refine device when assign on cpu (#5925)

* refine device when assign on cpu

* cherry-pick assign on cpu in mask postprocess

ppdet/modeling/post_process.py

@@ -154,6 +154,7 @@ class BBoxPostProcess(nn.Layer):
 
 @register
 class MaskPostProcess(object):
+    __shared__ = ['assign_on_cpu']
     """
     refer to:
     https://github.com/facebookresearch/detectron2/layers/mask_ops.py
@@ -161,9 +162,10 @@ class MaskPostProcess(object):
     Get Mask output according to the output from model
     """
 
-    def __init__(self, binary_thresh=0.5):
+    def __init__(self, binary_thresh=0.5, assign_on_cpu=False):
         super(MaskPostProcess, self).__init__()
         self.binary_thresh = binary_thresh
+        self.assign_on_cpu = assign_on_cpu
 
     def paste_mask(self, masks, boxes, im_h, im_w):
         """
@@ -210,6 +212,9 @@ class MaskPostProcess(object):
         if bbox_num == 1 and bboxes[0][0] == -1:
             return pred_result
 
+        if self.assign_on_cpu:
+            device = paddle.device.get_device()
+            paddle.set_device('cpu')
         # TODO: optimize chunk paste
         pred_result = []
         for i in range(bboxes.shape[0]):
@@ -220,6 +225,9 @@ class MaskPostProcess(object):
             pred_mask = paddle.cast(pred_mask, 'int32')
             pred_result.append(pred_mask)
         pred_result = paddle.concat(pred_result)
+        if self.assign_on_cpu:
+            paddle.set_device(device)
+
         return pred_result
 
 

ppdet/modeling/proposal_generator/target.py

@@ -74,9 +74,10 @@ def label_box(anchors,
               is_crowd=None,
               assign_on_cpu=False):
     if assign_on_cpu:
+        device = paddle.device.get_device()
         paddle.set_device("cpu")
         iou = bbox_overlaps(gt_boxes, anchors)
-        paddle.set_device("gpu")
+        paddle.set_device(device)
     else:
         iou = bbox_overlaps(gt_boxes, anchors)
     n_gt = gt_boxes.shape[0]