[while grad]Support pruning op in find_op_path about while sub-block when...

[while grad]Support pruning op in find_op_path about while sub-block when appending backward (#25330)

Prune OPs which are not related with loss in while sub-block when constructing backward OP path.

python/paddle/fluid/backward.py

@@ -797,32 +797,51 @@ def _append_backward_ops_with_checkpoints_(
     return program_stat, checkpoints_name, vars_should_be_hold, recompute_segments
 
 
-def _get_sub_block_path(sub_block, sub_block_op_desc, no_grad_set):
+def _get_sub_block_path(sub_block,
+                        sub_block_op_desc,
+                        no_grad_set,
+                        op_path_dict,
+                        sub_block_target_names=None):
     """
     Get output vars in subblock which will be assigned to parent block.
-    It is used to find the grad path in subblock
+    It is used to find the grad path in subblock.
+
+    Args:
+        sub_block(Block): The sub-block in which to get op path.
+        sub_block_op_desc: The op desc of the sub-block op such as 'while', 'conditional_block' and 'recurrent'.
+        no_grad_set(set): The set of no grad var name. no_grad_set will be changed.
+        op_path_dict(dict): op_path_dict will be changed.
+            key(int) block index
+            val(list) the op path of block(index)
+        sub_block_target_names(set): Target var names of sub-block.
+    Return:
+        The forward op path of sub-block corresponding to backward op.
     """
+
     assert sub_block_op_desc.has_attr(
         "sub_block") and sub_block.idx == sub_block_op_desc._block_attr_id(
             "sub_block")
-    # TODO(huihuangzheng): add support for recurrent op and while op
-    if sub_block_op_desc.type == "conditional_block":
-        sub_outputs = []
-        sub_assign_to_out_ops = []
-        for var in sub_block_op_desc.output_arg_names:
+    assert isinstance(sub_block_target_names, (set, type(None)))
+
+    if sub_block_target_names is None:
+        sub_block_target_names = sub_block_op_desc.output_arg_names
+
+    # TODO(huihuangzheng): add support for recurrent op.
+    if sub_block_op_desc.type in ["conditional_block", "while"]:
+        # Step1: get the output vars in sub-block
+        sub_outputs = [
+            sub_block._var_recursive(var) for var in sub_block_target_names
+        ]
+        for var in sub_block_target_names:
             for op_desc in sub_block.ops:
-                if op_desc.type == "assign" and var in op_desc.output_arg_names:
-                    sub_assign_to_out_ops.append(op_desc)
+                if var in op_desc.output_arg_names:
                     for name in op_desc.input_arg_names:
-                        if sub_block.has_var(name):
-                            sub_outputs.append(sub_block.var(name))
+                        sub_outputs.append(sub_block._var_recursive(name))
 
+        # Step2: find op path of sub-block
+        is_while = sub_block_op_desc.type in ["while"]
         sub_block_op_path = _find_op_path_(sub_block, sub_outputs, [],
-                                           no_grad_set)
-        # TODO better way than finding in list
-        for op_desc in sub_assign_to_out_ops:
-            if op_desc not in sub_block_op_path:
-                sub_block_op_path.append(op_desc)
+                                           no_grad_set, op_path_dict, is_while)
         return sub_block_op_path
     return sub_block.ops
 
@@ -846,7 +865,8 @@ def _append_backward_ops_(block,
                           no_grad_dict,
                           grad_to_var,
                           callbacks=None,
-                          input_grad_names_set=None):
+                          input_grad_names_set=None,
+                          op_path_dict=None):
     """
     Create all grad ops, and insert them into given block
 
@@ -864,6 +884,9 @@ def _append_backward_ops_(block,
         input_grad_names_set(set): this set is used to store the gradients' name which is
             generated by backward ops, and input_grad_names_set can help to prune the unnecessary
             backward ops.
+        op_path_dict(dict): op_path_dict will be changed.
+            key(int) block index
+            val(list) the op path of block(index)
     """
     if callbacks is not None:
         assert (isinstance(callbacks, list))
@@ -888,11 +911,10 @@ def _append_backward_ops_(block,
             # see follwing comments for why set None here.
             pre_input_grad_names_set = copy.copy(input_grad_names_set)
             input_grad_names_set = None
-            sub_block_path = _get_sub_block_path(sub_block, op,
-                                                 no_grad_dict[sub_block.idx])
+            sub_block_path = op_path_dict[op._block_attr_id("sub_block")]
             _append_backward_ops_(sub_block, sub_block_path, grad_sub_block,
                                   no_grad_dict, grad_to_var, callbacks,
-                                  input_grad_names_set)
+                                  input_grad_names_set, op_path_dict)
             input_grad_names_set = pre_input_grad_names_set
 
