Merge pull request #12160 from luotao1/vis_demo

add independent vis_demo for inference

Author: luotao1

paddle/contrib/inference/CMakeLists.txt

@@ -104,7 +104,3 @@ if (WITH_ANAKIN) # only needed in CI
         target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
      endif(WITH_TESTING)
 endif()
-
-if(WITH_TESTING)
-    add_subdirectory(demo)
-endif()

paddle/contrib/inference/demo/CMakeLists.txt

@@ -0,0 +1 @@
+data

paddle/contrib/inference/demo/README.md

@@ -0,0 +1,26 @@
+# Inference Demos
+
+There are several demos:
+
+- simple_on_word2vec: 
+  - Follow the C++ codes is in `simple_on_word2vec.cc`. 
+  - It is suitable for word2vec model.
+- vis_demo: 
+  - Follow the C++ codes is in `vis_demo.cc`. 
+  - It is suitable for mobilenet, se_resnext50 and ocr three models.
+  - Input data format:
+    - Each line contains a single record
+    - Each record's format is
+    ```
+    <space splitted floats as data>\t<space splitted ints as shape>
+    ```
+
+To build and execute the demos, simply run 
+```
+./run.sh $PADDLE_ROOT $TURN_ON_MKL $TEST_GPU_CPU
+```
+- It will build and execute the demos in both static and shared library.
+- `$PADDLE_ROOT`: paddle library path
+- `$TURN_ON_MKL`: use MKL or Openblas
+- `$TEST_GPU_CPU`: test both GPU/CPU mode or only CPU mode
+- NOTE: for simple_on_word2vec, must run `ctest -R test_word2vec -R` to obtain word2vec model at first.

paddle/contrib/inference/demo_ci/.gitignore

@@ -13,10 +13,30 @@ else
   use_gpu_list='false'
 fi
 
+# download vis_demo data
+function download() {
+  dir_name=$1
+  mkdir -p $dir_name
+  cd $dir_name
+  wget -q ${URL_ROOT}$dir_name.tar.gz
+  tar xzf *.tar.gz
+  cd ..
+}
+URL_ROOT=http://paddlemodels.bj.bcebos.com/inference-vis-demos%2F
+mkdir -p data
+cd data
+vis_demo_list='se_resnext50 ocr mobilenet'
+for vis_demo_name in $vis_demo_list; do
+  download $vis_demo_name
+done
+cd ..
+
+# compile and test the demo
 mkdir -p build
 cd build
 
 for WITH_STATIC_LIB in ON OFF; do
+  # -----simple_on_word2vec-----
   rm -rf *
   cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \
     -DWITH_MKL=$TURN_ON_MKL \
@@ -29,9 +49,30 @@ for WITH_STATIC_LIB in ON OFF; do
       --dirname=${PADDLE_ROOT}/build/python/paddle/fluid/tests/book/word2vec.inference.model \
       --use_gpu=$use_gpu
     if [ $? -ne 0 ]; then
-      echo "inference demo runs fail."
+      echo "simple_on_word2vec demo runs fail."
       exit 1
     fi
   done
+  # ---------vis_demo---------
+  rm -rf *
+  cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \
+    -DWITH_MKL=$TURN_ON_MKL \
+    -DDEMO_NAME=vis_demo \
+    -DWITH_GPU=$TEST_GPU_CPU \
+    -DWITH_STATIC_LIB=$WITH_STATIC_LIB
+  make -j
+  for use_gpu in false; do
+    for vis_demo_name in $vis_demo_list; do 
+      ./vis_demo \
+        --modeldir=../data/$vis_demo_name/model \
+        --data=../data/$vis_demo_name/data.txt \
+        --refer=../data/$vis_demo_name/result.txt \
+        --use_gpu=$use_gpu
+      if [ $? -ne 0 ]; then
+        echo "vis demo $vis_demo_name runs fail."
+        exit 1
+      fi
+    done
+  done
 done
 set +x

paddle/contrib/inference/demo_ci/README.md

@@ -16,7 +16,7 @@
 #include <string>
 #include <vector>
 
-#include "paddle/contrib/inference/paddle_inference_api.h"
+#include "contrib/inference/paddle_inference_api.h"
 
 namespace paddle {
 namespace demo {

paddle/contrib/inference/demo_ci/run.sh

@@ -18,26 +18,25 @@ limitations under the License. */
 
 #include <gflags/gflags.h>
 #include <glog/logging.h>  // use glog instead of PADDLE_ENFORCE to avoid importing other paddle header files.
-#include <gtest/gtest.h>
 #include <fstream>
 #include <iostream>
-#include "paddle/contrib/inference/demo/utils.h"
-#include "paddle/contrib/inference/paddle_inference_api.h"
+#include "paddle/fluid/platform/enforce.h"
+#include "utils.h"
 
