fix word2vec related inference unit-tests (#14203)

fe8f1785 · Tao Luo · Yan Chunwei · e1742050 · fe8f1785 · fe8f1785
9 changed file
--- a/paddle/fluid/inference/CMakeLists.txt
+++ b/paddle/fluid/inference/CMakeLists.txt
+if(WITH_TESTING)
+  include(test.cmake) # some generic cmake funtion for inference
+endif()
 # analysis and tensorrt must be added before creating static library,
 # otherwise, there would be undefined reference to them in static library.
 add_subdirectory(analysis)

--- a/paddle/fluid/inference/analysis/CMakeLists.txt
+++ b/paddle/fluid/inference/analysis/CMakeLists.txt
@@ -20,22 +20,17 @@ cc_test(test_node SRCS node_tester.cc DEPS analysis)
 cc_test(test_dot SRCS dot_tester.cc DEPS analysis)
 cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis paddle_fluid)

-function (inference_analysis_test TARGET)
-    if(WITH_TESTING)
-        set(options "")
-        set(oneValueArgs "")
-        set(multiValueArgs SRCS ARGS EXTRA_DEPS)
-        cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-        set(mem_opt "")
-        if(WITH_GPU)
-            set(mem_opt "--fraction_of_gpu_memory_to_use=0.5")
-        endif()
-        cc_test(${TARGET}
-                SRCS "${analysis_test_SRCS}"
-                DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS}
-                ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model ${mem_opt} ${analysis_test_ARGS})
-        set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec)
-    endif(WITH_TESTING)
+function(inference_analysis_test TARGET)
+  if(WITH_TESTING)
+     set(options "")
+     set(oneValueArgs "")
+     set(multiValueArgs SRCS ARGS EXTRA_DEPS)
+     cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+     inference_base_test(${TARGET}
+             SRCS ${analysis_test_SRCS}
+             DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS}
+             ARGS --inference_model_dir=${WORD2VEC_MODEL_DIR} ${analysis_test_ARGS})
+  endif()
 endfunction(inference_analysis_test)

 inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS paddle_inference_api)

--- a/paddle/fluid/inference/api/CMakeLists.txt
+++ b/paddle/fluid/inference/api/CMakeLists.txt
@@ -17,39 +17,12 @@ if(APPLE)
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move")
 endif(APPLE)

-
-set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager naive_executor ${GLOB_PASS_LIB}
-        )
+set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager naive_executor ${GLOB_PASS_LIB})

 if(WITH_GPU AND TENSORRT_FOUND)
    set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine analysis_predictor)
 endif()

-function(inference_api_test TARGET_NAME)
-    if (WITH_TESTING)
-        set(options "")
-        set(oneValueArgs SRC)
-        set(multiValueArgs ARGS)
-        cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-
-	if (WITH_GPU)
-		cc_test(${TARGET_NAME}
-			SRCS ${inference_test_SRC}
-			DEPS "${inference_deps}"
-			ARGS --dirname=${PYTHON_TESTS_DIR}/book/ --fraction_of_gpu_memory_to_use=0.15)
-        else()
-		cc_test(${TARGET_NAME}
-			SRCS ${inference_test_SRC}
-			DEPS "${inference_deps}"
-			ARGS --dirname=${PYTHON_TESTS_DIR}/book/)
-	endif()
-        if(inference_test_ARGS)
-            set_tests_properties(${TARGET_NAME}
-                    PROPERTIES DEPENDS "${inference_test_ARGS}")
-        endif()
-    endif(WITH_TESTING)
-endfunction(inference_api_test)
-
 cc_library(reset_tensor_array SRCS details/reset_tensor_array.cc DEPS lod_tensor scope)
 cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS reset_tensor_array lod_tensor scope)
 cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api analysis naive_executor zero_copy_tensor)
@@ -59,8 +32,11 @@ cc_test(test_paddle_inference_api
        SRCS api_tester.cc
        DEPS paddle_inference_api)

