optimized iou_similarity_op (#10231)

paddle/fluid/operators/iou_similarity_op.h

@@ -41,22 +41,24 @@ struct IOUSimilarityFunctor {
   IOUSimilarityFunctor(const T* x, const T* y, T* z, int cols)
       : x_(x), y_(y), z_(z), cols_(static_cast<size_t>(cols)) {}
 
-  inline HOSTDEVICE void operator()(size_t row_id) const {
+  inline HOSTDEVICE void operator()(size_t tid) const {
+    size_t row_id = tid / cols_;
+    size_t col_id = tid % cols_;
+
     T x_min1 = x_[row_id * 4];
     T y_min1 = x_[row_id * 4 + 1];
     T x_max1 = x_[row_id * 4 + 2];
     T y_max1 = x_[row_id * 4 + 3];
-    for (size_t i = 0; i < cols_; ++i) {
-      T x_min2 = y_[i * 4];
-      T y_min2 = y_[i * 4 + 1];
-      T x_max2 = y_[i * 4 + 2];
-      T y_max2 = y_[i * 4 + 3];
 
-      T sim = IOUSimilarity(x_min1, y_min1, x_max1, y_max1, x_min2, y_min2,
-                            x_max2, y_max2);
+    T x_min2 = y_[col_id * 4];
+    T y_min2 = y_[col_id * 4 + 1];
+    T x_max2 = y_[col_id * 4 + 2];
+    T y_max2 = y_[col_id * 4 + 3];
+
+    T sim = IOUSimilarity(x_min1, y_min1, x_max1, y_max1, x_min2, y_min2,
+                          x_max2, y_max2);
 
-      z_[row_id * cols_ + i] = sim;
-    }
+    z_[row_id * cols_ + col_id] = sim;
   }
   const T* x_;
   const T* y_;
@@ -81,7 +83,7 @@ class IOUSimilarityKernel : public framework::OpKernel<T> {
                                     out->mutable_data<T>(ctx.GetPlace()), y_n);
 
     platform::ForRange<DeviceContext> for_range(
-        static_cast<const DeviceContext&>(ctx.device_context()), x_n);
+        static_cast<const DeviceContext&>(ctx.device_context()), x_n * y_n);
     for_range(functor);
   }
 };  // namespace operators

python/paddle/fluid/tests/unittests/test_iou_similarity_op.py

@@ -14,6 +14,7 @@
 
 import unittest
 import numpy as np
+import numpy.random as random
 import sys
 import math
 from op_test import OpTest
@@ -25,14 +26,27 @@ class TestIOUSimilarityOp(OpTest):
 
     def setUp(self):
         self.op_type = "iou_similarity"
-        self.boxes1 = np.array(
-            [[4.0, 3.0, 7.0, 5.0], [5.0, 6.0, 10.0, 7.0]]).astype('float32')
-        self.boxes2 = np.array([[3.0, 4.0, 6.0, 8.0], [14.0, 14.0, 15.0, 15.0],
-                                [0.0, 0.0, 20.0, 20.0]]).astype('float32')
-        self.output = np.array(
-            [[2.0 / 16.0, 0, 6.0 / 400.0],
-             [1.0 / 16.0, 0.0, 5.0 / 400.0]]).astype('float32')
-
+        self.boxes1 = random.rand(2, 4).astype('float32')
+        self.boxes2 = random.rand(3, 4).astype('float32')
+        self.output = random.rand(2, 3).astype('float32')
+        for row in range(self.boxes1.shape[0]):
+            for col in range(self.boxes2.shape[0]):
+                xmin1, ymin1, xmax1, ymax1 = self.boxes1[row]
+                xmin2, ymin2, xmax2, ymax2 = self.boxes2[col]
+                area1 = (ymax1 - ymin1) * (xmax1 - xmin1)
+                area2 = (ymax2 - ymin2) * (xmax2 - xmin2)
+                inter_xmax = min(xmax1, xmax2)
+                inter_ymax = min(ymax1, ymax2)
+                inter_xmin = max(xmin1, xmin2)
+                inter_ymin = max(ymin1, ymin2)
+                inter_height = inter_ymax - inter_ymin
+                inter_width = inter_xmax - inter_xmin
+                inter_height = max(inter_height, 0)
+                inter_width = max(inter_width, 0)
+                inter_area = inter_width * inter_height
+                union_area = area1 + area2 - inter_area
+                sim_score = inter_area / union_area
+                self.output[row, col] = sim_score
         self.inputs = {'X': self.boxes1, 'Y': self.boxes2}
 
         self.outputs = {'Out': self.output}