[LITE][TEST] Enhance mobilenet test, mobile_light demo (#3245)

* Ehance mobilenetv1/v2, add std_dev computation. test=develop

* Ehance cxx mobile_light demo, add std_dev computation. test=develop

lite/api/mobilenetv1_test.cc

@@ -119,21 +119,21 @@ void TestModel(const std::vector<Place>& valid_places,
 
   // Get detailed result
   size_t output_tensor_num = predictor.GetOutputNames().size();
-  VLOG(1) << "output tesnor num:" << output_tensor_num;
+  VLOG(1) << "output tensor num:" << output_tensor_num;
 
   for (size_t tidx = 0; tidx < output_tensor_num; ++tidx) {
     auto* output_tensor = predictor.GetOutput(tidx);
     VLOG(1) << "============= output tensor " << tidx << " =============\n";
     auto out_dims = output_tensor->dims();
-    VLOG(1) << "out_dims:" << out_dims;
-
-    float sum = 0.f;
-    for (int i = 0; i < out_dims.production(); ++i) {
-      sum += output_tensor->data<float>()[i];
-    }
-    VLOG(1) << "out_dims.production():" << out_dims.production();
-    VLOG(1) << "output tensor sum value:" << sum;
-    VLOG(1) << "output tensor mean value:" << sum / out_dims.production();
+    auto out_data = output_tensor->data<float>();
+    auto out_mean = compute_mean<float>(out_data, out_dims.production());
+    auto out_std_dev = compute_standard_deviation<float>(
+        out_data, out_dims.production(), true, out_mean);
+
+    VLOG(1) << "output tensor dims:" << out_dims;
+    VLOG(1) << "output tensor elements num:" << out_dims.production();
+    VLOG(1) << "output tensor standard deviation:" << out_std_dev;
+    VLOG(1) << "output tensor mean value:" << out_mean;
 
     // print result
     for (int i = 0; i < out_dims.production(); ++i) {

lite/api/mobilenetv2_test.cc

@@ -121,21 +121,21 @@ void TestModel(const std::vector<Place>& valid_places,
 
   // Get detailed result
   size_t output_tensor_num = predictor.GetOutputNames().size();
-  VLOG(1) << "output tesnor num:" << output_tensor_num;
+  VLOG(1) << "output tensor num:" << output_tensor_num;
 
   for (size_t tidx = 0; tidx < output_tensor_num; ++tidx) {
     auto* output_tensor = predictor.GetOutput(tidx);
     VLOG(1) << "============= output tensor " << tidx << " =============\n";
     auto out_dims = output_tensor->dims();
-    VLOG(1) << "out_dims:" << out_dims;
-
-    float sum = 0.f;
-    for (int i = 0; i < out_dims.production(); ++i) {
-      sum += output_tensor->data<float>()[i];
-    }
-    VLOG(1) << "out_dims.production():" << out_dims.production();
-    VLOG(1) << "output tensor sum value:" << sum;
-    VLOG(1) << "output tensor mean value:" << sum / out_dims.production();
+    auto out_data = output_tensor->data<float>();
+    auto out_mean = compute_mean<float>(out_data, out_dims.production());
+    auto out_std_dev = compute_standard_deviation<float>(
+        out_data, out_dims.production(), true, out_mean);
+
+    VLOG(1) << "output tensor dims:" << out_dims;
+    VLOG(1) << "output tensor elements num:" << out_dims.production();
+    VLOG(1) << "output tensor standard deviation:" << out_std_dev;
+    VLOG(1) << "output tensor mean value:" << out_mean;
 
     // print result
     for (int i = 0; i < out_dims.production(); ++i) {

lite/api/test_helper.h

@@ -17,6 +17,7 @@
 #include <gflags/gflags.h>
 #include <sys/time.h>
 #include <time.h>
+#include <cmath>
 
 // for eval
 DEFINE_string(model_dir, "", "model dir");
@@ -43,5 +44,31 @@ inline double GetCurrentUS() {
   return 1e+6 * time.tv_sec + time.tv_usec;
 }
 
+template <typename T>
+double compute_mean(const T* in, const size_t length) {
+  double sum = 0.;
+  for (size_t i = 0; i < length; ++i) {
+    sum += in[i];
+  }
+  return sum / length;
+}
+
+template <typename T>
+double compute_standard_deviation(const T* in,
+                                  const size_t length,
+                                  bool has_mean = false,
+                                  double mean = 10000) {
+  if (!has_mean) {
+    mean = compute_mean<T>(in, length);
+  }
+
+  double variance = 0.;
+  for (size_t i = 0; i < length; ++i) {
+    variance += pow((in[i] - mean), 2);
+  }
+  variance /= length;
+  return sqrt(variance);
+}
+
 }  // namespace lite
 }  // namespace paddle

lite/demo/cxx/mobile_light/mobilenetv1_light_api.cc

@@ -14,6 +14,7 @@
 
 #include <sys/time.h>
 #include <time.h>
+#include <cmath>
 #include <iostream>
 #include <string>
 #include <vector>
@@ -36,6 +37,32 @@ std::string ShapePrint(const shape_t& shape) {
   return shape_str;
 }
 
+template <typename T>
+double compute_mean(const T* in, const size_t length) {
+  double sum = 0.;
+  for (size_t i = 0; i < length; ++i) {
+    sum += in[i];
+  }
+  return sum / length;
+}
+
+template <typename T>
+double compute_standard_deviation(const T* in,
+                                  const size_t length,
+                                  bool has_mean = false,
+                                  double mean = 10000) {
+  if (!has_mean) {
+    mean = compute_mean<T>(in, length);
+  }
+
+  double variance = 0.;
+  for (size_t i = 0; i < length; ++i) {
+    variance += pow((in[i] - mean), 2);
+  }
+  variance /= length;
+  return sqrt(variance);
+}
+
 inline double GetCurrentUS() {
   struct timeval time;
   gettimeofday(&time, NULL);
@@ -108,17 +135,17 @@ void RunModel(std::string model_dir,
         predictor->GetOutput(tidx);
     std::cout << "\n--- output tensor " << tidx << " ---" << std::endl;
     auto out_shape = output_tensor->shape();
-    std::cout << "out_shape(NCHW):" << ShapePrint(out_shape) << std::endl;
+    auto out_data = output_tensor->data<float>();
+    auto out_mean = compute_mean<float>(out_data, ShapeProduction(out_shape));
+    auto out_std_dev = compute_standard_deviation<float>(
+        out_data, ShapeProduction(out_shape), true, out_mean);
 
-    float sum = 0.f;
-    for (int i = 0; i < ShapeProduction(out_shape); ++i) {
-      sum += output_tensor->data<float>()[i];
-    }
+    std::cout << "output shape(NCHW):" << ShapePrint(out_shape) << std::endl;
     std::cout << "output tensor " << tidx
               << " elem num:" << ShapeProduction(out_shape) << std::endl;
-    std::cout << "output tensor " << tidx << " sum value:" << sum << std::endl;
     std::cout << "output tensor " << tidx
-              << " mean value:" << sum / ShapeProduction(out_shape)
+              << " standard deviation:" << out_std_dev << std::endl;
+    std::cout << "output tensor " << tidx << " mean value:" << out_mean
               << std::endl;
 
     // print output