feat: support `half_pixel_centers` arg for `resize` type ops

N/A
Signed-off-by: NLuxuhui <luxuhui@xiaomi.com>

mace/ops/opencl/cl/resize_bicubic.cl

@@ -16,6 +16,7 @@ __kernel void resize_bicubic_nocache(OUT_OF_RANGE_PARAMS
                                      __write_only image2d_t output,
                                      __private const float height_scale,
                                      __private const float width_scale,
+                                     __private const int half_pixel_centers,
                                      __private const int in_height,
                                      __private const int in_width,
                                      __private const int out_height) {
@@ -38,8 +39,10 @@ __kernel void resize_bicubic_nocache(OUT_OF_RANGE_PARAMS
   const int b = hb / out_height;
   const int h = hb - mul24(b, out_height);
 
-  const float h_in = h * height_scale;
-  const float w_in = w * width_scale;
+  const float h_in = half_pixel_centers ?
+      ((float)h + 0.5f) * height_scale - 0.5f : h * height_scale;
+  const float w_in = half_pixel_centers ?
+      ((float)w + 0.5f) * width_scale - 0.5f : w * width_scale;
 
   const int in_w_offset = mul24(ch_blk, in_width);
   const int in_h_offset = mul24(b, in_height);

mace/ops/opencl/cl/resize_bilinear.cl

@@ -6,6 +6,7 @@ __kernel void resize_bilinear_nocache(OUT_OF_RANGE_PARAMS
                                       __write_only image2d_t output,
                                       __private const float height_scale,
                                       __private const float width_scale,
+                                      __private const int half_pixel_centers,
                                       __private const int in_height,
                                       __private const int in_width,
                                       __private const int out_height) {
@@ -26,8 +27,10 @@ __kernel void resize_bilinear_nocache(OUT_OF_RANGE_PARAMS
   const int b = hb / out_height;
   const int h = hb - mul24(b, out_height);
 
-  const float h_in = h * height_scale;
-  const float w_in = w * width_scale;
+  const float h_in = half_pixel_centers ?
+      ((float)h + 0.5f) * height_scale - 0.5f : h * height_scale;
+  const float w_in = half_pixel_centers ?
+      ((float)w + 0.5f) * width_scale - 0.5f : w * width_scale;
   const int h_lower = max(0, (int) floor(h_in));
   const int h_upper = min(in_height - 1, h_lower + 1);
   const int w_lower = max(0, (int) floor(w_in));

mace/ops/opencl/cl/resize_nearest_neighbor.cl

@@ -7,6 +7,7 @@ __kernel void resize_nearest_neighbor_nocache(
     __write_only image2d_t output,
     __private const float height_scale,
     __private const float width_scale,
+    __private const int half_pixel_centers,
     __private const int in_height,
     __private const int in_width,
     __private const int out_height,
@@ -27,10 +28,15 @@ __kernel void resize_nearest_neighbor_nocache(
   const int b = hb / out_height;
   const int h = hb - mul24(b, out_height);
 
-  const int h_in = min((align_corner) ? (int) round(h * height_scale) :
-      (int) floor(h * height_scale), in_height - 1);
-  const int w_in = min((align_corner) ? (int) round(w * width_scale) :
-      (int) floor(w * width_scale), in_width - 1);
+  const float h_in_f = half_pixel_centers ?
+      ((float)h + 0.5f) * height_scale : h * height_scale;
+  const float w_in_f = half_pixel_centers ?
+      ((float)w + 0.5f) * width_scale : w * width_scale;
+
+  const int h_in = min((align_corner) ? (int) round(h_in_f) :
+      (int) floor(h_in_f), in_height - 1);
+  const int w_in = min((align_corner) ? (int) round(w_in_f) :
+      (int) floor(w_in_f), in_width - 1);
 
   const int in_w_offset = mul24(ch_blk, in_width);
   const int in_h_offset = mul24(b, in_height);

