[Memory]More memory optimization policy (#8690)

* add memopt level

* add opt level for image classification demo

* clean code

* add delete op

* clean code

* test machine translation demo

* clean code

* clean code

* skip fill constant with force cpu

* clean code

* clean code

* refine code

* clean code

* fix bug

paddle/fluid/framework/block_desc.cc

@@ -135,6 +135,14 @@ OpDesc *BlockDesc::PrependOp() {
   return ops_.front().get();
 }
 
+OpDesc *BlockDesc::InsertOp(size_t index) {
+  need_update_ = true;
+  auto it = ops_.begin() + index;
+  std::unique_ptr<OpDesc> new_op(new OpDesc(this));
+  it = ops_.insert(it, std::move(new_op));
+  return (*it).get();
+}
+
 void BlockDesc::RemoveOp(size_t s, size_t e) {
   if (ops_.begin() + s == ops_.end() || ops_.begin() + e == ops_.end()) {
     return;

paddle/fluid/framework/block_desc.h

@@ -87,6 +87,8 @@ class BlockDesc {
 
   OpDesc *PrependOp();
 
+  OpDesc *InsertOp(size_t index);
+
   void RemoveOp(size_t s, size_t e);
 
   std::vector<OpDesc *> AllOps() const;

paddle/fluid/framework/scope.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include <memory>  // for unique_ptr
 #include <mutex>   // for call_once
+#include <set>
 #include "glog/logging.h"
 #include "paddle/fluid/framework/threadpool.h"
 #include "paddle/fluid/string/printf.h"
@@ -102,6 +103,18 @@ void Scope::DeleteScope(Scope* scope) {
   }
 }
 
+void Scope::EraseVars(std::vector<std::string>& var_names) {
+  std::set<std::string> var_set(var_names.begin(), var_names.end());
+  for (auto it = vars_.begin(); it != vars_.end();) {
+    if (var_set.find(it->first) != var_set.end()) {
+      delete it->second;
+      it = vars_.erase(it);
+    } else {
+      ++it;
+    }
+  }
+}
+
 void Scope::Rename(const std::string& origin_name,
                    const std::string& new_name) const {
   auto origin_it = vars_.find(origin_name);

paddle/fluid/framework/scope.h

@@ -51,6 +51,8 @@ class Scope {
   /// Create a variable with a scope-unique name.
   Variable* Var(std::string* name = nullptr);
 
+  void EraseVars(std::vector<std::string>& var_names);
+
   /// Find a variable in the scope or any of its ancestors.  Returns
   /// nullptr if cannot find.
   Variable* FindVar(const std::string& name) const;

paddle/fluid/operators/delete_var_op.cc

@@ -0,0 +1,52 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+    http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+
+namespace paddle {
+namespace operators {
+class DeleteVarOp : public framework::OperatorBase {
+ public:
+  DeleteVarOp(const std::string &type, const framework::VariableNameMap &inputs,
+              const framework::VariableNameMap &outputs,
+              const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+  void RunImpl(const framework::Scope &scope,
+               const platform::Place &place) const override {
+    // get device context from pool
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+    auto &dev_ctx = *pool.Get(place);
+    dev_ctx.Wait();
+
+    auto delete_var_names = Inputs("X");
+    const_cast<framework::Scope &>(scope).EraseVars(delete_var_names);
+  }
+};
+
+class DeleteVarOpInfoMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  DeleteVarOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The input of delete op").AsDuplicable();
+    AddComment(R"DOC(
+Delete Operator.
+It should not be configured by users directly.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OPERATOR(delete_var, paddle::operators::DeleteVarOp,
+                  paddle::framework::EmptyGradOpMaker,
+                  paddle::operators::DeleteVarOpInfoMaker);

paddle/fluid/pybind/protobuf.cc

@@ -161,6 +161,8 @@ void BindBlockDesc(py::module &m) {
            py::return_value_policy::reference)
       .def("prepend_op", &BlockDesc::PrependOp,
            py::return_value_policy::reference)
+      .def("insert_op", &BlockDesc::InsertOp,
+           py::return_value_policy::reference)
       .def("remove_op", &BlockDesc::RemoveOp)
       .def("var",
            [](BlockDesc &self, py::bytes byte_name) {

python/paddle/fluid/__init__.py

@@ -37,7 +37,7 @@ from distribute_transpiler_simple import SimpleDistributeTranspiler
 from concurrency import (Go, make_channel, channel_send, channel_recv,
                          channel_close)
 import clip
-from memory_optimization_transpiler import memory_optimize
+from memory_optimization_transpiler import memory_optimize, release_memory
 import profiler
 import unique_name
 
