Changed seeding strategy and have one generator per row of output data

src/toolkits/object_detection/one_shot_object_detection/one_shot_object_detector.cpp

@@ -32,9 +32,11 @@ namespace one_shot_object_detection {
 namespace data_augmentation {
 
 flex_dict build_annotation(ParameterSampler &parameter_sampler,
-                           std::string label, size_t object_width,
-                           size_t object_height, long seed) {
-  parameter_sampler.sample(seed);
+                           std::string label, 
+                           size_t object_width, size_t object_height,
+                           size_t background_width, size_t background_height,
+                           size_t seed, size_t row_number) {
+  parameter_sampler.sample(background_width, background_height, seed, row_number);
 
   size_t original_top_left_x = 0;
   size_t original_top_left_y = 0;
@@ -118,20 +120,19 @@ flex_image create_rgba_flex_image(const flex_image &object_input) {
 }
 
 std::pair<flex_image, flex_dict>
-create_synthetic_image_from_background_and_starter(const flex_image &starter,
+create_synthetic_image_from_background_and_starter(ParameterSampler &parameter_sampler,
+                                                   const flex_image &starter,
                                                    const flex_image &background,
                                                    std::string &label,
                                                    size_t seed,
                                                    size_t row_number) {
-  ParameterSampler parameter_sampler =
-      ParameterSampler(background.m_width, background.m_height,
-                       (background.m_width - starter.m_width) / 2,
-                       (background.m_height - starter.m_height) / 2);
 
   // construct annotation dictionary from parameters
   flex_dict annotation =
-      build_annotation(parameter_sampler, label, starter.m_width,
-                       starter.m_height, seed + row_number);
+      build_annotation(parameter_sampler, label, 
+                       starter.m_width, starter.m_height,
+                       background.m_width, background.m_height,
+                       seed, row_number);
 
   if (background.get_image_data() == nullptr) {
     log_and_throw("Background image has null image data.");
@@ -149,14 +150,14 @@ create_synthetic_image_from_background_and_starter(const flex_image &starter,
                        reinterpret_cast<const boost::gil::rgba8_pixel_t *>(
                            rgba_flex_image.get_image_data()),
                        rgba_flex_image.m_channels *
-                           rgba_flex_image.m_width  // row length in bytes
+                           rgba_flex_image.m_width // row length in bytes
       );
 
   boost::gil::rgb8_image_t::const_view_t background_view = interleaved_view(
       background.m_width, background.m_height,
       reinterpret_cast<const boost::gil::rgb8_pixel_t *>(
           background.get_image_data()),
-      background.m_channels * background.m_width  // row length in bytes
+      background.m_channels * background.m_width // row length in bytes
   );
   flex_image synthetic_image = create_synthetic_image(
       starter_image_view, background_view, parameter_sampler);
@@ -209,23 +210,28 @@ gl_sframe augment_data(const gl_sframe &data,
    * Replacing the `for` with a `parallel_for` fails the export_coreml unit test
    * with an EXC_BAD_ACCESS in the function call to boost::gil::resample_pixels
    */
-  for (size_t segment_id = 0; segment_id < nsegments; segment_id++) {
-    size_t segment_start = (segment_id * backgrounds.size()) / nsegments;
-    size_t segment_end = ((segment_id + 1) * backgrounds.size()) / nsegments;
-    size_t row_number = segment_start;
-    for (const auto &background_ft :
-         backgrounds.range_iterator(segment_start, segment_end)) {
-      row_number++;
-      flex_image flex_background =
-          image_util::decode_image(background_ft.to<flex_image>());
-      for (const auto &row : decompressed_data.range_iterator()) {
-        // go through all the starter images and create augmented images for
-        // all starter images and the respective chunk of background images
-        const flex_image &object = row[image_column_index].get<flex_image>();
-        std::string label = row[target_column_index].to<flex_string>();
+  for (const auto &row : decompressed_data.range_iterator()) {
+    // go through all the starter images and create augmented images for
+    // all starter images and the respective chunk of background images
+    const flex_image &object = row[image_column_index].get<flex_image>();
+    std::string label = row[target_column_index].to<flex_string>();
+    ParameterSampler parameter_sampler = ParameterSampler(
+      object.m_width, object.m_height, 0, 0);
+
+    for (size_t segment_id = 0; segment_id < nsegments; segment_id++) {
+      size_t segment_start = (segment_id * backgrounds.size()) / nsegments;
+      size_t segment_end = ((segment_id + 1) * backgrounds.size()) / nsegments;
+      size_t row_number = segment_start;
+      for (const auto &background_ft :
+          backgrounds.range_iterator(segment_start, segment_end)) {
+        row_number++;
+        flex_image flex_background =
+            image_util::decode_image(background_ft.to<flex_image>());
+
         std::pair<flex_image, flex_dict> synthetic_row =
             create_synthetic_image_from_background_and_starter(
-                object, flex_background, label, seed, row_number);
+                parameter_sampler, object, flex_background,
+                label, seed, row_number);
         flex_image synthetic_image = synthetic_row.first;
         flex_dict annotation = synthetic_row.second;
         // write the synthetically generated image and the constructed

src/toolkits/object_detection/one_shot_object_detection/util/parameter_sampler.cpp

