Merge pull request #11 from rapidsai/branch-0.15

merge with upstream

CHANGELOG.md

@@ -5,7 +5,10 @@
 - PR #2287: Single-GPU TfidfTransformer implementation
 - PR #2289: QR SVD solver for MNMG PCA
 - PR #2312: column-major support for make_blobs
+- PR #2172: Initial support for auto-ARIMA
+- PR #2394: Adding cosine & correlation distance for KNN
 - PR #2392: PCA can accept sparse inputs, and sparse prim for computing covariance
+- PR #2465: Support pandas 1.0+
 
 ## Improvements
 - PR #2336: Eliminate `rmm.device_array` usage
@@ -14,24 +17,29 @@
 - PR #1945: enable clang tidy
 - PR #2339: umap performance improvements
 - PR #2308: Using fixture for Dask client to eliminate possiblity of not closing
-- PR #2310: Pinning ucx-py to 0.14 to make 0.15 CI pass
-- PR #1945: enable clang tidy
 - PR #2345: make C++ logger level definition to be the same as python layer
 - PR #2329: Add short commit hash to conda package name
 - PR #2363: Update threshold and make other changes for stress tests
 - PR #2371: Updating MBSGD tests to use larger batches
 - PR #2380: Pinning libcumlprims version to ease future updates
+- PR #2405: Remove references to deprecated RMM headers.
 - PR #2340: Import ARIMA in the root init file and fix the `test_fit_function` test
 - PR #2408: Install meta packages for dependencies
 - PR #2417: Move doc customization scripts to Jenkins
+- PR #2427: Moving MNMG decomposition to cuml
 - PR #2433: Add libcumlprims_mg to CMake
 - PR #2420: Add and set convert_dtype default to True in estimator fit methods
 - PR #2411: Refactor Mixin classes and use in classifier/regressor estimators
 - PR #2442: fix setting RAFT_DIR from the RAFT_PATH env var
-- PR #2411 Refactor Mixin classes and use in classifier/regressor estimators
+- PR #2469: Updating KNN c-api to document all arguments
 - PR #2453: Add CumlArray to API doc
 - PR #2440: Use Treelite Conda package
-- PR #2403 Support for input and output type consistency in logistic regression predict_proba
+- PR #2403: Support for input and output type consistency in logistic regression predict_proba
+- PR #2468: Add `_n_features_in_` attribute to all single GPU estimators that implement fit
+- PR #2483: Moving MNMG KNN to cuml
+- PR #2492: Adding additional assertions to mnmg nearest neighbors pytests
+- PR #2439: Update dask RF code to have print_detailed function
+- PR #2431: Match output of classifier predict with target dtype
 
 ## Bug Fixes
 - PR #2369: Update RF code to fix set_params memory leak
@@ -48,6 +56,12 @@
 - PR #2450: Update local build to use new gpuCI image
 - PR #2454: Mark RF memleak test as XFAIL, because we can't detect memleak reliably
 - PR #2455: Use correct field to store data type in `LabelEncoder.fit_transform`
+- PR #2475: Fix typo in build.sh
+- PR #2496: Fixing indentation for simulate_data in test_fil.py
+- PR #2494: Set QN regularization strength consistent with scikit-learn
+- PR #2486: Fix cupy input to kmeans init
+- PR #2497: Changes to accomodate cuDF unsigned categorical changes
+- PR #2507: Import `treelite.sklearn`
 
 # cuML 0.14.0 (03 Jun 2020)
 
@@ -221,7 +235,7 @@
 - PR #2305: Fixed race condition in DBScan
 - PR #2354: Fix broken links in README
 
-# cuML 0.13.0 (Date TBD)
+# cuML 0.13.0 (31 Mar 2020)
 
 ## New Features
 - PR #1777: Python bindings for entropy
@@ -340,7 +354,6 @@
 - PR #1950: Fix UMAP test failure
 
 
-
 # cuML 0.12.0 (04 Feb 2020)
 
