Merge Intel Caffe release_1.0.7.

CMakeLists.txt

@@ -47,6 +47,7 @@ caffe_option(PERFORMANCE_MONITORING "Build Caffe with PERFORMANCE_MONITORING " O
 
 # ---[ Dependencies
 include(cmake/MKLDNN.cmake)
+include(cmake/DLCP.cmake)
 include(cmake/Dependencies.cmake)
 
 # ---[ Flags

Makefile

@@ -71,11 +71,13 @@ endif
 
 	RETURN_STRING=$(shell ./external/mlsl/prepare_mlsl.sh)
 	MLSL_ROOT=$(firstword $(RETURN_STRING))
-	MLSL_LDFLAGS=$(lastword $(RETURN_STRING))	
+	MLSL_LDFLAGS=$(lastword $(RETURN_STRING))
 	COMMON_FLAGS += -DUSE_MLSL=1
 	LIBRARIES += mlsl
 	INCLUDE_DIRS += $(MLSL_ROOT)/intel64/include
 	LIBRARY_DIRS += $(MLSL_ROOT)/intel64/lib
+	IGNORE := $(shell bash -c "source $(MLSL_ROOT)/intel64/bin/mlslvars.sh; env | sed 's/=/:=/' | sed 's/^/export /' > make_mlsl_env")
+	include make_mlsl_env
 
 ifeq ($(CAFFE_PER_LAYER_TIMINGS), 1)
 	COMMON_FLAGS += -DCAFFE_PER_LAYER_TIMINGS
@@ -249,7 +251,13 @@ ifneq ($(CPU_ONLY), 1)
 	LIBRARIES := cudart cublas curand
 endif
 
-LIBRARIES += glog gflags protobuf m hdf5_hl hdf5
+LIBRARIES += glog gflags protobuf m
+ifeq ($(shell  find /usr/lib* /usr/local/lib* -name "libhdf5_serial.so" 2> /dev/null | grep serial >/dev/null  && echo serial) , serial)
+	LIBRARIES += hdf5_serial_hl hdf5_serial
+else
+	LIBRARIES += hdf5_hl hdf5
+endif
+
 BOOST_LIBRARIES += boost_system boost_filesystem boost_regex
 
 # handle IO dependencies
@@ -479,6 +487,11 @@ ifeq ($(WITH_PYTHON_LAYER), 1)
 	LIBRARIES += $(PYTHON_LIBRARIES)
 endif
 
+# Remove batch norm feature support
+ifeq ($(DISABLE_BN_FOLDING), 1)
+	COMMON_FLAGS += -DDISABLE_BN_FOLDING
+endif
+
 # Performance monitoring
 ifeq ($(PERFORMANCE_MONITORING), 1)
 	CXXFLAGS += -DPERFORMANCE_MONITORING
@@ -489,6 +502,8 @@ ifeq ($(USE_MKLDNN_AS_DEFAULT_ENGINE), 1)
 	CXXFLAGS += -DUSE_MKLDNN_AS_DEFAULT_ENGINE
 endif
 
+include Makefile.dlcp
+
 # BOOST configuration
 # detect support for custom boost version
 BOOST_LDFLAGS += $(foreach boost_lib,$(BOOST_LIBRARIES),-l$(boost_lib))
@@ -602,7 +617,7 @@ endif
 ##############################
 .PHONY: all lib test clean docs linecount lint lintclean tools examples $(DIST_ALIASES) \
 	py mat py$(PROJECT) mat$(PROJECT) proto runtest \
-	superclean supercleanlist supercleanfiles warn everything mkldnn mkldnn_clean
+	superclean supercleanlist supercleanfiles warn everything mkldnn mkldnn_clean dlcp dlcp_clean
 
 .DEFAULT_GOAL := all
 
@@ -757,14 +772,14 @@ $(ALL_BUILD_DIRS): | $(BUILD_DIR_LINK)
 
