Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into update-api-reference-1

benchmark/fluid/models/machine_translation.py

@@ -173,21 +173,6 @@ def seq_to_seq_net(embedding_dim, encoder_size, decoder_size, source_dict_dim,
         return avg_cost, feeding_list
 
 
-def to_lodtensor(data, place):
-    seq_lens = [len(seq) for seq in data]
-    cur_len = 0
-    lod = [cur_len]
-    for l in seq_lens:
-        cur_len += l
-        lod.append(cur_len)
-    flattened_data = np.concatenate(data, axis=0).astype("int64")
-    flattened_data = flattened_data.reshape([len(flattened_data), 1])
-    lod_t = core.LoDTensor()
-    lod_t.set(flattened_data, place)
-    lod_t.set_lod([lod])
-    return lod_t, lod[-1]
-
-
 def lodtensor_to_ndarray(lod_tensor):
     dims = lod_tensor.get_dims()
     ndarray = np.zeros(shape=dims).astype('float32')

benchmark/fluid/models/stacked_dynamic_lstm.py

@@ -125,18 +125,3 @@ def get_model(args):
         batch_size=args.batch_size)
 
     return loss, inference_program, adam, train_reader, test_reader, batch_acc
-
-
-def to_lodtensor(data, place):
-    seq_lens = [len(seq) for seq in data]
-    cur_len = 0
-    lod = [cur_len]
-    for l in seq_lens:
-        cur_len += l
-        lod.append(cur_len)
-    flattened_data = numpy.concatenate(data, axis=0).astype("int64")
-    flattened_data = flattened_data.reshape([len(flattened_data), 1])
-    res = fluid.LoDTensor()
-    res.set(flattened_data, place)
-    res.set_lod([lod])
-    return res

doc/fluid/api/gen_doc.sh

@@ -1,5 +1,5 @@
 #!/bin/bash
-python gen_doc.py layers --submodules control_flow device io nn ops tensor detection learning_rate_scheduler > layers.rst
+python gen_doc.py layers --submodules control_flow device io nn ops tensor detection learning_rate_scheduler metric > layers.rst
 
 for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer
 do

doc/fluid/api/initializer.rst

@@ -33,6 +33,13 @@ Xavier
     :members:
     :noindex:
 
+Bilinear
+--------
+
+..  autoclass:: paddle.fluid.initializer.Bilinear
+    :members:
+    :noindex:
+
 force_init_on_cpu
 -----------------
 
@@ -73,3 +80,10 @@ XavierInitializer
     :members:
     :noindex:
 
+BilinearInitializer
+-------------------
+
+..  autoclass:: paddle.fluid.initializer.BilinearInitializer
+    :members:
+    :noindex:
+

doc/fluid/api/io.rst

@@ -59,3 +59,39 @@ get_inference_program
 ..  autofunction:: paddle.fluid.io.get_inference_program
     :noindex:
 
+save_checkpoint
+---------------
+
+..  autofunction:: paddle.fluid.io.save_checkpoint
+    :noindex:
+
+load_checkpoint
+---------------
+
+..  autofunction:: paddle.fluid.io.load_checkpoint
+    :noindex:
+
+clean_checkpoint
+----------------
+
+..  autofunction:: paddle.fluid.io.clean_checkpoint
+    :noindex:
+
+load_persist_vars_without_grad
+------------------------------
+
+..  autofunction:: paddle.fluid.io.load_persist_vars_without_grad
+    :noindex:
+
+save_persist_vars_without_grad
+------------------------------
+
+..  autofunction:: paddle.fluid.io.save_persist_vars_without_grad
+    :noindex:
+
+get_latest_checkpoint_serial
+----------------------------
+
+..  autofunction:: paddle.fluid.io.get_latest_checkpoint_serial
+    :noindex:
+

doc/fluid/api/optimizer.rst

@@ -89,6 +89,13 @@ DecayedAdagradOptimizer
     :members:
     :noindex:
 
+RMSPropOptimizer
+----------------
+
+..  autoclass:: paddle.fluid.optimizer.RMSPropOptimizer
+    :members:
+    :noindex:
+
 Adadelta
 --------
 

doc/fluid/api/profiler.rst

@@ -23,3 +23,15 @@ profiler
 ..  autofunction:: paddle.fluid.profiler.profiler
     :noindex:
 
