diff --git a/ppdet/modeling/heads/mask_head.py b/ppdet/modeling/heads/mask_head.py
index 939debbaae129293551394b5571f7da158a0cccb..403d4ceebcf676f9c55c666e2969a33be5816e9c 100644
--- a/ppdet/modeling/heads/mask_head.py
+++ b/ppdet/modeling/heads/mask_head.py
@@ -80,7 +80,7 @@ class MaskFeat(nn.Layer):
         mask_conv.add_sublayer(
             'conv5_mask',
             nn.Conv2DTranspose(
-                in_channels=self.in_channel,
+                in_channels=self.out_channel if num_convs > 0 else self.in_channel,
                 out_channels=self.out_channel,
                 kernel_size=2,
                 stride=2,
diff --git a/ppdet/modeling/heads/sparsercnn_head.py b/ppdet/modeling/heads/sparsercnn_head.py
index 377cf27fc6e10ff62b06345c7f639f064c05e5e7..801ff04fb772aec568dd94d9d4916d2b778e88fe 100644
--- a/ppdet/modeling/heads/sparsercnn_head.py
+++ b/ppdet/modeling/heads/sparsercnn_head.py
@@ -242,6 +242,8 @@ class SparseRCNNHead(nn.Layer):
             loss_func="SparseRCNNLoss",
             roi_input_shape=None, ):
         super().__init__()
+        assert head_num_heads > 0, \
+            f'At least one RoI Head is required, but {head_num_heads}.'
 
         # Build RoI.
         box_pooler = self._init_box_pooler(roi_input_shape)
@@ -337,11 +339,11 @@ class SparseRCNNHead(nn.Layer):
         inter_class_logits = []
         inter_pred_bboxes = []
 
-        for rcnn_head in self.head_series:
+        for stage, rcnn_head in enumerate(self.head_series):
             class_logits, pred_bboxes, proposal_features = rcnn_head(
                 features, bboxes, proposal_features, self.box_pooler)
 
-            if self.return_intermediate:
+            if self.return_intermediate or stage == len(self.head_series) - 1:
                 inter_class_logits.append(class_logits)
                 inter_pred_bboxes.append(pred_bboxes)
             bboxes = pred_bboxes.detach()