diff --git a/configs/_base_/readers/mask_reader.yml b/configs/_base_/readers/mask_reader.yml
index c4908308ff5b9853996653b0dfbba7346a2a2f6a..07159041f1bcc67a733da01db3d5481853af62f2 100644
--- a/configs/_base_/readers/mask_reader.yml
+++ b/configs/_base_/readers/mask_reader.yml
@@ -12,7 +12,7 @@ TrainReader:
   - PadBatch: {pad_to_stride: 32, use_padded_im_info: false, pad_gt: true}
   batch_size: 1
   shuffle: true
-  drop_last: false
+  drop_last: true
 
 
 EvalReader:
diff --git a/ppdet/data/reader.py b/ppdet/data/reader.py
index 06e642acca17041a3d279f52945bd7068d4876f2..c279e5abd9b70119d5b898f0ba4614d369412691 100644
--- a/ppdet/data/reader.py
+++ b/ppdet/data/reader.py
@@ -135,7 +135,7 @@ class TrainReader(BaseDataLoader):
                  batch_transforms=None,
                  batch_size=1,
                  shuffle=True,
-                 drop_last=False,
+                 drop_last=True,
                  drop_empty=True,
                  num_classes=81,
                  with_background=True):
diff --git a/ppdet/data/source/coco.py b/ppdet/data/source/coco.py
index 49ac310256630400ed37777d2df141e36ea519ca..69e383ac795671ce700b39672b157d8dbe8098f8 100644
--- a/ppdet/data/source/coco.py
+++ b/ppdet/data/source/coco.py
@@ -129,10 +129,9 @@ class COCODataSet(DetDataset):
                     gt_bbox[i, :] = box['clean_bbox']
                     is_crowd[i][0] = box['iscrowd']
                     # check RLE format 
-                    if box['iscrowd'] == 1:
+                    if 'segmentation' in box and box['iscrowd'] == 1:
                         gt_poly[i] = [[0.0, 0.0], ]
-                        continue
-                    if 'segmentation' in box:
+                    elif 'segmentation' in box:
                         gt_poly[i] = box['segmentation']
 
                 if not any(gt_poly):
diff --git a/ppdet/data/source/dataset.py b/ppdet/data/source/dataset.py
index c8a2a9599688ccc2b8a442ca7f17750d7abf002d..f5ef1aad5f0982b7d5b1fbe623d3fe608a08e01c 100644
--- a/ppdet/data/source/dataset.py
+++ b/ppdet/data/source/dataset.py
@@ -22,6 +22,7 @@ except Exception:
 from paddle.io import Dataset
 from ppdet.core.workspace import register, serializable
 from ppdet.utils.download import get_dataset_path
+import copy
 
 
 @serializable
@@ -45,7 +46,7 @@ class DetDataset(Dataset):
 
     def __getitem__(self, idx):
         # data batch
-        roidb = self.roidbs[idx]
+        roidb = copy.deepcopy(self.roidbs[idx])
         # data augment
         roidb = self.transform(roidb)
         # data item 
diff --git a/ppdet/modeling/architecture/meta_arch.py b/ppdet/modeling/architecture/meta_arch.py
index ff6776269be857013c8e861821df9cfd7c6653aa..51d1852478c456deaec5203e785b46deed8ae138 100644
--- a/ppdet/modeling/architecture/meta_arch.py
+++ b/ppdet/modeling/architecture/meta_arch.py
@@ -32,8 +32,7 @@ class BaseArch(nn.Layer):
     def build_inputs(self, data, input_def):
         inputs = {}
         for i, k in enumerate(input_def):
-            v = paddle.to_tensor(data[i])
-            inputs[k] = v
+            inputs[k] = data[i]
         return inputs
 
     def model_arch(self):
diff --git a/tools/train.py b/tools/train.py
index 716b7b8a62b1f43199e98580aacdce9991fe6255..a59dd192d305ccfee320272e6363d32fc13b27ac 100755
--- a/tools/train.py
+++ b/tools/train.py
@@ -126,7 +126,7 @@ def run(FLAGS, cfg, place):
     model = create(cfg.architecture)
 
     # Optimizer
-    lr = create('LearningRate')(step_per_epoch / int(ParallelEnv().nranks))
+    lr = create('LearningRate')(step_per_epoch)
     optimizer = create('OptimizerBuilder')(lr, model.parameters())
 
     # Init Model & Optimzer