+start_profiler
+--------------
+
+..  autofunction:: paddle.fluid.profiler.start_profiler
+    :noindex:
+
+stop_profiler
+-------------
+
+..  autofunction:: paddle.fluid.profiler.stop_profiler
+    :noindex:
+

paddle/fluid/framework/executor.cc

@@ -406,6 +406,9 @@ void Executor::EnableMKLDNN(const ProgramDesc& program) {
       }
     }
   }
+#else
+  LOG(WARNING)
+      << "'MKLDNN' is not supported, Please re-compile with WITH_MKLDNN option";
 #endif
 }
 

paddle/fluid/framework/lod_tensor.cc

@@ -410,5 +410,38 @@ void LoDTensor::MergeLoDTensor(
   }
 }
 
+LoD ConvertToLengthBasedLoD(const LoD &offset_lod) {
+  LoD length_lod;
+  length_lod.reserve(offset_lod.size());
+  for (size_t lvl = 0; lvl < offset_lod.size(); ++lvl) {
+    std::vector<size_t> level;
+    if (offset_lod[lvl].size() > 0) {
+      level.reserve(offset_lod[lvl].size() - 1);
+    }
+    for (size_t idx = 0; idx < offset_lod[lvl].size() - 1; ++idx) {
+      level.push_back(offset_lod[lvl][idx + 1] - offset_lod[lvl][idx]);
+    }
+    length_lod.push_back(level);
+  }
+  return length_lod;
+}
+
+LoD ConvertToOffsetBasedLoD(const LoD &length_lod) {
+  LoD offset_lod;
+  offset_lod.reserve(length_lod.size());
+  for (size_t lvl = 0; lvl < length_lod.size(); ++lvl) {
+    std::vector<size_t> level;
+    level.reserve(length_lod[lvl].size() + 1);
+    size_t tmp = 0;
+    level.push_back(tmp);
+    for (size_t idx = 0; idx < length_lod[lvl].size(); ++idx) {
+      tmp += length_lod[lvl][idx];
+      level.push_back(tmp);
+    }
+    offset_lod.push_back(level);
+  }
+  return offset_lod;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/lod_tensor.h

@@ -226,5 +226,19 @@ extern void WriteToRecordIO(recordio::Writer* writer,
 extern std::vector<LoDTensor> ReadFromRecordIO(
     recordio::Scanner* scanner, const platform::DeviceContext& dev_ctx);
 
+/*
+ * Convert between length-based LoD and offset-based LoD.
+ * The implementation of LoDTensor class use offset-based LoD.
+ * However, we want to expose the more user-friendly length-based
+ * LoD to the Python side instead.
+ *
+ * Example:
+ * If offset_lod = [[0, 2, 3],[0, 3, 5, 9]]
+ * then length_lod = [[2, 1], [3, 2, 4]]
+ */
+LoD ConvertToLengthBasedLoD(const LoD& offset_lod);
+
+LoD ConvertToOffsetBasedLoD(const LoD& length_lod);
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/lod_tensor_test.cc

@@ -228,6 +228,38 @@ TEST(LoD, CheckAbsLoD) {
   ASSERT_FALSE(CheckAbsLoD(abs_lod0));
 }
 
+TEST(LoD, ConvertToLengthBasedLoD) {
+  LoD offset_lod;
+  offset_lod.push_back(std::vector<size_t>({0, 2}));
+  offset_lod.push_back(std::vector<size_t>({0, 1, 3}));
+  offset_lod.push_back(std::vector<size_t>({0, 2, 4, 5}));
+
+  LoD length_lod = ConvertToLengthBasedLoD(offset_lod);
+
+  LoD expected;
+  expected.push_back(std::vector<size_t>({2}));
+  expected.push_back(std::vector<size_t>({1, 2}));
+  expected.push_back(std::vector<size_t>({2, 2, 1}));
+
+  EXPECT_EQ(length_lod, expected);
+}
+
+TEST(LoD, ConvertToOffsetBasedLoD) {
+  LoD length_lod;
+  length_lod.push_back(std::vector<size_t>({2}));
+  length_lod.push_back(std::vector<size_t>({1, 2}));
+  length_lod.push_back(std::vector<size_t>({2, 2, 1}));
+
+  LoD offset_lod = ConvertToOffsetBasedLoD(length_lod);
+
+  LoD expected;
+  expected.push_back(std::vector<size_t>({0, 2}));
+  expected.push_back(std::vector<size_t>({0, 1, 3}));
+  expected.push_back(std::vector<size_t>({0, 2, 4, 5}));
+
+  EXPECT_EQ(offset_lod, expected);
+}
+
 template <typename T>
 static void TestRecordIO() {
   LoDTensor tensor;

