Temporarily fix bug for backward tanspiler when using parallel_do ope…

…rator.
PaddlePaddle · Mar 21, 2018 · 39e92b9 · 39e92b9
1 parent 7c041e4
commit 39e92b9
Show file tree

Hide file tree

Showing 2 changed files with 16 additions and 4 deletions.
diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py
@@ -307,16 +307,28 @@ def _append_backward_ops_(block,
             sub_block = program.block(op.block_attr("sub_block"))
             grad_sub_block = program.create_block()
             grad_sub_block.set_forward_block_idx(sub_block.idx)
+
+            all_vars = op.block.vars
+            target_vars = [all_vars[name] for name in op.output_arg_names]
+            no_grad_set = copy.copy(no_grad_dict[sub_block.idx])
+            new_no_grad_dict = _get_stop_gradients_(sub_block.program)
+            new_no_grad_dict[0].update(map(_append_grad_suffix_, no_grad_set))
+            block_no_grad_set = set(
+                map(_strip_grad_suffix_, new_no_grad_dict[0]))
+            op_path = _find_op_path_(sub_block, target_vars, [],
+                                     block_no_grad_set)
+            no_grad_dict[0].update(map(_append_grad_suffix_, block_no_grad_set))
+
             cb = _callback_lookup_(op)
             if cb is not None:
                 if callbacks is None:
                     new_callbacks = [cb]
                 else:
                     new_callbacks = callbacks + [_callback_lookup_(op)]
-                _append_backward_ops_(sub_block, sub_block.ops, grad_sub_block,
+                _append_backward_ops_(sub_block, op_path, grad_sub_block,
                                       no_grad_dict, grad_to_var, new_callbacks)
             else:
-                _append_backward_ops_(sub_block, sub_block.ops, grad_sub_block,
+                _append_backward_ops_(sub_block, op_path, grad_sub_block,
                                       no_grad_dict, grad_to_var, callbacks)
 
             program.rollback()

diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py
@@ -129,13 +129,11 @@ class number, M is number of bounding boxes. For each category
         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",
@@ -695,6 +693,8 @@ def _prior_box_(input,
             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):