 $(DYNAMIC_NAME): $(OBJS) | $(LIB_BUILD_DIR)
 	@ echo LD -o $@
-	$(Q)$(CXX) -shared -o $@ $(OBJS) $(VERSIONFLAGS) $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_SHARED_HARDENING_FLAGS) $(LDFLAGS)
+	$(Q)$(CXX) -shared -o $@ $(OBJS) $(VERSIONFLAGS) $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_SHARED_HARDENING_FLAGS) $(LDFLAGS)
 	@ cd $(BUILD_DIR)/lib; rm -f $(DYNAMIC_NAME_SHORT);   ln -s $(DYNAMIC_VERSIONED_NAME_SHORT) $(DYNAMIC_NAME_SHORT)
 
 $(STATIC_NAME): $(OBJS) | $(LIB_BUILD_DIR)
 	@ echo AR -o $@
 	$(Q)ar rcs $@ $(OBJS)
 
-$(BUILD_DIR)/%.o: %.cpp | mkldnn $(ALL_BUILD_DIRS)
+$(BUILD_DIR)/%.o: %.cpp | mkldnn dlcp $(ALL_BUILD_DIRS)
 	@ echo CXX $<
 	$(Q)$(CXX) $< $(CXX_HARDENING_FLAGS) $(CXXFLAGS) -c -o $@ 2> $@.$(WARNS_EXT) \
 		|| (cat $@.$(WARNS_EXT); exit 1)
@@ -789,7 +804,7 @@ $(TEST_ALL_BIN): $(TEST_MAIN_SRC) $(TEST_OBJS) $(GTEST_OBJS) \
 		| $(DYNAMIC_NAME) $(TEST_BIN_DIR)
 	@ echo CXX/LD -o $@ $<
 	$(Q)$(CXX) -std=c++11 $(TEST_MAIN_SRC) $(TEST_OBJS) $(GTEST_OBJS) \
-		-o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) $(LDFLAGS) -l$(LIBRARY_NAME) -Wl,-rpath,$(ORIGIN)/../lib
+		-o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) $(LDFLAGS) -l$(LIBRARY_NAME) -Wl,-rpath,$(ORIGIN)/../lib
 
 $(TEST_CU_BINS): $(TEST_BIN_DIR)/%.testbin: $(TEST_CU_BUILD_DIR)/%.o \
 	$(GTEST_OBJS) | $(DYNAMIC_NAME) $(TEST_BIN_DIR)
@@ -801,7 +816,7 @@ $(TEST_CXX_BINS): $(TEST_BIN_DIR)/%.testbin: $(TEST_CXX_BUILD_DIR)/%.o \
 	$(GTEST_OBJS) | $(DYNAMIC_NAME) $(TEST_BIN_DIR)
 	@ echo LD $<
 	$(Q)$(CXX) -std=c++11 $(TEST_MAIN_SRC) $< $(GTEST_OBJS) \
-		-o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) $(LDFLAGS) -l$(LIBRARY_NAME) -Wl,-rpath,$(ORIGIN)/../lib
+		-o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) $(LDFLAGS) -l$(LIBRARY_NAME) -Wl,-rpath,$(ORIGIN)/../lib
 
 # Target for extension-less symlinks to tool binaries with extension '*.bin'.
 $(TOOL_BUILD_DIR)/%: $(TOOL_BUILD_DIR)/%.bin | $(TOOL_BUILD_DIR)
@@ -810,12 +825,12 @@ $(TOOL_BUILD_DIR)/%: $(TOOL_BUILD_DIR)/%.bin | $(TOOL_BUILD_DIR)
 
 $(TOOL_BINS): %.bin : %.o | $(DYNAMIC_NAME)
 	@ echo CXX/LD -o $@
-	$(Q)$(CXX) $< -o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) -l$(LIBRARY_NAME) $(LDFLAGS) \
+	$(Q)$(CXX) $< -o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) -l$(LIBRARY_NAME) $(LDFLAGS) \
 		-Wl,-rpath,$(ORIGIN)/../lib
 
 $(EXAMPLE_BINS): %.bin : %.o | $(DYNAMIC_NAME)
 	@ echo CXX/LD -o $@
-	$(Q)$(CXX) $< -o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) -l$(LIBRARY_NAME) $(LDFLAGS) \
+	$(Q)$(CXX) $< -o $@ $(BOOST_LDFLAGS) $(LINKFLAGS) $(MKL_LDFLAGS) $(MKLDNN_LDFLAGS) $(DLCP_LDFLAGS) $(CXX_HARDENING_FLAGS) $(LINKER_EXEC_HARDENING_FLAGS) -l$(LIBRARY_NAME) $(LDFLAGS) \
 		-Wl,-rpath,$(ORIGIN)/../../lib
 
 proto: $(PROTO_GEN_CC) $(PROTO_GEN_HEADER)
@@ -833,7 +848,7 @@ $(PY_PROTO_BUILD_DIR)/%_pb2.py : $(PROTO_SRC_DIR)/%.proto \
 $(PY_PROTO_INIT): | $(PY_PROTO_BUILD_DIR)
 	touch $(PY_PROTO_INIT)
 