paddle/fluid/framework/scope.cc

@@ -43,48 +43,29 @@ Scope& Scope::NewScope() const {
 }
 
 Variable* Scope::Var(const std::string& name) {
-  // acquire the lock when new var under this scope
   std::unique_lock<std::mutex> lock(mutex_);
-  auto* v = FindVarLocally(name);
-  if (v != nullptr) return v;
-
-  v = new Variable();
-  vars_[name].reset(v);
-  VLOG(3) << "Create variable " << name;
-  v->name_ = &(vars_.find(name)->first);
-  return v;
+  return VarInternal(name);
 }
 
 Variable* Scope::Var(std::string* name) {
-  auto var_name = string::Sprintf("%p.%d", this, vars_.size());
+  std::unique_lock<std::mutex> lock(mutex_);
+  auto new_name = string::Sprintf("%p.%d", this, vars_.size());
   if (name != nullptr) {
-    *name = var_name;
+    *name = new_name;
   }
-  return Var(var_name);
+  return VarInternal(new_name);
 }
 
 Variable* Scope::FindVar(const std::string& name) const {
-  // acquire the lock when find var
   std::unique_lock<std::mutex> lock(mutex_);
   return FindVarInternal(name);
 }
 
-Variable* Scope::FindVarInternal(const std::string& name) const {
-  auto var = FindVarLocally(name);
-  if (var != nullptr) {
-    return var;
-  }
-  return (parent_ == nullptr) ? nullptr : parent_->FindVarInternal(name);
-}
-
 const Scope* Scope::FindScope(const Variable* var) const {
-  for (auto& kv : vars_) {
-    if (kv.second.get() == var) {
-      return this;
-    }
-  }
-  return (parent_ == nullptr) ? nullptr : parent_->FindScope(var);
+  std::unique_lock<std::mutex> lock(mutex_);
+  return FindScopeInternal(var);
 }
+
 void Scope::DropKids() {
   std::unique_lock<std::mutex> lock(mutex_);
   for (Scope* s : kids_) delete s;
@@ -92,6 +73,7 @@ void Scope::DropKids() {
 }
 
 std::vector<std::string> Scope::LocalVarNames() const {
+  std::unique_lock<std::mutex> lock(mutex_);
   std::vector<std::string> known_vars;
   known_vars.reserve(this->vars_.size());
   for (auto& p : vars_) {
@@ -127,6 +109,39 @@ void Scope::EraseVars(const std::vector<std::string>& var_names) {
 
 void Scope::Rename(const std::string& origin_name,
                    const std::string& new_name) const {
+  std::unique_lock<std::mutex> lock(mutex_);
+  RenameInternal(origin_name, new_name);
+}
+
+std::string Scope::Rename(const std::string& origin_name) const {
+  std::unique_lock<std::mutex> lock(mutex_);
+  auto new_name = string::Sprintf("%p.%d", this, vars_.size());
+  RenameInternal(origin_name, new_name);
+  return new_name;
+}
+
+Variable* Scope::VarInternal(const std::string& name) {
+  auto* v = FindVarLocally(name);
+  if (v != nullptr) return v;
+
+  v = new Variable();
+  vars_[name].reset(v);
+  VLOG(3) << "Create variable " << name;
+  v->name_ = &(vars_.find(name)->first);
+  return v;
+}
+
+const Scope* Scope::FindScopeInternal(const Variable* var) const {
+  for (auto& kv : vars_) {
+    if (kv.second.get() == var) {
+      return this;
+    }
+  }
+  return (parent_ == nullptr) ? nullptr : parent_->FindScope(var);
+}
+
+void Scope::RenameInternal(const std::string& origin_name,
+                           const std::string& new_name) const {
   auto origin_it = vars_.find(origin_name);
   PADDLE_ENFORCE(origin_it != vars_.end(),
                  "Cannot find original variable with name %s", origin_name);
@@ -137,10 +152,12 @@ void Scope::Rename(const std::string& origin_name,
   vars_.erase(origin_it);
 }
 
