Add python bindings (#69)

Co-authored-by: Daniel Mann <dmann@apptek.com>
Co-authored-by: Eugen Beck <ebeck@apptek.com>

.gitignore

@@ -22,6 +22,7 @@
 *.pyc
 *.o
 *.so
+/RASR.egg-info
 
 # For exec-files, the postfix is ".$(OS)-$(PROC)-$(COMPILE)[-$(SUFFIX)]".
 *.*-*-*
@@ -53,6 +54,9 @@
 .history*
 *.autosave
 
+# Singularity/Apptainer images
+*.sif
+
 # Broken NFS.
 .nfs*
 

Modules.make

@@ -123,6 +123,7 @@ TOOLS += CorpusStatistics
 TOOLS += FeatureExtraction
 TOOLS += FeatureStatistics
 TOOLS += Fsa
+TOOLS += LibRASR
 TOOLS += Lm
 TOOLS += SpeechRecognizer
 TOOLS += Xml

apptainer/2023-11-08_tensorflow-2.14_v1/image.def

@@ -9,6 +9,8 @@ Stage: build
     DEBIAN_FRONTEND=noninteractive apt install -y wget git unzip gzip libssl-dev lsb-release zsh \
         bison libxml2-dev libopenblas-dev libsndfile1-dev libcrypto++-dev libcppunit-dev \
         parallel xmlstarlet python3-lxml htop strace gdb sox python3-pip cmake ffmpeg vim
+
+    pip3 install pybind11==2.11.1
 
     # download the cache manager and place in /usr/local
     cd /usr/local

apptainer/2023-11-08_tensorflow-2.14_v1/makefiles/Modules.make

@@ -110,6 +110,9 @@ MODULES += MODULE_TEST
 # Tensorflow integration
 MODULES += MODULE_TENSORFLOW
 
+# ONNX integration
+# MODULES += MODULE_ONNX
+
 # define variables for the makefiles
 $(foreach module, $(MODULES), $(eval $(module) = 1))
 
@@ -120,6 +123,8 @@ TOOLS += CorpusStatistics
 TOOLS += FeatureExtraction
 TOOLS += FeatureStatistics
 TOOLS += Fsa
+TOOLS += LibRASR
+TOOLS += Lm
 TOOLS += SpeechRecognizer
 TOOLS += Xml
 

apptainer/2023-11-08_tensorflow-2.14_v1/makefiles/config/cc-gcc.make

@@ -22,6 +22,7 @@ CXX_MINOR = $(shell $(CXX) --version | head -n 1 | sed -e 's/.*[ \t]\([0-9]\)\.\
 # -----------------------------------------------------------------------------
 # compiler options
 DEFINES		+= -D_GNU_SOURCE
+CCFLAGS		+= -fPIC
 CCFLAGS		+= -pipe
 CCFLAGS		+= -funsigned-char
 CCFLAGS		+= -fno-exceptions

apptainer/2023-11-08_tensorflow-2.14_v1/makefiles/config/os-linux.make

@@ -144,6 +144,8 @@ LDFLAGS     += `${PYTHON_PATH}/bin/python3.11-config --ldflags --embed 2>/dev/nu
 LDFLAGS     += -Wl,-rpath -Wl,${PYTHON_PATH}/lib
 else
 INCLUDES    += `python3.11-config --includes 2>/dev/null`
+INCLUDES    += -I$(shell python3.11 -c 'import numpy as np; print(np.get_include())')
+INCLUDES    += `python3.11 -m pybind11 --includes 2>/dev/null`
 LDFLAGS     += `python3.11-config --ldflags --embed 2>/dev/null`
 # IF you want to use Python2 for whatever reason:
 # INCLUDES    += `pkg-config --cflags python`

config/cc-gcc.make

@@ -22,6 +22,7 @@ CXX_MINOR = $(shell $(CXX) --version | head -n 1 | sed -e 's/.*[ \t]\([0-9]\)\.\
 # -----------------------------------------------------------------------------
 # compiler options
 DEFINES		+= -D_GNU_SOURCE
+CCFLAGS		+= -fPIC
 CCFLAGS		+= -pipe
 CCFLAGS		+= -funsigned-char
 CCFLAGS		+= -fno-exceptions

config/os-linux.make

@@ -152,7 +152,9 @@ LDFLAGS     += `${PYTHON_PATH}/bin/python3-config --ldflags 2>/dev/null`
 LDFLAGS     += -Wl,-rpath -Wl,${PYTHON_PATH}/lib
 else
 INCLUDES    += `python3-config --includes 2>/dev/null`
-LDFLAGS     += `python3-config --ldflags 2>/dev/null`
+INCLUDES    += -I$(shell python3 -c 'import numpy as np; print(np.get_include())')
+INCLUDES    += `python3 -m pybind11 --includes 2>/dev/null`
+LDFLAGS     += `python3-config --ldflags --embed 2>/dev/null`
 # IF you want to use Python2 for whatever reason:
 # INCLUDES    += `pkg-config --cflags python`
 # LDFLAGS     += `pkg-config --libs python`