-clean: mkldnn_clean
+clean: mkldnn_clean dlcp_clean
 	@echo "Will download the new version of MKL2017 and MLSL when clean and prepare the environment."
 	@- $(RM) -rf $(ALL_BUILD_DIRS)
 	@- $(RM) -rf $(OTHER_BUILD_DIR)

Makefile.config.example

@@ -76,6 +76,8 @@ USE_MKLDNN_AS_DEFAULT_ENGINE := 1
 # already. 
 # BOOST_ROOT := 
 
+# Use remove batch norm optimization to boost inferrence
+DISABLE_BN_FOLDING := 0
 # Intel(r) Machine Learning Scaling Library (uncomment to build
 # with MLSL for multi-node training)
 # USE_MLSL :=1
@@ -115,6 +117,8 @@ BLAS := mkl
 # MATLAB_DIR := /usr/local
 # MATLAB_DIR := /Applications/MATLAB_R2012b.app
 
+SERIAL_HDF5_INCLUDE := /usr/include/hdf5/serial/
+
 # NOTE: this is required only if you will compile the python interface.
 # We need to be able to find Python.h and numpy/arrayobject.h.
 PYTHON_INCLUDE := /usr/include/python2.7 \
@@ -143,7 +147,7 @@ PYTHON_LIB := /usr/lib
 # WITH_PYTHON_LAYER := 1
 
 # Whatever else you find you need goes here.
-INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include
+INCLUDE_DIRS := $(PYTHON_INCLUDE) ${SERIAL_HDF5_INCLUDE} /usr/local/include
 LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib
 
 # If Homebrew is installed at a non standard location (for example your home directory) and you use it for general dependencies

Makefile.dlcp

@@ -0,0 +1,58 @@
+CAFFE_ROOTDIR := $(shell pwd)
+DLCP_IDEEPDIR := $(CAFFE_ROOTDIR)/external/ideep
+DLCP_ROOTDIR := $(DLCP_IDEEPDIR)/ideep/dlcp
+DLCP_INCLDIR := $(DLCP_ROOTDIR)/include
+DLCP_LIBDIR := $(DLCP_ROOTDIR)/lib
+IDEEP_GITHUB := https://github.com/intel/ideep.git
+DLCP_CXX := $(CXX)
+
+ifeq ("$(wildcard $(DLCP_INCLDIR)/dl_compression.h)", "")
+dlcp_download:
+	rm -rf $(DLCP_IDEEPDIR)
+	git clone $(IDEEP_GITHUB) -b dlcp $(DLCP_IDEEPDIR)
+
+dlcp_build: dlcp_download
+	make -C $(DLCP_ROOTDIR) -j$(shell cat /proc/cpuinfo |grep 'processor'|wc -l)
+else
+dlcp_download:
+dlcp_build:
+	make -C $(DLCP_ROOTDIR) -j$(shell cat /proc/cpuinfo |grep 'processor'|wc -l)
+endif
+
+dlcp_clean:
+	@rm -rf $(DLCP_IDEEPDIR)
+
+dlcproot_set:
+	$(eval COMMON_FLAGS += -DENABLE_WEIGHT_GRAD_COMPRESSION) 		\
+	$(eval INCLUDE_DIRS += $(DLCP_INCLDIR)) 		\
+	$(eval DLCP_LDFLAGS += -ldlcomp -L$(DLCPROOT)/lib) \
+	$(eval DLCP_LDFLAGS += -Wl,-rpath,$(DLCPROOT)/lib) 
+
+
+dlcproot_notset: dlcp_build
+	$(eval COMMON_FLAGS += -DENABLE_WEIGHT_GRAD_COMPRESSION) 		\
+	$(eval INCLUDE_DIRS += $(DLCP_INCLDIR)) \
+	$(eval DLCP_LDFLAGS += -ldlcomp) \
+	$(eval DLCP_LDFLAGS += -L$(DLCP_LIBDIR) -Wl,-rpath,${DLCP_LIBDIR})
+
+ifeq (,$(findstring icpc, $(DLCP_CXX)))
+ifneq (,$(findstring icpc,$(shell command -v icpc)))
+DLCP_CXX := icpc
+endif
+endif
+
+ifeq ($(USE_MLSL), 1)
+ifeq ($(origin DLCPROOT), environment)
+dlcp: dlcproot_set
+else
+ifneq (,$(findstring icpc,$(DLCP_CXX)))
+dlcp: dlcproot_notset
+else
+dlcp:
+	@echo "weight grad compression disabled because intel compiler not found"
+endif
+endif
+else
+dlcp:
+	@echo "disabling weight grad compression if USE_MLSL is not 1"
+endif

