change output shape to [2, layer_height, layer_width, num_priors, 4]

Author: wanghaox

paddle/operators/prior_box_op.cc

@@ -93,17 +93,12 @@ class PriorBoxOp : public framework::OperatorWithKernel {
     const int layer_height = input_dims[2];
     const int layer_width = input_dims[3];
 
-    std::vector<int64_t> dim_vec(3);
-    // Since all images in a batch has same height and width, we only need to
-    // generate one set of priors which can be shared across all images.
-    dim_vec[0] = 1;
-    // 2 channels. First channel stores the mean of each prior coordinate.
-    // Second channel stores the variance of each prior coordinate.
-    dim_vec[1] = 2;
-    dim_vec[2] = layer_width * layer_height * num_priors * 4;
-    PADDLE_ENFORCE_GT(dim_vec[2], 0,
-                      "output_dim[2] must larger than 0."
-                      "check your data dims");
+    std::vector<int64_t> dim_vec(5);
+    dim_vec[0] = 2;
+    dim_vec[1] = layer_height;
+    dim_vec[2] = layer_width;
+    dim_vec[3] = num_priors;
+    dim_vec[4] = 4;
     auto output_dim = framework::make_ddim(dim_vec);
     ctx->SetOutputDim("Out", output_dim);
   }
@@ -130,7 +125,8 @@ class PriorBoxOpMaker : public framework::OpProtoAndCheckerMaker {
              "the input image data of PriorBoxOp, The format is NCHW.");
     AddOutput("Out",
               "(Tensor, default Tensor<float>), the output prior boxes of "
-              "PriorBoxOp.");
+              "PriorBoxOp. The format is [2, layer_height, layer_width, "
+              "num_priors, 4]");
     AddAttr<std::vector<int>>("min_sizes", "(vector<int>) ",
                               "List of min sizes of generated prior boxes.");
     AddAttr<std::vector<int>>("max_sizes", "(vector<int>) ",

paddle/operators/prior_box_op.h

@@ -15,7 +15,6 @@ limitations under the License. */
 #pragma once
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/math/math_function.h"
-// #include "paddle/operators/strided_memcpy.h"
 
 namespace paddle {
 namespace operators {
@@ -94,50 +93,52 @@ class PriorBoxOpKernel : public framework::OpKernel<T> {
       num_priors += max_sizes.size();
     }
 
-    int dim = layer_height * layer_width * num_priors * 4;
-
     T* output_data = nullptr;
     framework::Tensor output_cpu;
+    framework::Tensor* output_tensor;
     out->mutable_data<T>(ctx.GetPlace());
     if (platform::is_gpu_place(ctx.GetPlace())) {
-      output_data =
-          output_cpu.mutable_data<T>(out->dims(), platform::CPUPlace());
+      output_cpu.mutable_data<T>(out->dims(), platform::CPUPlace());
+      output_tensor = &output_cpu;
     } else {
-      output_data = out->mutable_data<T>(ctx.GetPlace());
+      output_tensor = out;
     }
 
-    int idx = 0;
+    auto e_out = framework::EigenTensor<T, 5>::From(*output_tensor);
     for (int h = 0; h < layer_height; ++h) {
       for (int w = 0; w < layer_width; ++w) {
         float center_x = (w + offset) * step_width;
         float center_y = (h + offset) * step_height;
         float box_width, box_height;
+        int idx = 0;
         for (size_t s = 0; s < min_sizes.size(); ++s) {
           int min_size = min_sizes[s];
           // first prior: aspect_ratio = 1, size = min_size
           box_width = box_height = min_size;
           // xmin
-          output_data[idx++] = (center_x - box_width / 2.) / img_width;
+          e_out(0, h, w, idx, 0) = (center_x - box_width / 2.) / img_width;
           // ymin
-          output_data[idx++] = (center_y - box_height / 2.) / img_height;
+          e_out(0, h, w, idx, 1) = (center_y - box_height / 2.) / img_height;
           // xmax
-          output_data[idx++] = (center_x + box_width / 2.) / img_width;
+          e_out(0, h, w, idx, 2) = (center_x + box_width / 2.) / img_width;
           // ymax
-          output_data[idx++] = (center_y + box_height / 2.) / img_height;
+          e_out(0, h, w, idx, 3) = (center_y + box_height / 2.) / img_height;
 
