diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec
index f8c97b05bc08a26c1fa0bdcc2cd1abd932af158a..b1a380995065d7d06d1056afc1c552541194e54c 100644
--- a/paddle/fluid/API.spec
+++ b/paddle/fluid/API.spec
@@ -277,6 +277,7 @@ paddle.fluid.layers.has_nan (ArgSpec(args=['x'], varargs=None, keywords=None, de
 paddle.fluid.layers.isfinite (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '0a437011c3906079fd8947ed3e52d292'))
 paddle.fluid.layers.range (ArgSpec(args=['start', 'end', 'step', 'dtype'], varargs=None, keywords=None, defaults=None), ('document', '2ec937ede953ded2fdff2675883900bb'))
 paddle.fluid.layers.linspace (ArgSpec(args=['start', 'stop', 'num', 'dtype'], varargs=None, keywords=None, defaults=None), ('document', '495e21e9a848c2d075a102802fc67756'))
+paddle.fluid.layers.zeros_like (ArgSpec(args=['x', 'out'], varargs=None, keywords=None, defaults=(None,)), ('document', 'c7e4cfffc93ae89c8f6f53b6d650f923'))
 paddle.fluid.layers.While.__init__ (ArgSpec(args=['self', 'cond', 'is_test', 'name'], varargs=None, keywords=None, defaults=(False, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.layers.While.block (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.layers.Switch.__init__ (ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@@ -286,7 +287,11 @@ paddle.fluid.layers.increment (ArgSpec(args=['x', 'value', 'in_place'], varargs=
 paddle.fluid.layers.array_write (ArgSpec(args=['x', 'i', 'array'], varargs=None, keywords=None, defaults=(None,)), ('document', '40b6d15f4c86b2b09df340d7778ad713'))
 paddle.fluid.layers.create_array (ArgSpec(args=['dtype'], varargs=None, keywords=None, defaults=None), ('document', '2d4f20087080ba5105b55205ad5c5b6a'))
 paddle.fluid.layers.less_than (ArgSpec(args=['x', 'y', 'force_cpu', 'cond'], varargs=None, keywords=None, defaults=(None, None)), ('document', '067bbc799c66289ca8b8924c26b6673f'))
+paddle.fluid.layers.less_equal (ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', 'd6b173ae1a149e0bdfe7b8bf69285957'))
+paddle.fluid.layers.greater_than (ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', '2c9bd414caa6c615539018d27001b44c'))
+paddle.fluid.layers.greater_equal (ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', '62c667d24e7b07e166b47a53b61b2ff4'))
 paddle.fluid.layers.equal (ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', '80c29b1dc64718f0116de90d1ac88a77'))
+paddle.fluid.layers.not_equal (ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', '56148fb1024687a08e96af79bdc5c929'))
 paddle.fluid.layers.array_read (ArgSpec(args=['array', 'i'], varargs=None, keywords=None, defaults=None), ('document', 'dd68bead34dfbaf6b0a163fc1cc3c385'))
 paddle.fluid.layers.array_length (ArgSpec(args=['array'], varargs=None, keywords=None, defaults=None), ('document', 'ffb8b9578ec66db565b223d313aa82a2'))
 paddle.fluid.layers.IfElse.__init__ (ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@@ -319,6 +324,7 @@ paddle.fluid.layers.atan (ArgSpec(args=['x', 'name'], varargs=None, keywords=Non
 paddle.fluid.layers.tanh_shrink (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '1e521554b9fdda9061ec6d306f0709b7'))
 paddle.fluid.layers.softshrink (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '9eef31597bbafa2bd49691e072296e13'))
 paddle.fluid.layers.sqrt (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e9e27491c39ac74d0b1ffe506aec0ebb'))
+paddle.fluid.layers.rsqrt (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'c445467ebe58b3c0d7f0bba7795b6f56'))
 paddle.fluid.layers.abs (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '64650ac42cf82e9920cb0b172b1d29fd'))
 paddle.fluid.layers.ceil (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'c75d67dc5fe28f68e4cfffead4f698ad'))
 paddle.fluid.layers.floor (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '647b16c5da5ef909649ae02abb434973'))
@@ -331,13 +337,13 @@ paddle.fluid.layers.reciprocal (ArgSpec(args=['x', 'name'], varargs=None, keywor
 paddle.fluid.layers.square (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '48dfb45d773dbc30126c3a7f777de5ee'))
 paddle.fluid.layers.softplus (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '459c5781e9d1dd88283b7c5769d7872a'))
 paddle.fluid.layers.softsign (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '80846bcd4bd457207457a6d5411f4148'))
-paddle.fluid.layers.uniform_random (ArgSpec(args=['shape', 'dtype', 'min', 'max', 'seed'], varargs=None, keywords=None, defaults=('float32', -1.0, 1.0, 0)), ('document', '308b619af849caa82bbc31e897f5e641'))
+paddle.fluid.layers.uniform_random (ArgSpec(args=['shape', 'dtype', 'min', 'max', 'seed'], varargs=None, keywords=None, defaults=('float32', -1.0, 1.0, 0)), ('document', 'a8c4e972b7d6742c838a37abf407ed9a'))
 paddle.fluid.layers.hard_shrink (ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,)), ('document', 'c142f5884f3255e0d6075c286bbd531e'))
 paddle.fluid.layers.cumsum (ArgSpec(args=['x', 'axis', 'exclusive', 'reverse'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '944d7c03057f5fc88bc78acd4d82f926'))
 paddle.fluid.layers.thresholded_relu (ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,)), ('document', '90566ea449ea4c681435546e2f70610a'))
 paddle.fluid.layers.prior_box (ArgSpec(args=['input', 'image', 'min_sizes', 'max_sizes', 'aspect_ratios', 'variance', 'flip', 'clip', 'steps', 'offset', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, [1.0], [0.1, 0.1, 0.2, 0.2], False, False, [0.0, 0.0], 0.5, None, False)), ('document', '14cac0ee643fa6e026ad82aeeee75bd8'))
 paddle.fluid.layers.density_prior_box (ArgSpec(args=['input', 'image', 'densities', 'fixed_sizes', 'fixed_ratios', 'variance', 'clip', 'steps', 'offset', 'flatten_to_2d', 'name'], varargs=None, keywords=None, defaults=(None, None, None, [0.1, 0.1, 0.2, 0.2], False, [0.0, 0.0], 0.5, False, None)), ('document', 'a0d762bb08de9ce93bc780aa57cd5cd9'))
-paddle.fluid.layers.multi_box_head (ArgSpec(args=['inputs', 'image', 'base_size', 'num_classes', 'aspect_ratios', 'min_ratio', 'max_ratio', 'min_sizes', 'max_sizes', 'steps', 'step_w', 'step_h', 'offset', 'variance', 'flip', 'clip', 'kernel_size', 'pad', 'stride', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, 0.5, [0.1, 0.1, 0.2, 0.2], True, False, 1, 0, 1, None, False)), ('document', 'a6ab47a2fe681e52fabb7057ddf0efdd'))
+paddle.fluid.layers.multi_box_head (ArgSpec(args=['inputs', 'image', 'base_size', 'num_classes', 'aspect_ratios', 'min_ratio', 'max_ratio', 'min_sizes', 'max_sizes', 'steps', 'step_w', 'step_h', 'offset', 'variance', 'flip', 'clip', 'kernel_size', 'pad', 'stride', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, 0.5, [0.1, 0.1, 0.2, 0.2], True, False, 1, 0, 1, None, False)), ('document', 'fe9afaee481dd09f28866df22756466f'))
 paddle.fluid.layers.bipartite_match (ArgSpec(args=['dist_matrix', 'match_type', 'dist_threshold', 'name'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '3ddb9b966f193900193a95a3df77c3c1'))
 paddle.fluid.layers.target_assign (ArgSpec(args=['input', 'matched_indices', 'negative_indices', 'mismatch_value', 'name'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', 'c0b334f917828f95056f6ebe10907b1c'))
 paddle.fluid.layers.detection_output (ArgSpec(args=['loc', 'scores', 'prior_box', 'prior_box_var', 'background_label', 'nms_threshold', 'nms_top_k', 'keep_top_k', 'score_threshold', 'nms_eta'], varargs=None, keywords=None, defaults=(0, 0.3, 400, 200, 0.01, 1.0)), ('document', 'c33093a82a46e3091e789e5572588db1'))
@@ -352,7 +358,7 @@ paddle.fluid.layers.generate_mask_labels (ArgSpec(args=['im_info', 'gt_classes',
 paddle.fluid.layers.iou_similarity (ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '587845f60c5d97ffdf2dfd21da52eca1'))
 paddle.fluid.layers.box_coder (ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name', 'axis'], varargs=None, keywords=None, defaults=('encode_center_size', True, None, 0)), ('document', '032d0f4b7d8f6235ee5d91e473344f0e'))
 paddle.fluid.layers.polygon_box_transform (ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '0e5ac2507723a0b5adec473f9556799b'))
-paddle.fluid.layers.yolov3_loss (ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'gtscore', 'use_label_smooth', 'name'], varargs=None, keywords=None, defaults=(None, True, None)), ('document', '57fa96922e42db8f064c3fb77f2255e8'))
+paddle.fluid.layers.yolov3_loss (ArgSpec(args=['x', 'gt_box', 'gt_label', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'gt_score', 'use_label_smooth', 'name'], varargs=None, keywords=None, defaults=(None, True, None)), ('document', '059021025283ad1ee6f4d32228cf3e4e'))
 paddle.fluid.layers.yolo_box (ArgSpec(args=['x', 'img_size', 'anchors', 'class_num', 'conf_thresh', 'downsample_ratio', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '5566169a5ab993d177792c023c7fb340'))
 paddle.fluid.layers.box_clip (ArgSpec(args=['input', 'im_info', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '397e9e02b451d99c56e20f268fa03f2e'))
 paddle.fluid.layers.multiclass_nms (ArgSpec(args=['bboxes', 'scores', 'score_threshold', 'nms_top_k', 'keep_top_k', 'nms_threshold', 'normalized', 'nms_eta', 'background_label', 'name'], varargs=None, keywords=None, defaults=(0.3, True, 1.0, 0, None)), ('document', 'ca7d1107b6c5d2d6d8221039a220fde0'))
diff --git a/paddle/fluid/inference/tests/api/trt_models_tester.cc b/paddle/fluid/inference/tests/api/trt_models_tester.cc
index 98ce225a0476b38c021b0b81489f69d7953ae456..ec10e36c3b3707a88eebe116aaf3de454fc199b5 100644
--- a/paddle/fluid/inference/tests/api/trt_models_tester.cc
+++ b/paddle/fluid/inference/tests/api/trt_models_tester.cc
@@ -116,7 +116,7 @@ void compare_continuous_input(std::string model_dir, bool use_tensorrt) {
       reinterpret_cast<const PaddlePredictor::Config*>(&analysis_config);
   auto native_pred = CreateTestPredictor(config, false);
   auto analysis_pred = CreateTestPredictor(config, true);
-  for (int i = 0; i < 100; i++) {
+  for (int i = 0; i < 20; i++) {
     std::vector<std::vector<PaddleTensor>> inputs_all;
     if (!FLAGS_prog_filename.empty() && !FLAGS_param_filename.empty()) {
       SetFakeImageInput(&inputs_all, model_dir, true, FLAGS_prog_filename,
@@ -133,11 +133,13 @@ void compare_continuous_input(std::string model_dir, bool use_tensorrt) {
 TEST(TensorRT_mobilenet, compare) {
   std::string model_dir = FLAGS_infer_model + "/mobilenet";
   compare(model_dir, /* use_tensorrt */ true);
+  // Open it when need.
+  // profile(model_dir, /* use_analysis */ true, FLAGS_use_tensorrt);
 }
 
-TEST(TensorRT_resnet50, compare) {
+TEST(resnet50, compare_continuous_input) {
   std::string model_dir = FLAGS_infer_model + "/resnet50";
-  compare(model_dir, /* use_tensorrt */ true);
+  compare_continuous_input(model_dir, true);
 }
 
 TEST(TensorRT_resnext50, compare) {
@@ -145,24 +147,6 @@ TEST(TensorRT_resnext50, compare) {
   compare(model_dir, /* use_tensorrt */ true);
 }
 
-TEST(TensorRT_resnext50, profile) {
-  std::string model_dir = FLAGS_infer_model + "/resnext50";
-  // Set FLAGS_record_benchmark to true to record benchmark to file.
-  // FLAGS_record_benchmark=true;
-  FLAGS_model_name = "resnext50";
-  profile(model_dir, /* use_analysis */ true, FLAGS_use_tensorrt);
-}
-
-TEST(resnext50, compare_analysis_native) {
-  std::string model_dir = FLAGS_infer_model + "/resnext50";
-  compare(model_dir, false /*use tensorrt*/);
-}
-
-TEST(TensorRT_mobilenet, analysis) {
-  std::string model_dir = FLAGS_infer_model + "/" + "mobilenet";
-  compare(model_dir, false /* use_tensorrt */);
-}
-
 TEST(AnalysisPredictor, use_gpu) {
   std::string model_dir = FLAGS_infer_model + "/" + "mobilenet";
   AnalysisConfig config;
@@ -180,20 +164,5 @@ TEST(AnalysisPredictor, use_gpu) {
   }
 }
 
-TEST(TensorRT_mobilenet, profile) {
-  std::string model_dir = FLAGS_infer_model + "/" + "mobilenet";
-  profile(model_dir, true, false);
-}
-
-TEST(resnet50, compare_continuous_input) {
-  std::string model_dir = FLAGS_infer_model + "/resnet50";
-  compare_continuous_input(model_dir, true);
-}
-
-TEST(resnet50, compare_continuous_input_native) {
-  std::string model_dir = FLAGS_infer_model + "/resnet50";
-  compare_continuous_input(model_dir, false);
-}
-
 }  // namespace inference
 }  // namespace paddle
diff --git a/paddle/fluid/op_use_default_grad_op_maker.spec b/paddle/fluid/op_use_default_grad_op_maker.spec
index 21a25ce7d5e2bad172cf50cee6138ef4b44b07c1..403be1fc2c97a189a541c0c887eaadfe4266a124 100644
--- a/paddle/fluid/op_use_default_grad_op_maker.spec
+++ b/paddle/fluid/op_use_default_grad_op_maker.spec
@@ -18,7 +18,6 @@ gru
 hierarchical_sigmoid
 lrn
 lstm_unit
-lstmp
 max_pool2d_with_index
 max_pool3d_with_index
 maxout
diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc
index 1e5d63fc11d1d81350525e2b3390a3ae44f00f8d..348902c656cec1ea1eeaccc90feefd56d307111d 100644
--- a/paddle/fluid/operators/activation_op.cc
+++ b/paddle/fluid/operators/activation_op.cc
@@ -227,6 +227,15 @@ $out = \sqrt{x}$
 
 )DOC";
 
+UNUSED constexpr char RsqrtDoc[] = R"DOC(
+Rsqrt Activation Operator.
+
+Please make sure input is legal in case of numeric errors.
+
+$out = \frac{1}{\sqrt{x}}$
+
+)DOC";
+
 UNUSED constexpr char AbsDoc[] = R"DOC(
 Abs Activation Operator.
 