mace/ops/opencl/image/resize_bicubic.cc

@@ -85,6 +85,7 @@ MaceStatus ResizeBicubicKernel::Compute(
     kernel_.setArg(idx++, *(output->opencl_image()));
     kernel_.setArg(idx++, height_scale);
     kernel_.setArg(idx++, width_scale);
+    kernel_.setArg(idx++, static_cast<int32_t>(half_pixel_centers_));
     kernel_.setArg(idx++, static_cast<int32_t>(in_height));
     kernel_.setArg(idx++, static_cast<int32_t>(in_width));
     kernel_.setArg(idx++, static_cast<int32_t>(out_height));

mace/ops/opencl/image/resize_bicubic.h

@@ -64,9 +64,11 @@ inline std::vector<uint32_t> LocalWS(OpenCLRuntime *runtime,
 class ResizeBicubicKernel : public OpenCLResizeBicubicKernel {
  public:
   ResizeBicubicKernel(bool align_corners,
+                      bool half_pixel_centers,
                       const index_t out_height,
                       const index_t out_width)
       : align_corners_(align_corners),
+        half_pixel_centers_(half_pixel_centers),
         out_height_(out_height),
         out_width_(out_width) {}
 
@@ -77,6 +79,7 @@ class ResizeBicubicKernel : public OpenCLResizeBicubicKernel {
 
  private:
   bool align_corners_;
+  bool half_pixel_centers_;
   index_t out_height_;
   index_t out_width_;
   cl::Kernel kernel_;

mace/ops/opencl/image/resize_bilinear.cc

@@ -85,6 +85,7 @@ MaceStatus ResizeBilinearKernel::Compute(
     kernel_.setArg(idx++, *(output->opencl_image()));
     kernel_.setArg(idx++, height_scale);
     kernel_.setArg(idx++, width_scale);
+    kernel_.setArg(idx++, static_cast<int32_t>(half_pixel_centers_));
     kernel_.setArg(idx++, static_cast<int32_t>(in_height));
     kernel_.setArg(idx++, static_cast<int32_t>(in_width));
     kernel_.setArg(idx++, static_cast<int32_t>(out_height));

mace/ops/opencl/image/resize_bilinear.h

@@ -66,8 +66,9 @@ inline std::vector<uint32_t> LocalWS(OpenCLRuntime *runtime,
 
 class ResizeBilinearKernel : public OpenCLResizeBilinearKernel {
  public:
-  explicit ResizeBilinearKernel(bool align_corners)
-      : align_corners_(align_corners) {}
+  explicit ResizeBilinearKernel(bool align_corners, bool half_pixel_centers)
+      : align_corners_(align_corners),
+        half_pixel_centers_(half_pixel_centers) {}
 
   MaceStatus Compute(
       OpContext *context,
@@ -78,6 +79,7 @@ class ResizeBilinearKernel : public OpenCLResizeBilinearKernel {
 
  private:
   bool align_corners_;
+  bool half_pixel_centers_;
   cl::Kernel kernel_;
   uint32_t kwg_size_;
   std::vector<index_t> input_shape_;

mace/ops/opencl/image/resize_nearest_neighbor.cc

@@ -82,6 +82,7 @@ MaceStatus ResizeNearestNeighborKernel::Compute(
     kernel_.setArg(idx++, *(output->opencl_image()));
     kernel_.setArg(idx++, height_scale);
     kernel_.setArg(idx++, width_scale);
+    kernel_.setArg(idx++, static_cast<int32_t>(half_pixel_centers_));
     kernel_.setArg(idx++, static_cast<int32_t>(in_height));
     kernel_.setArg(idx++, static_cast<int32_t>(in_width));
     kernel_.setArg(idx++, static_cast<int32_t>(out_height));

mace/ops/opencl/image/resize_nearest_neighbor.h

@@ -66,8 +66,10 @@ inline std::vector<uint32_t> LocalWS(OpenCLRuntime *runtime,
 
 class ResizeNearestNeighborKernel : public OpenCLResizeNearestNeighborKernel {
  public:
-  explicit ResizeNearestNeighborKernel(bool align_corners)
-      : align_corners_(align_corners) {}
+  explicit ResizeNearestNeighborKernel(bool align_corners,
+                                       bool half_pixel_centers)
+      : align_corners_(align_corners),
+        half_pixel_centers_(half_pixel_centers) {}
 