 ## New Features

build.sh

@@ -116,7 +116,7 @@ if hasArg --singlegpu; then
     SINGLEGPU_PYTHON_FLAG="--singlegpu"
     SINGLEGPU_CPP_FLAG=ON
 fi
-if hasArg mgtests; then
+if hasArg cpp-mgtests; then
     BUILD_CPP_MG_TESTS=ON
 fi
 if hasArg --nvtx; then
@@ -149,7 +149,7 @@ fi
 
 ################################################################################
 # Configure for building all C++ targets
-if completeBuild || hasArg libcuml || hasArg prims || hasArg bench || hasArg prims-bench || hasArg cppdocs; then
+if completeBuild || hasArg libcuml || hasArg prims || hasArg bench || hasArg prims-bench || hasArg cppdocs || hasArg cpp-mgtests; then
     if (( ${BUILD_ALL_GPU_ARCH} == 0 )); then
         GPU_ARCH=""
         echo "Building for the architecture of the GPU in the system..."
@@ -185,7 +185,7 @@ MAKE_TARGETS=
 if hasArg libcuml; then
     MAKE_TARGETS="${MAKE_TARGETS}cuml++ cuml ml"
 fi
-if hasarg mgtests; then
+if hasArg cpp-mgtests; then
     MAKE_TARGETS="${MAKE_TARGETS} ml_mg"
 fi
 if hasArg prims; then
@@ -199,7 +199,7 @@ if hasArg prims-bench; then
 fi
 
 # If `./build.sh cuml` is called, don't build C/C++ components
-if completeBuild || hasArg libcuml || hasArg prims || hasArg bench; then
+if completeBuild || hasArg libcuml || hasArg prims || hasArg bench || hasArg cpp-mgtests; then
     # If there are no targets specified when calling build.sh, it will
     # just call `make -j`. This avoids a lot of extra printing
     cd ${LIBCUML_BUILD_DIR}

cpp/CMakeLists.txt

@@ -50,7 +50,7 @@ option(BUILD_CUML_CPP_LIBRARY "Build libcuml++ shared library" ON)
 
 option(BUILD_CUML_TESTS "Build cuML algorithm tests" ON)
 
-option(BUILD_CUML_MG_TESTS "Build cuML multigpu algorithm tests" ON)
+option(BUILD_CUML_MG_TESTS "Build cuML multigpu algorithm tests" OFF)
 
 option(BUILD_PRIMS_TESTS "Build ml-prim tests" ON)
 
@@ -156,10 +156,10 @@ if(SINGLEGPU)
   set(WITH_UCX OFF)
 endif(SINGLEGPU)
 
-if(NOT BUILD_CUML_MPI_COMMS AND NOT SINGLEGPU)
-  message(STATUS "Detected BUILD_CUML_MPI_COMMS set to OFF. Disabling BUILD_CUML_MG_TESTS")
-  set(BUILD_CUML_MG_TESTS OFF)
-endif(NOT BUILD_CUML_MPI_COMMS AND NOT SINGLEGPU)
+if(BUILD_CUML_MG_TESTS AND NOT SINGLEGPU)
+  message(STATUS "Detected BUILD_CUML_MG_TESTS set to ON. Enabling BUILD_CUML_MPI_COMMS")
+  set(BUILD_CUML_MPI_COMMS ON)
+endif(BUILD_CUML_MG_TESTS AND NOT SINGLEGPU)
 
 ##############################################################################
 # - Requirements -------------------------------------------------------------
@@ -295,7 +295,6 @@ include(cmake/Dependencies.cmake)
 
 set(CUML_INCLUDE_DIRECTORIES
   ${CUML_INCLUDE_DIR}
-  ${CUMLPRIMS_MG_INCLUDE_DIRS}
   ${CMAKE_CURRENT_SOURCE_DIR}/src
   ${CMAKE_CURRENT_SOURCE_DIR}/src_prims
   ${CMAKE_CURRENT_SOURCE_DIR}/test/prims
@@ -304,16 +303,33 @@ set(CUML_INCLUDE_DIRECTORIES
   ${CUTLASS_DIR}/src/cutlass
   ${CUB_DIR}/src/cub
   ${SPDLOG_DIR}/src/spdlog/include
-  ${RAFT_DIR}/cpp/include)
+  ${RAFT_DIR}/cpp/include
+  )
 