@@ -575,6 +584,7 @@ REGISTER_ACTIVATION_OP_MAKER(Gelu, GeluDoc);
 REGISTER_ACTIVATION_OP_MAKER(Tanh, TanhDoc);
 REGISTER_ACTIVATION_OP_MAKER(TanhShrink, TanhShrinkDoc);
 REGISTER_ACTIVATION_OP_MAKER(Sqrt, SqrtDoc);
+REGISTER_ACTIVATION_OP_MAKER(Rsqrt, RsqrtDoc);
 REGISTER_ACTIVATION_OP_MAKER(Abs, AbsDoc);
 REGISTER_ACTIVATION_OP_MAKER(Ceil, CeilDoc);
 REGISTER_ACTIVATION_OP_MAKER(Floor, FloorDoc);
@@ -586,6 +596,7 @@ REGISTER_ACTIVATION_OP_MAKER(Log, LogDoc);
 REGISTER_ACTIVATION_OP_MAKER(Square, SquareDoc);
 REGISTER_ACTIVATION_OP_MAKER(Softplus, SoftplusDoc);
 REGISTER_ACTIVATION_OP_MAKER(Softsign, SoftsignDoc);
+
 }  // namespace operators
 }  // namespace paddle
 
diff --git a/paddle/fluid/operators/activation_op.h b/paddle/fluid/operators/activation_op.h
index 915632a328feb99c021ec062a9b22a04623eff4a..1732f61582f79365d6872e15b9df1ee8f053903c 100644
--- a/paddle/fluid/operators/activation_op.h
+++ b/paddle/fluid/operators/activation_op.h
@@ -511,6 +511,26 @@ struct SqrtGradFunctor : public BaseActivationFunctor<T> {
   static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
+// rsqrt(x) = x^(-1/2)
+template <typename T>
+struct RsqrtFunctor : public BaseActivationFunctor<T> {
+  template <typename Device, typename X, typename Out>
+  void operator()(Device d, X x, Out out) const {
+    out.device(d) = x.rsqrt();
+  }
+};
+
+template <typename T>
+struct RsqrtGradFunctor : public BaseActivationFunctor<T> {
+  template <typename Device, typename X, typename Out, typename dOut,
+            typename dX>
+  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
+    dx.device(d) = static_cast<T>(-0.5) * dout * out * out * out;
+  }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
+};
+
 // ceil(x) = ceiling(x)
 template <typename T>
 struct CeilFunctor : public BaseActivationFunctor<T> {
@@ -1191,6 +1211,7 @@ struct SwishGradFunctor : public BaseActivationFunctor<T> {
   __macro(atan, Atan, AtanFunctor, AtanGradFunctor);                          \
   __macro(softshrink, SoftShrink, SoftShrinkFunctor, SoftShrinkGradFunctor);  \
   __macro(sqrt, Sqrt, SqrtFunctor, SqrtGradFunctor);                          \
+  __macro(rsqrt, Rsqrt, RsqrtFunctor, RsqrtGradFunctor);                      \
   __macro(abs, Abs, AbsFunctor, AbsGradFunctor);                              \
   __macro(ceil, Ceil, CeilFunctor, ZeroGradFunctor);                          \
   __macro(floor, Floor, FloorFunctor, ZeroGradFunctor);                       \
diff --git a/paddle/fluid/operators/affine_channel_op.cc b/paddle/fluid/operators/affine_channel_op.cc
index 268a5b894a95df8e27730879473b457a31e18cd6..27370a3c29a073f3ce6f01fd5aaf28b5ef1ca3a6 100644
--- a/paddle/fluid/operators/affine_channel_op.cc
+++ b/paddle/fluid/operators/affine_channel_op.cc
@@ -79,9 +79,13 @@ class AffineChannelOp : public framework::OperatorWithKernel {
                            : x_dims[x_dims.size() - 1]);
 
     PADDLE_ENFORCE_EQ(scale_dims.size(), 1UL);
-    PADDLE_ENFORCE_EQ(scale_dims[0], C);
     PADDLE_ENFORCE_EQ(b_dims.size(), 1UL);
-    PADDLE_ENFORCE_EQ(b_dims[0], C);
+    if (ctx->IsRuntime() || scale_dims[0] > 0) {
+      PADDLE_ENFORCE_EQ(scale_dims[0], C);
+    }
+    if (ctx->IsRuntime() || b_dims[0] > 0) {
+      PADDLE_ENFORCE_EQ(b_dims[0], C);
+    }
 
     ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
     ctx->ShareLoD("X", "Out");
diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc
index 494d26f58f23ad1e445bbe8d7f8ce1037e5aa598..0cc3e1f2b8350acb693ad7ba4f4dab270068723b 100644
--- a/paddle/fluid/operators/batch_norm_op.cc
+++ b/paddle/fluid/operators/batch_norm_op.cc
@@ -65,11 +65,22 @@ void BatchNormOp::InferShape(framework::InferShapeContext *ctx) const {
       (data_layout == DataLayout::kNCHW ? x_dims[1]
                                         : x_dims[x_dims.size() - 1]);
 
-  PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL);
-  PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], C);
-  PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1UL);
-  PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias")[0], C);
+  auto scale_dim = ctx->GetInputDim("Scale");
+  auto bias_dim = ctx->GetInputDim("Bias");
 
+  PADDLE_ENFORCE_EQ(scale_dim.size(), 1UL);
+  PADDLE_ENFORCE_EQ(scale_dim.size(), 1UL);
+
+  bool check = true;
+  if ((!ctx->IsRuntime()) && (framework::product(scale_dim) <= 0 ||
+                              framework::product(bias_dim) <= 0)) {
+    check = false;
+  }
+
+  if (check) {
+    PADDLE_ENFORCE_EQ(scale_dim[0], C);
+    PADDLE_ENFORCE_EQ(scale_dim[0], C);
+  }
   ctx->SetOutputDim("Y", x_dims);
   ctx->SetOutputDim("MeanOut", {C});
   ctx->SetOutputDim("VarianceOut", {C});
diff --git a/paddle/fluid/operators/concat_op.cc b/paddle/fluid/operators/concat_op.cc
index 1f71555180361a1522b7a1c8383fe128bc4edcd0..b1a6d66b80efdae3e78d7c3321a6107d2dd607aa 100644
--- a/paddle/fluid/operators/concat_op.cc
+++ b/paddle/fluid/operators/concat_op.cc
@@ -49,7 +49,15 @@ class ConcatOp : public framework::OperatorWithKernel {
     for (size_t i = 1; i < n; i++) {
       for (size_t j = 0; j < in_zero_dims_size; j++) {
         if (j == axis) {
-          out_dims[axis] += ins[i][j];
+          if (ctx->IsRuntime()) {
+            out_dims[axis] += ins[i][j];
+          } else {
+            if (ins[i][j] == -1) {
+              out_dims[axis] = -1;
+            } else {
+              out_dims[axis] += ins[i][j];
+            }
+          }
         } else {
           if (ctx->IsRuntime()) {
             // check all shape in run time
diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc
index 619e12e6ba7c73e46beafadd50770aedfb52c964..e1281602bf0d1bf25a2c4dfa32f495ed724d24eb 100644
--- a/paddle/fluid/operators/conv_op.cc
+++ b/paddle/fluid/operators/conv_op.cc
@@ -68,9 +68,14 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
 
   std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
   for (size_t i = 0; i < strides.size(); ++i) {
-    output_shape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
-                                          dilations[i], paddings[i],
-                                          strides[i]));
+    if ((!ctx->IsRuntime()) &&
+        (in_dims[i + 2] <= 0 || filter_dims[i + 2] <= 0)) {
+      output_shape.push_back(-1);
+    } else {
+      output_shape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
+                                            dilations[i], paddings[i],
+                                            strides[i]));
+    }
   }
   ctx->SetOutputDim("Output", framework::make_ddim(output_shape));
   ctx->ShareLoD("Input", "Output");
diff --git a/paddle/fluid/operators/conv_shift_op.cc b/paddle/fluid/operators/conv_shift_op.cc
index 08506ddd18ed35831702814e70962cb36ec958b1..fa4edb70b48e529102f11a1b0b9cac2110a33966 100644
--- a/paddle/fluid/operators/conv_shift_op.cc
+++ b/paddle/fluid/operators/conv_shift_op.cc
@@ -36,14 +36,17 @@ class ConvShiftOp : public framework::OperatorWithKernel {
     auto y_dims = ctx->GetInputDim("Y");
     PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
     PADDLE_ENFORCE_EQ(y_dims.size(), 2, "Input(Y)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(x_dims[0], y_dims[0],
-                      "The 1st dimension of Input(X) and Input(Y) should "
-                      "be equal.");
-    PADDLE_ENFORCE_EQ(y_dims[1] % 2, 1,
-                      "The 2nd dimension of Input(Y) should be odd.");
-    PADDLE_ENFORCE_LE(y_dims[1], x_dims[1],
-                      "The 2nd dimension of Input(Y) should be less than or "
-                      "equal to the 2nd dimension of Input(X).");
+    if (ctx->IsRuntime() || (x_dims[0] > 0 && y_dims[0] > 0))
+      PADDLE_ENFORCE_EQ(x_dims[0], y_dims[0],
+                        "The 1st dimension of Input(X) and Input(Y) should "
+                        "be equal.");
+    if (ctx->IsRuntime() || y_dims[1] > 0)
+      PADDLE_ENFORCE_EQ(y_dims[1] % 2, 1,
+                        "The 2nd dimension of Input(Y) should be odd.");
+    if (ctx->IsRuntime() || (x_dims[1] > 0 && y_dims[1] > 0))
+      PADDLE_ENFORCE_LE(y_dims[1], x_dims[1],
+                        "The 2nd dimension of Input(Y) should be less than or "
+                        "equal to the 2nd dimension of Input(X).");
     ctx->ShareDim("X", /*->*/ "Out");
     ctx->ShareLoD("X", /*->*/ "Out");
   }
diff --git a/paddle/fluid/operators/cos_sim_op.cc b/paddle/fluid/operators/cos_sim_op.cc
index 30ec74d8442d2f42510220b825988b340f79d0a2..93304ec6700b795c923f24a5d0663884b818b9b3 100644
--- a/paddle/fluid/operators/cos_sim_op.cc
+++ b/paddle/fluid/operators/cos_sim_op.cc
@@ -40,17 +40,27 @@ class CosSimOp : public framework::OperatorWithKernel {
     auto x_dims = ctx->GetInputDim("X");
     auto y_dims = ctx->GetInputDim("Y");
 
-    PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(),
-                      "Ranks of Input(X) and Input(Y) must be equal.");
-    PADDLE_ENFORCE_GE(x_dims.size(), 2,
-                      "Rank of Input(X) must not be less than 2.");
-    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 1, x_dims.size()),
-                      framework::slice_ddim(y_dims, 1, y_dims.size()),
-                      "All dimensions except the 1st of Input(X) and Input(Y) "
-                      "must be equal.");
-    PADDLE_ENFORCE(x_dims[0] == y_dims[0] || y_dims[0] == 1,
-                   "The 1st dimension of Input(Y) must be equal to Input(X) or"
-                   " just 1 (which will be broadcasted to match Input(X)).");
+    bool check = true;
+    if ((!ctx->IsRuntime()) &&
+        (framework::product(x_dims) <= 0 || framework::product(y_dims) <= 0)) {
+      check = false;
+    }
+
+    if (check) {
+      PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(),
+                        "Ranks of Input(X) and Input(Y) must be equal.");
+      PADDLE_ENFORCE_GE(x_dims.size(), 2,
+                        "Rank of Input(X) must not be less than 2.");
+      PADDLE_ENFORCE_EQ(
+          framework::slice_ddim(x_dims, 1, x_dims.size()),
+          framework::slice_ddim(y_dims, 1, y_dims.size()),
+          "All dimensions except the 1st of Input(X) and Input(Y) "
+          "must be equal.");
+      PADDLE_ENFORCE(
+          x_dims[0] == y_dims[0] || y_dims[0] == 1,
+          "The 1st dimension of Input(Y) must be equal to Input(X) or"
+          " just 1 (which will be broadcasted to match Input(X)).");
+    }
 