   MaceStatus Compute(
       OpContext *context,
@@ -78,6 +80,7 @@ class ResizeNearestNeighborKernel : public OpenCLResizeNearestNeighborKernel {
 
  private:
   bool align_corners_;
+  bool half_pixel_centers_;
   cl::Kernel kernel_;
   uint32_t kwg_size_;
   std::vector<index_t> input_shape_;

mace/ops/resize_bicubic.cc

@@ -56,11 +56,15 @@ inline int64_t Bound(int64_t val, int64_t limit) {
   return std::min<int64_t>(limit - 1ll, std::max<int64_t>(0ll, val));
 }
 
-inline void GetWeightsAndIndices(float scale, int64_t out_loc, int64_t limit,
+inline void GetWeightsAndIndices(float scale, bool half_pixel_centers,
+                                 int64_t out_loc, int64_t limit,
                                  std::vector<float> *weights,
                                  std::vector<int64_t> *indices) {
-  auto in_loc = static_cast<int64_t>(scale * out_loc);
-  const float delta = scale * out_loc - in_loc;
+  const float in = half_pixel_centers ?
+                   (static_cast<float>(out_loc) + 0.5f) * scale - 0.5f :
+                   out_loc * scale;
+  auto in_loc = static_cast<int64_t>(in);
+  const float delta = in - in_loc;
   const int64_t offset = lrintf(delta * common::utils::kTableSize);
   const float *coeffs_tab = GetCoeffsTable();
   *weights = {coeffs_tab[offset * 2 + 1],
@@ -87,6 +91,7 @@ inline void ResizeImage(const OpContext *context,
                         const index_t channels,
                         const float height_scale,
                         const float width_scale,
+                        const bool half_pixel_centers,
                         float *output) {
   utils::ThreadPool
       &thread_pool = context->device()->cpu_runtime()->thread_pool();
@@ -97,13 +102,13 @@ inline void ResizeImage(const OpContext *context,
       for (index_t y = start1; y < end1; y += step1) {
         std::vector<float> y_weights;
         std::vector<index_t> y_indices;
-        GetWeightsAndIndices(height_scale, y, in_height, &y_weights,
-                             &y_indices);
+        GetWeightsAndIndices(height_scale, half_pixel_centers, y, in_height,
+                             &y_weights, &y_indices);
         for (index_t x = 0; x < out_width; ++x) {
           std::vector<float> x_weights;
           std::vector<index_t> x_indices;
-          GetWeightsAndIndices(width_scale, x, in_width, &x_weights,
-                               &x_indices);
+          GetWeightsAndIndices(width_scale, half_pixel_centers, x, in_width,
+                               &x_weights, &x_indices);
 
           for (index_t c = 0; c < channels; ++c) {
             // Use a 4x4 patch to compute the interpolated output value at
@@ -139,6 +144,8 @@ class ResizeBicubicOp<DeviceType::CPU, float> : public Operation {
   explicit ResizeBicubicOp(OpConstructContext *context)
       : Operation(context),
         align_corners_(Operation::GetOptionalArg<bool>("align_corners", false)),
+        half_pixel_centers_(
+            Operation::GetOptionalArg<bool>("half_pixel_centers", false)),
         size_(Operation::GetRepeatedArgs<index_t>("size", {-1, -1})) {}
 
   MaceStatus Run(OpContext *context) override {
@@ -191,6 +198,7 @@ class ResizeBicubicOp<DeviceType::CPU, float> : public Operation {
                 channels,
                 height_scale,
                 width_scale,
+                half_pixel_centers_,
                 output_data);
 
     return MaceStatus::MACE_SUCCESS;
@@ -198,6 +206,7 @@ class ResizeBicubicOp<DeviceType::CPU, float> : public Operation {
 
  private:
   bool align_corners_;
+  bool half_pixel_centers_;
   std::vector<index_t> size_;
 };
 
