diff --git a/rtdetr_paddle/README.md b/rtdetr_paddle/README.md
index 072184f4..2be1e39d 100644
--- a/rtdetr_paddle/README.md
+++ b/rtdetr_paddle/README.md
@@ -108,6 +108,11 @@ path/to/custom/data/
 
 3. Modify model config [`pretrain_weights`](configs/rtdetr/_base_/rtdetr_r50vd.yml) to coco pretrained parameters url in model zoo.
 
+```bash
+# or modified in command line
+
+fleetrun --gpus=0,1,2,3 tools/train.py -c configs/rtdetr/rtdetr_r50vd_6x_coco.yml -o pretrain_weights=https://bj.bcebos.com/v1/paddledet/models/rtdetr_r50vd_6x_coco.pdparams --eval
+```
 </details>
 
 
@@ -171,10 +176,11 @@ trtexec --onnx=./rtdetr_r50vd_6x_coco.onnx \
 <details>
 <summary>1. Parameters and FLOPs </summary>
 
-1. Find and modify [paddle flops source code](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/hapi/dynamic_flops.py#L28)
+1. Find and modify paddle [`dynamic_flops.py` ](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/hapi/dynamic_flops.py#L28) source code in your local machine
 
 ```python
-# anaconda3/lib/python3.8/site-packages/paddle/hapi/dynamic_flops.py
+# eg. /path/to/anaconda3/lib/python3.8/site-packages/paddle/hapi/dynamic_flops.py
+
 def flops(net, input_size, inputs=None, custom_ops=None, print_detail=False):
     if isinstance(net, nn.Layer):
         # If net is a dy2stat model, net.forward is StaticFunction instance,
diff --git a/rtdetr_paddle/configs/rtdetr/_base_/rtdetr_r50vd.yml b/rtdetr_paddle/configs/rtdetr/_base_/rtdetr_r50vd.yml
index fc5fb3ad..7859dfb4 100644
--- a/rtdetr_paddle/configs/rtdetr/_base_/rtdetr_r50vd.yml
+++ b/rtdetr_paddle/configs/rtdetr/_base_/rtdetr_r50vd.yml
@@ -7,7 +7,7 @@ ema_decay_type: "exponential"
 ema_filter_no_grad: True
 hidden_dim: 256
 use_focal_loss: True
-eval_size: [640, 640]
+eval_size: [640, 640] # h, w
 
 
 DETR:
diff --git a/rtdetr_paddle/configs/rtdetr/_base_/rtdetr_reader.yml b/rtdetr_paddle/configs/rtdetr/_base_/rtdetr_reader.yml
index 1b6f86c8..b1a2a002 100644
--- a/rtdetr_paddle/configs/rtdetr/_base_/rtdetr_reader.yml
+++ b/rtdetr_paddle/configs/rtdetr/_base_/rtdetr_reader.yml
@@ -22,7 +22,7 @@ TrainReader:
 EvalReader:
   sample_transforms:
     - Decode: {}
-    - Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
+    - Resize: {target_size: [640, 640], keep_ratio: False, interp: 2} # target_size: (h, w)
     - NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
     - Permute: {}
   batch_size: 4
diff --git a/rtdetr_paddle/ppdet/engine/trainer.py b/rtdetr_paddle/ppdet/engine/trainer.py
index 364ca87d..c22b1af0 100644
--- a/rtdetr_paddle/ppdet/engine/trainer.py
+++ b/rtdetr_paddle/ppdet/engine/trainer.py
@@ -650,7 +650,8 @@ def setup_metrics_for_loader():
         for step_id, data in enumerate(tqdm(loader)):
             self.status['step_id'] = step_id
             # forward
-            outs = self.model(data)
+            with paddle.no_grad():
+                outs = self.model(data)
 
             outs['bbox'] = outs['bbox'].numpy()  # only in test mode
             shift_amount = data['st_pix']
@@ -798,10 +799,12 @@ def setup_metrics_for_loader():
         for step_id, data in enumerate(tqdm(loader)):
             self.status['step_id'] = step_id
             # forward
-            if hasattr(self.model, 'modelTeacher'):
-                outs = self.model.modelTeacher(data)
-            else:
-                outs = self.model(data)
+            with paddle.no_grad():
+                if hasattr(self.model, 'modelTeacher'):
+                    outs = self.model.modelTeacher(data)
+                else:
+                    outs = self.model(data)
+
             for _m in metrics:
                 _m.update(data, outs)
 
diff --git a/rtdetr_paddle/ppdet/utils/visualizer.py b/rtdetr_paddle/ppdet/utils/visualizer.py
index bd0ac307..406589da 100644
--- a/rtdetr_paddle/ppdet/utils/visualizer.py
+++ b/rtdetr_paddle/ppdet/utils/visualizer.py
@@ -127,12 +127,8 @@ def draw_bbox(image, im_id, catid2name, bboxes, threshold):
         # draw label
         text = "{} {:.2f}".format(catid2name[catid], score)
         # tw, th = draw.textsize(text)
-        
-        if int(PIL.__version__.split('.')[0]) < 10:
-            tw, th = draw.textsize(text)
-        else:
-            left, top, right, bottom = draw.textbbox((0, 0), text)
-            tw, th = right - left, bottom - top
+        left, top, right, bottom = draw.textbbox((0, 0), text)
+        tw, th = right - left, bottom - top
 
         draw.rectangle(
             [(xmin + 1, ymin - th), (xmin + tw + 1, ymin)], fill=color)