SINGA-346 Update cudnn from V5 to V7

Export numpy path by cmake; no need to export it into a env var now.

replace PYTHON3 with USE_PYTHON3

one bug left due to the unicode and byte string problem from py2 to py3.
which could be resolved by defining SWIG_PYTHON_2_UNICODE in *.i files.
It converts byte string in python2 into unicode and then to c++ std::string.

Author: Wang Wei

CMakeLists.txt

@@ -52,19 +52,21 @@ SET(SINGA_INCLUDE_DIR
     "${CMAKE_SOURCE_DIR}/include;${PROJECT_BINARY_DIR}")
 INCLUDE_DIRECTORIES(${SINGA_INCLUDE_DIR})
 
-#OPTION(USE_CBLAS "Use CBlas libs" ON)
+
 OPTION(USE_CUDA "Use Cuda libs" OFF)
+OPTION(ENABLE_TEST "Enable unit test" OFF)
+OPTION(USE_PYTHON "Generate py wrappers" ON)
+OPTION(USE_PYTHON3 "Python 3x" OFF)
+
 OPTION(USE_CUDNN "Use Cudnn libs" ON)
 OPTION(USE_OPENCV "Use opencv" OFF)
 OPTION(USE_LMDB "Use LMDB libs" OFF)
-OPTION(USE_PYTHON "Generate py wrappers" ON)
 OPTION(USE_JAVA "Generate java wrappers" OFF)
 OPTION(USE_OPENCL "Use OpenCL" OFF)
 OPTION(ENABLE_DIST "Enable distributed training" OFF)
-OPTION(ENABLE_TEST "Enable unit test" OFF)
 OPTION(DISABLE_WARNINGS "Disable warnings under windows" ON)
 OPTION(USE_MODULES "Compile dependent libs as submodules together with singa" OFF)
-#OPTION(USE_SHARED_LIBS "Use shared library" OFF)
+
 
 # TODO: remove all USE_CBLAS in codes
 SET(USE_CBLAS ON)
@@ -192,7 +194,7 @@ IF(PACKAGE)
 		SET(CORE_DEPENDENCIES "libgoogle-glog-dev, libprotobuf-dev, libopenblas-dev, libstdc++6, libc6")
 	ENDIF()
 
-	IF(PYTHON3)
+	IF(USE_PYTHON3)
 		SET(PYTHON_DEPENDENCIES "${CORE_DEPENDENCIES}, python3, python3-dev, python3-pip, python3-numpy, python3-pillow, python3-matplotlib")
 	ELSE()
 		SET(PYTHON_DEPENDENCIES "${CORE_DEPENDENCIES}, python-dev, libpython2.7, python-pip, python-numpy, python-pillow")
@@ -210,7 +212,7 @@ IF(PACKAGE)
 		SET(CPACK_DEBIAN_PACKAGE_DEPENDS ${PYTHON_DEPENDENCIES})
 		SET(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${PROJECT_SOURCE_DIR}/tool/debian/postinst" )
 		SET(CPACK_DEBIAN_PACKAGE_PREDEPENDS "ca-certificates")
-		IF(PYTHON3)
+		IF(USE_PYTHON3)
 			SET(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${PROJECT_SOURCE_DIR}/tool/debian/postinst" )
 			IF (USE_CUDA)
 				SET(CPACK_DEBIAN_PACKAGE_NAME "python3-singa-cuda")

cmake/Dependencies.cmake

@@ -124,14 +124,14 @@ ENDIF()
 #MESSAGE(STATUS "link lib : " ${SINGA_LINKER_LIBS})
 