@@ -6,6 +6,7 @@
  */
 
 #include <core/data/image/numeric_extension/perspective_projection.hpp>
+#include <limits>
 
 #include <toolkits/object_detection/one_shot_object_detection/util/parameter_sampler.hpp>
 
@@ -16,9 +17,9 @@
 namespace turi {
 namespace one_shot_object_detection {
 
-ParameterSampler::ParameterSampler(size_t width, size_t height, size_t dx,
-                                   size_t dy)
-    : width_(width), height_(height), dx_(dx), dy_(dy) {}
+ParameterSampler::ParameterSampler(size_t starter_width, size_t starter_height,
+                                   size_t dx, size_t dy)
+    : starter_width_(starter_width), starter_height_(starter_height), dx_(dx), dy_(dy) {}
 
 double deg_to_rad(double angle) { return angle * M_PI / 180.0; }
 
@@ -64,33 +65,31 @@ void ParameterSampler::set_warped_corners(
 /* Function to sample all the parameters needed to build a transform, and
  * then also build the transform.
  */
-void ParameterSampler::sample(long seed) {
+void ParameterSampler::sample(size_t background_width, size_t background_height,
+                              size_t seed, size_t row_number) {
   double theta_mean, phi_mean, gamma_mean;
-  std::srand(seed);
-  theta_mean = theta_means_[std::rand() % theta_means_.size()];
-  std::srand(seed + 1);
-  phi_mean = phi_means_[std::rand() % phi_means_.size()];
-  std::srand(seed + 2);
-  gamma_mean = gamma_means_[std::rand() % gamma_means_.size()];
+  std::seed_seq seed_seq = {static_cast<int>(seed), static_cast<int>(row_number)};
+  std::mt19937 engine(seed_seq);
+
+  std::uniform_int_distribution<int> index_distribution(0, INT_MAX);
+  theta_mean = theta_means_[index_distribution(engine) % theta_means_.size()];
+  phi_mean = phi_means_[index_distribution(engine) % phi_means_.size()];
+  gamma_mean = gamma_means_[index_distribution(engine) % gamma_means_.size()];
+
   std::normal_distribution<double> theta_distribution(theta_mean, angle_stdev_);
   std::normal_distribution<double> phi_distribution(phi_mean, angle_stdev_);
   std::normal_distribution<double> gamma_distribution(gamma_mean, angle_stdev_);
-  std::normal_distribution<double> focal_distribution((double)width_,
+  std::normal_distribution<double> focal_distribution((double)background_width,
                                                       focal_stdev_);
-  theta_generator_.seed(seed + 3);
-  theta_ = theta_distribution(theta_generator_);
-  phi_generator_.seed(seed + 4);
-  phi_ = phi_distribution(phi_generator_);
-  gamma_generator_.seed(seed + 5);
-  gamma_ = gamma_distribution(gamma_generator_);
-  focal_generator_.seed(seed + 6);
-  focal_ = focal_distribution(focal_generator_);
-  std::uniform_int_distribution<int> dz_distribution(std::max(width_, height_),
+  theta_ = theta_distribution(engine);
+  phi_ = phi_distribution(engine);
+  gamma_ = gamma_distribution(engine);
+  focal_ = focal_distribution(engine);
+  std::uniform_int_distribution<int> dz_distribution(std::max(background_width, background_height),
                                                      max_depth_);
-  dz_generator_.seed(seed + 7);
-  dz_ = focal_ + dz_distribution(dz_generator_);
+  dz_ = focal_ + dz_distribution(engine);
   transform_ = warp_perspective::get_transformation_matrix(
-      width_, height_, theta_, phi_, gamma_, dx_, dy_, dz_, focal_);
+      starter_width_, starter_height_, theta_, phi_, gamma_, dx_, dy_, dz_, focal_);
   warped_corners_.reserve(4);
 }
 

src/toolkits/object_detection/one_shot_object_detection/util/parameter_sampler.hpp

@@ -26,7 +26,7 @@ namespace one_shot_object_detection {
  */
 class ParameterSampler {
  public:
-  ParameterSampler(size_t width, size_t height, size_t dx, size_t dy);
+  ParameterSampler(size_t starter_width, size_t starter_height, size_t dx, size_t dy);
 
   /* Getters for all the parameters:
    * theta: rotation around the x axis.
@@ -54,22 +54,18 @@ class ParameterSampler {
   /* Function to sample all the parameters needed to build a transform, and
    * then also build the transform.
    */
-  void sample(long seed);
+  void sample(size_t background_width, size_t background_height,
+              size_t seed, size_t row_number);
 
  private:
-  size_t width_;
-  size_t height_;
+  size_t starter_width_;
+  size_t starter_height_;
   size_t max_depth_ = 13000;
   double angle_stdev_ = 20.0;
   double focal_stdev_ = 40.0;
   std::vector<double> theta_means_ = {-180.0, 0.0, 180.0};
   std::vector<double> phi_means_ = {-180.0, 0.0, 180.0};
   std::vector<double> gamma_means_ = {-180.0, -90.0, 0.0, 90.0, 180.0};
-  std::default_random_engine theta_generator_;
-  std::default_random_engine phi_generator_;
-  std::default_random_engine gamma_generator_;
-  std::default_random_engine dz_generator_;
-  std::default_random_engine focal_generator_;
   double theta_;
   double phi_;
   double gamma_;