cmake/DLCP.cmake

@@ -0,0 +1,29 @@
+function(Download_DLCP)
+  find_program(HAS_ICPC NAMES icpc DOC "Intel Compiler")
+  if (${CMAKE_CXX_COMPILER_ID} STREQUAL "Intel")
+    set(DLCP_CXX "${CMAKE_CXX_COMPILER}" PARENT_SCOPE)
+  elseif(HAS_ICPC)
+    set(DLCP_CXX "icpc" PARENT_SCOPE)
+  else()
+    message(WARNING "weight grad compression is disabled because intel compiler is not found")
+    return()  
+  endif()  
+
+  set(EXTERNAL_DIR ${CMAKE_SOURCE_DIR}/external)
+  set(DLCP_IDEEPDIR ${EXTERNAL_DIR}/ideep)
+  set(DLCP_ROOTDIR ${DLCP_IDEEPDIR}/ideep/dlcp)
+  set(DLCP_INCLDIR "${DLCP_ROOTDIR}/include" PARENT_SCOPE)
+  set(DLCP_LIBDIR ${DLCP_ROOTDIR}/lib PARENT_SCOPE)
+
+  # Download dl compression lib source code if it doesn't exist 
+  if (NOT EXISTS ${DLCP_INCLDIR}/dl_compression.h)
+    execute_process(COMMAND rm -rf ${DLCP_IDEEPDIR})
+    execute_process(COMMAND git clone https://github.com/intel/ideep.git -b dlcp ${DLCP_IDEEPDIR})
+  endif()
+
+  add_custom_target(DLCP_Build ALL
+                    COMMAND export DLCP_CXX=${DLCP_CXX}
+                    COMMAND make -j
+                    WORKING_DIRECTORY ${DLCP_ROOTDIR})
+
+endfunction(Download_DLCP)

cmake/Dependencies.cmake

@@ -132,6 +132,33 @@ if(USE_MLSL)
   endif()
 endif()
 
+# ---| DLCP
+if(USE_MLSL)
+  if(DEFINED ENV{DLCPROOT})
+    set(DLCPROOT_DIR $ENV{DLCPROOT})
+    if(NOT ${DLCPROOT_DIR} STREQUAL "")
+      set(DLCPROOT_INCLUDE_DIR "${DLCPROOT_DIR}/include/")
+      set(DLCPROOT_LIB_DIR "${DLCPROOT_DIR}/lib/")
+      if(EXISTS ${DLCPROOT_INCLUDE_DIR}/dl_compression.h AND EXISTS ${DLCPROOT_LIB_DIR}/libdlcomp.so)
+        message(STATUS "Found DLCP: ${DLCPROOT_DIR}")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DENABLE_WEIGHT_GRAD_COMPRESSION")
+        list(APPEND Caffe_LINKER_LIBS "${DLCPROOT_LIB_DIR}/libdlcomp.so")
+        include_directories(SYSTEM ${DLCP_INCLUDE_DIR})
+      else()
+        message(STATUS "DLCP not found. DLCP_INCLUDE_DIR = ${DLCPROOT_INCLUDE_DIR} DLCP_LIB_DIR = ${DLCPROOT_LIB_DIR}")
+        message(WARNING "weight grad compression is disabled.")
+      endif()
+    endif()
+  else()
+    Download_DLCP()
+    if(DLCP_CXX)
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DENABLE_WEIGHT_GRAD_COMPRESSION")
+      list(APPEND Caffe_LINKER_LIBS "${DLCP_LIBDIR}/libdlcomp.so")
+      include_directories(SYSTEM ${DLCP_INCLDIR})
+    endif()
+  endif()
+endif()
+
 # ---[ BLAS
 set(MKL_EXTERNAL "0")
 if(NOT APPLE)