+          idx++;
           if (max_sizes.size() > 0) {
             int max_size = max_sizes[s];
             // second prior: aspect_ratio = 1,
             // size = sqrt(min_size * max_size)
             box_width = box_height = sqrt(min_size * max_size);
             // xmin
-            output_data[idx++] = (center_x - box_width / 2.) / img_width;
+            e_out(0, h, w, idx, 0) = (center_x - box_width / 2.) / img_width;
             // ymin
-            output_data[idx++] = (center_y - box_height / 2.) / img_height;
+            e_out(0, h, w, idx, 1) = (center_y - box_height / 2.) / img_height;
             // xmax
-            output_data[idx++] = (center_x + box_width / 2.) / img_width;
+            e_out(0, h, w, idx, 2) = (center_x + box_width / 2.) / img_width;
             // ymax
-            output_data[idx++] = (center_y + box_height / 2.) / img_height;
+            e_out(0, h, w, idx, 3) = (center_y + box_height / 2.) / img_height;
+            idx++;
           }
 
           // rest of priors
@@ -149,40 +150,46 @@ class PriorBoxOpKernel : public framework::OpKernel<T> {
             box_width = min_size * sqrt(ar);
             box_height = min_size / sqrt(ar);
             // xmin
-            output_data[idx++] = (center_x - box_width / 2.) / img_width;
+            e_out(0, h, w, idx, 0) = (center_x - box_width / 2.) / img_width;
             // ymin
-            output_data[idx++] = (center_y - box_height / 2.) / img_height;
+            e_out(0, h, w, idx, 1) = (center_y - box_height / 2.) / img_height;
             // xmax
-            output_data[idx++] = (center_x + box_width / 2.) / img_width;
+            e_out(0, h, w, idx, 2) = (center_x + box_width / 2.) / img_width;
             // ymax
-            output_data[idx++] = (center_y + box_height / 2.) / img_height;
+            e_out(0, h, w, idx, 3) = (center_y + box_height / 2.) / img_height;
+            idx++;
           }
         }
       }
     }
 
     // clip the prior's coordidate such that it is within [0, 1]
     if (clip) {
-      for (int d = 0; d < dim; ++d) {
-        output_data[d] = std::min<T>(std::max<T>(output_data[d], 0.), 1.);
+      for (int h = 0; h < layer_height; ++h) {
+        for (int w = 0; w < layer_width; ++w) {
+          for (int i = 0; i < num_priors; ++i) {
+            for (int j = 0; j < 4; ++j) {
+              e_out(0, h, w, i, j) =
+                  std::min<T>(std::max<T>(e_out(0, h, w, i, j), 0.), 1.);
+            }
+          }
+        }
       }
-    }
 
-    // set the variance.
-    auto output_stride = framework::stride(out->dims());
-    output_data += output_stride[1];
-    if (variances.size() == 1) {
-      for (int i = 0; i < dim; ++i) {
-        output_data[i] = variances[0];
+      // set the variance.
+      auto output_stride = framework::stride(out->dims());
+      output_data += output_stride[1];
+      if (variances.size() == 1) {
+        variances.resize(4);
+        variances[1] = variances[0];
+        variances[2] = variances[0];
+        variances[3] = variances[0];
       }
-    } else {
-      int count = 0;
       for (int h = 0; h < layer_height; ++h) {
         for (int w = 0; w < layer_width; ++w) {
           for (int i = 0; i < num_priors; ++i) {
             for (int j = 0; j < 4; ++j) {
-              output_data[count] = variances[j];
-              ++count;
+              e_out(1, h, w, i, j) = variances[j];
             }
           }
         }

python/paddle/v2/fluid/tests/test_prior_box_op.py

@@ -81,33 +81,30 @@ def init_test_input(self):
              self.layer_h)).astype('float32')
 
     def init_test_output(self):
-        dim = self.layer_w * self.layer_h * self.num_priors * 4
-        out_dim = (1, 2, dim)
+        out_dim = (2, self.layer_h, self.layer_w, self.num_priors, 4)
         output = np.zeros(out_dim).astype('float32')
 
