
/* layer normalization for encoder-decoder attention */
LN* enDeAttLayerNorms;
/* dynamic layer history */
LayerHistory* history;
/* layer cache list */
Cache* selfAttCache;
......@@ -83,6 +86,12 @@ public:
/* the location of layer normalization */
bool preNorm;
/* add LN to the decoder output or not */
bool finalNorm;
/* reserve history for layers or not */
bool useHistory;
public:
/* constructor */
AttDecoder();
......@@ -96,10 +105,6 @@ public:
/* make the decoding network */
XTensor Make(XTensor& inputDec, XTensor& outputEnc, XTensor* mask,
XTensor* maskEncDec, int nstep, bool isTraining);
/* make the decoding network (pre norm) */
XTensor MakeFast(XTensor& inputDec, XTensor& outputEnc, XTensor* mask,
XTensor* maskEncDec, int nstep, bool isTraining);
};
}
......
source/sample/transformer/Encoder.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -31,11 +31,22 @@ namespace nmt
/* constructor */
AttEncoder::AttEncoder()
{
devID = -1;
selfAtt = NULL;
fnns = NULL;
attLayerNorms = NULL;
fnnLayerNorms = NULL;
encoderLayerNorm = NULL;
useHistory = false;
history = NULL;
dropoutP = 0.0;
eSize = -1;
finalNorm = false;
hSize = -1;
ignored = -1;
nlayer = -1;
preNorm = false;
vSize = -1;
}
/* de-constructor */
......@@ -45,8 +56,10 @@ AttEncoder::~AttEncoder()
delete[] fnns;
delete[] attLayerNorms;
delete[] fnnLayerNorms;
if (preNorm)
if (finalNorm)
delete encoderLayerNorm;
if (useHistory)
delete history;
}
/*
......@@ -62,31 +75,36 @@ void AttEncoder::InitModel(Config& config)
hSize = config.modelSize;
vSize = config.srcVocabSize;
preNorm = config.preNorm;
finalNorm = config.finalNorm;
useHistory = config.useHistory;
dropoutP = config.dropout;
CheckNTErrors(nlayer >= 1, "We have one encoding layer at least!");
CheckNTErrors(vSize > 1, "Set vocabulary size by \"-vsize\"");
/* embedding model */
embedder.InitModel(config);
selfAtt = new Attention[nlayer];
fnns = new FNN[nlayer];
attLayerNorms = new LN[nlayer];
fnnLayerNorms = new LN[nlayer];
if (preNorm)
if (finalNorm)
encoderLayerNorm = new LN;
if (useHistory)
history = new LayerHistory;
/* initialize the stacked layers */
embedder.InitModel(config);
for (int i = 0; i < nlayer; i++) {
selfAtt[i].InitModel(config);
fnns[i].InitModel(config);
attLayerNorms[i].InitModel(config);
fnnLayerNorms[i].InitModel(config);
}
if (preNorm)
if (finalNorm)
encoderLayerNorm->InitModel(config);
if (useHistory)
history->InitModel(config);
}
/*
......@@ -99,15 +117,25 @@ make the encoding network
*/
XTensor AttEncoder::Make(XTensor& input, XTensor* mask, XTensor& maskEncDec, bool isTraining)
{
XTensor x;
/* clear the history */
if (useHistory)
history->ClearHistory();
XTensor x;
x = embedder.Make(input, false, isTraining);
/* dropout */
if (isTraining && dropoutP > 0)
x = Dropout(x, dropoutP);
x = Dropout(x, dropoutP, /*inplace=*/true);
if (useHistory)
history->Add(x);
for (int i = 0; i < nlayer; i++) {
if (useHistory)
x = history->Pop();
XTensor att;
XTensor fnn;
XTensor res;
......@@ -123,10 +151,10 @@ XTensor AttEncoder::Make(XTensor& input, XTensor* mask, XTensor& maskEncDec, boo
/* dropout */
if (isTraining && dropoutP > 0)
att = Dropout(att, dropoutP);
att = Dropout(att, dropoutP, /*inplace=*/true);
/* residual connection */
res = Sum(att, x);
res = Sum(att, x, /*inplace=*/true);
/* layer normalization with post-norm for self-attn */
attnAfter = LayerNorm(res, attLayerNorms[i], preNorm, false, true);
......@@ -139,72 +167,27 @@ XTensor AttEncoder::Make(XTensor& input, XTensor* mask, XTensor& maskEncDec, boo
/* dropout */
if (isTraining && dropoutP > 0)
fnn = Dropout(fnn, dropoutP);
fnn = Dropout(fnn, dropoutP, /*inplace=*/true);
/* residual connection */
res = Sum(fnn, attnAfter);
res = Sum(fnn, attnAfter, /*inplace=*/true);
/* layer normalization with post-norm for fnn */
x = LayerNorm(res, fnnLayerNorms[i], preNorm, false, true);
}
if (preNorm)
return encoderLayerNorm->Make(x);
return x;
}
/*
make the encoding network
>> input - the input tensor of the encoder
>> mask - the mask that indicate each position is valid
>> maskEncDec - no use
>> isTraining - indicates whether the model is used for training
<< return - the output tensor of the encoder
*/
XTensor AttEncoder::MakeFast(XTensor& input, XTensor* mask, XTensor& maskEncDec, bool isTraining)
{
XTensor x;
x = embedder.Make(input, false, isTraining);
/* dropout */
if (isTraining && dropoutP > 0)
x = Dropout(x, dropoutP);
for (int i = 0; i < nlayer; i++) {
XTensor res;
res = x;
/* layer normalization with pre-norm for self-attn */
x = attLayerNorms[i].Make(x);
/* self attention */
x = selfAtt[i].Make(x, x, x, mask, isTraining, NULL, SELF_ATT);
/* dropout */
if (isTraining && dropoutP > 0)
x = Dropout(x, dropoutP);
/* residual connection */
x = Sum(res, x);
res = x;
if (useHistory)
history->Add(x);
}
/* layer normalization with pre-norm for fnn */
x = fnnLayerNorms[i].Make(x);
if (useHistory)
x = history->Pop();
/* fnn */
x = fnns[i].Make(x, isTraining);
/* clear the history while not training */
if (useHistory && !isTraining)
history->ClearHistory();
/* dropout */
if (isTraining && dropoutP > 0)
x = Dropout(x, dropoutP);
/* residual connection */
x = Sum(res, x);
}
x = encoderLayerNorm->Make(x);
if (finalNorm)
return encoderLayerNorm->Make(x);
return x;
}
......
source/sample/transformer/Encoder.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -27,6 +27,7 @@
#include "submodel/Attention.h"
#include "submodel/Embedding.h"
#include "submodel/LayerNorm.h"
#include "submodel/LayerHistory.h"
#include "../../network/XNet.h"
using namespace nts;
......@@ -89,9 +90,18 @@ public:
/* layer normalization for encoder */
LN* encoderLayerNorm;
/* dynamic layer history */
LayerHistory* history;
/* the location of layer normalization */
bool preNorm;
/* add LN to the encoder output or not */
bool finalNorm;
/* reserve history for layers or not */
bool useHistory;
public:
/* constructor */
AttEncoder();
......@@ -105,9 +115,6 @@ public:
/* make the encoding network */
XTensor Make(XTensor& input, XTensor* mask, XTensor& maskEncDec, bool isTraining);
/* make the encoding network */
XTensor MakeFast(XTensor& input, XTensor* mask, XTensor& maskEncDec, bool isTraining);
/* make the encoding network (wrapper) */
XTensor Make(XTensor& input, XTensor* mask, bool isTraining);
};
......