 #ifdef PADDLE_WITH_CUDA
 DECLARE_double(fraction_of_gpu_memory_to_use);
 #endif
-
-namespace paddle {
-namespace demo {
-
 DEFINE_string(modeldir, "", "Directory of the inference model.");
 DEFINE_string(refer, "", "path to reference result for comparison.");
 DEFINE_string(
     data,
     "",
     "path of data; each line is a record, format is "
     "'<space splitted floats as data>\t<space splitted ints as shape'");
+DEFINE_bool(use_gpu, false, "Whether use gpu.");
+
+namespace paddle {
+namespace demo {
 
 struct Record {
   std::vector<float> data;
@@ -47,7 +46,7 @@ struct Record {
 void split(const std::string& str, char sep, std::vector<std::string>* pieces);
 
 Record ProcessALine(const std::string& line) {
-  LOG(INFO) << "process a line";
+  VLOG(3) << "process a line";
   std::vector<std::string> columns;
   split(line, '\t', &columns);
   CHECK_EQ(columns.size(), 2UL)
@@ -65,8 +64,8 @@ Record ProcessALine(const std::string& line) {
   for (auto& s : shape_strs) {
     record.shape.push_back(std::stoi(s));
   }
-  LOG(INFO) << "data size " << record.data.size();
-  LOG(INFO) << "data shape size " << record.shape.size();
+  VLOG(3) << "data size " << record.data.size();
+  VLOG(3) << "data shape size " << record.shape.size();
   return record;
 }
 
@@ -78,20 +77,22 @@ void CheckOutput(const std::string& referfile, const PaddleTensor& output) {
   file.close();
 
   size_t numel = output.data.length() / PaddleDtypeSize(output.dtype);
-  LOG(INFO) << "predictor output numel " << numel;
-  LOG(INFO) << "reference output numel " << refer.data.size();
-  EXPECT_EQ(numel, refer.data.size());
+  VLOG(3) << "predictor output numel " << numel;
+  VLOG(3) << "reference output numel " << refer.data.size();
+  PADDLE_ENFORCE_EQ(numel, refer.data.size());
   switch (output.dtype) {
     case PaddleDType::INT64: {
       for (size_t i = 0; i < numel; ++i) {
-        EXPECT_EQ(static_cast<int64_t*>(output.data.data())[i], refer.data[i]);
+        PADDLE_ENFORCE_EQ(static_cast<int64_t*>(output.data.data())[i],
+                          refer.data[i]);
       }
       break;
     }
     case PaddleDType::FLOAT32:
       for (size_t i = 0; i < numel; ++i) {
-        EXPECT_NEAR(
-            static_cast<float*>(output.data.data())[i], refer.data[i], 1e-5);
+        PADDLE_ENFORCE_LT(
+            fabs(static_cast<float*>(output.data.data())[i] - refer.data[i]),
+            1e-5);
       }
       break;
   }
@@ -106,15 +107,15 @@ void Main(bool use_gpu) {
   config.prog_file = FLAGS_modeldir + "/__model__";
   config.use_gpu = use_gpu;
   config.device = 0;
-#ifdef PADDLE_WITH_CUDA
-  config.fraction_of_gpu_memory = FLAGS_fraction_of_gpu_memory_to_use;
-#endif
+  if (FLAGS_use_gpu) {
+    config.fraction_of_gpu_memory = 0.1;  // set by yourself
+  }
 
-  LOG(INFO) << "init predictor";
+  VLOG(3) << "init predictor";
   auto predictor =
       CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
 
-  LOG(INFO) << "begin to process data";
+  VLOG(3) << "begin to process data";
   // Just a single batch of data.
   std::string line;
   std::ifstream file(FLAGS_data);
@@ -129,21 +130,26 @@ void Main(bool use_gpu) {
       .data = PaddleBuf(record.data.data(), record.data.size() * sizeof(float)),
       .dtype = PaddleDType::FLOAT32};
 
-  LOG(INFO) << "run executor";
+  VLOG(3) << "run executor";
   std::vector<PaddleTensor> output;
   predictor->Run({input}, &output);
 
-  LOG(INFO) << "output.size " << output.size();
+  VLOG(3) << "output.size " << output.size();
   auto& tensor = output.front();
-  LOG(INFO) << "output: " << SummaryTensor(tensor);
+  VLOG(3) << "output: " << SummaryTensor(tensor);
 
   // compare with reference result
   CheckOutput(FLAGS_refer, tensor);
 }
 
-TEST(demo, vis_demo_cpu) { Main(false /*use_gpu*/); }
-#ifdef PADDLE_WITH_CUDA
-TEST(demo, vis_demo_gpu) { Main(true /*use_gpu*/); }
-#endif
 }  // namespace demo
 }  // namespace paddle
+
+int main(int argc, char** argv) {
+  google::ParseCommandLineFlags(&argc, &argv, true);
+  paddle::demo::Main(false /* use_gpu*/);
+  if (FLAGS_use_gpu) {
+    paddle::demo::Main(true /*use_gpu*/);
+  }
+  return 0;
+}