cmake/Misc.cmake

@@ -6,6 +6,10 @@ if(DEFINED CMAKE_BUILD_TYPE)
   set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS ${CMAKE_CONFIGURATION_TYPES})
 endif()
 
+if(DISABLE_BN_FOLDING)
+	message(STATUS "Bn folding is disabled!")
+	add_definitions("-DDISABLE_BN_FOLDING")
+endif()
 # --[ If user doesn't specify build type then assume release
 if("${CMAKE_BUILD_TYPE}" STREQUAL "")
   set(CMAKE_BUILD_TYPE Release)

docker/standalone/cpu-ubuntu/Dockerfile

@@ -23,6 +23,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
         python-dev \
         python-numpy \
         python-pip \
+        python-setuptools \
         python-scipy && \
     rm -rf /var/lib/apt/lists/*
 
@@ -32,6 +33,8 @@ WORKDIR $CAFFE_ROOT
 # FIXME: clone a specific git tag and use ARG instead of ENV once DockerHub supports this.
 ENV CLONE_TAG=master
 
+RUN pip install --upgrade pip
+
 RUN git clone -b ${CLONE_TAG} --depth 1 https://github.com/intel/caffe.git . && \
     for req in $(cat python/requirements.txt) pydot; do pip install $req; done && \
     mkdir build && cd build && \

examples/LRCN_activity_recognition/README.txt

@@ -5,3 +5,31 @@ Chicago
 
 Please see http://www.eecs.berkeley.edu/~lisa_anne/LRCN_video for detailed instructions on how to reimplement experiments and download pre-trained models.
 
+# Support Flow LRCN From Quadra_L
+- [Quadra_L](http://www.swcontest.net/index.php?a=show&m=news&aid=177)
+## Preparing
+- install protocol buffer 3.4.0, referring this link http://blog.csdn.net/twilightdream/article/details/72953338
+- sudo -H pip install --upgrade protobuf==3.1.0.post1
+- sudo apt-get install libhdf5-dev
+- sudo apt-get install python-h5py
+## Building
+- remember WITH_PYTHON_LAYER := 1 !
+- make all
+- make pycaffe
+
+## Training
+- download data and models from http://www.eecs.berkeley.edu/~lisa_anne/LRCN_video,
+- create flow frame using python instead of matlab like below
+    https://github.com/pathak22/pyflow
+    git clone https://github.com/pathak22/pyflow.git
+    cd pyflow/
+    python setup.py build_ext -i
+    python demo.py    # -viz option to visualize output
+- change my path like '/home/link/frames' in sequence_input_layer.py to your path and modify other relative path.
+- bash run_lstm_flow.sh, start training!
+
+
+
+
+
+

examples/LRCN_activity_recognition/sequence_input_layer.py

@@ -23,8 +23,8 @@
 import skimage.io
 import copy
 
-flow_frames = 'flow_images/'
-RGB_frames = 'frames/'
+flow_frames = '/home/link/flow_images/'
+RGB_frames = '/home/link/frames/'
 test_frames = 16 
 train_frames = 16
 test_buffer = 3
@@ -174,11 +174,11 @@ def setup(self, bottom, top):
     if self.flow:
       image_mean = [128, 128, 128]
 # TODO: No flow support in transformer currently
-#      self.transformer.set_is_flow('data_in', True)
+      self.transformer.set_is_flow('data_in', True)
     else:
       image_mean = [103.939, 116.779, 128.68]
 # TODO: No flow support in transformer currently
-#      self.transformer.set_is_flow('data_in', False)
+      self.transformer.set_is_flow('data_in', False)
     channel_mean = np.zeros((3,227,227))
     for channel_index, mean_val in enumerate(image_mean):
       channel_mean[channel_index, ...] = mean_val