source/sample/transformer/Model.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -72,20 +72,23 @@ void Model::InitModel(Config& config)
&config.tgtVocabSize, &config.nhead,
&config.maxRP, &config.shareAllEmbeddings,
&config.shareDecInputOutputWeight,
&config.maxPosLen
&config.maxPosition
};
FILE* modelFile = NULL;
/* read model configurations */
if (!config.isTraining) {
modelFile = fopen(config.modelFN, "rb");
if (!config.isTraining || strcmp(config.pretrainedModel, "") != 0) {
if (strcmp(config.pretrainedModel, "") != 0)
modelFile = fopen(config.pretrainedModel, "rb");
else
modelFile = fopen(config.modelFN, "rb");
CheckNTErrors(modelFile, "Failed to open the model file");
for (auto& meta : metaInfo) {
fread(meta, sizeof(int), 1, modelFile);
}
}
else {
if (config.isTraining) {
/* read the source and target vocab size */
FILE* trainF = fopen(config.trainFN, "rb");
CheckNTErrors(trainF, "Failed to open the training file");
......@@ -110,9 +113,10 @@ void Model::InitModel(Config& config)
decoder->InitModel(config);
/* load parameters */
if (!config.isTraining)
if (!config.isTraining || strcmp(config.pretrainedModel, "") != 0)
Read(modelFile);
else {
if (config.isTraining) {
TensorList params;
GetParams(params);
for (int i = 0; i < params.Size(); i++)
......@@ -220,6 +224,8 @@ void Model::MakeMT(XTensor& inputEnc, XTensor& inputDec, XTensor& output,
XTensor maskDec;
XTensor maskEncDec;
bool debug(false);
/* encoder mask */
MakeMTMaskEnc(paddingEnc, maskEnc);
......@@ -228,9 +234,25 @@ void Model::MakeMT(XTensor& inputEnc, XTensor& inputDec, XTensor& output,
encoding = MakeEncoder(inputEnc, &maskEnc, isTraining);
if (debug) {
LOG("after encoding:");
encoding.mem->ShowMemUsage(stderr);
}
decoding = MakeDecoder(inputDec, encoding, &maskDec, maskEncDec, isTraining);
if (debug) {
LOG("after decoding:");
encoding.mem->ShowMemUsage(stderr);
}
outputLayer->Make(decoding, output, true, true);
if (debug) {
LOG("after outputing:");
encoding.mem->ShowMemUsage(stderr);
exit(0);
}
}
/*
......@@ -265,9 +287,9 @@ void Model::MakeMTMask(XTensor& inputEnc, XTensor& inputDec,
dims[inputDec.order + 1] = inputEnc.GetDim(inputEnc.order - 1);
InitTensor(&maskEncDec, inputDec.order + 2, dims, X_FLOAT, paddingEnc.devID);
XTensor* maskEncDecTMPEnc = NewTensorBuf(paddingEnc.order + 1, dims + 1,
paddingEnc.dataType, paddingEnc.devID);
XTensor* maskEncDecTMPDec = NewTensorBuf(maskEncDecTMPEnc, paddingEnc.devID);
XTensor* maskEncDecTMPEnc = NewTensorBufV2(paddingEnc.order + 1, dims + 1,
paddingEnc.dataType, 1.0F, paddingEnc.devID, paddingEnc.mem);
XTensor* maskEncDecTMPDec = NewTensorBufV2(maskEncDecTMPEnc, paddingEnc.devID, paddingEnc.mem);
_Unsqueeze(&paddingEnc, maskEncDecTMPEnc, paddingEnc.order - 1, paddingDec.GetDim(-1));
_ScaleAndShiftMe(maskEncDecTMPEnc, 1e9F, -1e9F);
......@@ -283,14 +305,14 @@ void Model::MakeMTMask(XTensor& inputEnc, XTensor& inputDec,
dimsPadding[paddingEnc.order - 1] = paddingEnc.GetDim(-1);
dimsPadding[paddingEnc.order] = paddingEnc.GetDim(-1);
XTensor* padding2 = NewTensorBuf(paddingEnc.order + 1, dimsPadding, paddingEnc.dataType,
paddingEnc.devID);
XTensor* padding2 = NewTensorBufV2(paddingEnc.order + 1, dimsPadding, paddingEnc.dataType, 1.0F,
paddingEnc.devID, paddingEnc.mem);
for (int i = 0; i < padding2->order; i++)
dimsPadding[i + 1] = padding2->GetDim(i);
dimsPadding[0] = nhead;
XTensor* padding3 = NewTensorBuf(paddingEnc.order + 2, dimsPadding, paddingEnc.dataType, paddingEnc.devID);
XTensor* padding3 = NewTensorBufV2(paddingEnc.order + 2, dimsPadding, paddingEnc.dataType, 1.0F, paddingEnc.devID, paddingEnc.mem);
/* mask of the padding */
_Unsqueeze(&paddingEnc, padding2, paddingEnc.order - 1, paddingEnc.GetDim(-1));
......@@ -322,6 +344,7 @@ void Model::MakeMTMaskEnc(XTensor& paddingEnc, XTensor& maskEnc)
/* mask of the padding */
Unsqueeze(paddingEnc, padding2, paddingEnc.order - 1, paddingEnc.GetDim(-1));
Unsqueeze(padding2, maskEnc, 0, nhead);
ScaleAndShiftMe(maskEnc, 1e9F, -1e9F);
}
......@@ -355,6 +378,7 @@ void Model::MakeMTMaskDec(XTensor& paddingEnc, XTensor& paddingDec,
Unsqueeze(paddingEnc, maskEncDecTMP, paddingEnc.order - 1, paddingDec.GetDim(-1));
ScaleAndShiftMe(maskEncDecTMP, 1e9F, -1e9F);
Unsqueeze(maskEncDecTMP, maskEncDec, 0, dims[0]);
delete[] dims;
......@@ -369,6 +393,14 @@ void Model::GetParams(TensorList& list)
list.Clear();
/* encoder parameters */
if (encoder->useHistory) {
for (int i = 0; i < encoder->nlayer + 1; i++)
list.Add(&encoder->history->weights[i]);
for (int i = 0; i < encoder->nlayer; i++) {
list.Add(&encoder->history->layerNorms[i].weight);
list.Add(&encoder->history->layerNorms[i].bias);
}
}
for (int i = 0; i < encoder->nlayer; i++) {
list.Add(&encoder->selfAtt[i].weightQ);
list.Add(&encoder->selfAtt[i].weightK);
......@@ -384,18 +416,27 @@ void Model::GetParams(TensorList& list)
list.Add(&encoder->fnns[i].b1);
list.Add(&encoder->fnns[i].w2);
list.Add(&encoder->fnns[i].b2);
list.Add(&encoder->attLayerNorms[i].w);
list.Add(&encoder->attLayerNorms[i].b);
list.Add(&encoder->fnnLayerNorms[i].w);
list.Add(&encoder->fnnLayerNorms[i].b);
list.Add(&encoder->attLayerNorms[i].weight);
list.Add(&encoder->attLayerNorms[i].bias);
list.Add(&encoder->fnnLayerNorms[i].weight);
list.Add(&encoder->fnnLayerNorms[i].bias);
}
if (encoder->preNorm) {
list.Add(&encoder->encoderLayerNorm->w);
list.Add(&encoder->encoderLayerNorm->b);
if (encoder->finalNorm) {
list.Add(&encoder->encoderLayerNorm->weight);
list.Add(&encoder->encoderLayerNorm->bias);
}
if (isMT) {
/* decoder parameters */
if (decoder->useHistory) {
for (int i = 0; i < decoder->nlayer + 1; i++)
list.Add(&decoder->history->weights[i]);
for (int i = 0; i < decoder->nlayer; i++) {
list.Add(&decoder->history->layerNorms[i].weight);
list.Add(&decoder->history->layerNorms[i].bias);
}
}