         idx = 0
         for h in range(self.layer_h):
             for w in range(self.layer_w):
                 center_x = (w + self.offset) * self.step_w
                 center_y = (h + self.offset) * self.step_h
+                idx = 0
                 for s in range(len(self.min_sizes)):
                     min_size = self.min_sizes[s]
                     # first prior: aspect_ratio = 1, size = min_size
                     box_width = box_height = min_size
                     # xmin
-                    output[0, 0, idx] = (
+                    output[0, h, w, idx, 0] = (
                         center_x - box_width / 2.) / self.image_w
-                    idx += 1
                     # ymin
-                    output[0, 0, idx] = (
+                    output[0, h, w, idx, 1] = (
                         center_y - box_height / 2.) / self.image_h
-                    idx += 1
                     # xmax
-                    output[0, 0, idx] = (
+                    output[0, h, w, idx, 2] = (
                         center_x + box_width / 2.) / self.image_w
-                    idx += 1
                     # ymax
-                    output[0, 0, idx] = (
+                    output[0, h, w, idx, 3] = (
                         center_y + box_height / 2.) / self.image_h
                     idx += 1
 
@@ -117,19 +114,16 @@ def init_test_output(self):
                         # size = sqrt(min_size * max_size)
                         box_width = box_height = math.sqrt(min_size * max_size)
                         # xmin
-                        output[0, 0, idx] = (
+                        output[0, h, w, idx, 0] = (
                             center_x - box_width / 2.) / self.image_w
-                        idx += 1
                         # ymin
-                        output[0, 0, idx] = (
+                        output[0, h, w, idx, 1] = (
                             center_y - box_height / 2.) / self.image_h
-                        idx += 1
                         # xmax
-                        output[0, 0, idx] = (
+                        output[0, h, w, idx, 2] = (
                             center_x + box_width / 2.) / self.image_w
-                        idx += 1
                         # ymax
-                        output[0, 0, idx] = (
+                        output[0, h, w, idx, 3] = (
                             center_y + box_height / 2.) / self.image_h
                         idx += 1
 
@@ -141,37 +135,35 @@ def init_test_output(self):
                         box_width = min_size * math.sqrt(ar)
                         box_height = min_size / math.sqrt(ar)
                         # xmin
-                        output[0, 0, idx] = (
+                        output[0, h, w, idx, 0] = (
                             center_x - box_width / 2.) / self.image_w
-                        idx += 1
                         # ymin
-                        output[0, 0, idx] = (
+                        output[0, h, w, idx, 1] = (
                             center_y - box_height / 2.) / self.image_h
-                        idx += 1
                         # xmax
-                        output[0, 0, idx] = (
+                        output[0, h, w, idx, 2] = (
                             center_x + box_width / 2.) / self.image_w
-                        idx += 1
                         # ymax
-                        output[0, 0, idx] = (
+                        output[0, h, w, idx, 3] = (
                             center_y + box_height / 2.) / self.image_h
                         idx += 1
         # clip the prior's coordidate such that it is within[0, 1]
         if self.clip:
-            for d in range(dim):
-                output[0, 0, d] = min(max(output[0, 0, d], 0), 1)
-        # set the variance.
-        if len(self.variances) == 1:
-            for i in range(dim):
-                output[0, 1, i] = self.variances[0]
-        else:
-            count = 0
             for h in range(self.layer_h):
                 for w in range(self.layer_w):
                     for i in range(self.num_priors):
                         for j in range(4):
-                            output[0, 1, count] = self.variances[j]
-                            count += 1
+                            output[0, h, w, i, j] = min(
+                                max(output[0, h, w, i, j], 0), 1)
+        # set the variance.
+        for h in range(self.layer_h):
+            for w in range(self.layer_w):
+                for i in range(self.num_priors):
+                    for j in range(4):
+                        if len(self.variances) == 1:
+                            output[1, h, w, i, j] = self.variances[0]
+                        else:
+                            output[1, h, w, i, j] = self.variances[j]
         self.output = output.astype('float32')