 IF(USE_PYTHON)
-    IF(PYTHON3)
-        set(Python_ADDITIONAL_VERSIONS 3.6)
-        FIND_PACKAGE(PythonLibs 3 REQUIRED)
+    IF(USE_PYTHON3)
+        set(Python_ADDITIONAL_VERSIONS 3.6 3.5 3.4)        
         FIND_PACKAGE(PythonInterp 3 REQUIRED)
-	    FIND_PACKAGE(SWIG 3.0.10 REQUIRED)
-    ELSE()
-        FIND_PACKAGE(PythonLibs 2.7 REQUIRED)
+        FIND_PACKAGE(PythonLibs 3 REQUIRED)
+	    FIND_PACKAGE(SWIG 3.0.8 REQUIRED)
+    ELSE()        
         FIND_PACKAGE(PythonInterp 2.7 REQUIRED)
+        FIND_PACKAGE(PythonLibs 2.7 REQUIRED)
 	    FIND_PACKAGE(SWIG 3.0.8 REQUIRED)
     ENDIF()
 ENDIF()

doc/en/docs/installation.md

@@ -109,7 +109,7 @@ The following libraries are optional
     * `USE_CUDA=ON`, used if CUDA and cuDNN is available
     * `USE_PYTHON=ON`, used for compiling PySINGA
     * `USE_OPENCL=ON`, used for compiling with OpenCL support
-    * `PYTHON3=ON`, used for compiling with Python 3 support. (The default is Python 2)
+    * `USE_PYTHON3=ON`, used for compiling with Python 3 support. (The default is Python 2)
     * `PACKAGE=ON`, used for building the Debian package
 
 3. compile the code, e.g., `make`

python/CMakeLists.txt

@@ -94,7 +94,7 @@ file(MAKE_DIRECTORY ${CMAKE_BINARY_DIR}/python/singa/proto)
 file(MAKE_DIRECTORY ${CMAKE_BINARY_DIR}/python/rafiki)
 file(MAKE_DIRECTORY ${CMAKE_BINARY_DIR}/src/api)
 
-IF(PYTHON3)
+IF(USE_PYTHON3)
 	SET(SWIG_PYTHON3 "-py3")
 ELSE()
 	SET(SWIG_PYTHON3 "")
@@ -113,6 +113,12 @@ file(GLOB_RECURSE python_source_files RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} *.py)
 create_symlinks(${python_source_files})
 
 
+execute_process(
+    COMMAND ${PYTHON_EXECUTABLE} -c "from __future__ import print_function; import numpy; print(numpy.get_include())"
+    OUTPUT_VARIABLE NUMPY_INCLUDE_DIR)
+
+#message(status "numpy path ${NUMPY_INCLUDE_DIR}")
+
 IF(USE_CUDA)
 # remain this custom command to avoid cuda objs can't find
 ADD_CUSTOM_COMMAND(
@@ -123,7 +129,7 @@ ENDIF(USE_CUDA)
 
 ADD_LIBRARY(_singa_wrap SHARED $<TARGET_OBJECTS:singa_objects>  ${python_srcs} ${proto_pys} ${global_cuda_objs})
 TARGET_LINK_LIBRARIES(_singa_wrap ${SINGA_LINKER_LIBS} ${PYTHON_LIBRARIES})
-TARGET_INCLUDE_DIRECTORIES(_singa_wrap PRIVATE ${PYTHON_INCLUDE_DIRS})
+TARGET_INCLUDE_DIRECTORIES(_singa_wrap PRIVATE ${PYTHON_INCLUDE_DIRS} ${NUMPY_INCLUDE_DIR})
 SET_TARGET_PROPERTIES(_singa_wrap
 PROPERTIES PREFIX ""
 LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/python/singa

python/singa/layer.py

@@ -102,7 +102,7 @@ def __init__(self, name, conf=None, **kwargs):
             self.param_specs = []
         else:
             self.conf = conf
-            self.name = str(conf.name)
+            self.name = conf.name
             self.caffe_layer()
             self.param_specs = []
 
@@ -154,7 +154,7 @@ def caffe_layer(self):
         if self.conf.type == 'InnerProduct' or self.conf.type == 14:
             self.layer = _create_layer(engine, 'Dense')
         else:
-            self.layer = _create_layer(engine, str(self.conf.type))
+            self.layer = _create_layer(engine, self.conf.type)
 
     def get_output_sample_shape(self):
         '''Called after setup to get the shape of the output sample(s).

python/singa/net.py

@@ -18,6 +18,7 @@
 Nerual net class for constructing the nets using layers and providing access
 functions for net info, e.g., parameters.
 