     // resize tensor
     ctx->SetOutputDim("Out", {x_dims[0], 1});
diff --git a/paddle/fluid/operators/detection_map_op.cc b/paddle/fluid/operators/detection_map_op.cc
index e1d113f8542da8827b9e36e44fc1bac6c07c9257..554e50725ffa5fc30849dc62fe525d72c6561a8b 100644
--- a/paddle/fluid/operators/detection_map_op.cc
+++ b/paddle/fluid/operators/detection_map_op.cc
@@ -51,8 +51,10 @@ class DetectionMAPOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(label_dims.size(), 2,
                       "The rank of Input(Label) must be 2, "
                       "the shape is [N, 6].");
-    PADDLE_ENFORCE(label_dims[1] == 6 || label_dims[1] == 5,
-                   "The shape of Input(Label) is [N, 6] or [N, 5].");
+    if (ctx->IsRuntime() || label_dims[1] > 0) {
+      PADDLE_ENFORCE(label_dims[1] == 6 || label_dims[1] == 5,
+                     "The shape of Input(Label) is [N, 6] or [N, 5].");
+    }
 
     if (ctx->HasInput("PosCount")) {
       PADDLE_ENFORCE(ctx->HasInput("TruePos"),
diff --git a/paddle/fluid/operators/distributed_ops/split_byref_op.cc b/paddle/fluid/operators/distributed_ops/split_byref_op.cc
index d65e7ffe5a492fe5df038bb6bd469e09de6f95ca..43980107c14176f1751a3db2858c80cb65c764de 100644
--- a/paddle/fluid/operators/distributed_ops/split_byref_op.cc
+++ b/paddle/fluid/operators/distributed_ops/split_byref_op.cc
@@ -31,14 +31,16 @@ class SplitByrefOp : public framework::OperatorWithKernel {
     auto in_dims = ctx->GetInputDim("X");
     auto outs_names = ctx->Outputs("Out");
     size_t num = static_cast<size_t>(ctx->Attrs().Get<int>("num"));
-    std::vector<int> sections = static_cast<std::vector<int>>(
-        ctx->Attrs().Get<std::vector<int>>("sections"));
+    auto sections = ctx->Attrs().Get<std::vector<int>>("sections");
     const size_t outs_number = outs_names.size();
     std::vector<framework::DDim> outs_dims;
     outs_dims.reserve(outs_number);
 
     if (num > 0) {
-      int64_t in_axis_dim = in_dims[0];
+      int64_t in_axis_dim = 0;
+      if (ctx->IsRuntime()) {
+        in_axis_dim = in_dims[0];
+      }
       PADDLE_ENFORCE_EQ(in_axis_dim % num, 0,
                         "tensor split does not result"
                         " in an equal division");
diff --git a/paddle/fluid/operators/grid_sampler_op.cc b/paddle/fluid/operators/grid_sampler_op.cc
index 241184c6f4a19a1da0d6d75c5d4e2b372c14e9da..57a1fcd42da04a766ebd8713e3863f259b3784ac 100644
--- a/paddle/fluid/operators/grid_sampler_op.cc
+++ b/paddle/fluid/operators/grid_sampler_op.cc
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/grid_sampler_op.h"
+#include <memory>
 #include "paddle/fluid/framework/op_registry.h"
 #ifdef PADDLE_WITH_CUDA
 #include "paddle/fluid/platform/cudnn_helper.h"
@@ -40,10 +41,12 @@ class GridSampleOp : public framework::OperatorWithKernel {
                    "Input(X) of GridSampleOp should be 4-D Tensor.");
     PADDLE_ENFORCE(grid_dims.size() == 4,
                    "Input(Grid) of GridSampleOp should be 4-D Tensor.");
-    PADDLE_ENFORCE(grid_dims[3] == 2, "Input(Grid) dims[3] should be 2.");
-    PADDLE_ENFORCE_EQ(grid_dims[0], x_dims[0],
-                      "Input(X) and Input(Grid) dims[0] should be equal.");
+    if (ctx->IsRuntime() || grid_dims[3] > 0) {
+      PADDLE_ENFORCE(grid_dims[3] == 2, "Input(Grid) dims[3] should be 2.");
+    }
     if (ctx->IsRuntime()) {
+      PADDLE_ENFORCE_EQ(grid_dims[0], x_dims[0],
+                        "Input(X) and Input(Grid) dims[0] should be equal.");
       PADDLE_ENFORCE_EQ(
           grid_dims[1], x_dims[2],
           "Input(X) dims[2] and Input(Grid) dims[1] should be equal.");
diff --git a/paddle/fluid/operators/hierarchical_sigmoid_op.h b/paddle/fluid/operators/hierarchical_sigmoid_op.h
index 82c8171ca52ffb128df103f27bafbdba1e72e52f..7cfe0aabcb7f3ce86ccc3a9a1c54b3b60d384aa1 100644
--- a/paddle/fluid/operators/hierarchical_sigmoid_op.h
+++ b/paddle/fluid/operators/hierarchical_sigmoid_op.h
@@ -238,6 +238,8 @@ class HierarchicalSigmoidGradOpKernel : public framework::OpKernel<T> {
       zero(dev_ctx, w_grad, static_cast<T>(0.0));
       bit_code->MulGradWeight(pre_out_grad, w_grad, in);
     } else {
+      PADDLE_ENFORCE(path != nullptr,
+                     "Sparse mode should not be used without custom tree!");
       framework::Vector<int64_t> real_rows = PathToRows(*path);
       auto* w_grad =
           ctx.Output<framework::SelectedRows>(framework::GradVarName("W"));
diff --git a/paddle/fluid/operators/interpolate_op.cc b/paddle/fluid/operators/interpolate_op.cc
index 9f2e3ad4a5ac1786096c67154d5a9ef5ea62855c..900b0c636ddafc8c033560adf58d596eb696621f 100644
--- a/paddle/fluid/operators/interpolate_op.cc
+++ b/paddle/fluid/operators/interpolate_op.cc
@@ -45,9 +45,14 @@ class InterpolateOp : public framework::OperatorWithKernel {
       // round down
       out_h = static_cast<int>(dim_x[2] * scale);
       out_w = static_cast<int>(dim_x[3] * scale);
+      // protect when input shape is -1
+      out_h = out_h > 0 ? out_h : -1;
+      out_w = out_w > 0 ? out_w : -1;
     } else {
       out_h = ctx->Attrs().Get<int>("out_h");
       out_w = ctx->Attrs().Get<int>("out_w");
+      PADDLE_ENFORCE_GT(out_h, 0, "out_h should be greater than 0.");
+      PADDLE_ENFORCE_GT(out_w, 0, "out_w should be greater than 0.");
     }
 
     if (ctx->HasInput("OutSize") && ctx->IsRuntime()) {
@@ -58,6 +63,7 @@ class InterpolateOp : public framework::OperatorWithKernel {
       ctx->ShareLoD("X", "Out");
       return;
     }
+
     std::vector<int64_t> dim_out({dim_x[0], dim_x[1], out_h, out_w});
     ctx->SetOutputDim("Out", framework::make_ddim(dim_out));
   }
diff --git a/paddle/fluid/operators/kldiv_loss_op.cc b/paddle/fluid/operators/kldiv_loss_op.cc
index a43f22c0496f89943d2fd5110446f1aae6a99315..a7c5d6305b09afb93be0b3b8524a91bd53e719fe 100644
--- a/paddle/fluid/operators/kldiv_loss_op.cc
+++ b/paddle/fluid/operators/kldiv_loss_op.cc
@@ -35,8 +35,10 @@ class KLDivLossOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(dim_x.size(), dim_target.size(),
                       "Input(X) rank and Input(Target) rank should be same.");
     for (int i = 0; i < dim_x.size(); i++) {
-      PADDLE_ENFORCE_EQ(dim_x[i], dim_target[i],
-                        "Input(X) and Input(Target) should in same shape.");
+      if (ctx->IsRuntime() || (dim_x[i] > 0 && dim_target[i] > 0)) {
+        PADDLE_ENFORCE_EQ(dim_x[i], dim_target[i],
+                          "Input(X) and Input(Target) should in same shape.");
+      }
     }
 
     auto reduction = ctx->Attrs().Get<std::string>("reduction");
diff --git a/paddle/fluid/operators/lstmp_op.cc b/paddle/fluid/operators/lstmp_op.cc
index 2728aa8a4ee21a9e1fe3deddcdba4c35a6aba7bc..f31c177c92d0a9e4cc731c478ea8339b450f318a 100644
--- a/paddle/fluid/operators/lstmp_op.cc
+++ b/paddle/fluid/operators/lstmp_op.cc
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/lstmp_op.h"
+#include <memory>
 #include <string>
 
 namespace paddle {
@@ -45,6 +46,7 @@ class LSTMPOp : public framework::OperatorWithKernel {
                    "Output(BatchHidden) of LSTMP operator should not be null.");
 
     auto in_dims = ctx->GetInputDim("Input");
+
     PADDLE_ENFORCE_EQ(in_dims.size(), 2,
                       "Input(X)'s rank of LSTMP operator must be 2.");
 
@@ -269,13 +271,47 @@ Users can choose to use fully-connected operator before LSTMP operator.
   }
 };
 
+class LSTMPGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto* grad_op = new framework::OpDesc();
+    grad_op->SetType("lstmp_grad");
+    grad_op->SetInput("Weight", Input("Weight"));
+    grad_op->SetInput("ProjWeight", Input("ProjWeight"));
+    grad_op->SetInput("Bias", Input("Bias"));
+
+    grad_op->SetInput("Projection", Output("Projection"));
+    grad_op->SetInput("Cell", Output("Cell"));
+    grad_op->SetInput("BatchGate", Output("BatchGate"));
+    grad_op->SetInput("BatchCellPreAct", Output("BatchCellPreAct"));
+    grad_op->SetInput("BatchHidden", Output("BatchHidden"));
+    grad_op->SetInput("H0", Input("H0"));
+    grad_op->SetInput("C0", Input("C0"));
+
+    grad_op->SetInput(framework::GradVarName("Projection"),
+                      OutputGrad("Projection"));
+
+    grad_op->SetOutput(framework::GradVarName("Input"), InputGrad("Input"));
+    grad_op->SetOutput(framework::GradVarName("Weight"), InputGrad("Weight"));
+    grad_op->SetOutput(framework::GradVarName("ProjWeight"),
+                       InputGrad("ProjWeight"));
+    grad_op->SetOutput(framework::GradVarName("Bias"), InputGrad("Bias"));
+    grad_op->SetOutput(framework::GradVarName("H0"), InputGrad("H0"));
+    grad_op->SetOutput(framework::GradVarName("C0"), InputGrad("C0"));
+
+    grad_op->SetAttrMap(Attrs());
+    return std::unique_ptr<framework::OpDesc>(grad_op);
+  }
+};
+
 class LSTMPGradOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
   void InferShape(framework::InferShapeContext* ctx) const override {
-    PADDLE_ENFORCE(ctx->HasInput("Input"),
-                   "Input(Input) of LSTMP operator should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Projection"),
                    "Input(Projection) of LSTMP operator should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Cell"),
@@ -298,7 +334,8 @@ class LSTMPGradOp : public framework::OperatorWithKernel {
         ctx->SetOutputDim(g_name, ctx->GetInputDim(name));
     };
 
-    SetOutGradDim("Input");
+    ctx->SetOutputDim(framework::GradVarName("Input"),
+                      ctx->GetInputDim("BatchGate"));
     SetOutGradDim("Weight");
     SetOutGradDim("ProjWeight");
     SetOutGradDim("Bias");
@@ -310,7 +347,8 @@ class LSTMPGradOp : public framework::OperatorWithKernel {
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
-        ctx.Input<framework::LoDTensor>("Input")->type(), ctx.device_context());
+        ctx.Input<framework::LoDTensor>("BatchGate")->type(),
+        ctx.device_context());
   }
 };
 
@@ -318,8 +356,7 @@ class LSTMPGradOp : public framework::OperatorWithKernel {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OPERATOR(lstmp, ops::LSTMPOp, ops::LSTMPOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
+REGISTER_OPERATOR(lstmp, ops::LSTMPOp, ops::LSTMPOpMaker, ops::LSTMPGradMaker);
 REGISTER_OPERATOR(lstmp_grad, ops::LSTMPGradOp);
 REGISTER_OP_CPU_KERNEL(
     lstmp, ops::LSTMPKernel<paddle::platform::CPUDeviceContext, float>,
diff --git a/paddle/fluid/operators/lstmp_op.h b/paddle/fluid/operators/lstmp_op.h
index c7d6e4205f8862526904e4fa767a2f4c4a2d8481..36da882639a235f27b4e5a9e77bf0813ea9c0ee3 100644
--- a/paddle/fluid/operators/lstmp_op.h
+++ b/paddle/fluid/operators/lstmp_op.h
@@ -267,7 +267,6 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
   }
 
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto* input = ctx.Input<LoDTensor>("Input");
     auto* weight = ctx.Input<Tensor>("Weight");
     auto* proj_weight = ctx.Input<Tensor>("ProjWeight");
     auto* bias = ctx.Input<Tensor>("Bias");
@@ -323,7 +322,8 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
       ordered_c0_g.mutable_data<T>(c0_g->dims(), ctx.GetPlace());
     }
 
-    auto in_dims = input->dims();
+    // batch_gate dims equal to input dims
+    auto in_dims = batch_gate->dims();
     auto out_dims = cell_out->dims();
     framework::DDim proj_dims({in_dims[0], proj_weight->dims()[1]});
     int frame_size = static_cast<int>(in_dims[1] / 4);
diff --git a/paddle/fluid/operators/merge_lod_tensor_op.cc b/paddle/fluid/operators/merge_lod_tensor_op.cc
index da7fa1b81d601f4dd03d6716de601a4b1abc7fa0..5edc233f6f73262c3d1b803aae0089f5b15d403d 100644
--- a/paddle/fluid/operators/merge_lod_tensor_op.cc
+++ b/paddle/fluid/operators/merge_lod_tensor_op.cc
@@ -164,7 +164,9 @@ class MergeLoDTensorInferShape : public framework::InferShapeBase {
 
     auto mask_dim = context->GetInputDim("Mask");
     PADDLE_ENFORCE_EQ(mask_dim.size(), 2);
-    PADDLE_ENFORCE_EQ(mask_dim[1], 1);
+    if (context->IsRuntime() || mask_dim[1] > 0) {
+      PADDLE_ENFORCE_EQ(mask_dim[1], 1);
+    }
 
     context->SetOutputDim("Out", context->GetInputDim("InTrue"));
   }
diff --git a/paddle/fluid/operators/mkldnn/batch_norm_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/batch_norm_mkldnn_op.cc
index bddca232e6c8a2a7fde998877006e37ee6d3d0dc..911c4d22ee5cd84c0b42646a1d3e62a0d765732e 100644
--- a/paddle/fluid/operators/mkldnn/batch_norm_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/batch_norm_mkldnn_op.cc
@@ -39,13 +39,9 @@ struct bn_type_traits {
 
 class BatchNormMKLDNNHandler : public platform::MKLDNNHandler {
  public:
-  BatchNormMKLDNNHandler(
-      std::shared_ptr<batch_norm_fwd::primitive_desc> batch_norm_pd,
-      const platform::MKLDNNDeviceContext &dev_ctx, mkldnn::engine engine,
-      const std::string &base_key)
-      : platform::MKLDNNHandler(dev_ctx, engine, base_key) {
-    batch_norm_pd_ = batch_norm_pd;
-  }
+  BatchNormMKLDNNHandler(const platform::MKLDNNDeviceContext &dev_ctx,
+                         mkldnn::engine engine, const std::string &base_key)
+      : platform::MKLDNNHandler(dev_ctx, engine, base_key) {}
 
   std::shared_ptr<memory> AcquireScaleshiftMemoryFromPrimitive(void *ptr) {
     return this->AcquireMemoryFromPrimitive(
@@ -62,6 +58,26 @@ class BatchNormMKLDNNHandler : public platform::MKLDNNHandler {
         batch_norm_pd_->variance_primitive_desc(), ptr, "@variance_mem_p");
   }
 
