From 9d26f1a3dfd03bbce40c52f0195d2ebba9ced22b Mon Sep 17 00:00:00 2001
From: Yang Yang <yangyang62@baidu.com>
Date: Thu, 15 Feb 2018 02:08:33 +0000
Subject: [PATCH] callback to list of callbacks

---
 paddle/fluid/framework/executor.cc  |  3 +--
 python/paddle/v2/fluid/backward.py  | 33 +++++++++++++++++------------
 python/paddle/v2/fluid/optimizer.py |  2 +-
 3 files changed, 22 insertions(+), 16 deletions(-)

diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc
index 92b32b04d6..93a4883368 100644
--- a/paddle/fluid/framework/executor.cc
+++ b/paddle/fluid/framework/executor.cc
@@ -120,12 +120,11 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
 
   for (auto& op_desc : block.AllOps()) {
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
-    //    VLOG(3) << op->DebugStringEx(local_scope);
+    VLOG(3) << place_ << " " << op->DebugStringEx(local_scope);
 
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     platform::RecordEvent record_event(op->Type(), pool.Get(place_));
 
-    VLOG(3) << op->Type();
     op->Run(*local_scope, place_);
     if (FLAGS_benchmark) {
       VLOG(2) << "Memory used after operator " + op->Type() + " running: "
diff --git a/python/paddle/v2/fluid/backward.py b/python/paddle/v2/fluid/backward.py
index 682df3301b..4e494db93b 100644
--- a/python/paddle/v2/fluid/backward.py
+++ b/python/paddle/v2/fluid/backward.py
@@ -234,7 +234,6 @@ def _callback_lookup_(op):
                         "ncclInit",
                         {"parallel_scopes": self.parallel_scopes_name},
                         {"Communicator": ['nccl_com__do_not_change_']}, {})
-                    print(serialize_op_decs(op_desc))
                     block.program.global_block().desc.append_op().copy_from(
                         op_desc)
                     self.has_inserted_nccl_init = True
@@ -285,9 +284,7 @@ def _append_backward_ops_(block,
             val(str): corresponding forward variable name
         callback(callable object): a callable object used to decorate new generated grad ops
     """
-    if callbacks is None:
-        callbacks = []
-    else:
+    if callbacks is not None:
         assert (isinstance(callbacks, list))
         for cb in callbacks:
             if not hasattr(cb, '__call__'):
@@ -302,12 +299,17 @@ def _append_backward_ops_(block,
         if op.has_attr("sub_block"):
             sub_block = program.block(op.block_attr("sub_block"))
             grad_sub_block = program.create_block(parent_idx=sub_block.idx)
-            if callbacks is None:
-                callbacks = [_callback_lookup_(op)]
+            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,
+                                      no_grad_dict, grad_to_var, new_callbacks)
             else:
-                callbacks.append(_callback_lookup_(op))
-            _append_backward_ops_(sub_block, sub_block.ops, grad_sub_block,
-                                  no_grad_dict, grad_to_var, callbacks)
+                _append_backward_ops_(sub_block, sub_block.ops, grad_sub_block,
+                                      no_grad_dict, grad_to_var, callbacks)
             grad_sub_block_list.append(grad_sub_block.desc)
 
         # Getting op's corresponding grad_op
@@ -327,8 +329,10 @@ def _append_backward_ops_(block,
         new_op_desc = target_block.desc.append_op()
         new_op_desc.copy_from(op_desc)
         grad_to_var["__current_op_desc__"] = new_op_desc
-        for cb in callbacks:
-            cb(block=target_block, context=grad_to_var)
+        if callbacks is not None:
+            assert (isinstance(callbacks, list))
+            for cb in callbacks:
+                cb(block=target_block, context=grad_to_var)
 
 
 def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map):
@@ -408,7 +412,8 @@ def _get_stop_gradients_(program):
     return no_grad_dict
 
 
-def append_backward(loss, parameter_list=None, no_grad_set=None, callback=None):
+def append_backward(loss, parameter_list=None, no_grad_set=None,
+                    callbacks=None):
     """
     Append backward part to main_program
 
@@ -424,6 +429,8 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, callback=None):
         (list[(Variable,Variable)]): list of (parameter, gradient) pair.
     """
     assert isinstance(loss, framework.Variable)
+    if callbacks is not None:
+        isinstance(callbacks, list)
 
     program = loss.block.program
     if no_grad_set is None:
@@ -451,7 +458,7 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, callback=None):
     no_grad_dict[0].update(map(_append_grad_suffix_, block_no_grad_set))
 
     _append_backward_ops_(root_block, op_path, root_block, no_grad_dict,
-                          grad_to_var, callback)
+                          grad_to_var, callbacks)
 
     # Because calc_gradient may be called multiple times,
     # we need rename the internal gradient variables so that they have
diff --git a/python/paddle/v2/fluid/optimizer.py b/python/paddle/v2/fluid/optimizer.py
index 39391eb8e4..ecc42f6215 100644
--- a/python/paddle/v2/fluid/optimizer.py
+++ b/python/paddle/v2/fluid/optimizer.py
@@ -225,7 +225,7 @@ class Optimizer(object):
         `create_optimization_pass()` into one.
         """
         params_grads = append_backward(loss, parameter_list, no_grad_set,
-                                       error_clip_callback)
+                                       [error_clip_callback])
 
         params_grads = append_gradient_clip_ops(params_grads)
 
-- 
GitLab