From c9d4676bee930d8678c5c590e75c31e7a40a77be Mon Sep 17 00:00:00 2001
From: dengkaipeng <dengkaipeng@baidu.com>
Date: Tue, 19 Feb 2019 13:26:15 +0000
Subject: [PATCH] fix multi batch idx error. test=develop

---
 paddle/fluid/operators/detection/yolo_box_op.cu    | 14 +++++++-------
 .../fluid/tests/unittests/test_yolo_box_op.py      |  4 ++--
 2 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/paddle/fluid/operators/detection/yolo_box_op.cu b/paddle/fluid/operators/detection/yolo_box_op.cu
index fbe934c7e..a4513bd2f 100644
--- a/paddle/fluid/operators/detection/yolo_box_op.cu
+++ b/paddle/fluid/operators/detection/yolo_box_op.cu
@@ -22,14 +22,14 @@ using Tensor = framework::Tensor;
 
 template <typename T>
 __global__ void KeYoloBoxFw(const T* input, const int* imgsize, T* boxes,
-                            T* scores, const float conf_thresh,
-                            const int* anchors, const int h, const int w,
+                            T* scores, const float conf_thresh, const int* anchors, 
+			    const int n, const int h, const int w,
                             const int an_num, const int class_num,
                             const int box_num, int input_size) {
   int tid = blockIdx.x * blockDim.x + threadIdx.x;
   int stride = blockDim.x * gridDim.x;
   T box[4];
-  for (; tid < box_num; tid += stride) {
+  for (; tid < n * box_num; tid += stride) {
     int grid_num = h * w;
     int i = tid / box_num;
     int j = (tid % box_num) / grid_num;
@@ -99,12 +99,12 @@ class YoloBoxOpCUDAKernel : public framework::OpKernel<T> {
     set_zero(dev_ctx, boxes, static_cast<T>(0));
     set_zero(dev_ctx, scores, static_cast<T>(0));
 
-    int grid_dim = (n * box_num + 4 - 1) / 4;
-    grid_dim = grid_dim > 2 ? 2 : grid_dim;
+    int grid_dim = (n * box_num + 512 - 1) / 512;
+    grid_dim = grid_dim > 8 ? 8 : grid_dim;
 
-    KeYoloBoxFw<T><<<grid_dim, 4, 0, ctx.cuda_device_context().stream()>>>(
+    KeYoloBoxFw<T><<<grid_dim, 512, 0, ctx.cuda_device_context().stream()>>>(
 	input_data, imgsize_data, boxes_data, scores_data, conf_thresh, 
-	anchors_data, h, w, an_num, class_num, n * box_num, input_size);	
+	anchors_data, n, h, w, an_num, class_num, box_num, input_size);	
   }
 };
 
diff --git a/python/paddle/fluid/tests/unittests/test_yolo_box_op.py b/python/paddle/fluid/tests/unittests/test_yolo_box_op.py
index e28c05e3e..a1da4f64b 100644
--- a/python/paddle/fluid/tests/unittests/test_yolo_box_op.py
+++ b/python/paddle/fluid/tests/unittests/test_yolo_box_op.py
@@ -99,11 +99,11 @@ class TestYoloBoxOp(OpTest):
     def initTestCase(self):
         self.anchors = [10, 13, 16, 30, 33, 23]
         an_num = int(len(self.anchors) // 2)
-        self.batch_size = 1
+        self.batch_size = 32
         self.class_num = 2
         self.conf_thresh = 0.5
         self.downsample = 32
-        self.x_shape = (self.batch_size, an_num * (5 + self.class_num), 2, 2)
+        self.x_shape = (self.batch_size, an_num * (5 + self.class_num), 13, 13)
         self.imgsize_shape = (self.batch_size, 2)
 
 
-- 
GitLab