-set(CUML_LINK_LIBRARIES
+set(CUML_PUBLIC_LINK_LIBRARIES
   ${CUDA_cublas_LIBRARY}
   ${CUDA_curand_LIBRARY}
   ${CUDA_cusolver_LIBRARY}
   ${CUDA_CUDART_LIBRARY}
   ${CUDA_cusparse_LIBRARY}
   ${CUDA_nvgraph_LIBRARY}
-  ${CUMLPRIMS_MG_LIBRARIES})
+  )
+
+set(CUML_PRIVATE_LINK_LIBRARIES
+  ${Protobuf_LIBRARIES}
+  faisslib
+  treelite::treelite
+  treelite::treelite_runtime
+  )
+
+if(ENABLE_CUMLPRIMS_MG)
+  list(APPEND CUML_INCLUDE_DIRECTORIES
+       ${CUMLPRIMS_MG_INCLUDE_DIRS})
+
+  list(APPEND CUML_PRIVATE_LINK_LIBRARIES
+       CUMLPRIMS_MG::CUMLPRIMS_MG)
+
+endif(ENABLE_CUMLPRIMS_MG)
 
 ##############################################################################
 # - build libcuml++ shared library -------------------------------------------
@@ -352,25 +368,30 @@ if(BUILD_CUML_CPP_LIBRARY)
     src/svm/svc.cu
     src/svm/svr.cu
     src/svm/ws_util.cu
+    src/tsa/auto_arima.cu
     src/tsa/stationarity.cu
     src/tsne/tsne.cu
     src/tsvd/tsvd.cu
     src/umap/umap.cu)
 
   # mnmg components
 
-  # if(NOT SINGLEGPU)
-  #   target_sources(${CUML_CPP_TARGET}
-  #     PRIVATE src/kmeans/kmeans_mg.cu
-  #   )
-  # endif(NOT SINGLEGPU)
+  if(NOT SINGLEGPU)
+    target_sources(${CUML_CPP_TARGET}
+      PRIVATE
+        src/knn/knn_mg.cu
+        src/pca/pca_mg.cu
+        src/pca/sign_flip_mg.cu
+        src/tsvd/tsvd_mg.cu
+    )
+  endif(NOT SINGLEGPU)
 
   if(OPENMP_FOUND)
-    set(CUML_LINK_LIBRARIES ${CUML_LINK_LIBRARIES} ${OpenMP_CXX_LIB_NAMES} Threads::Threads)
+    set(CUML_PUBLIC_LINK_LIBRARIES ${CUML_PUBLIC_LINK_LIBRARIES} ${OpenMP_CXX_LIB_NAMES} Threads::Threads)
   endif(OPENMP_FOUND)
 
   if(NVTX)
-    set(CUML_LINK_LIBRARIES ${CUML_LINK_LIBRARIES} nvToolsExt)
+    set(CUML_PUBLIC_LINK_LIBRARIES ${CUML_PUBLIC_LINK_LIBRARIES} nvToolsExt)
     link_directories(${CUDA_TOOLKIT_ROOT_DIR}/lib64)
   endif(NVTX)
 
@@ -379,12 +400,10 @@ if(BUILD_CUML_CPP_LIBRARY)
 
   target_link_libraries(${CUML_CPP_TARGET}
     PUBLIC
-      ${CUML_LINK_LIBRARIES}
+      ${CUML_PUBLIC_LINK_LIBRARIES}
     PRIVATE
-      ${Protobuf_LIBRARIES}
-      faisslib
-      treelite::treelite
-      treelite::treelite_runtime)
+      ${CUML_PRIVATE_LINK_LIBRARIES}
+  )
   # If we export the libdmlc symbols, they can lead to weird crashes with other
   # libraries that use libdmlc. This just hides the symbols internally.
   target_link_options(${CUML_CPP_TARGET} PRIVATE "-Wl,--exclude-libs,libdmlc.a")

