polish the label_smooth (#17138)

* polish the label_smooth

test=develop

* polish code

test=develop

paddle/fluid/operators/detection/yolov3_loss_op.h

@@ -282,8 +282,9 @@ class Yolov3LossKernel : public framework::OpKernel<T> {
     T label_pos = 1.0;
     T label_neg = 0.0;
     if (use_label_smooth) {
-      label_pos = 1.0 - 1.0 / static_cast<T>(class_num);
+      T smooth_weight = std::min(1.0 / static_cast<T>(class_num), 1.0 / 40);
-      label_neg = 1.0 / static_cast<T>(class_num);
+      label_pos = 1.0 - smooth_weight;
+      label_neg = smooth_weight;
     }
 
     const T* input_data = input->data<T>();
@@ -437,8 +438,9 @@ class Yolov3LossGradKernel : public framework::OpKernel<T> {
     T label_pos = 1.0;
     T label_neg = 0.0;
     if (use_label_smooth) {
-      label_pos = 1.0 - 1.0 / static_cast<T>(class_num);
+      T smooth_weight = std::min(1.0 / static_cast<T>(class_num), 1.0 / 40);
-      label_neg = 1.0 / static_cast<T>(class_num);
+      label_pos = 1.0 - smooth_weight;
+      label_neg = smooth_weight;
     }
 
     const T* input_data = input->data<T>();

python/paddle/fluid/tests/unittests/test_yolov3_loss_op.py

@@ -81,8 +81,9 @@ def YOLOv3Loss(x, gtbox, gtlabel, gtscore, attrs):
     x = x.reshape((n, mask_num, 5 + class_num, h, w)).transpose((0, 1, 3, 4, 2))
     loss = np.zeros((n)).astype('float32')
 
-    label_pos = 1.0 - 1.0 / class_num if use_label_smooth else 1.0
+    smooth_weight = min(1.0 / class_num, 1.0 / 40)
-    label_neg = 1.0 / class_num if use_label_smooth else 0.0
+    label_pos = 1.0 - smooth_weight if use_label_smooth else 1.0
+    label_neg = smooth_weight if use_label_smooth else 0.0
 
     pred_box = x[:, :, :, :, :4].copy()
     grid_x = np.tile(np.arange(w).reshape((1, w)), (h, 1))