+  std::shared_ptr<batch_norm_fwd::primitive_desc>
+  AcquireBatchNormPrimitiveDescriptor(const batch_norm_fwd::desc &bn_fwd_desc,
+                                      const mkldnn::engine &engine) {
+    const std::string key_batch_norm_fwd_pd = key_ + "@bn_fwd_pd";
+    auto batch_norm_pd =
+        std::static_pointer_cast<batch_norm_fwd::primitive_desc>(
+            dev_ctx_.GetBlob(key_batch_norm_fwd_pd));
+
+    if (batch_norm_pd == nullptr) {
+      batch_norm_pd_.reset(
+          new batch_norm_fwd::primitive_desc(bn_fwd_desc, engine));
+      dev_ctx_.SetBlob(key_batch_norm_fwd_pd, batch_norm_pd_);
+    } else {
+      batch_norm_pd_ = batch_norm_pd;
+      is_reusing_ = true;
+    }
+
+    return batch_norm_pd_;
+  }
+
   std::shared_ptr<batch_norm_fwd> AcquireTestTrainingBatchNormFwd(
       std::shared_ptr<memory> src_memory,
       std::shared_ptr<memory> scaleshift_memory,
@@ -213,7 +229,7 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     const std::string key = BatchNormMKLDNNHandler::GetHash(
         src_tz, epsilon, flags, global_stats, input_format,
         ctx.op().Output("SavedMean"));
-    const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd";
+    BatchNormMKLDNNHandler handler(dev_ctx, mkldnn_engine, key);
 
     auto user_src_md = platform::MKLDNNMemDesc(
         {src_tz}, platform::MKLDNNGetDataType<T>(), input_format);
@@ -222,13 +238,9 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>;
     auto batch_norm_fwd_desc =
         bn_fwd_types::op_desc{propagation, user_src_md, epsilon, flags};
-    auto batch_norm_fwd_pd = std::make_shared<batch_norm_fwd::primitive_desc>(
-        batch_norm_fwd_desc, mkldnn_engine);
-    // Save conv_pd/src_memory/weights_memory for backward pass
-    dev_ctx.SetBlob(key_batch_norm_fwd_pd, batch_norm_fwd_pd);
 
-    BatchNormMKLDNNHandler handler(batch_norm_fwd_pd, dev_ctx, mkldnn_engine,
-                                   key);
+    auto batch_norm_fwd_pd = handler.AcquireBatchNormPrimitiveDescriptor(
+        batch_norm_fwd_desc, mkldnn_engine);
 
     auto src_memory =
         handler.AcquireSrcMemory(user_src_md, to_void_cast(x_data));
diff --git a/paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
index 5e4d79f1c35af42f662711ae9d8bfc650bab2b4f..faf518005c8cb0958dd5b0bbfc5c6fc4b3c2b582 100644
--- a/paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
@@ -144,7 +144,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     const std::string key = platform::ConvMKLDNNHandler::GetHash(
         src_tz, weights_tz, strides, paddings, dilations, groups,
         ctx.op().Input("Input") + ctx.op().Input("Filter"));
-    const std::string key_conv_pd = key + "@conv_pd";
 
     std::vector<primitive> pipeline;
 
@@ -183,6 +182,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto dst_md = platform::MKLDNNMemDesc(
         dst_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
 
+    platform::ConvMKLDNNHandler handler(dev_ctx, mkldnn_engine, key);
+
     // create a conv primitive descriptor and save it for usage in backward
     std::shared_ptr<mkldnn::convolution_forward::primitive_desc> conv_pd;
     auto fwd_prop_kind = is_test ? mkldnn::prop_kind::forward_inference
@@ -191,18 +192,14 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
       bias_tz = paddle::framework::vectorize2int(bias->dims());
       auto bias_md = platform::MKLDNNMemDesc(
           bias_tz, platform::MKLDNNGetDataType<T>(), memory::format::x);
-      conv_pd = ConvFwdPrimitiveDesc(
+      conv_pd = handler.AcquireConvolutionPrimitiveDescriptor(
           src_md, weights_md, bias_md, dst_md, strides, paddings, mkldnn_engine,
           fuse_relu, fuse_residual_conn, fwd_prop_kind);
     } else {
-      conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides,
-                                     paddings, mkldnn_engine, fuse_relu,
-                                     fuse_residual_conn, fwd_prop_kind);
+      conv_pd = handler.AcquireConvolutionPrimitiveDescriptor(
+          src_md, weights_md, boost::none, dst_md, strides, paddings,
+          mkldnn_engine, fuse_relu, fuse_residual_conn, fwd_prop_kind);
     }
-    // Save conv_pd/src_memory/weights_memory for backward pass
-    if (!is_test) dev_ctx.SetBlob(key_conv_pd, conv_pd);
-
-    platform::ConvMKLDNNHandler handler(conv_pd, dev_ctx, mkldnn_engine, key);
 
     // create mkldnn memory from input tensors (data/weights)
     auto user_src_memory_p =
@@ -633,31 +630,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
   }
 
  private:
-  mkldnn::primitive_attr CreatePostOps(bool fuse_relu,
-                                       bool fuse_residual_conn) const {
-    mkldnn::primitive_attr conv_attr;
-    mkldnn::post_ops post_operations;
-    // Fusion with Elementwise layer relies on adding a sum post-operation with
-    // the scale parameter. It is assumed that when fuse_residual_connection is
-    // true, the output tensor contains the data coming from residual
-    // connection. The result of this post_op is:
-    // Output = scale * Output + Conv_Out.
-    if (fuse_residual_conn) {
-      post_operations.append_sum(1.0f);
-    }
-    // Fusion with ReLU layer is executed through the PostOps feature. Create a
-    // PostOps object and configure it to execute an eltwise relu operation.
-    if (fuse_relu) {
-      constexpr float scale = 1.0f;
-      constexpr float negative_slope = 0.0f;
-      constexpr float placeholder = 0.0f;
-      post_operations.append_eltwise(scale, mkldnn::algorithm::eltwise_relu,
-                                     negative_slope, placeholder);
-    }
-    conv_attr.set_post_ops(post_operations);
-    return conv_attr;
-  }
-
   mkldnn::primitive_attr CreatePostOps(
       bool fuse_relu, bool fuse_residual_conn,
       const std::vector<float> output_shift_scale, float sum_scale) const {
@@ -679,30 +651,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     return conv_attr;
   }
 
-  std::unique_ptr<mkldnn::convolution_forward::primitive_desc>
-  ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights,
-                       const memory::desc& dst, const std::vector<int>& strides,
-                       const std::vector<int>& paddings,
-                       const mkldnn::engine& engine, const bool fuse_relu,
-                       const bool fuse_residual_conn,
-                       mkldnn::prop_kind fwd_prop_kind) const {
-    memory::dims stride_dims = strides;
-    memory::dims padding_dims = paddings;
-
-    auto conv_desc = mkldnn::convolution_forward::desc(
-        fwd_prop_kind, mkldnn::convolution_direct, src, weights, dst,
-        stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero);
-
-    mkldnn::primitive_attr conv_attr =
-        CreatePostOps(fuse_relu, fuse_residual_conn);
-
-    auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
-        conv_desc, conv_attr, engine);
-
-    return std::unique_ptr<mkldnn::convolution_forward::primitive_desc>(
-        p_conv_pd);
-  }
-
   std::unique_ptr<mkldnn::convolution_forward::primitive_desc>
   ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights,
                        const memory::desc& dst, const std::vector<int>& strides,
@@ -731,31 +679,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         p_conv_pd);
   }
 
-  std::unique_ptr<mkldnn::convolution_forward::primitive_desc>
-  ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights,
-                       const memory::desc& bias, const memory::desc& dst,
-                       const std::vector<int>& strides,
-                       const std::vector<int>& paddings,
-                       const mkldnn::engine& engine, const bool fuse_relu,
-                       const bool fuse_residual_conn,
-                       mkldnn::prop_kind fwd_prop_kind) const {
-    memory::dims stride_dims = strides;
-    memory::dims padding_dims = paddings;
-
-    auto conv_desc = mkldnn::convolution_forward::desc(
-        fwd_prop_kind, mkldnn::convolution_direct, src, weights, bias, dst,
-        stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero);
-
-    mkldnn::primitive_attr conv_attr =
-        CreatePostOps(fuse_relu, fuse_residual_conn);
-
-    auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
-        conv_desc, conv_attr, engine);
-
-    return std::unique_ptr<mkldnn::convolution_forward::primitive_desc>(
-        p_conv_pd);
-  }
-
   std::unique_ptr<mkldnn::convolution_forward::primitive_desc>
   ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights,
                        const memory::desc& bias, const memory::desc& dst,
diff --git a/paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc
index 317d4cebe26b81ff03c212e6328233d5152ed1b4..30d2469eeaf6938f1f93730b8b645ca2cfe97364 100644
--- a/paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc
@@ -12,6 +12,7 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
+#include "boost/optional.hpp"
 #include "paddle/fluid/framework/data_layout_transform.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/memory/malloc.h"
@@ -124,7 +125,6 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     const std::string key = platform::ConvTransposeMKLDNNHandler::GetHash(
         src_tz, weights_tz, strides, paddings, dilations, groups,
         ctx.op().Output("Output"));
-    const std::string key_conv_transpose_pd = key + "@conv_transpose_pd";
 
     std::vector<mkldnn::primitive> pipeline;
 
@@ -153,6 +153,7 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto dst_md = platform::MKLDNNMemDesc(
         dst_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
 
+    platform::ConvTransposeMKLDNNHandler handler(dev_ctx, mkldnn_engine, key);
     // create a deconv(conv transpose) primitive descriptor and save it for
     // usage in backward
     std::shared_ptr<mkldnn::deconvolution_forward::primitive_desc>
@@ -163,19 +164,14 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
       bias_tz = paddle::framework::vectorize2int(bias->dims());
       auto bias_md = platform::MKLDNNMemDesc(
           bias_tz, platform::MKLDNNGetDataType<T>(), mkldnn::memory::format::x);
-      conv_transpose_pd = ConvTransposeFwdPrimitiveDesc(
+      conv_transpose_pd = handler.AcquireConvolutionPrimitiveDescriptor(
           src_md, weights_md, bias_md, dst_md, strides, paddings, mkldnn_engine,
-          fuse_relu, fwd_prop_kind);
+          fuse_relu, false, fwd_prop_kind);
     } else {
-      conv_transpose_pd = ConvTransposeFwdPrimitiveDesc(
-          src_md, weights_md, dst_md, strides, paddings, mkldnn_engine,
-          fuse_relu, fwd_prop_kind);
+      conv_transpose_pd = handler.AcquireConvolutionPrimitiveDescriptor(
+          src_md, weights_md, boost::none, dst_md, strides, paddings,
+          mkldnn_engine, fuse_relu, false, fwd_prop_kind);
     }
-    // Save conv_pd/src_memory/weights_memory for backward pass
-    if (!is_test) dev_ctx.SetBlob(key_conv_transpose_pd, conv_transpose_pd);
-
-    platform::ConvTransposeMKLDNNHandler handler(conv_transpose_pd, dev_ctx,
-                                                 mkldnn_engine, key);
 
     // create mkldnn memory from input tensors (data/weights)
     auto user_src_memory_p = handler.AcquireSrcMemory(
@@ -224,70 +220,6 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     output->set_layout(DataLayout::kMKLDNN);
     output->set_format(platform::GetMKLDNNFormat(*dst_memory_p));
   }
-
- private:
-  mkldnn::primitive_attr CreatePostOps(bool fuse_relu) const {
-    mkldnn::primitive_attr conv_attr;
-    mkldnn::post_ops post_operations;
-    // Fusion with ReLU layer is executed through the PostOps feature. Create a
-    // PostOps object and configure it to execute an eltwise relu operation.
-    if (fuse_relu) {
-      constexpr float scale = 1.0f;
-      constexpr float negative_slope = 0.0f;
-      constexpr float placeholder = 0.0f;
-      post_operations.append_eltwise(scale, mkldnn::algorithm::eltwise_relu,
-                                     negative_slope, placeholder);
-    }
-    conv_attr.set_post_ops(post_operations);
-    return conv_attr;
-  }
-
-  std::unique_ptr<mkldnn::deconvolution_forward::primitive_desc>
-  ConvTransposeFwdPrimitiveDesc(
-      const mkldnn::memory::desc& src, const mkldnn::memory::desc& weights,
-      const mkldnn::memory::desc& dst, const std::vector<int>& strides,
-      const std::vector<int>& paddings, const mkldnn::engine& engine,
-      const bool fuse_relu, mkldnn::prop_kind fwd_prop_kind) const {
-    mkldnn::memory::dims stride_dims = {strides[0], strides[1]};
-    mkldnn::memory::dims padding_dims = {paddings[0], paddings[1]};
-
-    auto deconv_desc = mkldnn::deconvolution_forward::desc(
-        fwd_prop_kind, mkldnn::deconvolution_direct, src, weights, dst,
-        stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero);
-
-    mkldnn::primitive_attr deconv_attr = CreatePostOps(fuse_relu);
-
-    auto p_conv_transpose_pd =
-        new mkldnn::deconvolution_forward::primitive_desc(deconv_desc,
-                                                          deconv_attr, engine);
-
-    return std::unique_ptr<mkldnn::deconvolution_forward::primitive_desc>(
-        p_conv_transpose_pd);
-  }
-
-  std::unique_ptr<mkldnn::deconvolution_forward::primitive_desc>
-  ConvTransposeFwdPrimitiveDesc(
-      const mkldnn::memory::desc& src, const mkldnn::memory::desc& weights,
-      const mkldnn::memory::desc& bias, const mkldnn::memory::desc& dst,
-      const std::vector<int>& strides, const std::vector<int>& paddings,
-      const mkldnn::engine& engine, const bool fuse_relu,
-      mkldnn::prop_kind fwd_prop_kind) const {
-    mkldnn::memory::dims stride_dims = {strides[0], strides[1]};
-    mkldnn::memory::dims padding_dims = {paddings[0], paddings[1]};
-
-    auto deconv_desc = mkldnn::deconvolution_forward::desc(
-        fwd_prop_kind, mkldnn::deconvolution_direct, src, weights, bias, dst,
-        stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero);
-
-    mkldnn::primitive_attr deconv_attr = CreatePostOps(fuse_relu);
-
-    auto p_conv_transpose_pd =
-        new mkldnn::deconvolution_forward::primitive_desc(deconv_desc,
-                                                          deconv_attr, engine);
-
-    return std::unique_ptr<mkldnn::deconvolution_forward::primitive_desc>(
-        p_conv_transpose_pd);
-  }
 };
 
 }  // namespace operators
diff --git a/paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc
index dc1176f0848b93dd6872f676c3a71dab4f3455fd..1b3f33d345f4e0fafd7ad5da41eec052ac2dc504 100644
--- a/paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc
@@ -34,12 +34,9 @@ using platform::to_void_cast;
 
 class SoftmaxMKLDNNHandler : public platform::MKLDNNHandler {
  public:
-  SoftmaxMKLDNNHandler(
-      std::shared_ptr<mkldnn::softmax_forward::primitive_desc> softmax_pd,
-      const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine,
-      const std::string& base_key)
-      : platform::MKLDNNHandler(dev_ctx, engine, base_key),
-        softmax_pd_(softmax_pd) {}
+  SoftmaxMKLDNNHandler(const platform::MKLDNNDeviceContext& dev_ctx,
+                       mkldnn::engine engine, const std::string& base_key)
+      : platform::MKLDNNHandler(dev_ctx, engine, base_key) {}
 
