Fix bug for backward tanspiler when using parallel_do operator. (#9282)

* Temporarily fix bug for backward tanspiler when using parallel_do operator.

* Fix bug for backward tanspiler when using parallel_do operator

paddle/fluid/operators/box_coder_op.cc

@@ -126,6 +126,7 @@ width and height.
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(box_coder, ops::BoxCoderOp, ops::BoxCoderOpMaker);
+REGISTER_OPERATOR(box_coder, ops::BoxCoderOp, ops::BoxCoderOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
 REGISTER_OP_CPU_KERNEL(box_coder, ops::BoxCoderKernel<float>,
                        ops::BoxCoderKernel<double>);

paddle/fluid/operators/detection_map_op.cc

@@ -188,8 +188,8 @@ The general steps are as follows. First, calculate the true positive and
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(detection_map, ops::DetectionMAPOp,
-                             ops::DetectionMAPOpMaker);
+REGISTER_OPERATOR(detection_map, ops::DetectionMAPOp, ops::DetectionMAPOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
 REGISTER_OP_CPU_KERNEL(
     detection_map, ops::DetectionMAPOpKernel<paddle::platform::CPUPlace, float>,
     ops::DetectionMAPOpKernel<paddle::platform::CPUPlace, double>);

paddle/fluid/operators/iou_similarity_op.cc

@@ -87,8 +87,9 @@ $$
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(iou_similarity, ops::IOUSimilarityOp,
-                             ops::IOUSimilarityOpMaker);
+REGISTER_OPERATOR(iou_similarity, ops::IOUSimilarityOp,
+                  ops::IOUSimilarityOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
 
 REGISTER_OP_CPU_KERNEL(
     iou_similarity,

paddle/fluid/operators/mine_hard_examples_op.cc

@@ -324,8 +324,9 @@ MatchIndices elements with value -1.
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(mine_hard_examples, ops::MineHardExamplesOp,
-                             ops::MineHardExamplesOpMaker);
+REGISTER_OPERATOR(mine_hard_examples, ops::MineHardExamplesOp,
+                  ops::MineHardExamplesOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
 
 REGISTER_OP_CPU_KERNEL(
     mine_hard_examples,

paddle/fluid/operators/prior_box_op.cc

@@ -168,7 +168,9 @@ https://arxiv.org/abs/1512.02325.
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(prior_box, ops::PriorBoxOp, ops::PriorBoxOpMaker);
+REGISTER_OPERATOR(prior_box, ops::PriorBoxOp, ops::PriorBoxOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+
 REGISTER_OP_CPU_KERNEL(
     prior_box, ops::PriorBoxOpKernel<paddle::platform::CPUPlace, float>,
     ops::PriorBoxOpKernel<paddle::platform::CPUPlace, double>);

paddle/fluid/operators/target_assign_op.cc

@@ -153,8 +153,8 @@ template struct NegTargetAssignFunctor<platform::CPUDeviceContext, float,
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(target_assign, ops::TargetAssignOp,
-                             ops::TargetAssignOpMaker);
+REGISTER_OPERATOR(target_assign, ops::TargetAssignOp, ops::TargetAssignOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
 REGISTER_OP_CPU_KERNEL(
     target_assign,
     ops::TargetAssignKernel<paddle::platform::CPUDeviceContext, int, float>,

python/paddle/fluid/layers/detection.py

@@ -129,13 +129,11 @@ def detection_output(loc,
         prior_box_var=prior_box_var,
         target_box=loc,
         code_type='decode_center_size')
-
     old_shape = scores.shape
     scores = ops.reshape(x=scores, shape=(-1, old_shape[-1]))
     scores = nn.softmax(input=scores)
     scores = ops.reshape(x=scores, shape=old_shape)
     scores = nn.transpose(scores, perm=[0, 2, 1])
-
     nmsed_outs = helper.create_tmp_variable(dtype=decoded_box.dtype)
     helper.append_op(
         type="multiclass_nms",
@@ -475,6 +473,7 @@ def ssd_loss(location,
     # 2. Compute confidence for mining hard examples
     # 2.1. Get the target label based on matched indices
     gt_label = ops.reshape(x=gt_label, shape=gt_label.shape + (1, ))
+    gt_label.stop_gradient = True
     target_label, _ = target_assign(
         gt_label, matched_indices, mismatch_value=background_label)
     # 2.2. Compute confidence loss.
@@ -482,10 +481,12 @@ def ssd_loss(location,
     confidence = __reshape_to_2d(confidence)
     target_label = tensor.cast(x=target_label, dtype='int64')
     target_label = __reshape_to_2d(target_label)
+    target_label.stop_gradient = True
     conf_loss = nn.softmax_with_cross_entropy(confidence, target_label)
 
     # 3. Mining hard examples
     conf_loss = ops.reshape(x=conf_loss, shape=(num, num_prior))
+    conf_loss.stop_gradient = True
     neg_indices = helper.create_tmp_variable(dtype='int32')
     dtype = matched_indices.dtype
     updated_matched_indices = helper.create_tmp_variable(dtype=dtype)
@@ -695,6 +696,8 @@ def multi_box_head(inputs,
             outputs={"Boxes": box,
                      "Variances": var},
             attrs=attrs, )
+        box.stop_gradient = True
+        var.stop_gradient = True
         return box, var
 
     def _reshape_with_axis_(input, axis=1):