diff --git a/ppdet/modeling/losses/pisa_utils.py b/ppdet/modeling/losses/pisa_utils.py
index 6d6345037eeaa740caeddc1cd3b1ea2bf0c1ca88..431eaf99d515d1dcf4cac174f702102f4383661d 100644
--- a/ppdet/modeling/losses/pisa_utils.py
+++ b/ppdet/modeling/losses/pisa_utils.py
@@ -59,7 +59,7 @@ def get_isr_p_func(pos_iou_thresh=0.25, bias=0, k=2):
             weight_factor = np.clip(max_l_num - l_max_iou_rank, 0.,
                                     None) / max_l_num
             weight_factor = np.power(bias + (1 - bias) * weight_factor, k)
-            pos_weights[l_inds] *= weight_factor * 1.2
+            pos_weights[l_inds] *= weight_factor
         pos_weights = pos_weights / np.mean(pos_weights)
         all_pos_weights[pos_mask] = pos_weights
 
diff --git a/ppdet/modeling/losses/yolo_loss.py b/ppdet/modeling/losses/yolo_loss.py
index 5f31a498d617af3722bfcac6a4cc18c45c4b954c..b2f0191e2a5d35c38b7efa6b741971b9d5ad1625 100644
--- a/ppdet/modeling/losses/yolo_loss.py
+++ b/ppdet/modeling/losses/yolo_loss.py
@@ -142,21 +142,24 @@ class YOLOv3Loss(object):
             iou = self._calc_iou(output, target, gt_box, anchors, batch_size,
                                  num_classes, downsample, scale_x_y)
 
-            # sorted_iou, sorted_gt_inds = fluid.layers.argsort(iou, axis=-1, descending=True)
-            # max_iou = sorted_iou[:, :, 0:1]
-            # gt_inds = fluid.layers.cast(sorted_gt_inds[:, :, 0:1], dtype='float32')
-            # pred_cls = fluid.layers.argmax(cls, axis=-1)
-            # pred_cls = fluid.layers.reshape(pred_cls, [batch_size, -1, 1])
-            # pred_cls = fluid.layers.cast(pred_cls, dtype='float32')
-            # isr_p_input = fluid.layers.concat([max_iou, gt_inds, pred_cls], axis=-1)
-            # isr_p = get_isr_p_func()
-            # pos_weights = fluid.layers.zeros_like(max_iou)
-            # fluid.layers.py_func(isr_p, isr_p_input, pos_weights)
-            #
-            # tobj_shape = fluid.layers.shape(tobj)
-            # pos_weights = fluid.layers.reshape(pos_weights, (-1, an_num, tobj_shape[2],
-            #                                            tobj_shape[3]))
-            # tobj = tobj * pos_weights
+            sorted_iou, sorted_gt_inds = fluid.layers.argsort(
+                iou, axis=-1, descending=True)
+            max_iou = sorted_iou[:, :, 0:1]
+            gt_inds = fluid.layers.cast(
+                sorted_gt_inds[:, :, 0:1], dtype='float32')
+            pred_cls = fluid.layers.argmax(cls, axis=-1)
+            pred_cls = fluid.layers.reshape(pred_cls, [batch_size, -1, 1])
+            pred_cls = fluid.layers.cast(pred_cls, dtype='float32')
+            isr_p_input = fluid.layers.concat(
+                [max_iou, gt_inds, pred_cls], axis=-1)
+            isr_p = get_isr_p_func()
+            pos_weights = fluid.layers.zeros_like(max_iou)
+            fluid.layers.py_func(isr_p, isr_p_input, pos_weights)
+
+            tobj_shape = fluid.layers.shape(tobj)
+            pos_weights = fluid.layers.reshape(pos_weights, (
+                -1, an_num, tobj_shape[2], tobj_shape[3]))
+            tobj = tobj * pos_weights
 
             # isr_tobj = tobj * pos_weights
             # loss_cls = fluid.layers.sigmoid_cross_entropy_with_logits(cls, tcls)