@@ -209,12 +218,14 @@ class ResizeBicubicOp<DeviceType::GPU, float> : public Operation {
       : Operation(context) {
     bool align_corners = Operation::GetOptionalArg<bool>(
         "align_corners", false);
+    bool half_pixel_centers = Operation::GetOptionalArg<bool>(
+        "half_pixel_centers", false);
     std::vector<index_t> size = Operation::GetRepeatedArgs<index_t>(
         "size", {-1, -1});
     MACE_CHECK(size.size() == 2);
     if (context->GetOpMemoryType() == MemoryType::GPU_IMAGE) {
       kernel_ = make_unique<opencl::image::ResizeBicubicKernel>(
-          align_corners, size[0], size[1]);
+          align_corners, half_pixel_centers, size[0], size[1]);
     } else {
       MACE_NOT_IMPLEMENTED;
     }

mace/ops/resize_bilinear.cc

@@ -40,11 +40,13 @@ inline void ComputeInterpolationWeights(
     const index_t out_size,
     const index_t in_size,
     const float scale,
+    bool half_pixel_centers,
     CachedInterpolation *interpolation) {
   interpolation[out_size].lower = 0;
   interpolation[out_size].upper = 0;
   for (index_t i = out_size - 1; i >= 0; --i) {
-    const float in = i * scale;
+    const float in = half_pixel_centers ?
+                     (static_cast<float>(i) + 0.5f) * scale - 0.5f : i * scale;
     interpolation[i].lower = static_cast<index_t>(in);
     interpolation[i].upper = std::min(interpolation[i].lower + 1, in_size - 1);
     interpolation[i].lerp = in - interpolation[i].lower;
@@ -183,7 +185,9 @@ class ResizeBilinearOp<DeviceType::CPU, T> : public Operation {
         align_corners_(Operation::GetOptionalArg<bool>("align_corners", false)),
         size_(Operation::GetRepeatedArgs<index_t>("size", {-1, -1})),
         height_scale_(Operation::GetOptionalArg<float>("height_scale", 0)),
-        width_scale_(Operation::GetOptionalArg<float>("width_scale", 0)) {}
+        width_scale_(Operation::GetOptionalArg<float>("width_scale", 0)),
+        half_pixel_centers_(
+            Operation::GetOptionalArg<bool>("half_pixel_centers", false)) {}
 
   MaceStatus Run(OpContext *context) override {
     MACE_UNUSED(context);
@@ -237,8 +241,10 @@ class ResizeBilinearOp<DeviceType::CPU, T> : public Operation {
     std::vector<CachedInterpolation> xs(out_width + 1);
 
     // Compute the cached interpolation weights on the x and y dimensions.
-    ComputeInterpolationWeights(out_height, in_height, height_scale, ys.data());
-    ComputeInterpolationWeights(out_width, in_width, width_scale, xs.data());
+    ComputeInterpolationWeights(out_height, in_height, height_scale,
+                                half_pixel_centers_, ys.data());
+    ComputeInterpolationWeights(out_width, in_width, width_scale,
+                                half_pixel_centers_, xs.data());
 
     ResizeImageNCHW(context,
                     input_data,
@@ -260,6 +266,7 @@ class ResizeBilinearOp<DeviceType::CPU, T> : public Operation {
   std::vector<index_t> size_;
   float height_scale_;
   float width_scale_;
+  bool half_pixel_centers_;
 };
 
 #ifdef MACE_ENABLE_QUANTIZE
@@ -271,7 +278,9 @@ class ResizeBilinearOp<DeviceType::CPU, uint8_t> : public Operation {
         align_corners_(Operation::GetOptionalArg<bool>("align_corners", false)),
         size_(Operation::GetRepeatedArgs<index_t>("size", {-1, -1})),
         height_scale_(Operation::GetOptionalArg<float>("height_scale", 0)),
-        width_scale_(Operation::GetOptionalArg<float>("width_scale", 0)) {}
+        width_scale_(Operation::GetOptionalArg<float>("width_scale", 0)),
+        half_pixel_centers_(
+            Operation::GetOptionalArg<bool>("half_pixel_centers", false)) {}
 