   SoftmaxMKLDNNHandler(
       std::shared_ptr<mkldnn::softmax_forward::primitive_desc> softmax_pd,
@@ -54,6 +51,26 @@ class SoftmaxMKLDNNHandler : public platform::MKLDNNHandler {
     key_ += "-BWD";
   }
 
+  std::shared_ptr<softmax_forward::primitive_desc>
+  AcquireSoftmaxPrimitiveDescriptor(const softmax_forward::desc& softmax_desc,
+                                    const mkldnn::engine& engine) {
+    const std::string key_softmax_pd = key_ + "@softmax_pd";
+
+    auto softmax_pd = std::static_pointer_cast<softmax_forward::primitive_desc>(
+        dev_ctx_.GetBlob(key_softmax_pd));
+
+    if (softmax_pd == nullptr) {
+      softmax_pd_.reset(
+          new softmax_forward::primitive_desc(softmax_desc, engine));
+      dev_ctx_.SetBlob(key_softmax_pd, softmax_pd_);
+    } else {
+      softmax_pd_ = softmax_pd;
+      is_reusing_ = true;
+    }
+
+    return softmax_pd_;
+  }
+
   std::shared_ptr<mkldnn::softmax_forward> AcquireSoftmax(
       std::shared_ptr<mkldnn::memory> dst_memory_p,
       std::shared_ptr<mkldnn::memory> src_memory_p) {
@@ -138,19 +155,18 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
     // Generate keys for storing/retriving primitives for this operator
     const std::string key =
         platform::MKLDNNHandler::GetHash(softmax_tz, ctx.op().Output("Out"));
-    const std::string key_softmax_pd = key + "@softmax_pd";
 
+    SoftmaxMKLDNNHandler handler(dev_ctx, mkldnn_engine, key);
     // Currently only NC data format is supported
     auto softmax_md = MKLDNNMemDesc(
         {softmax_tz}, platform::MKLDNNGetDataType<T>(), memory::format::nc);
     // Normalization is made after innermost dimension eg. C out of NC
     auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring,
                                               softmax_md, 1 /*dim: C*/);
-    auto softmax_pd = std::make_shared<mkldnn::softmax_forward::primitive_desc>(
-        softmax_desc, mkldnn_engine);
-    dev_ctx.SetBlob(key_softmax_pd, softmax_pd);
 
-    SoftmaxMKLDNNHandler handler(softmax_pd, dev_ctx, mkldnn_engine, key);
+    auto softmax_pd =
+        handler.AcquireSoftmaxPrimitiveDescriptor(softmax_desc, mkldnn_engine);
+
     auto softmax_src_memory_p =
         handler.AcquireSrcMemory(softmax_md, to_void_cast<T>(input_data));
     auto softmax_dst_memory_p =
diff --git a/paddle/fluid/operators/pad2d_op.cc b/paddle/fluid/operators/pad2d_op.cc
index 9731aefa95c5243e29ace87ad8c35d5b01904e60..1e8ba5922aa96ac40798d103868c839242ac1e55 100644
--- a/paddle/fluid/operators/pad2d_op.cc
+++ b/paddle/fluid/operators/pad2d_op.cc
@@ -483,8 +483,10 @@ class Pad2dOp : public framework::OperatorWithKernel {
       PADDLE_ENFORCE_EQ(
           paddings_dim.size(), 1,
           "Size of Input(Paddings)'s dimension should be equal to 1.");
-      PADDLE_ENFORCE_EQ(paddings_dim[0], 4,
-                        "Shape of Input(Paddings) should be equal to [4].");
+      if (ctx->IsRuntime()) {
+        PADDLE_ENFORCE_EQ(paddings_dim[0], 4,
+                          "Shape of Input(Paddings) should be equal to [4].");
+      }
       out_dims[1] = x_dim[1];
       out_dims[2] = x_dim[2];
       out_dims[3] = x_dim[3];
@@ -504,11 +506,7 @@ class Pad2dOp : public framework::OperatorWithKernel {
     }
 
     ctx->SetOutputDim("Out", framework::make_ddim(out_dims));
-    if (out_dims[0] == x_dim[0]) {
-      // Only pass LoD when the first dimension is equal between
-      // output and input.
-      ctx->ShareLoD("X", /*->*/ "Out");
-    }
+    ctx->ShareLoD("X", /*->*/ "Out");
   }
 
  protected:
diff --git a/paddle/fluid/operators/roi_align_op.cc b/paddle/fluid/operators/roi_align_op.cc
index 7bb10ce063109dbd8520430d2b32ac9370ef8d25..d0dd861af7be80ede75b9d14867087ec687fc1da 100644
--- a/paddle/fluid/operators/roi_align_op.cc
+++ b/paddle/fluid/operators/roi_align_op.cc
@@ -37,9 +37,11 @@ class ROIAlignOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(rois_dims.size() == 2,
                    "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)"
                    "given as [[x1, y1, x2, y2], ...].");
-    PADDLE_ENFORCE(rois_dims[1] == 4,
-                   "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)"
-                   "given as [[x1, y1, x2, y2], ...].");
+    if (ctx->IsRuntime()) {
+      PADDLE_ENFORCE(rois_dims[1] == 4,
+                     "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)"
+                     "given as [[x1, y1, x2, y2], ...].");
+    }
     int pooled_height = ctx->Attrs().Get<int>("pooled_height");
     int pooled_width = ctx->Attrs().Get<int>("pooled_width");
     float spatial_scale = ctx->Attrs().Get<float>("spatial_scale");
diff --git a/paddle/fluid/operators/row_conv_op.cc b/paddle/fluid/operators/row_conv_op.cc
index 81aabdd0061b3940f23d4731d55fc5cbe5817004..7e9611679ba9a988f40973aaa37f04bcfa48f1ad 100644
--- a/paddle/fluid/operators/row_conv_op.cc
+++ b/paddle/fluid/operators/row_conv_op.cc
@@ -45,9 +45,12 @@ class RowConvOp : public framework::OperatorWithKernel {
     auto filter_dims = ctx->GetInputDim("Filter");
     PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
     PADDLE_ENFORCE_EQ(filter_dims.size(), 2, "Input(Y)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(
-        x_dims[1], filter_dims[1],
-        "The 2nd dimension of Input(X) and Input(Filter) should be same.");
+    if (ctx->IsRuntime() || (x_dims[1] > 0 && filter_dims[1] > 0)) {
+      PADDLE_ENFORCE_EQ(
+          x_dims[1], filter_dims[1],
+          "The 2nd dimension of Input(X) and Input(Filter) should be same.");
+    }
+
     ctx->SetOutputDim("Out", x_dims);
     ctx->ShareLoD("X", "Out");
   }
diff --git a/paddle/fluid/operators/sample_logits_op.cc b/paddle/fluid/operators/sample_logits_op.cc
index 9793118b7004477659b73a128b015f4fb4c0bea9..8ce2d52273d7cc3d523e5d77c2c79b9989b9227f 100644
--- a/paddle/fluid/operators/sample_logits_op.cc
+++ b/paddle/fluid/operators/sample_logits_op.cc
@@ -12,6 +12,7 @@ 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/operators/sample_logits_op.h"
+#include <memory>
 #include "paddle/fluid/operators/math/sample_prob.h"
 
 namespace paddle {
@@ -59,6 +60,10 @@ class SampleLogitsOpMaker : public framework::OpProtoAndCheckerMaker {
         "(Tensor, default: Tensor<float>), A 2-D tensor with shape [N, NT + S]."
         "The probabilites of sampled positive and negtive labels.")
         .AsIntermediate();
+    AddOutput("LogitsDim", "Store dim information of Logits for gradient op")
+        .AsIntermediate();
+    AddOutput("LabelsDim", "Store dim information of Logits for gradient op")
+        .AsIntermediate();
     AddOutput("SampledLogits",
               "(Tensor, default: Tensor<float>), A 2-D tensor with shape"
               "[N, NT + S]. The outputs value of sampled logits, which will be"
@@ -120,6 +125,10 @@ class SampleLogitsOp : public framework::OperatorWithKernel {
                    "Output(SampledLogits) should be not null.");
     PADDLE_ENFORCE(ctx->HasOutput("SampledLabels"),
                    "Output(SampledLabels) should be not null.");
+    PADDLE_ENFORCE(ctx->HasOutput("LogitsDim"),
+                   "Output(LogitsDim) should be not null.");
+    PADDLE_ENFORCE(ctx->HasOutput("LabelsDim"),
+                   "Output(LabelsDim) should be not null.");
 
     auto logits_dims = ctx->GetInputDim("Logits");
     auto labels_dims = ctx->GetInputDim("Labels");
@@ -139,6 +148,15 @@ class SampleLogitsOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("Probabilities", {logits_dims[0], num_sampled_classes});
     ctx->SetOutputDim("SampledLogits", {logits_dims[0], num_sampled_classes});
     ctx->SetOutputDim("SampledLabels", {logits_dims[0], labels_dims[1]});
+
+    // append 0 to shape variable to avoid optimized by memory optimize pass
+    auto logits_dim_vec = framework::vectorize(logits_dims);
+    logits_dim_vec.push_back(0);
+    ctx->SetOutputDim("LogitsDim", framework::make_ddim(logits_dim_vec));
+
+    auto labels_dim_vec = framework::vectorize(labels_dims);
+    labels_dim_vec.push_back(0);
+    ctx->SetOutputDim("LabelsDim", framework::make_ddim(labels_dim_vec));
   }
 
  protected:
@@ -157,28 +175,27 @@ class SampleLogitsOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
   void InferShape(framework::InferShapeContext* ctx) const override {
-    PADDLE_ENFORCE(ctx->HasInput("Logits"),
-                   "Input(Logits) should not be null.");
-    PADDLE_ENFORCE(ctx->HasInput("Labels"),
-                   "Input(Labels) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput("LogitsDim"),
+                   "Input(LogitsDim) should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("LabelsDim"),
+                   "Input(LabelsDim) should be not null.");
     PADDLE_ENFORCE(ctx->HasInput("Samples"),
                    "Input(Samples) should be not null.");
-    PADDLE_ENFORCE(ctx->HasInput("SampledLogits"),
-                   "Input(SampledLogits) should be not null.");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("SampledLogits")),
                    "Input(SampledLogits@Grad) should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Logits")),
                    "Output(Logits@Grad) should be not null.");
 
-    auto logit_dims = ctx->GetInputDim("Logits");
-    auto label_dims = ctx->GetInputDim("Labels");
-    PADDLE_ENFORCE_EQ(label_dims.size(), 2UL,
+    auto logits_dims = ctx->GetInputDim("LogitsDim");
+    logits_dims = framework::DDim(logits_dims.Get(), logits_dims.size() - 1);
+    auto labels_dims = ctx->GetInputDim("LabelsDim");
+    labels_dims = framework::DDim(labels_dims.Get(), labels_dims.size() - 1);
+    PADDLE_ENFORCE_EQ(labels_dims.size(), 2UL,
                       "The label should be a 2-D tensor.");
-    PADDLE_ENFORCE_EQ(logit_dims.size(), 2UL,
+    PADDLE_ENFORCE_EQ(logits_dims.size(), 2UL,
                       "The logits should be a 2-D tensor.");
 
-    ctx->SetOutputDim(framework::GradVarName("Logits"),
-                      ctx->GetInputDim("Logits"));
+    ctx->SetOutputDim(framework::GradVarName("Logits"), logits_dims);
   }
 
  protected:
@@ -201,10 +218,9 @@ class SampleLogitsGradMaker : public framework::SingleGradOpDescMaker {
   std::unique_ptr<framework::OpDesc> Apply() const override {
     auto* grad_op = new framework::OpDesc();
     grad_op->SetType("sample_logits_grad");
-    grad_op->SetInput("Logits", Input("Logits"));
-    grad_op->SetInput("Labels", Input("Labels"));
+    grad_op->SetInput("LogitsDim", Output("LogitsDim"));
+    grad_op->SetInput("LabelsDim", Output("LabelsDim"));
     grad_op->SetInput("Samples", Output("Samples"));
-    grad_op->SetInput("SampledLogits", Output("SampledLogits"));
     grad_op->SetInput(framework::GradVarName("SampledLogits"),
                       OutputGrad("SampledLogits"));
     grad_op->SetOutput(framework::GradVarName("Logits"), InputGrad("Logits"));
diff --git a/paddle/fluid/operators/scatter_op.cc b/paddle/fluid/operators/scatter_op.cc
index 8e0e3bd6054018852b242d1dba5c250394ed81ce..68ad223b3c311bec5968eb18b50f15e9da84e6d3 100644
--- a/paddle/fluid/operators/scatter_op.cc
+++ b/paddle/fluid/operators/scatter_op.cc
@@ -42,10 +42,6 @@ class ScatterOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("Updates")[0],
                       ctx->GetInputDim("Ids")[0],
                       "Updates and Ids should have same batch-size.");
-    framework::DDim data_dim(updates_dims);
-    for (int i = 1; i < data_dim.size(); ++i) {
-      PADDLE_ENFORCE_EQ(data_dim[i], updates_dims[i]);
-    }
     ctx->SetOutputDim("Out", ref_dims);
   }
 
diff --git a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc
index 5c92588cc1d073612d2f6a7b315edf16cc14bedd..1c2726454f3d1fb8545e5d3260e59fcafbcb2aee 100644
--- a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc
+++ b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc
@@ -34,15 +34,22 @@ class SigmoidCrossEntropyWithLogitsOp : public framework::OperatorWithKernel {
 
     auto x_dims = ctx->GetInputDim("X");
     auto labels_dims = ctx->GetInputDim("Label");
-    PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(labels_dims.size(), 2,
-                      "Input(Label)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(x_dims[0], labels_dims[0],
-                      "The 1st dimension of Input(X) and Input(Label) should "
-                      "be equal.");
-    PADDLE_ENFORCE_EQ(x_dims[1], labels_dims[1],
-                      "The 2nd dimension of Input(X) and Input(Label) should "
-                      "be equal.");
+
+    int rank = x_dims.size();
+    PADDLE_ENFORCE_EQ(rank, labels_dims.size(),
+                      "Input(X) and Input(Label) shall have the same rank.");
+    bool check = true;
+    if ((!ctx->IsRuntime()) && (framework::product(x_dims) <= 0 ||
+                                framework::product(labels_dims) <= 0)) {
+      check = false;
+    }
+
+    if (check) {
+      PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank),
+                        framework::slice_ddim(labels_dims, 0, rank),
+                        "Input(X) and Input(Label) shall have the same shape "
+                        "except the last dimension.");
+    }
 