-std::string Scope::Rename(const std::string& origin_name) const {
-  auto var_name = string::Sprintf("%p.%d", this, vars_.size());
-  Rename(origin_name, var_name);
-  return var_name;
+Variable* Scope::FindVarInternal(const std::string& name) const {
+  auto var = FindVarLocally(name);
+  if (var != nullptr) {
+    return var;
+  }
+  return (parent_ == nullptr) ? nullptr : parent_->FindVar(name);
 }
 
 Variable* Scope::FindVarLocally(const std::string& name) const {

paddle/fluid/framework/scope.h

@@ -88,12 +88,20 @@ class Scope {
   // Call Scope::NewScope for a sub-scope.
   explicit Scope(Scope const* parent) : parent_(parent) {}
 
+  // Called by Var.
+  Variable* VarInternal(const std::string& name);
+
+  // Called by FindScope.
+  const Scope* FindScopeInternal(const Variable* var) const;
+
+  // Called by Rename.
+  void RenameInternal(const std::string& origin_name,
+                      const std::string& new_name) const;
+
   // Called by FindVar recursively.
-  // Caller doesn't own the returned Variable.
   Variable* FindVarInternal(const std::string& name) const;
 
   // Called by FindVarInternal and Var.
-  // Caller doesn't own the returned Variable.
   Variable* FindVarLocally(const std::string& name) const;
 
   // Scope in `kids_` are owned by this class.

paddle/fluid/inference/tests/book/test_inference_nlp.cc

@@ -29,6 +29,7 @@ DEFINE_string(data_file, "", "File of input index data.");
 DEFINE_int32(repeat, 100, "Running the inference program repeat times");
 DEFINE_bool(prepare_vars, true, "Prepare variables before executor");
 DEFINE_int32(num_threads, 1, "Number of threads should be used");
+DECLARE_bool(use_mkldnn);
 
 inline double GetCurrentMs() {
   struct timeval time;
@@ -103,9 +104,9 @@ void ThreadRunInfer(
     const int tid, paddle::framework::Scope* scope,
     const std::vector<std::vector<const paddle::framework::LoDTensor*>>& jobs) {
   // maybe framework:ProgramDesc is not thread-safe
+  paddle::platform::CPUPlace place;
+  paddle::framework::Executor executor(place);
   auto& sub_scope = scope->NewScope();
-  auto place = paddle::platform::CPUPlace();
-  auto executor = paddle::framework::Executor(place);
   auto inference_program =
       paddle::inference::Load(&executor, scope, FLAGS_model_path);
 
@@ -182,8 +183,8 @@ TEST(inference, nlp) {
     stop_ms = GetCurrentMs();
   } else {
     // 1. Define place, executor, scope
-    auto place = paddle::platform::CPUPlace();
-    auto executor = paddle::framework::Executor(place);
+    paddle::platform::CPUPlace place;
+    paddle::framework::Executor executor(place);
 
     // 2. Initialize the inference_program and load parameters
     std::unique_ptr<paddle::framework::ProgramDesc> inference_program;