   MaceStatus Run(OpContext *context) override {
     MACE_UNUSED(context);
@@ -325,8 +334,10 @@ class ResizeBilinearOp<DeviceType::CPU, uint8_t> : public Operation {
     std::vector<CachedInterpolation> xs(out_width + 1);
 
     // Compute the cached interpolation weights on the x and y dimensions.
-    ComputeInterpolationWeights(out_height, in_height, height_scale, ys.data());
-    ComputeInterpolationWeights(out_width, in_width, width_scale, xs.data());
+    ComputeInterpolationWeights(out_height, in_height, height_scale,
+                                half_pixel_centers_, ys.data());
+    ComputeInterpolationWeights(out_width, in_width, width_scale,
+                                half_pixel_centers_, xs.data());
 
     ResizeImageNHWC(context,
                     input_data,
@@ -348,6 +359,7 @@ class ResizeBilinearOp<DeviceType::CPU, uint8_t> : public Operation {
   std::vector<index_t> size_;
   float height_scale_;
   float width_scale_;
+  bool half_pixel_centers_;
 };
 #endif  // MACE_ENABLE_QUANTIZE
 
@@ -362,8 +374,11 @@ class ResizeBilinearOp<DeviceType::GPU, float> : public Operation {
         width_scale_(Operation::GetOptionalArg<float>("width_scale", 0))  {
     bool align_corners = Operation::GetOptionalArg<bool>(
         "align_corners", false);
+    bool half_pixel_centers = Operation::GetOptionalArg<bool>(
+        "half_pixel_centers", false);
     if (context->GetOpMemoryType() == MemoryType::GPU_IMAGE) {
-      kernel_ = make_unique<opencl::image::ResizeBilinearKernel>(align_corners);
+      kernel_ = make_unique<opencl::image::ResizeBilinearKernel>(
+          align_corners, half_pixel_centers);
     } else {
       MACE_NOT_IMPLEMENTED;
     }

mace/ops/resize_nearest_neighbor.cc

@@ -37,7 +37,8 @@ inline void ResizeImageNCHW(const OpContext *context,
                             const index_t channels,
                             const float height_scale,
                             const float width_scale,
-                            bool align_corners,
+                            const bool align_corners,
+                            const bool half_pixel_centers,
                             T *output) {
   utils::ThreadPool
       &thread_pool = context->device()->cpu_runtime()->thread_pool();
@@ -52,14 +53,20 @@ inline void ResizeImageNCHW(const OpContext *context,
         T *channel_output_ptr =
             output + (b * channels + c) * out_height * out_width;
         for (index_t y = 0; y < out_height; ++y) {
+          const float in_f_y = half_pixel_centers ?
+                               (static_cast<float>(y) + 0.5f) * height_scale :
+                               y * height_scale;
           const index_t in_y = std::min(
-              (align_corners) ? static_cast<index_t>(roundf(y * height_scale))
-                              : static_cast<index_t>(floorf(y * height_scale)),
+              (align_corners) ? static_cast<index_t>(roundf(in_f_y))
+                              : static_cast<index_t>(floorf(in_f_y)),
               in_height - 1);
           for (int x = 0; x < out_width; ++x) {
+            const float in_f_x = half_pixel_centers ?
+                                 (static_cast<float>(x) + 0.5f) * width_scale :
+                                 x * width_scale;
             const index_t in_x = std::min(
-                (align_corners) ? static_cast<index_t>(roundf(x * width_scale))
-                                : static_cast<index_t>(floorf(x * width_scale)),
+                (align_corners) ? static_cast<index_t>(roundf(in_f_x))
+                                : static_cast<index_t>(floorf(in_f_x)),
                 in_width - 1);
             channel_output_ptr[y * out_width + x] =
                 channel_input_ptr[in_y * in_width + in_x];
@@ -79,6 +86,8 @@ class ResizeNearestNeighborOp<DeviceType::CPU, T> : public Operation {
   explicit ResizeNearestNeighborOp(OpConstructContext *context)
       : Operation(context),
         align_corners_(Operation::GetOptionalArg<bool>("align_corners", false)),
+        half_pixel_centers_(
+            Operation::GetOptionalArg<bool>("half_pixel_centers", false)),
         height_scale_(Operation::GetOptionalArg<float>("height_scale", 0)),
         width_scale_(Operation::GetOptionalArg<float>("width_scale", 0)) {}
 