-inference_api_test(test_api_impl SRC api_impl_tester.cc
-                    ARGS test_word2vec test_image_classification)
+if(WITH_TESTING)
+  inference_base_test(test_api_impl SRCS api_impl_tester.cc DEPS ${inference_deps}
+                      ARGS --word2vec_dirname=${WORD2VEC_MODEL_DIR} --book_dirname=${PYTHON_TESTS_DIR}/book)
+  set_tests_properties(test_api_impl PROPERTIES DEPENDS test_image_classification)
+endif()
 cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_predictor ${inference_deps} paddle_inference_api
        ARGS --dirname=${PYTHON_TESTS_DIR}/book)

@@ -68,8 +44,10 @@ if(WITH_GPU AND TENSORRT_FOUND)
 cc_library(paddle_inference_tensorrt_subgraph_engine
        SRCS api_tensorrt_subgraph_engine.cc
        DEPS paddle_inference_api analysis tensorrt_engine paddle_inference_api paddle_fluid_api tensorrt_converter zero_copy_tensor_dummy)
-
-inference_api_test(test_api_tensorrt_subgraph_engine SRC api_tensorrt_subgraph_engine_tester.cc ARGS test_word2vec)
+  if(WITH_TESTING)
+    inference_base_test(test_api_tensorrt_subgraph_engine SRCS api_tensorrt_subgraph_engine_tester.cc DEPS ${inference_deps}
+                      ARGS --dirname=${WORD2VEC_MODEL_DIR})
+  endif()
 endif()

 if (WITH_ANAKIN AND WITH_MKL) # only needed in CI

--- a/paddle/fluid/inference/api/api_impl_tester.cc
+++ b/paddle/fluid/inference/api/api_impl_tester.cc
@@ -22,12 +22,14 @@ limitations under the License. */
 #include "paddle/fluid/inference/tests/test_helper.h"

 #ifdef __clang__
-#define ACC_DIFF 4e-2
+#define ACC_DIFF 4e-3
 #else
-#define ACC_DIFF 1e-2
+#define ACC_DIFF 1e-3
 #endif

-DEFINE_string(dirname, "", "Directory of the inference model.");
+DEFINE_string(word2vec_dirname, "",
+              "Directory of the word2vec inference model.");
+DEFINE_string(book_dirname, "", "Directory of the book inference model.");

 namespace paddle {

@@ -49,7 +51,7 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) {

 NativeConfig GetConfig() {
  NativeConfig config;
-  config.model_dir = FLAGS_dirname + "/word2vec.inference.model";
+  config.model_dir = FLAGS_word2vec_dirname;
  LOG(INFO) << "dirname  " << config.model_dir;
  config.fraction_of_gpu_memory = 0.15;
 #ifdef PADDLE_WITH_CUDA
@@ -116,7 +118,7 @@ void MainImageClassification(bool use_gpu) {
  NativeConfig config = GetConfig();
  config.use_gpu = use_gpu;
  config.model_dir =
-      FLAGS_dirname + "/image_classification_resnet.inference.model";
+      FLAGS_book_dirname + "/image_classification_resnet.inference.model";

  const bool is_combined = false;
  std::vector<std::vector<int64_t>> feed_target_shapes =
@@ -220,7 +222,7 @@ void MainThreadsImageClassification(bool use_gpu) {
  NativeConfig config = GetConfig();
  config.use_gpu = use_gpu;
  config.model_dir =
-      FLAGS_dirname + "/image_classification_resnet.inference.model";
+      FLAGS_book_dirname + "/image_classification_resnet.inference.model";

  auto main_predictor = CreatePaddlePredictor<NativeConfig>(config);
  std::vector<framework::LoDTensor> jobs(num_jobs);

--- a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc
+++ b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc
@@ -29,13 +29,13 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) {

  //# 1. Create PaddlePredictor with a config.
  NativeConfig config0;
-  config0.model_dir = FLAGS_dirname + "word2vec.inference.model";
+  config0.model_dir = FLAGS_dirname;
  config0.use_gpu = true;
  config0.fraction_of_gpu_memory = 0.3;
  config0.device = 0;