     ctx->ShareDim("X", /*->*/ "Out");
     ctx->ShareLoD("X", /*->*/ "Out");
@@ -65,23 +72,24 @@ class SigmoidCrossEntropyWithLogitsGradOp
     auto x_dims = ctx->GetInputDim("X");
     auto labels_dims = ctx->GetInputDim("Label");
     auto dout_dims = ctx->GetInputDim(framework::GradVarName("Out"));
-    PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(labels_dims.size(), 2,
-                      "Input(Label)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(dout_dims.size(), 2,
-                      "Input(Out@Grad)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(x_dims[0], labels_dims[0],
-                      "The 1st dimension of Input(X) and Input(Label) should "
-                      "be equal.");
-    PADDLE_ENFORCE_EQ(x_dims[1], labels_dims[1],
-                      "The 2nd dimension of Input(X) and Input(Label) should "
-                      "be equal.");
-    PADDLE_ENFORCE_EQ(x_dims[0], dout_dims[0],
-                      "The 1st dimension of Input(X) and Input(Out@Grad) "
-                      "should be equal.");
-    PADDLE_ENFORCE_EQ(x_dims[1], dout_dims[1],
-                      "The 2nd dimension of Input(X) and Input(Out@Grad) "
-                      "should be equal.");
+
+    int rank = x_dims.size();
+    bool check = true;
+    if ((!ctx->IsRuntime()) && (framework::product(x_dims) <= 0 ||
+                                framework::product(labels_dims) <= 0)) {
+      check = false;
+    }
+
+    if (check) {
+      PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 0, rank),
+                        framework::slice_ddim(labels_dims, 0, rank),
+                        "Input(X) and Input(Label) shall have the same shape.");
+
+      PADDLE_ENFORCE_EQ(
+          framework::slice_ddim(x_dims, 0, rank),
+          framework::slice_ddim(dout_dims, 0, rank),
+          "Input(X) and Input(Out@Grad) shall have the same shape.");
+    }
 
     ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
   }
diff --git a/paddle/fluid/operators/spectral_norm_op.cc b/paddle/fluid/operators/spectral_norm_op.cc
index 04f659a465a345653d251cbe6703309c804fe614..ec5ee487729d0650983d553dbffe14b63c16b26a 100644
--- a/paddle/fluid/operators/spectral_norm_op.cc
+++ b/paddle/fluid/operators/spectral_norm_op.cc
@@ -56,13 +56,19 @@ class SpectralNormOp : public framework::OperatorWithKernel {
     }
     auto dim_u = ctx->GetInputDim("U");
     auto dim_v = ctx->GetInputDim("V");
-    PADDLE_ENFORCE_EQ(dim_u[0], h,
-                      "Input(U) dims[0] should be equal to "
-                      "Input(Weight) dims[Attr(dim)]");
-    PADDLE_ENFORCE_EQ(
-        dim_v[0], w,
-        "Input(V) dims[0] should be equal to "
-        "the product of Input(Weight) dims except dims[Attr(dim)]");
+
+    if (ctx->IsRuntime() || (dim_u[0] > 0 && h > 0)) {
+      PADDLE_ENFORCE_EQ(dim_u[0], h,
+                        "Input(U) dims[0] should be equal to "
+                        "Input(Weight) dims[Attr(dim)]");
+    }
+
+    if (ctx->IsRuntime() || (dim_v[0] > 0 && w > 0)) {
+      PADDLE_ENFORCE_EQ(
+          dim_v[0], w,
+          "Input(V) dims[0] should be equal to "
+          "the product of Input(Weight) dims except dims[Attr(dim)]");
+    }
 
     ctx->SetOutputDim("Out", dim_weight);
     ctx->ShareLoD("Weight", /*->*/ "Out");
diff --git a/paddle/fluid/operators/sum_op.cc b/paddle/fluid/operators/sum_op.cc
index 1391148ccf5d13082cb31ef2e143249e8ef95bfc..67f7510e874d4b3dcb857510e42cbfa7081becfe 100644
--- a/paddle/fluid/operators/sum_op.cc
+++ b/paddle/fluid/operators/sum_op.cc
@@ -65,7 +65,21 @@ class SumOp : public framework::OperatorWithKernel {
       if (framework::product(in_dim) == 0) {
         in_dim = x_dim;
       } else {
-        PADDLE_ENFORCE_EQ(in_dim, x_dim, "Input tensors must have same shape");
+        if (ctx->IsRuntime()) {
+          PADDLE_ENFORCE_EQ(in_dim, x_dim,
+                            "Input tensors must have same shape");
+        } else {
+          PADDLE_ENFORCE_EQ(in_dim.size(), x_dim.size(),
+                            "Input tensors must have same shape size");
+          // if in_dim or x_dim has -1, not check equal
+          for (int i = 0; i < x_dim.size(); ++i) {
+            if (x_dim[i] == -1 || in_dim[i] == -1) {
+              continue;
+            }
+            PADDLE_ENFORCE_EQ(in_dim[i], x_dim[i],
+                              "Input tensors must have same shape if not -1");
+          }
+        }
       }
     }
     ctx->SetOutputDim("Out", in_dim);
diff --git a/paddle/fluid/operators/unpool_op.cc b/paddle/fluid/operators/unpool_op.cc
index 11e505d6df3beda7053c59b66a29ec2badde3b75..86b4c06a27cc63fca8ec077cb3044ffe9415e01d 100644
--- a/paddle/fluid/operators/unpool_op.cc
+++ b/paddle/fluid/operators/unpool_op.cc
@@ -99,10 +99,15 @@ class UnpoolOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(in_x_dims.size() == 4,
                    "Unpooling intput must be of 4-dimensional.");
     PADDLE_ENFORCE_EQ(in_x_dims, in_y_dims);
+
     std::vector<int64_t> output_shape({in_x_dims[0], in_x_dims[1]});
     for (size_t i = 0; i < ksize.size(); ++i) {
-      output_shape.push_back(UnpoolOutputSize(in_x_dims[i + 2], ksize[i],
-                                              paddings[i], strides[i]));
+      if (!ctx->IsRuntime() && in_x_dims[i + 2] <= 0) {
+        output_shape.push_back(-1);
+      } else {
+        output_shape.push_back(UnpoolOutputSize(in_x_dims[i + 2], ksize[i],
+                                                paddings[i], strides[i]));
+      }
     }
     ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
   }
diff --git a/paddle/fluid/platform/mkldnn_reuse.h b/paddle/fluid/platform/mkldnn_reuse.h
index ecaad4ec070fe60a522839e0718c424a441dec0b..ba3a82b4b07f4dcb3f0037e398c146ab167d7b57 100644
--- a/paddle/fluid/platform/mkldnn_reuse.h
+++ b/paddle/fluid/platform/mkldnn_reuse.h
@@ -16,6 +16,7 @@ limitations under the License. */
 #include <memory>
 #include <string>
 #include <vector>
+#include "boost/optional.hpp"
 #include "paddle/fluid/framework/data_layout_transform.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/platform/mkldnn_helper.h"
@@ -395,9 +396,28 @@ class TransposeMKLDNNHandler : public MKLDNNHandler {
   std::vector<int> logical_axis_;
 };
 
+template <typename T>
+struct convolutional_algorithm;
+
+template <>
+struct convolutional_algorithm<mkldnn::convolution_forward> {
+  static constexpr mkldnn::algorithm T = mkldnn::algorithm::convolution_direct;
+};
+
+template <>
+struct convolutional_algorithm<mkldnn::deconvolution_forward> {
+  static constexpr mkldnn::algorithm T =
+      mkldnn::algorithm::deconvolution_direct;
+};
+
 template <class forward_t, class backward_data_t, class backward_weights_t>
 class ConvMKLDNNTemplateHandler : public MKLDNNHandler {
  public:
+  ConvMKLDNNTemplateHandler(const platform::MKLDNNDeviceContext& dev_ctx,
+                            mkldnn::engine engine, const std::string& base_key)
+      : platform::MKLDNNHandler(dev_ctx, engine, base_key) {}
+
+  // TODO(jczaja): remove after conv int8 is adapted
   ConvMKLDNNTemplateHandler(
       std::shared_ptr<typename forward_t::primitive_desc> conv_pd,
       const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine,
@@ -542,6 +562,73 @@ class ConvMKLDNNTemplateHandler : public MKLDNNHandler {
                                scale_data, mask);
   }
 
+  mkldnn::primitive_attr CreatePostOps(bool fuse_relu,
+                                       bool fuse_residual_conn = false) const {
+    mkldnn::primitive_attr conv_attr;
+    mkldnn::post_ops post_operations;
+    // Fusion with Elementwise layer relies on adding a sum post-operation with
+    // the scale parameter. It is assumed that when fuse_residual_connection is
+    // true, the output tensor contains the data coming from residual
+    // connection. The result of this post_op is:
+    // Output = scale * Output + Conv_Out.
+    if (fuse_residual_conn) {
+      post_operations.append_sum(1.0f);
+    }
+    // Fusion with ReLU layer is executed through the PostOps feature. Create a
+    // PostOps object and configure it to execute an eltwise relu operation.
+    if (fuse_relu) {
+      constexpr float scale = 1.0f;
+      constexpr float negative_slope = 0.0f;
+      constexpr float placeholder = 0.0f;
+      post_operations.append_eltwise(scale, mkldnn::algorithm::eltwise_relu,
+                                     negative_slope, placeholder);
+    }
+    conv_attr.set_post_ops(post_operations);
+    return conv_attr;
+  }
+
+  std::shared_ptr<typename forward_t::primitive_desc>
+  AcquireConvolutionPrimitiveDescriptor(
+      const mkldnn::memory::desc& src, const mkldnn::memory::desc& weights,
+      boost::optional<const mkldnn::memory::desc&> bias,
+      const mkldnn::memory::desc& dst, const std::vector<int>& strides,
+      const std::vector<int>& paddings, const mkldnn::engine& engine,
+      const bool fuse_relu, const bool fuse_residual_conn,
+      mkldnn::prop_kind fwd_prop_kind) {
+    const std::string key_conv_pd = key_ + "@conv_pd";
+
+    auto conv_pd = std::static_pointer_cast<typename forward_t::primitive_desc>(
+        dev_ctx_.GetBlob(key_conv_pd));
+
+    if (conv_pd == nullptr) {
+      mkldnn::memory::dims stride_dims = strides;
+      mkldnn::memory::dims padding_dims = paddings;
+
+      auto conv_desc =
+          bias ? typename forward_t::desc(
+                     fwd_prop_kind, convolutional_algorithm<forward_t>::T, src,
+                     weights, *bias, dst, stride_dims, padding_dims,
+                     padding_dims, mkldnn::padding_kind::zero)
+               : typename forward_t::desc(
+                     fwd_prop_kind, convolutional_algorithm<forward_t>::T, src,
+                     weights, dst, stride_dims, padding_dims, padding_dims,
+                     mkldnn::padding_kind::zero);
+
+      mkldnn::primitive_attr conv_attr =
+          CreatePostOps(fuse_relu, fuse_residual_conn);
+
+      conv_pd_.reset(
+          new typename forward_t::primitive_desc(conv_desc, conv_attr, engine));
+      // Save conv_pd/src_memory/weights_memory for backward pass
+      dev_ctx_.SetBlob(key_conv_pd, conv_pd_);
+    } else {
+      conv_pd_ = conv_pd;
+      is_reusing_ = true;
+    }
+
+    return conv_pd_;
+  }
+
   std::shared_ptr<forward_t> AcquireConvolution(
       std::shared_ptr<mkldnn::memory> src_memory_p,
       std::shared_ptr<mkldnn::memory> weights_memory_p,
diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh
index 7bb713493182239b2fd17f7b7fb496afdc9b8e6c..ccfcb13db6e7339b6770242d8beb60152d2a25ef 100755
--- a/paddle/scripts/paddle_build.sh
+++ b/paddle/scripts/paddle_build.sh
@@ -446,7 +446,8 @@ function assert_api_spec_approvals() {
         BRANCH="develop"
     fi
 