for (int i = 0; i < decoder->nlayer; i++) {
list.Add(&decoder->selfAtt[i].weightQ);
list.Add(&decoder->selfAtt[i].weightK);
......@@ -407,8 +448,8 @@ void Model::GetParams(TensorList& list)
list.Add(&decoder->selfAtt[i].RPEmbK);
list.Add(&decoder->selfAtt[i].weightO);
list.Add(&decoder->selfAtt[i].biasO);
list.Add(&decoder->selfAttLayerNorms[i].w);
list.Add(&decoder->selfAttLayerNorms[i].b);
list.Add(&decoder->selfAttLayerNorms[i].weight);
list.Add(&decoder->selfAttLayerNorms[i].bias);
list.Add(&decoder->enDeAtt[i].weightQ);
list.Add(&decoder->enDeAtt[i].weightK);
list.Add(&decoder->enDeAtt[i].weightV);
......@@ -417,18 +458,18 @@ void Model::GetParams(TensorList& list)
list.Add(&decoder->enDeAtt[i].biasV);
list.Add(&decoder->enDeAtt[i].weightO);
list.Add(&decoder->enDeAtt[i].biasO);
list.Add(&decoder->enDeAttLayerNorms[i].w);
list.Add(&decoder->enDeAttLayerNorms[i].b);
list.Add(&decoder->enDeAttLayerNorms[i].weight);
list.Add(&decoder->enDeAttLayerNorms[i].bias);
list.Add(&decoder->fnns[i].w1);
list.Add(&decoder->fnns[i].b1);
list.Add(&decoder->fnns[i].w2);
list.Add(&decoder->fnns[i].b2);
list.Add(&decoder->fnnLayerNorms[i].w);
list.Add(&decoder->fnnLayerNorms[i].b);
list.Add(&decoder->fnnLayerNorms[i].weight);
list.Add(&decoder->fnnLayerNorms[i].bias);
}
if (decoder->preNorm) {
list.Add(&decoder->decoderLayerNorm->w);
list.Add(&decoder->decoderLayerNorm->b);
if (decoder->finalNorm) {
list.Add(&decoder->decoderLayerNorm->weight);
list.Add(&decoder->decoderLayerNorm->bias);
}
}
......@@ -490,7 +531,7 @@ void Model::Read(FILE* file)
TensorList params;
GetParams(params);
LOG("params count: %lu", (unsigned long)params.Size());
LOG("params count: %zd", params.Size());
int size = 0;
for (int i = 0; i < params.Size(); i++) {
size += params[i]->unitNum;
......
source/sample/transformer/Model.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/NMT.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -36,22 +36,31 @@ int NMTMain(int argc, const char** argv)
/* load configurations */
Config config(argc, argv);
srand(1);
srand(config.seed);
/* training */
if (strcmp(config.trainFN, "") != 0) {
Model model;
model.InitModel(config);
TensorList params;
model.GetParams(params);
int count = 0;
for (int i = 0; i < params.count; i++){
count += params[i]->unitNum;
}
LOG("number of parameters: %d", count);
Trainer trainer;
trainer.Init(config);
trainer.Train(config.trainFN, config.validFN, config.modelFN, &model);
}
/* translating */
if (strcmp(config.testFN, "") != 0 && strcmp(config.outputFN, "") != 0) {
else if (strcmp(config.testFN, "") != 0 && strcmp(config.outputFN, "") != 0) {
/* disable grad flow */
/* disable gradient flow */
DISABLE_GRAD;
Model model;
......
source/sample/transformer/NMT.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/Utility.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -63,18 +63,23 @@ Config::Config(int argc, const char** argv)
LoadParamInt(argsNum, args, "nhead", &nhead, 4);
LoadParamInt(argsNum, args, "enclayer", &nEncLayer, 6);
LoadParamInt(argsNum, args, "declayer", &nDecLayer, 6);
LoadParamInt(argsNum, args, "maxrp", &maxRP, 8);
LoadParamInt(argsNum, args, "maxrp", &maxRP, -1);
LoadParamInt(argsNum, args, "embsize", &embSize, 512);
LoadParamInt(argsNum, args, "modelsize", &modelSize, 512);
LoadParamInt(argsNum, args, "maxpos", &maxPosLen, 1024);
LoadParamInt(argsNum, args, "maxpos", &maxPosition, 1024);
LoadParamInt(argsNum, args, "maxsrclen", &maxSrcLen, 1024);
LoadParamInt(argsNum, args, "maxtgtlen", &maxTgtLen, 1024);
LoadParamInt(argsNum, args, "fnnhidden", &fnnHiddenSize, modelSize * 2);
LoadParamInt(argsNum, args, "vsize", &srcVocabSize, 10152);
LoadParamInt(argsNum, args, "vsizetgt", &tgtVocabSize, 10152);
LoadParamInt(argsNum, args, "padid", &padID, 1);
LoadParamInt(argsNum, args, "startid", &startID, 2);
LoadParamInt(argsNum, args, "endid", &endID, 2);
LoadParamBool(argsNum, args, "rpr", &useRPR, false);
LoadParamInt(argsNum, args, "unkid", &unkID, 3);
LoadParamBool(argsNum, args, "rpr", &useRPR, maxRP > 0);
LoadParamBool(argsNum, args, "prenorm", &preNorm, true);
LoadParamBool(argsNum, args, "finalnorm", &finalNorm, true);
LoadParamBool(argsNum, args, "dlcl", &useHistory, false);
// TODO: refactor the parameters type to support weight sharing during training
LoadParamInt(argsNum, args, "shareemb", &shareAllEmbeddings, 0);
......@@ -86,9 +91,12 @@ Config::Config(int argc, const char** argv)
/* options for training */
LoadParamString(argsNum, args, "train", trainFN, "");
LoadParamString(argsNum, args, "valid", validFN, "");
LoadParamString(argsNum, args, "pretrain", pretrainedModel, "");
LoadParamInt(argsNum, args, "dev", &devID, 0);
LoadParamInt(argsNum, args, "seed", &seed, 1);
LoadParamInt(argsNum, args, "log", &logInterval, 100);
LoadParamInt(argsNum, args, "wbatch", &wBatchSize, 4096);
LoadParamInt(argsNum, args, "sbatch", &sBatchSize, 8);
LoadParamInt(argsNum, args, "sbatch", &sBatchSize, 16);
isTraining = (strcmp(trainFN, "") == 0) ? false : true;
LoadParamBool(argsNum, args, "mt", &isMT, true);
LoadParamFloat(argsNum, args, "dropout", &dropout, 0.3F);
......@@ -117,7 +125,7 @@ Config::Config(int argc, const char** argv)
LoadParamBool(argc, args, "smallbatch", &isSmallBatch, true);
LoadParamBool(argc, args, "bigbatch", &isBigBatch, false);
LoadParamBool(argc, args, "randbatch", &isRandomBatch, false);
LoadParamInt(argc, args, "bucketsize", &bucketSize, wBatchSize * 10);
LoadParamInt(argc, args, "bucketsize", &bucketSize, wBatchSize * 1);
/* options for translating */
LoadParamString(argsNum, args, "test", testFN, "");
......@@ -241,8 +249,6 @@ void ShowParams(int argc, char** argv)
fprintf(stderr, "\n");
}
#define MAX_WORD_NUM 120
/*
split string by delimiter, this will return indices of all sub-strings
>> s - the original string
......