  MixedRTConfig config1;
-  config1.model_dir = FLAGS_dirname + "word2vec.inference.model";
+  config1.model_dir = FLAGS_dirname;
  config1.use_gpu = true;
  config1.fraction_of_gpu_memory = 0.3;
  config1.device = 0;

--- a/paddle/fluid/inference/api/demo_ci/run.sh
+++ b/paddle/fluid/inference/api/demo_ci/run.sh
@@ -62,7 +62,7 @@ for WITH_STATIC_LIB in ON OFF; do
    -DWITH_GPU=$TEST_GPU_CPU \
    -DWITH_STATIC_LIB=$WITH_STATIC_LIB
  make -j
-  word2vec_model=${PADDLE_ROOT}'/build/python/paddle/fluid/tests/book/word2vec.inference.model'
+  word2vec_model=$DATA_DIR'/word2vec/word2vec.inference.model'
  if [ -d $word2vec_model ]; then
    for use_gpu in $use_gpu_list; do
      ./simple_on_word2vec \

--- a/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc
+++ b/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc
@@ -70,12 +70,8 @@ void Main(bool use_gpu) {
    // The outputs' buffers are in CPU memory.
    for (size_t i = 0; i < std::min(static_cast<size_t>(5), num_elements);
         i++) {
-      // Here will result random fail, for that the model is trained by CI, the
-      // train phase is not stable, so the result will be random.
-      // TODO(Superjomn) will restore after the model is upload.
-      // CHECK_NEAR(static_cast<float*>(outputs.front().data.data())[i],
-      // result[i],
-      // 0.001);
+      CHECK_NEAR(static_cast<float*>(outputs.front().data.data())[i], result[i],
+                 0.001);
    }
  }
 }

--- a/paddle/fluid/inference/test.cmake
+++ b/paddle/fluid/inference/test.cmake
+set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com" CACHE STRING "inference download url")
+set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING
+    "A path setting inference demo download directories.")
+function (inference_download install_dir url filename)
+    message(STATUS "Download inference test stuff from ${url}/${filename}")
+    execute_process(COMMAND bash -c "mkdir -p ${install_dir}")
+    execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}/${filename}")
+    message(STATUS "finish downloading ${filename}")
+endfunction()
+
+function (inference_download_and_uncompress install_dir url filename)
+    inference_download(${install_dir} ${url} ${filename})
+    execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}")
+endfunction()
+
+set(WORD2VEC_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/word2vec")
+if (NOT EXISTS ${WORD2VEC_INSTALL_DIR})
+    inference_download_and_uncompress(${WORD2VEC_INSTALL_DIR} ${INFERENCE_URL} "word2vec.inference.model.tar.gz")
+endif()
+set(WORD2VEC_MODEL_DIR "${WORD2VEC_INSTALL_DIR}/word2vec.inference.model")
+
+function (inference_base_test TARGET)
+   set(options "")
+   set(oneValueArgs "")
+   set(multiValueArgs SRCS ARGS DEPS)
+   cmake_parse_arguments(base_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+   if(WITH_GPU)
+       set(mem_opt "--fraction_of_gpu_memory_to_use=0.5")
+   endif()
+   cc_test(${TARGET} SRCS ${base_test_SRCS} DEPS ${base_test_DEPS} ARGS ${mem_opt} ${base_test_ARGS})
+endfunction()
--- a/paddle/fluid/inference/tests/api/CMakeLists.txt
+++ b/paddle/fluid/inference/tests/api/CMakeLists.txt
-set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com")
-set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING
-    "A path setting inference demo download directories.")
 set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor)
-function (inference_download install_dir url filename)
-    message(STATUS "Download inference test stuff from ${url}/${filename}")
-    execute_process(COMMAND bash -c "mkdir -p ${install_dir}")
-    execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}/${filename}")
-    message(STATUS "finish downloading ${filename}")
-endfunction()
-
-function (inference_download_and_uncompress install_dir url filename)
-    inference_download(${install_dir} ${url} ${filename})
-    execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}")
-endfunction()

 function(download_model_and_data install_dir model_name data_name)
    if (NOT EXISTS ${install_dir})