paddle/fluid/operators/activation_op.cc

@@ -19,18 +19,18 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-#define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT)               \
-  class OP_NAME##OpMaker                                                \
-      : public ::paddle::framework::OpProtoAndCheckerMaker {            \
-   public:                                                              \
-    void Make() override {                                              \
-      AddInput("X", "Input of " #OP_NAME " operator");                  \
-      AddOutput("Out", "Output of " #OP_NAME " operator").Reuse("X");   \
-      AddAttr<bool>("use_mkldnn",                                       \
-                    "(bool, default false) Only used in mkldnn kernel") \
-          .SetDefault(false);                                           \
-      AddComment(OP_COMMENT);                                           \
-    }                                                                   \
+#define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT)             \
+  class OP_NAME##OpMaker                                              \
+      : public ::paddle::framework::OpProtoAndCheckerMaker {          \
+   public:                                                            \
+    void Make() override {                                            \
+      AddInput("X", "Input of " #OP_NAME " operator");                \
+      AddOutput("Out", "Output of " #OP_NAME " operator").Reuse("X"); \
+      AddAttr<bool>("use_mkldnn",                                     \
+                    "(default false) Only used in mkldnn kernel")     \
+          .SetDefault(false);                                         \
+      AddComment(OP_COMMENT);                                         \
+    }                                                                 \
   }
 
 #define REGISTER_ACTIVATION_OP_GRAD_MAKER(OP_NAME, KERNEL_TYPE)              \
@@ -112,7 +112,7 @@ $$out = \frac{1}{1 + e^{-x}}$$
 __attribute__((unused)) constexpr char LogSigmoidDoc[] = R"DOC(
 Logsigmoid Activation Operator
 
-$$out = \log \frac{1}{1 + e^{-x}}$$
+$$out = \\log \\frac{1}{1 + e^{-x}}$$
 
 )DOC";
 
@@ -196,7 +196,7 @@ $out = [x]$
 __attribute__((unused)) constexpr char ReciprocalDoc[] = R"DOC(
 Reciprocal Activation Operator.
 
-$$out = \frac{1}{x}$$
+$$out = \\frac{1}{x}$$
 
 )DOC";
 

paddle/fluid/operators/crop_op.h

@@ -52,7 +52,7 @@ static std::vector<int> GetOffsets(const framework::ExecutionContext& ctx) {
   } else {
     res = ctx.Attr<std::vector<int>>("offsets");
     PADDLE_ENFORCE_EQ(
-        rank, res.size(),
+        rank, static_cast<int>(res.size()),
         "Offsets size should be equal to dimension size of input tensor.");
   }
   return res;

paddle/fluid/operators/detection/box_coder_op.cc

@@ -106,23 +106,36 @@ class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker {
               "and M represents the number of deocded boxes.");
 
     AddComment(R"DOC(
-Bounding Box Coder Operator.
+
+Bounding Box Coder.
+
 Encode/Decode the target bounding box with the priorbox information.
+
 The Encoding schema described below:
-ox = (tx - px) / pw / pxv
-oy = (ty - py) / ph / pyv
-ow = log(abs(tw / pw)) / pwv 
-oh = log(abs(th / ph)) / phv 
+
+    ox = (tx - px) / pw / pxv
+
+    oy = (ty - py) / ph / pyv
+
+    ow = log(abs(tw / pw)) / pwv 
+
+    oh = log(abs(th / ph)) / phv 
+
 The Decoding schema described below:
-ox = (pw * pxv * tx * + px) - tw / 2
-oy = (ph * pyv * ty * + py) - th / 2
-ow = exp(pwv * tw) * pw + tw / 2
-oh = exp(phv * th) * ph + th / 2
-where tx, ty, tw, th denote the target box's center coordinates, width and
-height respectively. Similarly, px, py, pw, ph denote the priorbox's(anchor)
-center coordinates, width and height. pxv, pyv, pwv, phv denote the variance
-of the priorbox and ox, oy, ow, oh denote the encoded/decoded coordinates,
-width and height.
+
+    ox = (pw * pxv * tx * + px) - tw / 2
+
+    oy = (ph * pyv * ty * + py) - th / 2
+
+    ow = exp(pwv * tw) * pw + tw / 2
+
+    oh = exp(phv * th) * ph + th / 2
+
+where `tx`, `ty`, `tw`, `th` denote the target box's center coordinates, width
+and height respectively. Similarly, `px`, `py`, `pw`, `ph` denote the
+priorbox's (anchor) center coordinates, width and height. `pxv`, `pyv`, `pwv`,
+`phv` denote the variance of the priorbox and `ox`, `oy`, `ow`, `oh` denote the
+encoded/decoded coordinates, width and height.
 )DOC");
   }
 };