source/sample/transformer/Utility.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -73,9 +73,18 @@ public:
/* path to the validation file */
char validFN[1024];
/* path to the pre-trained model */
char pretrainedModel[1024];
/* device id */
int devID;
/* random seed */
int seed;
/* interval step for logging */
int logInterval;
/* beam size */
int beamSize;
......@@ -104,7 +113,13 @@ public:
int modelSize;
/* the maximum length in positional embedding */
int maxPosLen;
int maxPosition;
/* the maximum length for the source sequence */
int maxSrcLen;
/* the maximum length for the target sequence */
int maxTgtLen;
/* the dimension of fnn hidden layer */
int fnnHiddenSize;
......@@ -118,6 +133,9 @@ public:
/* the padding id */
int padID;
/* the unk id */
int unkID;
/* start symbol */
int startID;
......@@ -127,6 +145,12 @@ public:
/* indicates whether the model uses pre-norm */
bool preNorm;
/* add LN to the encoder/decoder output or not */
bool finalNorm;
/* reserve history for encoder/decoder layers or not */
bool useHistory;
/* indicates whether the model is running for machine translation */
bool isMT;
......@@ -139,10 +163,10 @@ public:
/* indicates whether the model is running with FP16 data type */
bool useFP16;
/* indicates whether we use the RPR attention */
/* use the RPR attention or not */
bool useRPR;
/* indicates whether we train the model */
/* train the model or not */
bool isTraining;
/* dropout rate for the model */
......
source/sample/transformer/submodel/Attention.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -29,10 +29,14 @@ namespace nmt
/* constructor */
Attention::Attention()
{
devID = -1;
nhead = -1;
dk = -1;
dv = -1;
d = -1;
dropoutP = 0.0;
maxRP = -1;
useRPR = false;
}
/* de-constructor */
......@@ -82,17 +86,17 @@ void Attention::InitModel(Config& config)
biasQ.SetZeroAll();
biasO.SetZeroAll();
biasK.SetDataRand(-(DTYPE)sqrt(6.0F / d), (DTYPE)sqrt(6.0F / d));
biasV.SetDataRand(-(DTYPE)sqrt(6.0F / d), (DTYPE)sqrt(6.0F / d));
biasK.SetDataRandn(-(DTYPE)sqrt(6.0F / d), (DTYPE)sqrt(6.0F / d));
biasV.SetDataRandn(-(DTYPE)sqrt(6.0F / d), (DTYPE)sqrt(6.0F / d));
}
/*
make the network
>> k - keys, B * L * H for encoders, B * 1 * H for decoders
>> k - keys, B * L * H
where B = batch size, L = sequence length,
and H = vector size of each position
>> q - queries, B * L * H
>> v - values, B * L * H for encoders, B * 1 * H for decoders
>> q - queries, B * L * H for encoders, B * 1 * H for decoders during inference
>> v - values, B * L * H
>> mask - as it is
>> isTraining - indicates whether the model is used for training
>> cache - decoder cache
......@@ -100,7 +104,7 @@ make the network
<< return - multi-attention result
*/
XTensor Attention::Make(XTensor& k, XTensor& q, XTensor& v, XTensor* mask,
bool isTraining, Cache* cache, int attType)
bool isTraining, Cache* cache, int attType)
{
const bool isEnc = (!cache) ? true : false;
......@@ -159,7 +163,7 @@ make the attention network given keys, queries and values (after linear transfor
>> isTraining - indicates whether the model is used for training
*/
XTensor Attention::MakeAttention(XTensor& k, XTensor& q, XTensor& v,
XTensor* mask, bool isTraining)
XTensor* mask, bool isTraining)
{
XTensor kheads;
XTensor qheads;
......@@ -188,9 +192,9 @@ XTensor Attention::MakeAttention(XTensor& k, XTensor& q, XTensor& v,
dot = BMMul(qheads, X_NOTRANS, kheads, X_TRANS);
if (mask)
dot = dot + *mask;
dot = Sum(dot, *mask, /*inplace=*/true);
dot = Linear(dot, 1.0F / (float)sqrt((float)dk / nhead));
dot = Linear(dot, 1.0F / (float)sqrt((float)dk / nhead), 0.0F, true);
scalar = Softmax(dot, -1);
......@@ -244,7 +248,7 @@ XTensor Attention::MakeRPRAttention(XTensor& k, XTensor& q, XTensor& v,
XTensor embMatrix, relativeKey;
/* generate the relative emb index (L_q, L_kv) */
embMatrix = GetRPEmbedding(lenQ, lenKV, maxRP, isEnc || isTraining);
embMatrix = GetRPEmbedding(lenQ, lenKV, maxRP, isEnc || isTraining, isTraining);
/* generate the relative key from the RPEmbK (L_q, L_kv, H/K) */
relativeKey = Gather(RPEmbK, embMatrix);
......@@ -255,13 +259,13 @@ XTensor Attention::MakeRPRAttention(XTensor& k, XTensor& q, XTensor& v,
relativeKey = ConvertDataType(relativeKey, X_FLOAT);
}
float scaling = (float)sqrt(d / nhead);
float scaling = float(sqrt(d / nhead));
qheads = ScaleAndShift(qheads, 1.0F / scaling);
dot = RPDotProduct(qheads, kheads, relativeKey, true);
if (mask)
dot = dot + *mask;
dot = Sum(dot, *mask, /*inplace=*/true);
/* softmax */
scalar = Softmax(dot, -1);
......@@ -287,12 +291,14 @@ generate relative position embeddings
>> lenQ - the length of query
>> lenKV - the length of key and value
>> maxRelativeLen - the maximum length of relative position
>> isEnc - indicates whether it is in the encoder
*/
XTensor Attention::GetRPEmbedding(const int lenQ, const int lenKV,
const int maxRelativeLen, const bool isEnc)
XTensor Attention::GetRPEmbedding(int lenQ, int lenKV,
int maxRelativeLen, bool isEnc, bool isTraining)
{
XTensor range;
XTensor embMatrix;
InitTensor1D(&range, lenKV, X_INT, devID);
int* index = new int[lenKV];