             program._rollback()
@@ -1358,7 +1380,10 @@ def append_backward(loss,
 
         block_no_grad_set = set(
             map(_strip_grad_suffix_, no_grad_dict[block_idx]))
-        op_path = _find_op_path_(block, [loss], [], block_no_grad_set)
+
+        op_path_dict = dict()
+        op_path = _find_op_path_(block, [loss], [], block_no_grad_set,
+                                 op_path_dict)
 
         no_grad_vars = _find_no_grad_vars(block, op_path, [loss],
                                           block_no_grad_set)
@@ -1371,13 +1396,13 @@ def append_backward(loss,
         # For double backward, input_grad_names is used for filtering
         # some non-used gradients op(s).
 
-        # Todo(liym27): need a better design.
+        # TODO(liym27): need a better design.
         # not support double grad in control flow sub-block now.
         if not is_in_control_flow:
             if program._appending_grad_times > 1:
                 input_grad_names_set = set([_append_grad_suffix_(loss.name)])
 
-        # Todo: support _append_backward_ops_with_checkpoints_ in
+        # TODO: support _append_backward_ops_with_checkpoints_ in
         #  sub-block (control flow)
         if checkpoints != None and \
                 isinstance(checkpoints, list) and \
@@ -1400,7 +1425,8 @@ def append_backward(loss,
                 no_grad_dict,
                 grad_to_var,
                 callbacks,
-                input_grad_names_set=input_grad_names_set)
+                input_grad_names_set=input_grad_names_set,
+                op_path_dict=op_path_dict)
 
     grad_info_map = dict()
 
@@ -1508,13 +1534,14 @@ def _get_output_names(cur_block, targets):
     """
 
     block = targets[0].block if targets else cur_block
-    prog = cur_block.program
-    if _is_ancestor_block(block, cur_block):
-        return set()
-
     current_output_names = set([out.name for out in targets])
 
-    # if `cur_block` is an ancestor of `targets[0].block`, run while loop
+    # 1. If `targets` in cur_block or the ancestral block of `cur_block`
+    if block.idx == cur_block.idx or _is_ancestor_block(block, cur_block):
+        return current_output_names
+
+    # 2. If `cur_block` is an ancestor of `targets[0].block`, run while loop
+    prog = cur_block.program
     while block.idx != cur_block.idx:
         assert block.parent_idx != -1
         parent_block = prog.block(block.parent_idx)
@@ -1554,12 +1581,32 @@ def _find_no_grad_vars(block, op_path, targets, no_grad_set):
     return set(no_grad_var)
 
 
-def _find_op_path_(block, outputs, inputs, no_grad_set):
+def _find_op_path_(block,
+                   targets,
+                   inputs,
+                   no_grad_set,
+                   op_path_dict=None,
+                   is_while=False):
     """
-    no_grad_set will also be changed
+    It is used to find the grad path in `block`.
+
+    Args:
+        block(Block): The block in which to get op path.
+        targets(list[Variable]): The target variables.
+        inputs(list[Variable]): The input variables.
+        no_grad_set(set): The set of no grad var name. no_grad_set will be changed.
+        op_path_dict(dict): op_path_dict will be changed. op_path_dict will be changed.
+            key(int) block index
+            val(list) the op path of block(index)
+        is_while(bool): Whether or not `block` is while block
+    Return:
+        The forward op path of block corresponding to backward op.
     """
+
     input_names = set([inp.name for inp in inputs])
-    output_names = _get_output_names(block, outputs)
+    output_names = _get_output_names(block, targets)
+    if op_path_dict is None:
+        op_path_dict = dict()
 
     relevant_op_flags = [True] * len(block.ops)
 
@@ -1576,6 +1623,15 @@ def _find_op_path_(block, outputs, inputs, no_grad_set):
                 relevant_op_flags[i] = False
 
     for i, op in reversed(list(enumerate(block.ops))):
+        if op.has_attr("sub_block"):
+            sub_block_id = op._block_attr_id("sub_block")
+            sub_block = block.program.block(sub_block_id)
+            sub_block_target_names = output_names & set(op.output_arg_names)
+            sub_block_path = _get_sub_block_path(sub_block, op,
+                                                 set(), op_path_dict,
+                                                 sub_block_target_names)
+            op_path_dict[sub_block_id] = sub_block_path
+
         if _some_in_set_(
                 op.desc.output_arg_names(),
                 output_names) and core.has_non_empty_grad_op_maker(op.type):
@@ -1585,6 +1641,14 @@ def _find_op_path_(block, outputs, inputs, no_grad_set):
         else:
             relevant_op_flags[i] = False
 