setup.py

@@ -0,0 +1,40 @@
+"""Setup.py inspired by https://thomastrapp.com/posts/building-a-pypi-package-for-a-modern-cpp-project/"""
+import os
+from setuptools import setup, dist
+
+PACKAGE_NAME = "librasr" # keyword to import
+LIBRASR_SUBDIR = os.path.join("src", "Tools", "LibRASR")
+LIBRASR_SO = "librasr.so"
+
+assert os.path.exists(os.path.join(LIBRASR_SUBDIR, LIBRASR_SO)), \
+    "Could not find librasr shared library. Did you compile it correctly?"
+class BinaryDistribution(dist.Distribution):
+    """Helper class: We assume that librasr.so has already been compiled."""
+    def has_ext_modules(self) -> bool:
+        return True
+
+# get README content for long description
+this_directory = os.path.abspath(os.path.dirname(__file__))
+with open(os.path.join(this_directory, 'README')) as f:
+    long_description = f.read()
+
+setup(
+    name='RASR',
+
+    # include shared library and link to lower case package name
+    packages=[PACKAGE_NAME],
+    package_dir={PACKAGE_NAME: LIBRASR_SUBDIR},
+    package_data={PACKAGE_NAME: [LIBRASR_SO]},
+    include_package_data=True,
+
+    description="RASR as a python module.",
+    long_description=long_description,
+    long_description_content_type="text/markdown",
+
+    # use custom distribution class to export librasr.so
+    distclass=BinaryDistribution,
+
+    version='0.0.1',
+    url='https://github.com/rwth-i6/rasr',
+    author_email='rwthasr@i6.informatik.rwth-aachen.de',
+)

src/Core/Configuration.cc

@@ -169,7 +169,7 @@ void Configuration::Resource::writeUsage(XmlWriter& os) const {
  * Central storage place for all resources.
  */
 
-class Configuration::ResourceDataBase {
+class Configuration::ResourceDataBase : public ReferenceCounted {
 private:
     std::set<Resource> resources;
     Resource           noResource_;
@@ -401,24 +401,19 @@ bool Configuration::isWellFormedParameterName(const std::string& s) {
 }
 
 Configuration::Configuration()
-        : db_(0),
-          isDataBaseOwner_(true),
+        : db_(new ResourceDataBase),
           selection_("UNNAMED"),
           name_("UNNAMED") {
-    // define new root configuration
-    db_ = new ResourceDataBase;
 }
 
 Configuration::Configuration(const Configuration& c)
         : db_(c.db_),
-          isDataBaseOwner_(false),
           selection_(c.selection_),
           name_(c.name_) {
 }
 
 Configuration::Configuration(const Configuration& c, const std::string& add_selection)
-        : db_(c.db_),
-          isDataBaseOwner_(false) {
+        : db_(c.db_) {
     require(add_selection.length());
     require(add_selection.find(resource_separation_char) == std::string::npos);
     require(add_selection.find(resource_wildcard_char) == std::string::npos);
@@ -428,16 +423,9 @@ Configuration::Configuration(const Configuration& c, const std::string& add_sele
     ensure(isWellFormedParameterName(selection_));
 }
 
-Configuration::~Configuration() {
-    if (isDataBaseOwner_)
-        delete db_;
-}
+Configuration::~Configuration() {}
 
 Configuration& Configuration::operator=(const Configuration& c) {
-    if (isDataBaseOwner_) {
-        delete db_;
-        isDataBaseOwner_ = false;
-    }
     db_        = c.db_;
     selection_ = c.selection_;
     name_      = c.name_;

src/Core/Configuration.hh

@@ -25,6 +25,7 @@
 #include <string.h>
 #include <sys/stat.h>
 #include "Assertions.hh"
+#include "ReferenceCounting.hh"
 #include "Types.hh"
 #include "XmlStream.hh"
 
@@ -112,10 +113,9 @@ public:
     class ResourceDataBase;
 
 private:
-    ResourceDataBase* db_;
-    bool              isDataBaseOwner_;
-    std::string       selection_;
-    std::string       name_;
+    Ref<ResourceDataBase> db_;
+    std::string           selection_;
+    std::string           name_;
 
     void warning(const std::string& filename, int lineNumber,
                  const std::string& description);