+
 Example usages::
 
     from singa import net as ffnet
@@ -175,11 +176,11 @@ def param_names(self):
         return [spec.name for spec in self.param_specs()]
 
     def train(self, x, y):
-	'''Run BP for one iteration.
-	This method is deprecated. It is only kept for backward compatibility.
-	The name of this method is confusing since it does not update parameters.
-	Please use backprob() instead.
-	The back progagation algorithm computes gradients but it does not train.
+        '''Run BP for one iteration.
+        This method is deprecated. It is only kept for backward compatibility.
+        The name of this method is confusing since it does not update parameters.
+        Please use backprob() instead.
+        The back progagation algorithm computes gradients but it does not train.
         '''
         return backprob(x, y)
 
@@ -295,7 +296,8 @@ def forward(self, flag, x, output=[], freeze=None):
                 dictionary: layer name -> output tensor(s)
         '''
         if self.ordered_layers is None:
-            self.ordered_layers = self.topo_sort(self.layers, self.src_of_layer)
+            self.ordered_layers = self.topo_sort(
+                self.layers, self.src_of_layer)
         if type(x) is dict:
             input_of_layer = x
         else:
@@ -326,7 +328,7 @@ def forward(self, flag, x, output=[], freeze=None):
                 outs = output_of_layer[src.name]
                 if type(outs) == list:
                     assert len(outs) > 0, \
-                            'the output from layer %s is empty' % src.name
+                        'the output from layer %s is empty' % src.name
                     inputs.append(outs[0])
                     outs.pop(0)
                     if len(outs) == 0:
@@ -406,7 +408,7 @@ def backward(self, dy, output=[], freeze=None):
                 outputs = output_of_layer[dst.name]
                 if type(outputs) == list:
                     assert len(outputs) > 0, \
-                            'the gradient from layer %s is empty' % dst.name
+                        'the gradient from layer %s is empty' % dst.name
                     inputs.append(outputs[0])
                     outputs.pop(0)
                 else:
@@ -418,7 +420,7 @@ def backward(self, dy, output=[], freeze=None):
             outs, pgrads = cur.backward(kTrain, inputs)
             if verbose:
                 disp_src = '+'.join(
-                        [dst.name for dst in self.dst_of_layer[cur.name]])
+                    [dst.name for dst in self.dst_of_layer[cur.name]])
                 disp_src += '-->' + cur.name
                 if type(outs) is list:
                     print('%s: %s' % (disp_src,
@@ -477,7 +479,7 @@ def get_name(name):
             if version < 1101:
                 idx = name.rfind('/')
                 assert idx > 0, '/ must be in the parameter name'
-                name = name[:idx] + '_' + name[idx+1:]
+                name = name[:idx] + '_' + name[idx + 1:]
             return name
 
         if use_pickle:
@@ -515,5 +517,5 @@ def get_name(name):
             except AssertionError as err:
                 print('Error from copying values for param: %s' % name)
                 print(('shape of param vs checkpoint',
-                      val.shape, params[name].shape))
+                       val.shape, params[name].shape))
                 raise err

python/singa/optimizer.py

@@ -206,7 +206,7 @@ def __init__(self, lr=None, momentum=None, weight_decay=None,
         if self.momentum is not None:
             conf.momentum = self.momentum
         conf.type = 'sgd'
-        self.opt = singa.CreateOptimizer('SGD'.encode())
+        self.opt = singa.CreateOptimizer('SGD')
         self.opt.Setup(conf.SerializeToString())
 
     def apply_with_lr(self, epoch, lr, grad, value, name, step=-1):
@@ -216,7 +216,7 @@ def apply_with_lr(self, epoch, lr, grad, value, name, step=-1):
             epoch, value, grad, name, step)
         if name is not None and name in self.learning_rate_multiplier:
             lr = lr * self.learning_rate_multiplier[name]
-        self.opt.Apply(epoch, lr, name.encode(), grad.data,
+        self.opt.Apply(epoch, lr, name, grad.data,
                        value.data)
         return value
 