-    API_FILES=("paddle/fluid/API.spec"
+    API_FILES=("CMakeLists.txt"
+               "paddle/fluid/API.spec"
                "paddle/fluid/op_use_default_grad_op_maker.spec"
                "python/paddle/fluid/parallel_executor.py"
                "paddle/fluid/framework/operator.h"
@@ -469,24 +470,29 @@ function assert_api_spec_approvals() {
       echo "checking ${API_FILE} change, PR: ${GIT_PR_ID}, changes: ${API_CHANGE}"
       if [ ${API_CHANGE} ] && [ "${GIT_PR_ID}" != "" ]; then
           # NOTE: per_page=10000 should be ok for all cases, a PR review > 10000 is not human readable.
-          # approval_user_list: velconia 1979255,panyx0718 2887803,XiaoguangHu01 46782768,chengduoZH 30176695,Xreki 12538138,luotao1 6836917,sneaxiy 32832641,tensor-tang 21351065,jacquesqiao 3048612,typhoonzero 13348433,shanyi15 35982308. 
+          # approval_user_list: velconia 1979255,XiaoguangHu01 46782768,chengduoZH 30176695,Xreki 12538138,luotao1 6836917,sneaxiy 32832641,tensor-tang 21351065,jacquesqiao 3048612,typhoonzero 13348433,shanyi15 35982308. 
           if [ "$API_FILE" == "paddle/fluid/API.spec" ];then
             APPROVALS=`curl -H "Authorization: token ${GITHUB_API_TOKEN}" https://api.github.com/repos/PaddlePaddle/Paddle/pulls/${GIT_PR_ID}/reviews?per_page=10000 | \
-            python ${PADDLE_ROOT}/tools/check_pr_approval.py 2 2887803 35982308 46782768 30176695`
+            python ${PADDLE_ROOT}/tools/check_pr_approval.py 2 35982308 46782768 30176695`
             if [ "${APPROVALS}" == "TRUE" ];then
               APPROVALS=`curl -H "Authorization: token ${GITHUB_API_TOKEN}" https://api.github.com/repos/PaddlePaddle/Paddle/pulls/${GIT_PR_ID}/reviews?per_page=10000 | \
               python ${PADDLE_ROOT}/tools/check_pr_approval.py 1 35982308`
             fi
+          elif [ "$API_FILE" == "CMakeLists.txt" ];then
+            APPROVALS=`curl -H "Authorization: token ${GITHUB_API_TOKEN}" https://api.github.com/repos/PaddlePaddle/Paddle/pulls/${GIT_PR_ID}/reviews?per_page=10000 | \
+            python ${PADDLE_ROOT}/tools/check_pr_approval.py 1 6836917 46782768 30176695`
           else
             APPROVALS=`curl -H "Authorization: token ${GITHUB_API_TOKEN}" https://api.github.com/repos/PaddlePaddle/Paddle/pulls/${GIT_PR_ID}/reviews?per_page=10000 | \
-            python ${PADDLE_ROOT}/tools/check_pr_approval.py 1 2887803 1979255 21351065 3048612 13348433 46782768 30176695 12538138 6836917 32832641`
+            python ${PADDLE_ROOT}/tools/check_pr_approval.py 1 1979255 21351065 3048612 13348433 46782768 30176695 12538138 6836917 32832641`
           fi
           echo "current pr ${GIT_PR_ID} got approvals: ${APPROVALS}"
           if [ "${APPROVALS}" == "FALSE" ]; then
             if [ "$API_FILE" == "paddle/fluid/API.spec" ];then
-              echo "You must have one RD (panyx0718 or chengduoZH or XiaoguangHu01) and one PM (shanyi15) approval for the api change! ${API_FILE}"
+              echo "You must have one RD (chengduoZH or XiaoguangHu01) and one PM (shanyi15) approval for the api change! ${API_FILE}"
+            elif [ "$API_FILE" == "CMakeLists.txt" ];then
+              echo "You must have one RD (luotao1 or chengduoZH or XiaoguangHu01) approval for the cmakelist change! ${API_FILE}"
             else
-              echo "You must have one RD (velconia,panyx0718,XiaoguangHu01,chengduoZH,Xreki,luotao1,sneaxiy,tensor-tang,jacquesqiao,typhoonzero) approval for the api change! ${API_FILE}"
+              echo "You must have one RD (velconia,XiaoguangHu01,chengduoZH,Xreki,luotao1,sneaxiy,tensor-tang,jacquesqiao,typhoonzero) approval for the api change! ${API_FILE}"
             fi
             exit 1
           fi
@@ -496,10 +502,10 @@ function assert_api_spec_approvals() {
     HAS_CONST_CAST=`git diff -U0 upstream/$BRANCH |grep -o -m 1 "const_cast" || true`
     if [ ${HAS_CONST_CAST} ] && [ "${GIT_PR_ID}" != "" ]; then
         APPROVALS=`curl -H "Authorization: token ${GITHUB_API_TOKEN}" https://api.github.com/repos/PaddlePaddle/Paddle/pulls/${GIT_PR_ID}/reviews?per_page=10000 | \
-        python ${PADDLE_ROOT}/tools/check_pr_approval.py 1 2887803 1979255 21351065 3048612 13348433 46782768 30176695 12538138 6836917 32832641`
+        python ${PADDLE_ROOT}/tools/check_pr_approval.py 1 1979255 21351065 3048612 13348433 46782768 30176695 12538138 6836917 32832641`
         echo "current pr ${GIT_PR_ID} got approvals: ${APPROVALS}"
         if [ "${APPROVALS}" == "FALSE" ]; then
-            echo "You must have one RD (velconia,panyx0718,XiaoguangHu01,chengduoZH,Xreki,luotao1,sneaxiy,tensor-tang,jacquesqiao,typhoonzero) approval for the api change! ${API_FILE}"
+            echo "You must have one RD (velconia,XiaoguangHu01,chengduoZH,Xreki,luotao1,sneaxiy,tensor-tang,jacquesqiao,typhoonzero) approval for the api change! ${API_FILE}"
             exit 1
         fi
     fi
diff --git a/python/paddle/fluid/dygraph/layers.py b/python/paddle/fluid/dygraph/layers.py
index 39e06e3486cd5479f69cbdb67811f03bd9646123..6b78e2abb32ab3f134fc7bae6fbd203a2730cf66 100644
--- a/python/paddle/fluid/dygraph/layers.py
+++ b/python/paddle/fluid/dygraph/layers.py
@@ -48,6 +48,12 @@ class Layer(core.Layer):
 
         self._helper = LayerObjectHelper(self._full_name)
 
+    def train(self):
+        framework._dygraph_tracer()._train_mode()
+
+    def eval(self):
+        framework._dygraph_tracer()._eval_mode()
+
     def full_name(self):
         """Full name for this layers.
 
@@ -254,6 +260,12 @@ class PyLayer(core.PyLayer):
     def __init__(self):
         super(PyLayer, self).__init__()
 
+    def train(self):
+        framework._dygraph_tracer()._train_mode()
+
+    def eval(self):
+        framework._dygraph_tracer()._eval_mode()
+
     @classmethod
     def _do_forward(cls, inputs):
         return cls._to_tuple(cls.forward(inputs))
diff --git a/python/paddle/fluid/dygraph/tracer.py b/python/paddle/fluid/dygraph/tracer.py
index 94e212b139b2b375aa9f5252d396e90235ba33c1..ee37ffab2cb7521b83108a40febcfe88cab28633 100644
--- a/python/paddle/fluid/dygraph/tracer.py
+++ b/python/paddle/fluid/dygraph/tracer.py
@@ -24,7 +24,9 @@ __all__ = ['Tracer']
 
 
 def release_op(op):
-    del framework._dygraph_tracer()._ops[op._trace_id]
+    del framework._dygraph_tracer()._ops[op._trace_id].inputs
+    del framework._dygraph_tracer()._ops[op._trace_id].outputs
+    del framework._dygraph_tracer()._ops[op._trace_id].backward_refs
 
 
 class Tracer(core.Tracer):
@@ -38,6 +40,7 @@ class Tracer(core.Tracer):
         self._ops = defaultdict()
         self._vars = defaultdict()
         self._trace_id = 0
+        self._train_mode = True
 
     def trace_var(self, name, var):
         self._vars[name] = var
@@ -46,15 +49,57 @@ class Tracer(core.Tracer):
         return list((item for name, item in six.iteritems(self._vars)
                      if isinstance(item, framework.Parameter)))
 
-    def trace_op(self, op, stop_gradient=False):
+    def trace_op(self, op, inputs, outputs, stop_gradient=False):
+        # TODO(minqiyang): remove this line after we take apart all
+        # backward grads and forward variables
+        if self._train_mode:
+            op.inputs = inputs
+            inps = defaultdict(list)
+            for k, vars in six.iteritems(inputs):
+                if isinstance(vars, framework.Variable):
+                    inps[k].append(vars._ivar)
+                elif isinstance(vars, list) or isinstance(vars, tuple):
+                    for var in vars:
+                        inps[k].append(var._ivar)
+
+            op.outputs = outputs
+            outs = defaultdict(list)
+            for k, vars in six.iteritems(outputs):
+                if isinstance(vars, framework.Variable):
+                    outs[k].append(vars._ivar)
+                elif isinstance(vars, list) or isinstance(vars, tuple):
+                    for var in vars:
+                        outs[k].append(var._ivar)
+        else:
+            inps = defaultdict(list)
+            for k, vars in six.iteritems(inputs):
+                if isinstance(vars, framework.Variable):
+                    op.previous_ops.append(vars.op)
+                    inps[k].append(vars._ivar)
+                elif isinstance(vars, list) or isinstance(vars, tuple):
+                    for var in vars:
+                        op.previous_ops.append(var.op)
+                        inps[k].append(var._ivar)
+
+            op.outputs = outputs
+            outs = defaultdict(list)
+            for k, vars in six.iteritems(outputs):
+                if isinstance(vars, framework.Variable):
+                    vars.op = op
+                    outs[k].append(vars._ivar)
+                elif isinstance(vars, list) or isinstance(vars, tuple):
+                    for var in vars:
+                        var.op = op
+                        outs[k].append(var._ivar)
+
         # record op's trace id
         op.iop._trace_id = self._trace_id
 
-        backward_refs = self.trace(op.iop, op.inputs, op.outputs, op.attrs,
+        backward_refs = self.trace(op.iop, inps, outs, op.attrs,
                                    framework._current_expected_place(),
                                    stop_gradient)
 
-        if not stop_gradient:
+        if not stop_gradient and self._train_mode:
             self._trace_id += 1
             self._ops[op.iop._trace_id] = op
 
@@ -65,10 +110,16 @@ class Tracer(core.Tracer):
                 # TODO(minqiyang): remove all inputs and outputs after separate
                 # var and grad
                 op.backward_refs = defaultdict(list)
-                for k, v in six.iteritems(op.inputs):
+                for k, v in six.iteritems(inputs):
                     if k in backward_refs:
-                        op.backward_refs[k] = op.inputs[k]
+                        op.backward_refs[k] = inputs[k]
 
-                for k, v in six.iteritems(op.outputs):
+                for k, v in six.iteritems(outputs):
                     if k in backward_refs:
-                        op.backward_refs[k] = op.outputs[k]
+                        op.backward_refs[k] = outputs[k]
+
+    def _train_mode(self):
+        self._train_mode = True
+
+    def _eval_mode(self):
+        self._train_mode = False
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index c05e5fb9e3a46e721c20fd9288b89009e32afcbe..535909d710eb352bfc90ba42cc3b1894491e9ede 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -411,6 +411,7 @@ class Variable(object):
                     if persistable else False)
             if persistable:
                 _dygraph_tracer().trace_var(name, self)
+            self.op = None
         else:
             self.error_clip = error_clip
 
@@ -939,24 +940,7 @@ class Operator(object):
                 raise ValueError(
                     "`type` to initialized an Operator can not be None.")
             self.iop = core.OpBase(type)
-
-            # TODO(minqiyang): remove these lines after we take apart all
-            # backward grads and forward variables
-            self.inputs = defaultdict(list)
-            if inputs is not None:
-                for k, v in six.iteritems(inputs):
-                    if isinstance(v, Variable):
-                        self.inputs[k].append(v._ivar)
-                    elif isinstance(v, list) or isinstance(v, tuple):
-                        self.inputs[k].extend([var._ivar for var in v])
-
-            self.outputs = defaultdict(list)
-            if outputs is not None:
-                for k, v in six.iteritems(outputs):
-                    if isinstance(v, Variable):
-                        self.outputs[k].append(v._ivar)
-                    elif isinstance(v, list) or isinstance(v, tuple):
-                        self.outputs[k].extend([var._ivar for var in v])
+            self.previous_ops = []
 
             self.attrs = attrs if attrs else {}
         else:
@@ -1647,15 +1631,18 @@ class Block(object):
                 block=self,
                 desc=None,
                 type=kwargs.get("type", None),
-                inputs=kwargs.get("inputs", None),
-                outputs=kwargs.get("outputs", None),
-                attrs=kwargs.get("attrs", None))
+                inputs=None,
+                outputs=None,
+                attrs=kwargs.get("attrs", {}))
 
             # record ops in tracer rather than blocks
             #
             # TODO(minqiyang): add op stop_gradient support in static mode too.
             # currently, we only support stop_gradient in dygraph mode.
-            _dygraph_tracer().trace_op(op, kwargs.get("stop_gradient", False))
+            _dygraph_tracer().trace_op(op,
+                                       kwargs.get("inputs", {}),
+                                       kwargs.get("outputs", {}),
+                                       kwargs.get("stop_gradient", False))
         else:
             op_desc = self.desc.append_op()
             op = Operator(
@@ -1719,10 +1706,14 @@ class Block(object):
                 self,
                 None,
                 type=kwargs.get("type", None),
-                inputs=kwargs.get("inputs", None),
-                outputs=kwargs.get("outputs", None),
-                attrs=kwargs.get("attrs", None))
-            _dygraph_tracer().trace_op(op, kwargs.get("stop_gradient", False))
+                inputs=None,
+                outputs=None,
+                attrs=kwargs.get("attrs", {}))
+
+            _dygraph_tracer().trace_op(op,
+                                       kwargs.get("inputs", {}),
+                                       kwargs.get("outputs", {}),
+                                       kwargs.get("stop_gradient", False))
         else:
             op_desc = self.desc._prepend_op()
             op = Operator(
diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py
index f8f461853f34a09eb2317f6ac93ad385cca3609f..2df63d723e6ce91d3819c5e4301b9d5682158d79 100644
--- a/python/paddle/fluid/layers/control_flow.py
+++ b/python/paddle/fluid/layers/control_flow.py
@@ -29,7 +29,8 @@ from functools import reduce
 
 __all__ = [
     'While', 'Switch', 'increment', 'array_write', 'create_array', 'less_than',
-    'equal', 'array_read', 'array_length', 'IfElse', 'DynamicRNN', 'StaticRNN',
+    'less_equal', 'greater_than', 'greater_equal', 'equal', 'not_equal',
+    'array_read', 'array_length', 'IfElse', 'DynamicRNN', 'StaticRNN',
     'reorder_lod_tensor_by_rank', 'Print', 'is_empty'
 ]
 
@@ -189,6 +190,7 @@ def Print(input,
             'print_tensor_lod': print_tensor_lod,
             'print_phase': print_phase.upper()
         })
+    return input
 
 
 class BlockGuard(object):
@@ -971,6 +973,114 @@ def less_than(x, y, force_cpu=None, cond=None):
     return cond
 