@@ -63,6 +63,7 @@ __all__ = framework.__all__ + executor.__all__ + concurrency.__all__ + [
     'SimpleDistributeTranspiler',
     'DistributeTranspiler',
     'memory_optimize',
+    'release_memory',
     'profiler',
     'unique_name',
 ]

python/paddle/fluid/backward.py

@@ -457,7 +457,8 @@ def append_backward(loss, parameter_list=None, no_grad_set=None,
         "Out": [_append_grad_suffix_(loss.name)]
     }, {"shape": [1],
         "value": 1.0,
-        "dtype": loss.dtype})
+        "dtype": loss.dtype,
+        "force_cpu": False})
     root_block.desc.append_op().copy_from(op_desc)
 
     block_no_grad_set = set(map(_strip_grad_suffix_, no_grad_dict[0]))

python/paddle/fluid/memory_optimization_transpiler.py

@@ -29,7 +29,10 @@ dtype_to_size = {
     core.VarDesc.VarType.BOOL: 1
 }
 
-sub_block_ops = ["while", "while_grad", "parallel_do", "parallel_do_grad"]
+sub_block_ops = [
+    "while", "while_grad", "parallel_do", "parallel_do_grad",
+    "conditional_block", "conditional_block_grad"
+]
 
 PRINT_LOG = False
 
@@ -122,36 +125,82 @@ class ControlFlowGraph(object):
         else:
             return block_desc.find_var_recursive(str(var_name))
 
-    def memory_optimize(self):
-        def check_var_validity(block_desc, x, is_forward):
-            if str(x) == "@EMPTY@":
-                return False
-            if not self._has_var(block_desc, x, is_forward):
-                return False
-            if self._find_var(block_desc, x, is_forward).persistable():
-                return False
-            if self._find_var(
-                    block_desc, x,
-                    is_forward).type() != core.VarDesc.VarType.LOD_TENSOR:
-                return False
-            if x in self._skip_opt:
-                return False
-            if not self._find_var(block_desc, x, is_forward).shape():
-                return False
-            return True
+    def _check_var_validity(self, block_desc, x, is_forward):
+        if str(x) == "@EMPTY@":
+            return False
+        if not self._has_var(block_desc, x, is_forward):
+            return False
+        if self._find_var(block_desc, x, is_forward).persistable():
+            return False
+        if self._find_var(block_desc, x,
+                          is_forward).type() != core.VarDesc.VarType.LOD_TENSOR:
+            return False
+        if x in self._skip_opt:
+            return False
+        if not self._find_var(block_desc, x, is_forward).shape():
+            return False
+        return True
 
+    def _update_skip_opt_set(self):
+        for i in range(self.op_size):
+            op = self._ops[i]
+            if op.type() == "fill_constant" and op.attr("force_cpu") == True:
+                self._skip_opt.update(op.output_arg_names())
+
+    def release_memory(self):
         self._build_graph()
         self._dataflow_analyze()
+        self._update_skip_opt_set()
+        fwd_id = 0
+        bwd_id = 0
+        for i in range(self.op_size):
+            op = self._ops[i]
+            if op.type() in sub_block_ops:
+                continue
+            block_desc = op.block()
+            is_forward = i < self._forward_num
+            in_diff, out_diff = self._get_diff(self._live_in[i],
+                                               self._live_out[i])
+            can_optimize = filter(
+                lambda x: self._check_var_validity(block_desc, x, is_forward),
+                in_diff)
+            if can_optimize:
+                index = i + fwd_id + 1 if is_forward else i - self._forward_num + bwd_id + 1
+                delete_op = block_desc.insert_op(index)
+                delete_op.set_type("delete_var")
+                delete_op.set_input("X", can_optimize)
+                if is_forward:
+                    fwd_id += 1
+                else:
+                    bwd_id += 1
+
+    def memory_optimize(self, level=0):
+        def compare_shape(x_shape, cache_shape, opt_level):
+            if opt_level == 0:
+                return x_shape == cache_shape
+            if opt_level == 1:
+                if (x_shape[0] == -1) ^ (cache_shape[0] == -1):
+                    return False
+                x_size = abs(reduce(lambda x, y: x * y, x_shape))
+                cache_size = abs(reduce(lambda x, y: x * y, cache_shape))
+                if x_size <= cache_size:
+                    return True
+            return False
+
+        self._build_graph()
+        self._dataflow_analyze()
+        self._update_skip_opt_set()
         self.pool = []
         for i in range(self.op_size):
             op = self._ops[i]
             if op.type() in sub_block_ops:
                 continue
             block_desc = op.block()
+            self.current_block_desc = block_desc
             is_forward = i < self._forward_num
             if self.pool:
                 defs_can_optimize = filter(
-                    lambda x: check_var_validity(block_desc, x, is_forward),
+                    lambda x: self._check_var_validity(block_desc, x, is_forward),
                     self._defs[i])
                 out_pair = [
                     (x, self._find_var(block_desc, x, is_forward).shape())
@@ -164,7 +213,7 @@ class ControlFlowGraph(object):
                     for index, cache_pair in enumerate(self.pool):
                         cache_var = cache_pair[0]
                         cache_shape = cache_pair[1]
-                        if x_shape == cache_shape:
+                        if compare_shape(x_shape, cache_shape, level):
                             if self._has_var(block_desc, cache_var, is_forward):
                                 x_dtype = self._find_var(block_desc, x,
                                                          is_forward).dtype()
@@ -196,7 +245,7 @@ class ControlFlowGraph(object):
             in_diff, out_diff = self._get_diff(self._live_in[i],
                                                self._live_out[i])
             can_optimize = filter(
-                lambda x: check_var_validity(block_desc, x, is_forward),
+                lambda x: self._check_var_validity(block_desc, x, is_forward),
                 in_diff)
             if can_optimize:
                 for var_name in can_optimize:
@@ -270,7 +319,8 @@ def _get_cfgs(input_program):
         ([block_desc.op(i) for i in range(op_size)], op_size, set()))
 