@@ -235,7 +235,7 @@ def __init__(self, lr=None, momentum=0.9, weight_decay=None,
         if self.momentum is not None:
             conf.momentum = momentum
         conf.type = 'nesterov'
-        self.opt = singa.CreateOptimizer('Nesterov'.encode())
+        self.opt = singa.CreateOptimizer('Nesterov')
         self.opt.Setup(conf.SerializeToString())
 
     def apply_with_lr(self, epoch, lr, grad, value, name, step=-1):
@@ -246,7 +246,7 @@ def apply_with_lr(self, epoch, lr, grad, value, name, step=-1):
             epoch, value, grad, name, step)
         if name is not None and name in self.learning_rate_multiplier:
             lr = lr * self.learning_rate_multiplier[name]
-        self.opt.Apply(epoch, lr, name.encode(), grad.data,
+        self.opt.Apply(epoch, lr, name, grad.data,
                        value.data)
         return value
 
@@ -268,7 +268,7 @@ def __init__(self, rho=0.9, epsilon=1e-8, lr=None, weight_decay=None,
         conf = model_pb2.OptimizerConf()
         conf.rho = rho
         conf.delta = epsilon
-        self.opt = singa.CreateOptimizer('RMSProp'.encode())
+        self.opt = singa.CreateOptimizer('RMSProp')
         self.opt.Setup(conf.SerializeToString())
 
     def apply_with_lr(self, epoch, lr, grad, value, name, step=-1):
@@ -279,7 +279,7 @@ def apply_with_lr(self, epoch, lr, grad, value, name, step=-1):
             epoch, value, grad, name, step)
         if name is not None and name in self.learning_rate_multiplier:
             lr = lr * self.learning_rate_multiplier[name]
-        self.opt.Apply(step, lr,  name.encode(), grad.data,
+        self.opt.Apply(step, lr,  name, grad.data,
                        value.data)
         return value
 
@@ -300,7 +300,7 @@ def __init__(self, epsilon=1e-8, lr=None, weight_decay=None, lr_gen=None,
         conf = model_pb2.OptimizerConf()
         conf.delta = epsilon
         conf.type = 'adagrad'
-        self.opt = singa.CreateOptimizer('AdaGrad'.encode())
+        self.opt = singa.CreateOptimizer('AdaGrad')
         self.opt.Setup(conf.SerializeToString())
 
     def apply_with_lr(self, epoch, lr, grad, value, name, step=-1):
@@ -311,7 +311,7 @@ def apply_with_lr(self, epoch, lr, grad, value, name, step=-1):
             epoch, value, grad, name, step)
         if name is not None and name in self.learning_rate_multiplier:
             lr = lr * self.learning_rate_multiplier[name]
-        self.opt.Apply(epoch, lr,  name.encode(), grad.data,
+        self.opt.Apply(epoch, lr,  name, grad.data,
                        value.data)
         return value
 

tool/conda/README.md

@@ -24,7 +24,7 @@ this folder should include a subfolder `include/cudnn.h` for the header file, an
 
 After exporting the environment variables, execute the following command to compile Singa and package it
 
-    conda-build .
+    conda-build .  --python 3.6  (or 2.7)
 
 You will see the package path from the screen output.
 

tool/conda/build.sh

@@ -16,26 +16,28 @@
 #
 
 # to compile swig api files which depdend on numpy.i
-export CPLUS_INCLUDE_PATH=`python -c "from __future__ import print_function; import numpy; print(numpy.get_include())"`:$CPLUS_INCLUDE_PATH
+# export CPLUS_INCLUDE_PATH=`python -c "from __future__ import print_function; import numpy; print(numpy.get_include())"`:$CPLUS_INCLUDE_PATH
 