......@@ -313,11 +319,19 @@ XTensor Attention::GetRPEmbedding(const int lenQ, const int lenKV,
embMatrix = Unsqueeze(range, 0, lenQ);
}
//ClipMe(embMatrix, -float(maxRelativeLen), float(maxRelativeLen));
embMatrix = Clip(embMatrix, -float(maxRelativeLen), float(maxRelativeLen));
embMatrix = ScaleAndShift(embMatrix, 1.0F, float(maxRelativeLen));
ClipMe(embMatrix, -float(maxRelativeLen), float(maxRelativeLen));
ScaleAndShiftMe(embMatrix, 1.0F, float(maxRelativeLen));
delete[] index;
/* disable gradient flow */
if (isTraining) {
XTensor copyEmbMatrix;
InitTensor(&copyEmbMatrix, &embMatrix);
_CopyValues(&embMatrix, &copyEmbMatrix);
return copyEmbMatrix;
}
return embMatrix;
}
......@@ -351,6 +365,7 @@ XTensor Attention::RPDotProduct(XTensor& x, XTensor& y, XTensor& z, const bool i
XTensor context;
context = BMMul(x, y);
int newDims[]{ headNum, batchSize, context.GetDim(1), context.GetDim(2) };
context = Reshape(context, 4, newDims);
......@@ -358,7 +373,7 @@ XTensor Attention::RPDotProduct(XTensor& x, XTensor& y, XTensor& z, const bool i
xTrans = Transpose(x, 0, 1);
XTensor relative;
relative = MatrixMulBatched(xTrans, X_NOTRANS, z, transposeFlag);
relative = BMMul(xTrans, X_NOTRANS, z, transposeFlag);
XTensor relativeTrans;
relativeTrans = Transpose(relative, 0, 1);
......@@ -367,7 +382,7 @@ XTensor Attention::RPDotProduct(XTensor& x, XTensor& y, XTensor& z, const bool i
relativeTrans = Reshape(relativeTrans, 4, splitDims);
return context + relativeTrans;
return Sum(context, relativeTrans);
}
/* constructor */
......
source/sample/transformer/submodel/Attention.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -93,10 +93,6 @@ public:
/* bias for V */
XTensor biasV;
XTensor wBig;
XTensor bBig;
/* RPR emb */
XTensor RPEmbK;
......@@ -148,7 +144,7 @@ public:
XTensor* mask, bool isTraining, bool isEnc);
/* generate relative position embeddings */
XTensor GetRPEmbedding(const int lenQ, const int lenKV, const int maxRelativeLen, const bool isEnc);
XTensor GetRPEmbedding(int lenQ, int lenKV, int maxRelativeLen, bool isEnc, bool isTraining);
/* relative position-aware dot-product attention inner calculation */
XTensor RPDotProduct(XTensor& x, XTensor& y, XTensor& z, const bool is_key);
......
source/sample/transformer/submodel/CommonModules.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/submodel/CommonModules.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/submodel/Embedding.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -29,8 +29,10 @@ namespace nmt
/* constructor */
Embedder::Embedder()
{
d = -1;
devID = -1;
vSize = -1;
eSize = -1;
maxLength = -1;
}
......@@ -50,7 +52,7 @@ void Embedder::InitModel(Config& config, bool isEnc)
d = config.modelSize;
padIdx = config.padID;
eSize = config.embSize;
maxLength = config.maxPosLen;
maxLength = config.maxPosition;
vSize = (isEnc) ? config.srcVocabSize : config.tgtVocabSize;
InitTensor2D(&w, vSize, eSize, X_FLOAT, devID);
......@@ -59,6 +61,10 @@ void Embedder::InitModel(Config& config, bool isEnc)
DTYPE v = 1.0F / (float)sqrt((float)eSize);
w.SetDataRandn(0, v);
for (int i = 0; i < eSize; i++) {
w.Set2D(0.0F, padIdx, i);
}
/* create the positional embedding matrix */
MakePosEmbedding(maxLength);
}
......@@ -138,13 +144,13 @@ XTensor Embedder::Make(XTensor& input, bool isDec, bool isTraining, int nstep)
posEmbedding = Unsqueeze(embTMP, 0, input.GetDim(0));
/* then we make word embeddings */
//w.enableGrad = false;
wordEmbedding = Gather(w, input);
wordEmbedding = Linear(wordEmbedding, (float)sqrt((float)eSize));
wordEmbedding = Linear(wordEmbedding, (float)sqrt((float)eSize), 0.0F, true);
/* we sum over the two embeddings */
SumMe(wordEmbedding, posEmbedding);
return wordEmbedding;
}
......
source/sample/transformer/submodel/Embedding.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/submodel/FNN.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -31,9 +31,11 @@ namespace nmt
/* constructor */
FNN::FNN()
{
dropoutP = 0.0F;
inSize = -1;
outSize = -1;
hSize = -1;
devID = -1;
}
/* de-constructor */
......@@ -66,8 +68,8 @@ void FNN::InitModel(Config& config)
_SetDataFanInOut(&w1, scale);
_SetDataFanInOut(&w2, scale);
w1.SetDataRand(-(DTYPE)sqrt(6.0F / inSize), (DTYPE)sqrt(6.0F / inSize));
w2.SetDataRand(-(DTYPE)sqrt(6.0F / hSize), (DTYPE)sqrt(6.0F / hSize));
//w1.SetDataRand(-(DTYPE)sqrt(6.0F / inSize), (DTYPE)sqrt(6.0F / inSize));
//w2.SetDataRand(-(DTYPE)sqrt(6.0F / hSize), (DTYPE)sqrt(6.0F / hSize));
b1.SetZeroAll();
b2.SetZeroAll();
......@@ -85,9 +87,9 @@ XTensor FNN::Make(XTensor& input, bool isTraining)
/* t1 = max(0, x * w1 + b1) */
t1 = Rectify(MulAndShift(input, w1, b1));
if (isTraining && dropoutP > 0)
t1 = Dropout(t1, dropoutP);
t1 = Dropout(t1, dropoutP, /*inplace=*/true);
/* result = t1 * w2 + b2 */
return MulAndShift(t1, w2, b2);
......
source/sample/transformer/submodel/FNN.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/submodel/GLU.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -30,6 +30,7 @@ namespace nmt
/* constructor */
GLU::GLU()
{
devID = -1;
inSize = -1;
outSize = -1;
hSize = -1;
......