paddle/fluid/operators/elementwise_mul_op.cc

@@ -15,7 +15,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/elementwise_mul_op.h"
 #include "paddle/fluid/operators/elementwise_op.h"
 namespace ops = paddle::operators;
-REGISTER_ELEMWISE_OP(elementwise_mul, "Mul", "Out = X \\odot\\ Y");
+REGISTER_ELEMWISE_OP(elementwise_mul, "Mul", "Out = X \\\\odot Y");
 REGISTER_OP_CPU_KERNEL(
     elementwise_mul,
     ops::ElementwiseMulKernel<paddle::platform::CPUDeviceContext, float>,

paddle/fluid/operators/gaussian_random_batch_size_like_op.cc

@@ -36,11 +36,12 @@ class GaussianRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker {
   void Apply() override {
     AddAttr<float>("mean",
                    "(float, default 0.0) "
-                   "mean of random tensor.")
+                   "The mean (or center) of the gaussian distribution.")
         .SetDefault(.0f);
     AddAttr<float>("std",
                    "(float, default 1.0) "
-                   "std of random tensor.")
+                   "The standard deviation (std, or spread) of the "
+                   "gaussian distribution.")
         .SetDefault(1.0f);
     AddAttr<int>("seed",
                  "(int, default 0) "
@@ -55,9 +56,11 @@ class GaussianRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker {
         .SetDefault(framework::proto::VarType::FP32);
 
     AddComment(R"DOC(
-GaussianRandom Operator.
 
 Used to initialize tensors with gaussian random generator.
+The defalut mean of the distribution is 0. and defalut standard
+deviation (std) of the distribution is 1.. Uers can set mean and std
+by input arguments.
 )DOC");
   }
 };

paddle/fluid/operators/nce_op.cc

@@ -128,8 +128,10 @@ class NCEOpMaker : public framework::OpProtoAndCheckerMaker {
                               "user should avoid setting this attribute.")
         .SetDefault({});
     AddComment(R"DOC(
-Compute and return the noise-contrastive estimation training loss.
-See [Noise-contrastive estimation: A new estimation principle for unnormalized statistical models](http://www.jmlr.org/proceedings/papers/v9/gutmann10a/gutmann10a.pdf).
+Compute and return the noise-contrastive estimation training loss. See 
+`Noise-contrastive estimation: A new estimation principle for unnormalized 
+statistical models 
+ <http://www.jmlr.org/proceedings/papers/v9/gutmann10a/gutmann10a.pdf>`_.
 By default this operator uses a uniform distribution for sampling.
 )DOC");
   }