 # to let cmake use the dependent libs installed by conda, including python
 export CMAKE_PREFIX_PATH=$PREFIX:$CMAKE_PREFIX_PATH
 export CMAKE_INCLUDE_PATH=$PREFIX/include:$CMAKE_INCLUDE_PATH
 export CMAKE_LIBRARY_PATH=$PREFIX/lib:$CMAKE_LIBRARY_PATH
 
-mkdir build
-cd build
+
 USE_CUDA=OFF
 if [ -z ${CUDNN_PATH+x} ]; then
 	USE_CUDA=ON
 	cp $CUDNN_PATH/include $PREFIX/include 
 	cp -P $CUDNN_PATH/lib64/libcudnn.so* $PREFIX/lib/
 fi
 
-PYTHON3=OFF
+USE_PYTHON3=OFF
 # PY3K is set by conda
-if  [ "$PY3K" == "1" ]; then PYTHON3=ON; fi
-echo "PY3K = $PY3K"
-cmake -DCMAKE_INSTALL_PREFIX=$PREFIX -DUSE_CUDA=$USE_CUDA -DPYTHON3=$PYTHON3 ..
+if  [ "$PY3K" == "1" ]; then USE_PYTHON3=ON; fi
+
+
+mkdir build
+cd build
+cmake -DCMAKE_INSTALL_PREFIX=$PREFIX -DUSE_CUDA=$USE_CUDA -DUSE_PYTHON3=$USE_PYTHON3 ..
 make
 make install

tool/docker/README.md

@@ -2,36 +2,23 @@
 
 ## Availabe images
 
-TO BE UPDATED.
 
-| Tag | OS version | devel/runtime | Device|CUDA/CUDNN|
-|:----|:-----------|:--------------|:------|:---------|
-|runtime| Ubuntu16.04|runtime|CPU|-|
-|runtime| Ubuntu16.04|runtime|CPU|-|
-|runtime-cuda| Ubuntu16.04|runtime|GPU|CUDA8.0+CUDNN5|
-|devel| Ubuntu16.04|devel|CPU|-|
-|devel-cuda| Ubuntu16.04|devel|GPU|CUDA8.0+CUDNN5|
+| Tag | OS version | devel/runtime | Device|CUDA/CUDNN|Python|
+|:----|:-----------|:--------------|:------|:---------|:-----|
+|runtime| Ubuntu16.04|runtime|CPU|-|3.6|
+|conda-cuda9.0| Ubuntu16.04|devel|GPU|CUDA9.0+CUDNN7.1.2|3.6|
+|cuda9.0-py2| Ubuntu16.04|devel|GPU|CUDA9.0+CUDNN7.1.2|2.7|
+|cuda9.0-py3| Ubuntu16.04|devel|GPU|CUDA9.0+CUDNN7.1.2|3.6|
+
+runtime and conda-xxx image has installed miniconda3;
+cudaxxx images have installed all depedent libs using apt-get.
 
 ## Usage
 
     docker pull nusdbsystem/singa:<Tag>
     docker run -it nusdbsystem/singa:<Tag> /bin/bash
+    nvidia-docker run -it nusdbsystem/singa:<Tag> /bin/bash
 
-* For the *devel* images, the container has a `incubator-singa` folder in the root directory,
-which has the latest SINGA code. The code has been compiled into `incubator-singa/build` directory and PySINGA has been installed.
-* For the *runtime* images, the container has only installed the PySINGA.
-
-## Tag naming style
 
-    singa:devel|runtime[-OS][-CUDA|OPENCL][-CUDNN]
 
-* devel: development images with all dependent libs' header files installed and SINGA's source code;
-* runtime: the minimal images which can run SINGA programs.
-* OS: ubuntu, ubuntu14.04, centos, centos6
-* CUDA: cuda, cuda8.0, cuda7.0
-* CUDNN: cudnn, cudnn5, cudnn4
-* OPENCL: opencl, opencl1.2
 
-By default, if the version is not included in the tag, the latest stable version is used.
-The default OS is ubuntu. The version is the latest stable version (e.g., 16.04 for now).
-For -cuda version, the **cudnn** is included by default. Their versions are also the latest stable version, i.e., cuda-8.0 and cudnn-5 for now.