+    if is_while:
+        # If block is while block, dealing with op specifically again.
+        # TODO(liym27): Consider special types of ops.
+        for i, op in reversed(list(enumerate(block.ops))):
+            if relevant_op_flags[i] == False \
+                    and _some_in_set_(op.desc.output_arg_names(),output_names):
+                relevant_op_flags[i] = True
+
     op_path = [
         block.ops[i] for i in range(len(block.ops)) if relevant_op_flags[i]
     ]
@@ -1688,7 +1752,10 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
             raise "input must be in the same program as targets"
 
     block_no_grad_set = set(map(_strip_grad_suffix_, no_grad_dict[0]))
-    op_path = _find_op_path_(block, targets, inputs, block_no_grad_set)
+
+    op_path_dict = dict()
+    op_path = _find_op_path_(block, targets, inputs, block_no_grad_set,
+                             op_path_dict)
     no_grad_dict[0].update(list(map(_append_grad_suffix_, block_no_grad_set)))
     grad_to_var = dict()
     grad_info_map = dict()
@@ -1698,7 +1765,8 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
         block,
         no_grad_dict,
         grad_to_var,
-        input_grad_names_set=input_grad_names_set)
+        input_grad_names_set=input_grad_names_set,
+        op_path_dict=op_path_dict)
 
     # Because calc_gradient may be called multiple times,
     # we need rename the internal gradient variables so that they have

python/paddle/fluid/tests/unittests/test_while_loop_op.py

@@ -199,10 +199,16 @@ class TestApiWhileLoop_Backward(unittest.TestCase):
         def cond(i, x):
             return layers.less_than(i, eleven)
 
-        def body(i, x):
+        def body(j, x):
+            # TODO: In while block, if the var created in parent block
+            # participates in the calculation of gradient, the result of gradient
+            # is incorrect because each step scope always returns the same value
+            # generated by last step.
+            # Here we call `assign` op in while block to avoid this bug, and working on fixing it in next PR.
+            i = layers.assign(j)
             x = layers.elementwise_mul(x=i, y=i)
-            i = layers.increment(i)
-            return [i, x]
+            j = layers.increment(j)
+            return [j, x]
 
         main_program = Program()
         startup_program = Program()
@@ -232,7 +238,48 @@ class TestApiWhileLoop_Backward(unittest.TestCase):
                             'x': feed_x},
                       fetch_list=[mean.name, i.grad_name])
         self.assertTrue(np.allclose(np.asarray(res[0]), data))
-        self.assertTrue(np.allclose(np.asarray(res[1]), i_grad))
+        self.assertTrue(
+            np.allclose(np.asarray(res[1]), i_grad),
+            msg=" \nres = \n{} \n\n ans = \n{}".format(res[1], i_grad))
+
+    def test_while_loop_backward2(self):
+        def cond(i, x):
+            return i < 5
+
+        def body(i, x):
+            x = x + i
+            i = i + 1
+            return [i, x]
+
+        main_program = Program()
+        startup_program = Program()
+        with fluid.program_guard(main_program, startup_program):
+            i = fluid.data(name='i', shape=[1], dtype='float32')
+            i.stop_gradient = False
+            x = fluid.data(name='x', shape=[1], dtype='float32')
+            x.stop_gradient = False
+
+            out = layers.while_loop(cond, body, [i, x])
+            mean = layers.mean(out[1])
+            append_backward(mean)
+
+        place = fluid.CUDAPlace(0) if core.is_compiled_with_cuda(
+        ) else fluid.CPUPlace()
+        exe = fluid.Executor(place)
+
+        feed_i = np.ones(1).astype('float32')
+        feed_x = np.ones(1).astype('float32')
+        data = np.asarray([11]).astype('float32')
+        i_grad = np.asarray([1]).astype('float32')
+
+        res = exe.run(main_program,
+                      feed={'i': feed_i,
+                            'x': feed_x},
+                      fetch_list=[mean.name, i.grad_name])
+        self.assertTrue(np.allclose(np.asarray(res[0]), data))
+        self.assertTrue(
+            np.allclose(np.asarray(res[1]), i_grad),
+            msg=" \nres = \n{} \n\n ans = \n{}".format(res[1], i_grad))
 
 
 class TestApiWhileLoop_NestedWithBackwardAndLoDTensorArray(unittest.TestCase):