paddle/fluid/pybind/pybind.cc

@@ -144,28 +144,74 @@ PYBIND11_PLUGIN(core) {
   py::class_<LoDTensor, Tensor>(m, "LoDTensor")
       .def_buffer(
           [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
-      .def(
-          "__init__",
-          [](LoDTensor &instance, const std::vector<std::vector<size_t>> &lod) {
-            LoD new_lod;
-            new_lod.reserve(lod.size());
-            std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
-            new (&instance) LoDTensor(new_lod);
-          })
+      .def("__init__",
+           [](LoDTensor &instance, const std::vector<std::vector<size_t>>
+                                       &recursive_sequence_lengths) {
+             LoD new_lod;
+             new_lod.reserve(recursive_sequence_lengths.size());
+             std::copy(recursive_sequence_lengths.begin(),
+                       recursive_sequence_lengths.end(),
+                       std::back_inserter(new_lod));
+             LoD new_offset_lod = ConvertToOffsetBasedLoD(new_lod);
+             PADDLE_ENFORCE(
+                 CheckLoD(new_offset_lod, -1),
+                 "the provided recursive_sequence_lengths info is invalid");
+             new (&instance) LoDTensor(new_offset_lod);
+           })
       .def("__init__", [](LoDTensor &instance) { new (&instance) LoDTensor(); })
       .def("set_lod",
            [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
+             // the input lod is offset-based level-of-detail info
+             LOG(WARNING)
+                 << "set_lod is deprecated and will be removed by 9.2018, "
+                    "please switch to set_recursive_sequence_lengths.";
              LoD new_lod;
              new_lod.reserve(lod.size());
              std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
+             PADDLE_ENFORCE(CheckLoD(new_lod, vectorize(self.dims()).front()),
+                            "the provided lod info is invalid");
              self.set_lod(new_lod);
            })
-      .def("lod", [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
-        auto lod = self.lod();
-        std::vector<std::vector<size_t>> new_lod;
-        new_lod.reserve(lod.size());
-        std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
-        return new_lod;
+      .def("set_recursive_sequence_lengths",
+           [](LoDTensor &self, const std::vector<std::vector<size_t>>
+                                   &recursive_sequence_lengths) {
+             // the input recursive_sequence_lengths is length-based
+             // level-of-detail info
+             LoD new_lod;
+             new_lod.reserve(recursive_sequence_lengths.size());
+             std::copy(recursive_sequence_lengths.begin(),
+                       recursive_sequence_lengths.end(),
+                       std::back_inserter(new_lod));
+             LoD new_offset_lod = ConvertToOffsetBasedLoD(new_lod);
+             PADDLE_ENFORCE(
+                 CheckLoD(new_offset_lod, vectorize(self.dims()).front()),
+                 "the provided recursive_sequence_lengths info is invalid");
+             self.set_lod(new_offset_lod);
+           })
+      .def("lod",
+           [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
+             // output the offset-based lod info
+             LOG(WARNING) << "lod is deprecated and will be removed by 9.2018, "
+                             "please switch to recursive_sequence_lengths.";
+             LoD lod = self.lod();
+             std::vector<std::vector<size_t>> new_lod;
+             new_lod.reserve(lod.size());
+             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
+             return new_lod;
+           })
+      .def("recursive_sequence_lengths",
+           [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
+             // output the length-based lod info
+             LoD lod = ConvertToLengthBasedLoD(self.lod());
+             std::vector<std::vector<size_t>> new_lod;
+             new_lod.reserve(lod.size());
+             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
+             return new_lod;
+           })
+      .def("has_valid_recursive_sequence_lengths", [](LoDTensor &self) -> bool {
+        // Check that the lod info is valid and match the outermost
+        // dimension of the LoDTensor data
+        return CheckLoD(self.lod(), vectorize(self.dims()).front());
       });
 
   py::class_<SelectedRows>(m, "SelectedRows")

paddle/testing/paddle_gtest_main.cc

@@ -31,6 +31,7 @@ int main(int argc, char** argv) {
       strdup("--tryfromenv=fraction_of_gpu_memory_to_use,use_pinned_memory"));
 #else
   new_argv.push_back(strdup("--tryfromenv=use_pinned_memory,use_mkldnn"));
+  new_argv.push_back(strdup("--undefok=use_mkldnn"));
 #endif
   int new_argc = static_cast<int>(new_argv.size());
   char** new_argv_address = new_argv.data();

python/paddle/fluid/data_feeder.py

@@ -47,7 +47,7 @@ class DataToLoDTensorConverter(object):
         self.lod = []
 
         for i in six.range(lod_level):
-            self.lod.append([0])
+            self.lod.append([])
 
     def feed(self, data):
         self._feed_impl_(data, self.lod, self.lod_level)
@@ -56,8 +56,7 @@ class DataToLoDTensorConverter(object):
         if lod_level == 0:
             self.data.append(data)
         else:
-            cur_lod_len = len(data)
-            lod[0].append(lod[0][-1] + cur_lod_len)
+            lod[0].append(len(data))
             for each_data in data:
                 self._feed_impl_(each_data, lod[1:], lod_level - 1)
 
@@ -66,7 +65,7 @@ class DataToLoDTensorConverter(object):
         t = core.LoDTensor()
         t.set(arr, self.place)
         if self.lod_level > 0:
-            t.set_lod(self.lod)
+            t.set_recursive_sequence_lengths(self.lod)
         return t
 
 

python/paddle/fluid/initializer.py

@@ -15,11 +15,13 @@
 import framework
 import numpy as np
 import contextlib
+from framework import convert_np_dtype_to_dtype_
+from core import VarDesc
 