cpp/README.md

@@ -34,11 +34,11 @@ Current cmake offers the following configuration options:
 | --- | --- | --- | --- |
 | BUILD_CUML_CPP_LIBRARY | [ON, OFF]  | ON  | Enable/disable building libcuml++ shared library. Setting this variable to `OFF` sets the variables BUILD_CUML_TESTS, BUILD_CUML_MG_TESTS and BUILD_CUML_EXAMPLES to `OFF` |
 | BUILD_CUML_TESTS | [ON, OFF]  | ON  |  Enable/disable building cuML algorithm test executable `ml_test`.  |
-| BUILD_CUML_MG_TESTS | [ON, OFF]  | ON  |  Enable/disable building cuML algorithm test executable `ml_mg_test`. Requires MPI installed and turning BUILD_CUML_MPI_COMMS to ON. |
+| BUILD_CUML_MG_TESTS | [ON, OFF]  | ON  |  Enable/disable building cuML algorithm test executable `ml_mg_test`. Requires MPI to be installed. When enabled, BUILD_CUML_MPI_COMMS will be automatically set to ON. |
 | BUILD_PRIMS_TESTS | [ON, OFF]  | ON  | Enable/disable building cuML algorithm test executable `prims_test`.  |
 | BUILD_CUML_STD_COMMS | [ON, OFF] | ON | Enable/disable building cuML NCCL+UCX communicator for running multi-node multi-GPU algorithms. Note that UCX support can also be enabled/disabled (see below). The standard communicator and MPI communicator are not mutually exclusive and can both be installed at the same time. |
 | WITH_UCX | [ON, OFF] | OFF | Enable/disable UCX support in the standard cuML communicator. Algorithms requiring point-to-point messaging will not work when this is disabled. This flag is ignored if BUILD_CUML_STD_COMMS is set to OFF. |
-| BUILD_CUML_MPI_COMMS | [ON, OFF] | OFF | Enable/disable building cuML MPI+NCCL communicator for running multi-node multi-GPU C++ tests. MPI communicator and STD communicator are not mutually exclusive and can both be installed at the same time. If OFF, it overrides BUILD_CUML_MG_TESTS to be OFF as well. |
+| BUILD_CUML_MPI_COMMS | [ON, OFF] | OFF | Enable/disable building cuML MPI+NCCL communicator for running multi-node multi-GPU C++ tests. MPI communicator and STD communicator may both be installed at the same time. If OFF, it overrides BUILD_CUML_MG_TESTS to be OFF as well. |
 | SINGLEGPU | [ON, OFF] | OFF | Disable all mnmg components. Disables building of all multi-GPU algorithms and all comms library components. Removes libcumlprims, UCX-py and NCCL dependencies. Overrides values of  BUILD_CUML_MG_TESTS, BUILD_CUML_STD_COMMS, WITH_UCX and BUILD_CUML_MPI_COMMS. |
 | BUILD_CUML_EXAMPLES | [ON, OFF]  | ON  | Enable/disable building cuML C++ API usage examples.  |
 | BUILD_CUML_BENCH | [ON, OFF] | ON | Enable/disable building oc cuML C++ benchark.  |

cpp/examples/dbscan/dbscan_example.cpp

@@ -28,7 +28,6 @@
 #endif  //HAVE_CUB
 
 #ifdef HAVE_RMM
-#include <rmm/rmm.h>
 #include <cuml/common/rmmAllocatorAdapter.hpp>
 #endif  //HAVE_RMM
 

cpp/examples/kmeans/kmeans_example.cpp

@@ -28,7 +28,6 @@
 #endif  //HAVE_CUB
 
 #ifdef HAVE_RMM
-#include <rmm/rmm.h>
 #include <cuml/common/rmmAllocatorAdapter.hpp>
 #endif  // HAVE_RMM