source/sample/transformer/submodel/GLU.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/submodel/LayerHistory.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -16,6 +16,7 @@
/*
* $Created by: Bei Li (libei_neu@outlook.com) 2020-02-03
* $Modified by: Chi Hu (huchinlp@gmail.com) 2020-12-10
*/
#include "Embedding.h"
......@@ -23,6 +24,7 @@
#include "LayerHistory.h"
#include "../Utility.h"
#include "../../../tensor/core/CHeader.h"
#include "../../../tensor/XName.h"
#define SAFE_DELETE(x) do{ if((x) != NULL){delete (x); (x) = NULL;} } while(false)
#define SAFE_DELETE_ARRAY(x) do{ if((x) != NULL) {delete [] (x); (x)=NULL;} } while(false)
......@@ -34,16 +36,20 @@ namespace nmt
LayerHistory::LayerHistory()
{
d = -1;
devID = -1;
count = -1;
weight = NULL;
nlayer = -1;
weights = NULL;
history = NULL;
layerNorms = NULL;
}
/* de-constructor */
LayerHistory::~LayerHistory()
{
history.Clear();
delete history;
delete[] layerNorms;
delete[] weights;
}
/*
......@@ -56,7 +62,20 @@ void LayerHistory::InitModel(Config& config)
d = config.modelSize;
nlayer = config.nEncLayer;
InitTensor2D(&weight, nlayer + 1, nlayer + 1, X_FLOAT, devID);
/* the triangle weight matrices for dlcl
layer 0: [1, 0, ..., 0]
layer 1: [0.5, 0.5, ..., 0]
layer 2: [0.33, 0.33, 0.33, ..., 0] */
weights = new XTensor[nlayer + 1];
for (int i = 0; i < nlayer + 1; i++) {
InitTensor1D(&(weights[i]), i + 1, X_FLOAT, devID);
float* data = new float[i + 1];
for (int j = 0; j < i + 1; j++) {
data[j] = 1.0F / float(i + 1);
}
weights[i].SetData(data, i + 1);
delete[] data;
}
layerNorms = new LN[nlayer];
......@@ -68,59 +87,88 @@ void LayerHistory::InitModel(Config& config)
/*
the Add operation
>> tensor - the previous layer output. It might be of size B * L * H
where B = batch size, L = sequence length,
and H = vector size of each position
>> layer - the previous layer output. It might be of size B * L * H
where B = batch size, L = sequence length,
and H = vector size of each position
*/
void LayerHistory::Add(XTensor& tensor)
void LayerHistory::Add(XTensor& layer)
{
/* the embedding is not normed */
count += 1;
if (history.Size() == 0) {
//sample_ = tensor;
history.Add(&tensor);
if (history->count == 0) {
history->Add(layer);
return;
}
XTensor ln = layerNorms[count - 2].Make(tensor);
history.Add(&ln);
layer = layerNorms[count - 2].Make(layer);
history->Add(layer);
}
/*
generate the weight sum vector of all previous layer output in the history as the layer input
calculate the weighted sum of previous layers
the result for the i-th layer is:
result = sum(layers[0...i] * weight[i][0...i])
shape of the result: B * L * H
*/
XTensor LayerHistory::Pop()
{
/* the number of layer output in the history */
int size = (int)history.Size();
TensorList historyList;
for (int i = 0; i < size; i++)
historyList.Add(history[i]);
TensorList list;
for (int i = 0; i < history->count; i++) {
list.Add(&(history->list[i]));
}
XTensor stack;
stack = Merge(list, 0);
//Stack(list, 0);
/* we need stack the tensor along the first dim*/
XTensor stackTensor = Stack(historyList, 0);
int dimSize[MAX_TENSOR_DIM_NUM];
for (int i = 0; i < stack.order + 1; i++)
dimSize[i + 1] = stack.dimSize[i];
dimSize[0] = int(list.Size());
dimSize[1] /= dimSize[0];
stack = Reshape(stack, stack.order + 1, dimSize);
XTensor interWeight;
XTensor res;
res = MultiplyDim(stack, weights[list.Size() - 1], 0);
InitTensor2D(&interWeight, 1, weight.dimSize[1], DEFAULT_DTYPE, devID);
XTensor layerWeight;
InitTensor1D(&layerWeight, size, DEFAULT_DTYPE, devID);
return ReduceSum(res, 0);
}
_SelectRange(&weight, &interWeight, 0, size - 1, size);
interWeight.Reshape(interWeight.unitNum);
_SelectRange(&interWeight, &layerWeight, 0, 0, size);
MultiplyDimMe(stackTensor, layerWeight, 0);
/* clear the history */
void LayerHistory::ClearHistory(bool reset)
{
if(history != NULL)
delete history;
if(reset)
history = new History;
else
history = NULL;
count = 0;
}
XTensor result;
ReduceSum(stackTensor, result, 0);
/* initialize the history */
History::History()
{
count = 0;
}
return result;
/* delete the history */
History::~History()
{
for (int i = 0; i < MAX_LAYER_NUM; i++) {
list[i].DestroyData();
XLink::ClearOutgoing(&list[i]);
XLink::ClearIncoming(&list[i]);
if (list[i].grad != NULL)
delete list[i].grad;
}
}
void LayerHistory::ClearHistory()
/* append a layer to the history */
void History::Add(XTensor& layer)
{
history.Clear();
list[count] = std::move(layer);
XLink::ClearOutgoing(&layer);
XLink::ClearIncoming(&layer);
count++;
}
}
\ No newline at end of file
source/sample/transformer/submodel/LayerHistory.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -27,16 +27,42 @@
#include "../../../tensor/function/FHeader.h"
using namespace nts;
using namespace std;
namespace nmt
{
/*
multi-head attention
y(Q, K, V) = cat(head_1, head_2, ..., head_n)
where head_i = Attention(Q * w_i^Q, K * w_i^K, V * w_i^V)
attention(Q, K, V) = softmax(Q * K^T/d_k^0.5) V
d_k = dimension size of K
#define MAX_LAYER_NUM 50
/*
the class of history list
*/
class History {
public:
/* number of elements in the list */
int count;
/* the history list */
XTensor list[MAX_LAYER_NUM];
public:
/* contructor */
History();
/* de-contructor */
~History();
/* append a layer to the list */
void Add(XTensor& layer);
};
/*
the class of layer history
it generates the weighted sum of previous layers
the result for the i-th layer is:
res = sum(layers[0...i] * weight[i][0...i])
*/
class LayerHistory
{
......@@ -44,8 +70,8 @@ public:
/* device id */
int devID;
/* the triangle weight matrix for dlcl */
XTensor weight;
/* the triangle weight matrices for dlcl */
XTensor* weights;
/* hidden size */
int d;
......@@ -57,7 +83,7 @@ public:
int count;
/* a history to store the value of intimidate layers */
TensorList history;
History* history;
/* layer normalization for each intimidate layer */
LN* layerNorms;
......@@ -79,7 +105,7 @@ public:
XTensor Pop();
/* clean the history*/
void ClearHistory();
void ClearHistory(bool reset=true);
};
}
......