     sub_block_pair = [("while", "while_grad"), ("parallel_do",
-                                                "parallel_do_grad")]
+                                                "parallel_do_grad"),
+                      ("conditional_block", "conditional_block_grad")]
 
     ops_list.extend(_process_sub_block_pair(pdesc, sub_block_pair))
 
@@ -281,9 +331,15 @@ def _get_cfgs(input_program):
     return cfgs
 
 
-def memory_optimize(input_program, print_log=False):
+def memory_optimize(input_program, print_log=False, level=0):
     global PRINT_LOG
     PRINT_LOG = print_log
     cfgs = _get_cfgs(input_program)
     for cfg in cfgs:
-        cfg.memory_optimize()
+        cfg.memory_optimize(level)
+
+
+def release_memory(input_program):
+    cfgs = _get_cfgs(input_program)
+    for cfg in cfgs:
+        cfg.release_memory()

python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py

@@ -50,6 +50,7 @@ sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.01)
 sgd_optimizer.minimize(avg_cost)
 
 fluid.memory_optimize(fluid.default_main_program(), print_log=True)
+# fluid.release_memory(fluid.default_main_program())
 
 BATCH_SIZE = 200
 
@@ -69,8 +70,6 @@ exe.run(fluid.default_startup_program())
 
 PASS_NUM = 100
 for pass_id in range(PASS_NUM):
-    fluid.io.save_persistables(exe, "./fit_a_line.model/")
-    fluid.io.load_persistables(exe, "./fit_a_line.model/")
     for data in train_reader():
         avg_loss_value, = exe.run(fluid.default_main_program(),
                                   feed=feeder.feed(data),

python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py

@@ -125,9 +125,10 @@ opts = optimizer.minimize(avg_cost)
 batch_size = fluid.layers.create_tensor(dtype='int64')
 batch_acc = fluid.layers.accuracy(input=predict, label=label, total=batch_size)
 
-fluid.memory_optimize(fluid.default_main_program())
+# fluid.memory_optimize(fluid.default_main_program(), level=0)
+fluid.release_memory(fluid.default_main_program())
 
-BATCH_SIZE = 128
+BATCH_SIZE = 16
 PASS_NUM = 1
 
 # fix the order of training data
@@ -159,8 +160,7 @@ for pass_id in range(PASS_NUM):
         print("loss:" + str(loss) + " acc:" + str(acc) + " pass_acc:" + str(
             pass_acc))
         # this model is slow, so if we can train two mini batch, we think it works properly.
-
-        if i > 2:
+        if i > 0:
             exit(0)
         if math.isnan(float(loss)):
             sys.exit("got NaN loss, training failed.")

python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py

@@ -105,7 +105,8 @@ def main():
     optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4)
     optimizer.minimize(avg_cost)
 
-    fluid.memory_optimize(fluid.default_main_program())
+    # fluid.memory_optimize(fluid.default_main_program())
+    fluid.release_memory(fluid.default_main_program())
 
     # fix the order of training data
     train_data = paddle.batch(

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

@@ -98,6 +98,9 @@ class TestLearningRateDecay(unittest.TestCase):
         exe = fluid.Executor(place)
 
         exe.run(fluid.default_startup_program())
+
+        fluid.memory_optimize(fluid.default_main_program())
+
         for step in range(10):
             lr_val, = exe.run(fluid.default_main_program(),
                               feed={},