@@ -144,12 +153,14 @@ class ResizeNearestNeighborOp<DeviceType::CPU, T> : public Operation {
                     height_scale,
                     width_scale,
                     align_corners_,
+                    half_pixel_centers_,
                     output_data);
     return MaceStatus::MACE_SUCCESS;
   }
 
  private:
-  bool align_corners_;
+  const bool align_corners_;
+  const bool half_pixel_centers_;
   float height_scale_;
   float width_scale_;
 };
@@ -164,9 +175,11 @@ class ResizeNearestNeighborOp<DeviceType::GPU, float> : public Operation {
         width_scale_(Operation::GetOptionalArg<float>("width_scale", 0)) {
     bool align_corners = Operation::GetOptionalArg<bool>(
         "align_corners", false);
+    bool half_pixel_centers = Operation::GetOptionalArg<bool>(
+        "half_pixel_centers", false);
     if (context->GetOpMemoryType() == MemoryType::GPU_IMAGE) {
       kernel_ = make_unique<opencl::image::ResizeNearestNeighborKernel>(
-          align_corners);
+          align_corners, half_pixel_centers);
     } else {
       MACE_NOT_IMPLEMENTED;
     }

mace/python/tools/mace_engine_factory.h.jinja2

@@ -29,6 +29,7 @@ namespace mace {
 namespace {{tag}} {
 
 extern const unsigned char *LoadModelData();
+extern int64_t GetModelSize();
 
 extern const std::shared_ptr<NetDef> CreateNet();
 
@@ -77,17 +78,16 @@ __attribute__((deprecated)) MaceStatus CreateMaceEngineFromCode(
 {% if embed_model_data %}
   (void)model_data_file;
   const unsigned char * model_data;
-  const int64_t model_size;
 {% endif %}
   MaceStatus status = MaceStatus::MACE_SUCCESS;
   switch (model_name_map[model_name]) {
 {% for i in range(model_tags |length) %}
-    case {{ i }}:
+    case {{ i }}: {
       net_def = mace::{{model_tags[i]}}::CreateNet();
       engine->reset(new mace::MaceEngine(config));
 {% if embed_model_data %}
       model_data = mace::{{model_tags[i]}}::LoadModelData();
-      model_size = mace::{{model_tags[i]}}::GetModelSize();
+      const int64_t model_size = mace::{{model_tags[i]}}::GetModelSize();
       status = (*engine)->Init(net_def.get(), input_nodes, output_nodes,
                                model_data, model_size);
 {% else %}
@@ -95,6 +95,7 @@ __attribute__((deprecated)) MaceStatus CreateMaceEngineFromCode(
                                model_data_file);
 {% endif %}
       break;
+    }
 {% endfor %}
    default:
      status = MaceStatus::MACE_INVALID_ARGS;
@@ -118,19 +119,19 @@ MaceStatus CreateMaceEngineFromCode(
   std::shared_ptr<NetDef> net_def;
 {% if embed_model_data %}
   const unsigned char * model_data;
-  const int64_t model_size;
   (void)model_weights_data;
+  (void)model_weights_data_size;
 {% endif %}
 