source/sample/transformer/submodel/LayerNorm.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -51,12 +51,12 @@ void LN::InitModel(Config& config)
d = config.modelSize;
InitTensor1D(&w, d, X_FLOAT, devID);
InitTensor1D(&b, d, X_FLOAT, devID);
w.SetDataRand(1.0F, 1.0F);
b.SetZeroAll();
InitTensor1D(&weight, d, X_FLOAT, devID);
InitTensor1D(&bias, d, X_FLOAT, devID);
weight.SetDataRand(1.0F, 1.0F);
bias.SetZeroAll();
w.SetDataFixed(1);
weight.SetDataFixed(1);
}
/*
......@@ -104,7 +104,11 @@ XTensor LN::Make(XTensor& input)
}
/* result = x' * w + b */
return xn * w + b;
xn = xn * weight;
xn = Sum(xn, bias, true);
return xn;
}
}
\ No newline at end of file
source/sample/transformer/submodel/LayerNorm.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -39,10 +39,10 @@ public:
int devID;
/* the transformation matrix w */
XTensor w;
XTensor weight;
/* the bias term b */
XTensor b;
XTensor bias;
/* dimension size of the model */
int d;
......
source/sample/transformer/submodel/NNUtil.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/submodel/NNUtil.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/submodel/Output.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -33,6 +33,7 @@ Output::Output()
devID = -1;
vSize = -1;
hSize = -1;
padIdx = -1;
}
/* de-constructor */
......@@ -49,11 +50,15 @@ void Output::InitModel(Config& config)
devID = config.devID;
hSize = config.modelSize;
vSize = config.tgtVocabSize;
padIdx = config.padID;
InitTensor2D(&w, vSize, hSize, X_FLOAT, devID);
DTYPE v = 1.0F / (float)sqrt((float)hSize);
w.SetDataRandn(0, v);
for (int i = 0; i < hSize; i++) {
w.Set2D(0.0F, padIdx, i);
}
}
/*
......
source/sample/transformer/submodel/Output.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -43,6 +43,9 @@ public:
/* vector size of the linear transformation */
int hSize;
/* the padding index */
int padIdx;
/* transformation matrix */
XTensor w;
......
source/sample/transformer/translate/DataSet.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -32,19 +32,19 @@ using namespace nmt;
namespace nts {
/* sort the output by id (in ascending order) */
/* sort the input by length (in descending order) */
void DataSet::SortInput() {
sort(inputBuffer.items, inputBuffer.items + inputBuffer.count,
[](Example* a, Example* b) {
return a->values.count > b->values.count;
sort(inputBuffer.begin(), inputBuffer.end(),
[](const Example& a, const Example& b) {
return a.values.size() > b.values.size();
});
}
/* sort the input by length (in descending order) */
/* sort the output by id (in ascending order) */
void DataSet::SortOutput() {
sort(outputBuffer.items, outputBuffer.items + outputBuffer.count,
[](Result* a, Result* b) {
return a->id < b->id;
sort(outputBuffer.begin(), outputBuffer.end(),
[](const Example& a, const Example& b) {
return a.id < b.id;
});
}
......@@ -54,43 +54,43 @@ load data from the file to the buffer
void DataSet::LoadDataToBuffer()
{
string line;
inputBuffer.Clear();
inputBuffer.clear();
bufferUsed = 0;
int id = 0;
const string tokenDelimiter = " ";
while (getline(*fp, line)) {
IntList values;
vector<int> values;
/* load words and transform them to ids */
auto indices = SplitToPos(line, tokenDelimiter);
UInt64List indices = SplitToPos(line, tokenDelimiter);
/* reserve the first 120 words if the input is too long */
size_t maxLen = indices.Size() > MAX_WORD_NUM ? MAX_WORD_NUM : indices.Size();
/* reserve the first maxSrcLen words if the input is too long */
int maxLen = int(indices.Size()) > maxSrcLen ? maxSrcLen : int(indices.Size());
for (size_t i = 0; i < maxLen; i++) {
size_t offset = (i != (indices.Size() - 1)) ?
(size_t)indices[(int)i + 1] - (size_t)indices[(int)i] - tokenDelimiter.size()
: line.size() - (size_t)indices[(int)i];
string word = line.substr((size_t)indices[(int)i], offset);
for (int i = 0; i < maxLen; i++) {
auto offset = (i != (int(indices.Size()) - 1)) ?
indices[i + 1] - indices[i] - tokenDelimiter.size()
: line.size() - indices[i];
string word = line.substr(indices[i], offset);
if (srcVocab.word2id.find(word) == srcVocab.word2id.end())
values.Add(UNK);
values.emplace_back(unkID);
else
values.Add(srcVocab.word2id.at(word));
values.emplace_back(srcVocab.word2id.at(word));
}
/* make sure that the sequence ends with EOS */
if (values.Size() != 0 && values[-1] != EOS)
values.Add(EOS);
Example* example = new Example;
example->id = id;
example->values = values;
if (values.Size() != 0)
inputBuffer.Add(example);
if (values.size() != 0 && values.back() != endID)
values.emplace_back(endID);
Example example;
example.id = id;
example.values = values;
if (values.size() != 0)
inputBuffer.emplace_back(example);
else
emptyLines.Add(id);
emptyLines.emplace_back(id);
id++;
}
fp->close();
......@@ -109,23 +109,16 @@ load a mini-batch to the device (for translating)
>> devID - the device id, -1 for the CPU
<< indices of the sentences
*/
UInt64List DataSet::LoadBatch(XTensor* batchEnc, XTensor* paddingEnc,
int minSentBatch, int batchSize, int devID)
UInt64List DataSet::LoadBatch(XTensor* batchEnc, XTensor* paddingEnc, int minSentBatch, int batchSize, int devID)
{
int realBatchSize = minSentBatch;
/* get the maximum sentence length in a mini-batch */
int maxLen = (int)inputBuffer[(int)bufferUsed]->values.Size();
/* dynamic batching for sentences */
//while ((realBatchSize < (inputBuffer.Size() - bufferUsed))
// && (realBatchSize * maxLen < batchSize)) {
// realBatchSize++;
//}
int maxLen = int(inputBuffer[bufferUsed].values.size());
/* real batch size */
if ((inputBuffer.Size() - bufferUsed) < realBatchSize) {
realBatchSize = (int)(inputBuffer.Size() - bufferUsed);
if ((int(inputBuffer.size()) - bufferUsed) < realBatchSize) {
realBatchSize = int(inputBuffer.size()) - bufferUsed;
}
CheckNTErrors(maxLen != 0, "invalid length");
......@@ -134,25 +127,25 @@ UInt64List DataSet::LoadBatch(XTensor* batchEnc, XTensor* paddingEnc,
float* paddingValues = new float[realBatchSize * maxLen];
for (int i = 0; i < realBatchSize * maxLen; i++) {
batchValues[i] = PAD;
batchValues[i] = padID;
paddingValues[i] = 1.0F;
}
size_t curSrc = 0;
int curSrc = 0;
/* right padding */
UInt64List infos;
size_t totalLength = 0;
int totalLength = 0;
for (size_t i = 0; i < (size_t)realBatchSize; ++i) {
infos.Add(inputBuffer[(int)(bufferUsed + i)]->id);
totalLength += inputBuffer[(int)(bufferUsed + i)]->values.Size();
for (int i = 0; i < realBatchSize; ++i) {
infos.Add(inputBuffer[bufferUsed + i].id);
totalLength += int(inputBuffer[bufferUsed + i].values.size());
curSrc = maxLen * i;