 __all__ = [
-    'Constant', 'Uniform', 'Normal', 'Xavier', 'force_init_on_cpu',
+    'Constant', 'Uniform', 'Normal', 'Xavier', 'Bilinear', 'force_init_on_cpu',
     'init_on_cpu', 'ConstantInitializer', 'UniformInitializer',
-    'NormalInitializer', 'XavierInitializer'
+    'NormalInitializer', 'XavierInitializer', 'BilinearInitializer'
 ]
 
 _force_init_on_cpu_ = False
@@ -422,6 +424,101 @@ class MSRAInitializer(Initializer):
         return op
 
 
+class BilinearInitializer(Initializer):
+    """Implements the bilinear initializer.
+
+    This initializer can be used in transposed convolution operator to
+    act as upsampling. Users can upsample a feature map with shape of
+    (B, C, H, W) by any integer factor. The usage is:
+  
+    >>>  factor = 2
+    >>>  w_attr = ParamAttr(learning_rate=0., regularizer=L2Decay(0.),
+    >>>                     initializer=Bilinear())
+    >>>  conv_up = fluid.layers.conv2d_transpose(
+    >>>      input,
+    >>>      num_filters=C,
+    >>>      output_size=None,
+    >>>      filter_size=2 * factor - factor % 2,
+    >>>      padding=ceil((factor - 1) / 2.),
+    >>>      stride=factor,
+    >>>      groups=C,
+    >>>      param_attr=w_attr,
+    >>>      bias_attr=False)
+
+
+    Where, `num_filters=C` and `groups=C` means this is channel-wise tranposed
+    convolution. The filter shape will be (C, 1, K, K) where K is `filer_size`,
+    This initializer will set a (K, K) interpolation kernel for every channel
+    of the filter identically. The resulting shape of the output feature map
+    will be (B, C, factor * H, factor * W). Note that the learning rate and the
+    weight decay are set to 0 in order to keep coefficient values of bilinear
+    interpolation unchanged during training. 
+    """
+
+    def __init__(self):
+        """Constructor for BilinearInitializer.
+        """
+        super(BilinearInitializer, self).__init__()
+
+    def __call__(self, var, block):
+        """Add biliear initialization ops for a variable
+
+        Args:
+            var (Variable): Variable that needs to be initialized.
+            block (Block): The block in which initialization ops should
+                           be added.
+
+        Returns:
+            the initialization op
+
+        Raises:
+            ValueError: If type of `var` and `block` is not right.
+                        If the shape of `var` size is not 4 and
+                        var.shape[2] != var.shape[3].
+        """
+        if not isinstance(var, framework.Variable):
+            raise ValueError("var must be framework.Variable.")
+
+        if not isinstance(block, framework.Block):
+            raise ValueError("block must be framework.Block.")
+
+        shape = var.shape
+        if len(shape) != 4:
+            raise ValueError("the length of shape must be 4.")
+        if shape[2] != shape[3]:
+            raise ValueError("shape[2] must be equal to shape[3].")
+
+        weight = np.zeros(np.prod(var.shape), dtype='float32')
+        size = shape[3]
+        # factor
+        f = np.ceil(size / 2.)
+        # center
+        c = (2 * f - 1 - f % 2) / (2. * f)
+        for i in range(np.prod(shape)):
+            x = i % size
+            y = (i / size) % size
+            weight[i] = (1 - abs(x / f - c)) * (1 - abs(y / f - c))
+        weight = np.reshape(weight, shape)
+
+        if var.dtype == VarDesc.VarType.FP32:
+            value_name = "fp32_values"
+            values = [float(v) for v in weight.flat]
+        else:
+            raise ValueError("Unsupported dtype %s", input.dtype)
+        if np.prod(shape) > 1024 * 1024:
+            raise ValueError("The size of input is too big. ")
+        op = block.append_op(
+            type='assign_value',
+            outputs={'Out': [var]},
+            attrs={
+                'dtype': var.dtype,
+                'shape': list(shape),
+                value_name: values
+            })
+        var.op = op
+        return op
+
+
 # We short the class name, since users will use the initializer with the package
 # name. The sample code:
 #
@@ -436,3 +533,4 @@ Uniform = UniformInitializer
 Normal = NormalInitializer
 Xavier = XavierInitializer
 MSRA = MSRAInitializer
+Bilinear = BilinearInitializer