   MaceStatus status = MaceStatus::MACE_SUCCESS;
   switch (model_name_map[model_name]) {
 {% for i in range(model_tags |length) %}
-    case {{ i }}:
+    case {{ i }}: {
       net_def = mace::{{model_tags[i]}}::CreateNet();
       engine->reset(new mace::MaceEngine(config));
 {% if embed_model_data %}
       model_data = mace::{{model_tags[i]}}::LoadModelData();
-      model_size = mace::{{model_tags[i]}}::GetModelSize();
+      const int64_t model_size = mace::{{model_tags[i]}}::GetModelSize();
       status = (*engine)->Init(net_def.get(), input_nodes, output_nodes,
                                model_data, model_size);
 {% else %}
@@ -138,6 +139,7 @@ MaceStatus CreateMaceEngineFromCode(
                                model_weights_data, model_weights_data_size);
 {% endif %}
       break;
+    }
 {% endfor %}
    default:
      status = MaceStatus::MACE_INVALID_ARGS;

tools/python/template/mace_engine_factory.h.jinja2

@@ -29,6 +29,7 @@ namespace mace {
 namespace {{tag}} {
 
 extern const unsigned char *LoadModelData();
+extern int64_t GetModelSize();
 
 extern const std::shared_ptr<NetDef> CreateNet();
 
@@ -77,17 +78,16 @@ __attribute__((deprecated)) MaceStatus CreateMaceEngineFromCode(
 {% if embed_model_data %}
   (void)model_data_file;
   const unsigned char * model_data;
-  const int64_t model_size;
 {% endif %}
   MaceStatus status = MaceStatus::MACE_SUCCESS;
   switch (model_name_map[model_name]) {
 {% for i in range(model_tags |length) %}
-    case {{ i }}:
+    case {{ i }}: {
       net_def = mace::{{model_tags[i]}}::CreateNet();
       engine->reset(new mace::MaceEngine(config));
 {% if embed_model_data %}
       model_data = mace::{{model_tags[i]}}::LoadModelData();
-      model_size = mace::{{model_tags[i]}}::GetModelSize();
+      const int64_t model_size = mace::{{model_tags[i]}}::GetModelSize();
       status = (*engine)->Init(net_def.get(), input_nodes, output_nodes,
                                model_data, model_size);
 {% else %}
@@ -95,6 +95,7 @@ __attribute__((deprecated)) MaceStatus CreateMaceEngineFromCode(
                                model_data_file);
 {% endif %}
       break;
+    }
 {% endfor %}
    default:
      status = MaceStatus::MACE_INVALID_ARGS;
@@ -118,19 +119,19 @@ MACE_API MaceStatus CreateMaceEngineFromCode(
   std::shared_ptr<NetDef> net_def;
 {% if embed_model_data %}
   const unsigned char * model_data;
-  const int64_t model_size;
   (void)model_weights_data;
+  (void)model_weights_data_size;
 {% endif %}
 
   MaceStatus status = MaceStatus::MACE_SUCCESS;
   switch (model_name_map[model_name]) {
 {% for i in range(model_tags |length) %}
-    case {{ i }}:
+    case {{ i }}: {
       net_def = mace::{{model_tags[i]}}::CreateNet();
       engine->reset(new mace::MaceEngine(config));
 {% if embed_model_data %}
       model_data = mace::{{model_tags[i]}}::LoadModelData();
-      model_size = mace::{{model_tags[i]}}::GetModelSize();
+      const int64_t model_size = mace::{{model_tags[i]}}::GetModelSize();
       status = (*engine)->Init(net_def.get(), input_nodes, output_nodes,
                                model_data, model_size);
 {% else %}
@@ -138,6 +139,7 @@ MACE_API MaceStatus CreateMaceEngineFromCode(
                                model_weights_data, model_weights_data_size);
 {% endif %}
       break;
+    }
 {% endfor %}
    default:
      status = MaceStatus::MACE_INVALID_ARGS;

tools/python/transform/base_converter.py

@@ -225,6 +225,7 @@ class MaceKeyword(object):
     mace_align_corners_str = 'align_corners'
     mace_height_scale_str = 'height_scale'
     mace_width_scale_str = 'width_scale'
+    mace_half_pixel_centers_str = 'half_pixel_centers'
     mace_space_batch_block_shape_str = 'block_shape'
     mace_space_depth_block_size_str = 'block_size'
     mace_constant_value_str = 'constant_value'