for (size_t j = 0; j < inputBuffer[(int)(bufferUsed + i)]->values.Size(); j++)
batchValues[(int)(curSrc++)] = (int)inputBuffer[(int)(bufferUsed + i)]->values[(int)j];
for (int j = 0; j < int(inputBuffer[bufferUsed + i].values.size()); j++)
batchValues[curSrc++] = inputBuffer[bufferUsed + i].values[j];
while (curSrc < maxLen * (i + 1))
paddingValues[(int)(curSrc++)] = 0;
paddingValues[curSrc++] = 0;
}
infos.Add(totalLength);
......@@ -201,7 +194,7 @@ void DataSet::Init(const char* dataFile, const char* srcVocabFN, const char* tgt
/* check if the buffer is empty */
bool DataSet::IsEmpty() {
if (bufferUsed < inputBuffer.Size())
if (bufferUsed < inputBuffer.size())
return false;
return true;
}
......@@ -211,12 +204,11 @@ void DataSet::DumpRes(const char* ofn)
{
ofstream ofile(ofn, ios::out);
for (int t = 0; t < outputBuffer.Size(); t++) {
auto res = outputBuffer[t];
for (int i = 0; i < res->res.Size(); i++) {
if (res->res[i] < 4)
for (const auto& tgtSent : outputBuffer) {
for (const auto& w : tgtSent.values) {
if (w < 4)
break;
ofile << tgtVocab.id2word[res->res[i]] << " ";
ofile << tgtVocab.id2word[w] << " ";
}
ofile << "\n";
}
......@@ -229,14 +221,6 @@ DataSet::~DataSet()
{
/* release the file */
delete fp;
/* release the input buffer */
for (int i = 0; i < inputBuffer.Size(); i++)
delete inputBuffer[i];
/* release the output buffer */
for (int i = 0; i < outputBuffer.Size(); i++)
delete outputBuffer[i];
}
}
\ No newline at end of file
source/sample/transformer/translate/DataSet.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -31,40 +31,41 @@
#include "../../../tensor/XTensor.h"
#include "../../../tensor/XGlobal.h"
#define MAX_WORD_NUM 120
using namespace std;
namespace nts {
/* the struct of tokenized input */
struct Example {
int id;
IntList values;
};
/* the struct of tokenized output */
struct Result {
int id;
IntList res;
vector<int> values;
public:
Example() {
id = 0;
}
};
/* A `DataSet` is associated with a file which contains variable length data.*/
struct DataSet {
public:
/* the data buffer */
InputBufferType inputBuffer;
vector<Example> inputBuffer;
/* a list of empty line number */
IntList emptyLines;
vector<int> emptyLines;
/* the result buffer */
OutputBufferType outputBuffer;
vector<Example> outputBuffer;
/* the pointer to file stream */
ifstream* fp;
/* size of used data in buffer */
size_t bufferUsed;
int bufferUsed;
/* the source vocabulary */
Vocab srcVocab;
......@@ -72,6 +73,21 @@ public:
/* the target vocabulary */
Vocab tgtVocab;
/* the maximum length of an input sequence */
int maxSrcLen;
/* the padding id */
int padID;
/* the unk id */
int unkID;
/* start symbol */
int startID;
/* end symbol */
int endID;
public:
/* sort the input by length */
......@@ -84,8 +100,7 @@ public:
void LoadDataToBuffer();
/* generate a mini-batch */
UInt64List LoadBatch(XTensor* batchEnc, XTensor* paddingEnc,
int sBatch, int wBatch, int devID);
UInt64List LoadBatch(XTensor* batchEnc, XTensor* paddingEnc, int minSentBatch, int batchSize, int devID);
/* initialization function */
void Init(const char* dataFile, const char* srcVocabFN, const char* tgtVocabFN);
......
source/sample/transformer/translate/LengthPenalty.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -42,7 +42,7 @@ float LengthPenalizer::GNMT(float length, float alpha)
base = (length + 5.0F) / (1.0F + 5.0F);
lp = pow(base, alpha);
lp = float(pow(base, alpha));
return lp;
}
......
source/sample/transformer/translate/LengthPenalty.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/translate/Predictor.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/translate/Predictor.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/translate/Search.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -727,6 +727,7 @@ GreedySearch::GreedySearch()
endSymbolNum = 0;
endSymbols = new int[32];
startSymbol = -1;
scalarMaxLength = -1;
}
/* de-constructor */
......
source/sample/transformer/translate/Search.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/translate/Translator.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......@@ -54,6 +54,12 @@ void Translator::Init(Config& config)
sentBatch = config.sBatchSize;
wordBatch = config.wBatchSize;
batchLoader.maxSrcLen = config.maxSrcLen;
batchLoader.unkID = config.unkID;
batchLoader.padID = config.padID;
batchLoader.startID = config.startID;
batchLoader.endID = config.endID;
if (beamSize > 1) {
LOG("translating with beam search (%d)", beamSize);
seacher = new BeamSearch();
......@@ -123,12 +129,12 @@ void Translator::Translate(const char* ifn, const char* sfn,
XTensor score;
((BeamSearch*)seacher)->Search(model, batchEnc, paddingEnc, output, score);
}
for (int i = 0; i < indices.Size() - 1; ++i) {
Result* res = new Result;
res->id = int(indices[i]);
res->res = output[i];
batchLoader.outputBuffer.Add(res);
Example res;
res.id = int(indices[i]);
for (int j = 0; j < output[i].Size(); j++)
res.values.emplace_back(output[i][j]);
batchLoader.outputBuffer.emplace_back(std::move(res));
}
delete[] output;
......@@ -142,17 +148,17 @@ void Translator::Translate(const char* ifn, const char* sfn,
double elapsed = GetClockSec() - batchStart;
batchStart = GetClockSec();
LOG("elapsed=%.1fs, sentence=%f, sword=%.1fw/s",
elapsed, float(sentCount) / float(batchLoader.inputBuffer.Size()),
elapsed, float(sentCount) / float(batchLoader.inputBuffer.size()),
double(wc) / elapsed);
wc = 0;
}
}
/* append empty lines to the result */
for (int i = 0; i < batchLoader.emptyLines.Size(); i++) {
Result* emptyRes = new Result;
emptyRes->id = batchLoader.emptyLines[i];
batchLoader.outputBuffer.Add(emptyRes);
for (const auto& empty: batchLoader.emptyLines) {
Example emptyRes;
emptyRes.id = empty;
batchLoader.outputBuffer.emplace_back(emptyRes);
}
double startDump = GetClockSec();
......@@ -166,7 +172,7 @@ void Translator::Translate(const char* ifn, const char* sfn,
double elapsed = GetClockSec() - startDump;
LOG("translation completed (word=%d, sent=%zu)",
wordCountTotal, batchLoader.inputBuffer.Size() + batchLoader.emptyLines.Size());
wordCountTotal, batchLoader.outputBuffer.size() + batchLoader.emptyLines.size());
}
/*
......
source/sample/transformer/translate/Translator.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/translate/Vocab.cpp
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
source/sample/transformer/translate/Vocab.h
/* NiuTrans.NMT - an open-source neural machine translation system.
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2020 NiuTrans Research. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论