src/Python/AllophoneStateFsaBuilder.cc

@@ -0,0 +1,61 @@
+#include "AllophoneStateFsaBuilder.hh"
+
+#include <memory>
+#include <vector>
+
+#undef ensure // macro duplication in pybind11/numpy.h
+#include <pybind11/pybind11.h>
+#include <pybind11/numpy.h>
+
+#include <Core/Application.hh>
+#include <Core/Component.hh>
+#include <Core/Archive.hh>
+#include <Bliss/CorpusDescription.hh>
+#include <Nn/AllophoneStateFsaExporter.hh>
+
+namespace py = pybind11;
+
+struct BuildSegmentToOrthMapVisitor : public Bliss::CorpusVisitor {
+    BuildSegmentToOrthMapVisitor()
+            : Bliss::CorpusVisitor(), map_(new Core::StringHashMap<std::string>()) {}
+
+    virtual void visitSpeechSegment(Bliss::SpeechSegment* s) {
+        (*map_)[s->fullName()] = s->orth();
+    }
+
+    std::shared_ptr<Core::StringHashMap<std::string>> map_;
+};
+
+static std::shared_ptr<Core::StringHashMap<std::string>> build_segment_to_orth_map(Core::Configuration const& config) {
+    Bliss::CorpusDescription     corpus(config);
+    BuildSegmentToOrthMapVisitor visitor;
+    corpus.accept(&visitor);
+    return visitor.map_;
+}
+
+AllophoneStateFsaBuilder::AllophoneStateFsaBuilder(const Core::Configuration& c)
+        : Core::Component(c),
+          allophoneStateFsaExporter_(nullptr),
+          segmentToOrthMap_(nullptr) {
+    // init exporter and orth map
+    allophoneStateFsaExporter_ = std::make_shared<Nn::AllophoneStateFsaExporter>(select("alignment-fsa-exporter"));
+    segmentToOrthMap_ = build_segment_to_orth_map(select("corpus"));
+}
+
+py::tuple AllophoneStateFsaBuilder::buildBySegmentName(const std::string& segmentName) {
+    auto iter = segmentToOrthMap_->find(segmentName);
+    if (iter == segmentToOrthMap_->end()) {
+        throw std::invalid_argument("Could not find segment with name " + segmentName);
+    }
+    return buildByOrthography(iter->second);
+}
+
+py::tuple AllophoneStateFsaBuilder::buildByOrthography(const std::string& orthography) {
+    Nn::AllophoneStateFsaExporter::ExportedAutomaton automaton = allophoneStateFsaExporter_->exportFsaForOrthography(orthography);
+    return py::make_tuple( 
+        automaton.num_states,
+        automaton.num_edges,
+        py::array_t<u32>(automaton.edges.size(), automaton.edges.data()),
+        py::array_t<f32>(automaton.weights.size(), automaton.weights.data())
+    );
+}

src/Python/AllophoneStateFsaBuilder.hh

@@ -0,0 +1,28 @@
+#ifndef _PYTHON_ALLOPHONESTATEFSABUILDER_HH
+#define _PYTHON_ALLOPHONESTATEFSABUILDER_HH
+
+#include <pybind11/pybind11.h>
+
+#include <Core/Application.hh>
+#include <Core/Archive.hh>
+#include <Core/Configuration.hh>
+#include <Nn/AllophoneStateFsaExporter.hh>
+
+namespace py = pybind11;
+
+class AllophoneStateFsaBuilder : public Core::Component {
+public:
+    AllophoneStateFsaBuilder(const Core::Configuration& c);
+    ~AllophoneStateFsaBuilder() = default;
+
+    py::tuple buildBySegmentName(const std::string& segmentName);
+
+    py::tuple buildByOrthography(const std::string& orthography);
+
+private:
+    std::shared_ptr<Nn::AllophoneStateFsaExporter>    allophoneStateFsaExporter_;
+    std::shared_ptr<Core::StringHashMap<std::string>> segmentToOrthMap_;
+
+};
+
+#endif // _PYTHON_ALLOPHONESTATEFSABUILDER_HH

src/Python/Configuration.cc

@@ -0,0 +1,8 @@
+#include "Configuration.hh"
+
+#include <Core/Configuration.hh>
+
+PyConfiguration::PyConfiguration()
+        : Core::Configuration() {
+    setSelection("lib-rasr");
+}

src/Python/Configuration.hh

@@ -0,0 +1,11 @@
+#ifndef _PYTHON_CONFIGURATION_HH
+#define _PYTHON_CONFIGURATION_HH
+
+#include <Core/Configuration.hh>
+
+class PyConfiguration : public Core::Configuration {
+public:
+    PyConfiguration();
+};
+
+#endif // _PYTHON_CONFIGURATION_HH

src/Python/Makefile

@@ -6,14 +6,21 @@ include $(TOPDIR)/Makefile.cfg
 
 CCFLAGS		+= -fexceptions
 
+ifdef MODULE_TENSORFLOW
+CXXFLAGS += $(TF_CXXFLAGS)
+LDFLAGS  += $(TF_LDFLAGS)
+endif
+
 # -----------------------------------------------------------------------------
 
 SUBDIRS 	=
 TARGETS		= libSprintPython.$(a) check$(exe)
 
 LIBPYTHON_O    = \
-		$(OBJDIR)/Init.o		\
-		$(OBJDIR)/Numpy.o		\
+		$(OBJDIR)/AllophoneStateFsaBuilder.o \
+		$(OBJDIR)/Configuration.o            \
+		$(OBJDIR)/Init.o                     \
+		$(OBJDIR)/Numpy.o                    \
 		$(OBJDIR)/Utilities.o
 
 CHECK_O		= $(OBJDIR)/check.o 			\