tools/python/transform/tensorflow_converter.py

@@ -202,8 +202,8 @@ class TensorflowConverter(base_converter.ConverterInterface):
     }
 
     pad_type = {
-        'CONSTANT':  PadType.CONSTANT,
-        'REFLECT':   PadType.REFLECT,
+        'CONSTANT': PadType.CONSTANT,
+        'REFLECT': PadType.REFLECT,
         'SYMMETRIC': PadType.SYMMETRIC
     }
 
@@ -267,7 +267,8 @@ class TensorflowConverter(base_converter.ConverterInterface):
             TFOpType.Reshape.name: self.convert_reshape,
             TFOpType.ResizeBicubic.name: self.convert_resize_bicubic,
             TFOpType.ResizeBilinear.name: self.convert_resize_bilinear,
-            TFOpType.ResizeNearestNeighbor.name: self.convert_resize_nearest_neighbor,  # noqa
+            TFOpType.ResizeNearestNeighbor.name:
+                self.convert_resize_nearest_neighbor,
             TFOpType.ReverseV2.name: self.convert_reverse,
             TFOpType.Select.name: self.convert_select,
             TFOpType.Shape.name: self.convert_shape,
@@ -715,6 +716,19 @@ class TensorflowConverter(base_converter.ConverterInterface):
         op = self.convert_general_op(tf_op)
         op.type = MaceOp.Softmax.name
 
+    def add_resize_args(self, op, tf_op):
+        align_corners_arg = op.arg.add()
+        align_corners_arg.name = MaceKeyword.mace_align_corners_str
+        align_corners_arg.i = tf_op.get_attr(tf_align_corners)
+        try:
+            half_pixel_centers = tf_op.get_attr('half_pixel_centers')
+            half_pixel_centers_arg = op.arg.add()
+            half_pixel_centers_arg.name = \
+                MaceKeyword.mace_half_pixel_centers_str
+            half_pixel_centers_arg.i = half_pixel_centers
+        except ValueError:
+            pass
+
     def convert_resize_bicubic(self, tf_op):
         op = self.convert_general_op(tf_op)
         op.type = MaceOp.ResizeBicubic.name
@@ -725,9 +739,7 @@ class TensorflowConverter(base_converter.ConverterInterface):
         size_value = tf_op.inputs[1].eval().astype(np.int32)
         size_arg.ints.extend(size_value)
         self._skip_tensor.add(tf_op.inputs[1].name)
-        align_corners_arg = op.arg.add()
-        align_corners_arg.name = MaceKeyword.mace_align_corners_str
-        align_corners_arg.i = tf_op.get_attr(tf_align_corners)
+        self.add_resize_args(op, tf_op)
 
     def convert_resize_bilinear(self, tf_op):
         op = self.convert_general_op(tf_op)
@@ -739,17 +751,12 @@ class TensorflowConverter(base_converter.ConverterInterface):
         size_value = tf_op.inputs[1].eval().astype(np.int32)
         size_arg.ints.extend(size_value)
         self._skip_tensor.add(tf_op.inputs[1].name)
-        align_corners_arg = op.arg.add()
-        align_corners_arg.name = MaceKeyword.mace_align_corners_str
-        align_corners_arg.i = tf_op.get_attr(tf_align_corners)
+        self.add_resize_args(op, tf_op)
 
     def convert_resize_nearest_neighbor(self, tf_op):
         op = self.convert_general_op(tf_op)
         op.type = MaceOp.ResizeNearestNeighbor.name
-
-        align_corners_arg = op.arg.add()
-        align_corners_arg.name = MaceKeyword.mace_align_corners_str
-        align_corners_arg.i = tf_op.get_attr(tf_align_corners)
+        self.add_resize_args(op, tf_op)
 
     def convert_space_batch(self, tf_op):
         op = self.convert_general_op(tf_op)