 
+@templatedoc()
+def less_equal(x, y, cond=None):
+    """
+    This layer returns the truth value of :math:`x <= y` elementwise, which is equivalent to the overloaded operator `<=`.
+
+    Args:
+        x(Variable): First operand of *less_equal*
+        y(Variable): Second operand of *less_equal*
+        cond(Variable|None): Optional output variable to store the result of *less_equal*
+
+    Returns:
+        Variable: The tensor variable storing the output of *less_equal*.
+
+    Examples:
+        .. code-block:: python
+
+          out = fluid.layers.less_equal(x=label, y=limit)
+    """
+    helper = LayerHelper("less_equal", **locals())
+    if cond is None:
+        cond = helper.create_variable_for_type_inference(dtype='bool')
+        cond.stop_gradient = True
+
+    attrs = dict()
+    if force_init_on_cpu():
+        attrs['force_cpu'] = force_init_on_cpu()
+
+    helper.append_op(
+        type='less_equal',
+        inputs={'X': [x],
+                'Y': [y]},
+        outputs={'Out': [cond]},
+        attrs=attrs)
+    return cond
+
+
+@templatedoc()
+def greater_than(x, y, cond=None):
+    """
+    This layer returns the truth value of :math:`x > y` elementwise, which is equivalent to the overloaded operator `>`.
+
+    Args:
+        x(Variable): First operand of *greater_than*
+        y(Variable): Second operand of *greater_than*
+        cond(Variable|None): Optional output variable to store the result of *greater_than*
+
+    Returns:
+        Variable: The tensor variable storing the output of *greater_than*.
+
+    Examples:
+        .. code-block:: python
+
+          out = fluid.layers.greater_than(x=label, y=limit)
+    """
+    helper = LayerHelper("greater_than", **locals())
+    if cond is None:
+        cond = helper.create_variable_for_type_inference(dtype='bool')
+        cond.stop_gradient = True
+
+    attrs = dict()
+    if force_init_on_cpu():
+        attrs['force_cpu'] = force_init_on_cpu()
+
+    helper.append_op(
+        type='greater_than',
+        inputs={'X': [x],
+                'Y': [y]},
+        outputs={'Out': [cond]},
+        attrs=attrs)
+    return cond
+
+
+@templatedoc()
+def greater_equal(x, y, cond=None):
+    """
+    This layer returns the truth value of :math:`x >= y` elementwise, which is equivalent to the overloaded operator `>=`.
+
+    Args:
+        x(Variable): First operand of *greater_equal*
+        y(Variable): Second operand of *greater_equal*
+        cond(Variable|None): Optional output variable to store the result of *greater_equal*
+
+    Returns:
+        Variable: The tensor variable storing the output of *greater_equal*.
+
+    Examples:
+        .. code-block:: python
+
+          out = fluid.layers.greater_equal(x=label, y=limit)
+    """
+    helper = LayerHelper("greater_equal", **locals())
+    if cond is None:
+        cond = helper.create_variable_for_type_inference(dtype='bool')
+        cond.stop_gradient = True
+
+    attrs = dict()
+    if force_init_on_cpu():
+        attrs['force_cpu'] = force_init_on_cpu()
+
+    helper.append_op(
+        type='greater_equal',
+        inputs={'X': [x],
+                'Y': [y]},
+        outputs={'Out': [cond]},
+        attrs=attrs)
+    return cond
+
+
 def equal(x, y, cond=None):
     """
     This layer returns the truth value of :math:`x == y` elementwise.
@@ -999,6 +1109,34 @@ def equal(x, y, cond=None):
     return cond
 
 
+def not_equal(x, y, cond=None):
+    """
+    This layer returns the truth value of :math:`x != y` elementwise, which is equivalent to the overloader operator `!=`.
+
+    Args:
+        x(Variable): First operand of *not_equal*
+        y(Variable): Second operand of *not_equal*
+        cond(Variable|None): Optional output variable to store the result of *not_equal*
+
+    Returns:
+        Variable: The tensor variable storing the output of *not_equal*.
+
+    Examples:
+        .. code-block:: python
+
+          out = fluid.layers.not_equal(x=label, y=limit)
+    """
+    helper = LayerHelper("not_equal", **locals())
+    if cond is None:
+        cond = helper.create_variable_for_type_inference(dtype='bool')
+        cond.stop_gradient = True
+
+    helper.append_op(
+        type='not_equal', inputs={'X': [x],
+                                  'Y': [y]}, outputs={'Out': [cond]})
+    return cond
+
+
 def array_read(array, i):
     """
     This function performs the operation to read the data in as an
diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py
index 0a1ddbc1dba51692e75fa76856dd689b77ab9f35..09a573e8381e7402f6f617e6045110a93ba4f589 100644
--- a/python/paddle/fluid/layers/detection.py
+++ b/python/paddle/fluid/layers/detection.py
@@ -509,14 +509,14 @@ def polygon_box_transform(input, name=None):
 
 @templatedoc(op_type="yolov3_loss")
 def yolov3_loss(x,
-                gtbox,
-                gtlabel,
+                gt_box,
+                gt_label,
                 anchors,
                 anchor_mask,
                 class_num,
                 ignore_thresh,
                 downsample_ratio,
-                gtscore=None,
+                gt_score=None,
                 use_label_smooth=True,
                 name=None):
     """
@@ -524,12 +524,12 @@ def yolov3_loss(x,
 
     Args:
         x (Variable): ${x_comment}
-        gtbox (Variable): groud truth boxes, should be in shape of [N, B, 4],
+        gt_box (Variable): groud truth boxes, should be in shape of [N, B, 4],
                           in the third dimenstion, x, y, w, h should be stored 
                           and x, y, w, h should be relative value of input image.
                           N is the batch number and B is the max box number in 
                           an image.
-        gtlabel (Variable): class id of ground truth boxes, shoud be in shape
+        gt_label (Variable): class id of ground truth boxes, shoud be in shape
                             of [N, B].
         anchors (list|tuple): ${anchors_comment}
         anchor_mask (list|tuple): ${anchor_mask_comment}
@@ -537,7 +537,7 @@ def yolov3_loss(x,
         ignore_thresh (float): ${ignore_thresh_comment}
         downsample_ratio (int): ${downsample_ratio_comment}
         name (string): the name of yolov3 loss. Default None.
-        gtscore (Variable): mixup score of ground truth boxes, shoud be in shape
+        gt_score (Variable): mixup score of ground truth boxes, shoud be in shape
                             of [N, B]. Default None.
         use_label_smooth (bool): ${use_label_smooth_comment}
 
@@ -558,13 +558,13 @@ def yolov3_loss(x,
       .. code-block:: python
 
           x = fluid.layers.data(name='x', shape=[255, 13, 13], dtype='float32')
-          gtbox = fluid.layers.data(name='gtbox', shape=[6, 4], dtype='float32')
-          gtlabel = fluid.layers.data(name='gtlabel', shape=[6], dtype='int32')
-          gtscore = fluid.layers.data(name='gtscore', shape=[6], dtype='float32')
+          gt_box = fluid.layers.data(name='gt_box', shape=[6, 4], dtype='float32')
+          gt_label = fluid.layers.data(name='gt_label', shape=[6], dtype='int32')
+          gt_score = fluid.layers.data(name='gt_score', shape=[6], dtype='float32')
           anchors = [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326]
           anchor_mask = [0, 1, 2]
-          loss = fluid.layers.yolov3_loss(x=x, gtbox=gtbox, gtlabel=gtlabel,
-                                          gtscore=gtscore, anchors=anchors, 
+          loss = fluid.layers.yolov3_loss(x=x, gt_box=gt_box, gt_label=gt_label,
+                                          gt_score=gt_score, anchors=anchors, 
                                           anchor_mask=anchor_mask, class_num=80,
                                           ignore_thresh=0.7, downsample_ratio=32)
     """
@@ -572,11 +572,11 @@ def yolov3_loss(x,
 
     if not isinstance(x, Variable):
         raise TypeError("Input x of yolov3_loss must be Variable")
-    if not isinstance(gtbox, Variable):
+    if not isinstance(gt_box, Variable):
         raise TypeError("Input gtbox of yolov3_loss must be Variable")
-    if not isinstance(gtlabel, Variable):
+    if not isinstance(gt_label, Variable):
         raise TypeError("Input gtlabel of yolov3_loss must be Variable")
-    if gtscore is not None and not isinstance(gtscore, Variable):
+    if gt_score is not None and not isinstance(gt_score, Variable):
         raise TypeError("Input gtscore of yolov3_loss must be Variable")
     if not isinstance(anchors, list) and not isinstance(anchors, tuple):
         raise TypeError("Attr anchors of yolov3_loss must be list or tuple")
@@ -602,11 +602,11 @@ def yolov3_loss(x,
 
     inputs = {
         "X": x,
-        "GTBox": gtbox,
-        "GTLabel": gtlabel,
+        "GTBox": gt_box,
+        "GTLabel": gt_label,
     }
-    if gtscore:
-        inputs["GTScore"] = gtscore
+    if gt_score:
+        inputs["GTScore"] = gt_score
 
     attrs = {
         "anchors": anchors,
@@ -1542,7 +1542,7 @@ def multi_box_head(inputs,
         .. code-block:: python
 
           mbox_locs, mbox_confs, box, var = fluid.layers.multi_box_head(
-            inputs=[conv1, conv2, conv3, conv4, conv5, conv5],
+            inputs=[conv1, conv2, conv3, conv4, conv5, conv6],
             image=images,
             num_classes=21,
             min_ratio=20,
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
index e0e32bb6737c1bdde8bee2708afdebb186dc18f6..2bac9dd9a46b1b291d7ee39876f7b60d2d5e298b 100644
--- a/python/paddle/fluid/layers/nn.py
+++ b/python/paddle/fluid/layers/nn.py
@@ -5721,12 +5721,21 @@ def hsigmoid(input,
         raise ValueError(
             "num_classes must not be less than 2 with default tree")
 
+    if (not is_custom) and (is_sparse):
+        print("Sparse mode should not be used without custom tree")
+        is_sparse = False
+
+    if (not is_custom) and ((path_table is not None) or
+                            (path_code is not None)):
+        raise ValueError(
+            "only num_classes should be passed without custom tree")
+
     if (is_custom) and (path_code is None):
-        raise ValueError("path_code should not be None with costum tree")
+        raise ValueError("path_code should not be None with custom tree")
     elif (is_custom) and (path_table is None):
-        raise ValueError("path_table should not be None with costum tree")
+        raise ValueError("path_table should not be None with custom tree")
     elif (is_custom) and (num_classes is None):
-        raise ValueError("num_classes should not be None with costum tree")
+        raise ValueError("num_classes should not be None with custom tree")
     else:
         pass
 
@@ -6269,6 +6278,8 @@ def sampled_softmax_with_cross_entropy(logits,
     sampled_label = helper.create_variable_for_type_inference(dtype='int64')
     sampled_softlabel = helper.create_variable_for_type_inference(
         dtype=logits.dtype)
+    logits_dim = helper.create_variable_for_type_inference(dtype=logits.dtype)
+    labels_dim = helper.create_variable_for_type_inference(dtype=label.type)
 
     helper.append_op(
         type='sample_logits',
@@ -6282,7 +6293,9 @@ def sampled_softmax_with_cross_entropy(logits,
             'Samples': samples,
             'Probabilities': probabilities,
             'SampledLabels': sampled_label,
-            'SampledLogits': sampled_logits
+            'SampledLogits': sampled_logits,
+            'LogitsDim': logits_dim,
+            'LabelsDim': labels_dim
         },
         attrs={
             'use_customized_samples': use_customized_samples,
diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py
index f018bb8af8cc9f7ed965c86d5aff40352014c393..f06c0abaf901950d072cf857696bc7479fb7b52f 100644
--- a/python/paddle/fluid/layers/ops.py
+++ b/python/paddle/fluid/layers/ops.py
@@ -27,6 +27,7 @@ __activations_noattr__ = [
     'tanh_shrink',
     'softshrink',
     'sqrt',
+    'rsqrt',
     'abs',
     'ceil',
     'floor',
@@ -81,8 +82,8 @@ def uniform_random(shape, dtype='float32', min=-1.0, max=1.0, seed=0):
 
     Examples:
         .. code-block:: python
-
-        result = fluid.layers.uniform_random(shape=[32, 784])
+     
+            result = fluid.layers.uniform_random(shape=[32, 784])
     """
 
     locals_var = locals().keys()
diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py
index 03ebd41fa00c69bfce66d325e32fc9aeb25a2486..d1681580bebc454d26be518180b649bfb3c76e4e 100644
--- a/python/paddle/fluid/layers/tensor.py
+++ b/python/paddle/fluid/layers/tensor.py
@@ -28,7 +28,7 @@ __all__ = [
     'tensor_array_to_tensor', 'concat', 'sums', 'assign',
     'fill_constant_batch_size_like', 'fill_constant', 'argmin', 'argmax',
     'argsort', 'ones', 'zeros', 'reverse', 'has_inf', 'has_nan', 'isfinite',
-    'range', 'linspace'
+    'range', 'linspace', 'zeros_like'
 ]
 
 
@@ -853,3 +853,34 @@ def linspace(start, stop, num, dtype):
                 'Num': num},
         outputs={'Out': [out]})
     return out
+
+
+def zeros_like(x, out=None):
+    """
+    **zeros_like**
+
+    This function creates a zeros tensor which has identical shape and dtype 
+    with `x`.
+
+    Args:
+        x(Variable): The input tensor which specifies shape and dtype.
+        out(Variable): The output tensor.
+
+    Returns:
+        Variable: The tensor variable storing the output.
+
+    Examples:
+        .. code-block:: python
+
+          x = fluid.layers.data(name='x', dtype='float32', shape=[3], append_batch_size=False)
+          data = fluid.layers.zeros_like(x) # [0.0, 0.0, 0.0]
+
+    """
+
+    helper = LayerHelper("zeros_like", **locals())
+    if out is None:
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
+    helper.append_op(
+        type='fill_zeros_like', inputs={'X': [x]}, outputs={'Out': [out]})
+    out.stop_gradient = True
+    return out
diff --git a/python/paddle/fluid/tests/book/high-level-api/cifar10_small_test_set.py b/python/paddle/fluid/tests/book/high-level-api/cifar10_small_test_set.py
index 48c0f3d3611547308b5d4460748d3aab765f5805..6f24ec45aa6f27814e489b8dce49fe69f62d4f10 100644
--- a/python/paddle/fluid/tests/book/high-level-api/cifar10_small_test_set.py
+++ b/python/paddle/fluid/tests/book/high-level-api/cifar10_small_test_set.py
@@ -88,3 +88,19 @@ def train10(batch_size=None):
         paddle.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5),
         'data_batch',
         batch_size=batch_size)
+
+
+def test10(batch_size=None):
+    """
+    CIFAR-10 test set creator.
+
+    It returns a reader creator, each sample in the reader is image pixels in
+    [0, 1] and label in [0, 9].
+
+    :return: Test reader creator.
+    :rtype: callable
+    """
+    return reader_creator(
+        paddle.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5),
+        'test_batch',
+        batch_size=batch_size)
diff --git a/python/paddle/fluid/tests/book/high-level-api/test_image_classification_vgg_new_api.py b/python/paddle/fluid/tests/book/high-level-api/test_image_classification_vgg_new_api.py
index 82294d4b26fe64e6cddc81f9ba3480caf5b51620..0a27aa0fcfece36f1a8ae5ad0477d75a15fd88da 100644
--- a/python/paddle/fluid/tests/book/high-level-api/test_image_classification_vgg_new_api.py
+++ b/python/paddle/fluid/tests/book/high-level-api/test_image_classification_vgg_new_api.py
@@ -89,9 +89,11 @@ def train(use_cuda, train_program, parallel, params_dirname):
             cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10),
         batch_size=BATCH_SIZE,
         drop_last=False)
-
+    # Use only part of the test set data validation program
     test_reader = paddle.batch(
-        paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE, drop_last=False)
+        cifar10_small_test_set.test10(BATCH_SIZE),
+        batch_size=BATCH_SIZE,
+        drop_last=False)
 
     def event_handler(event):
         if isinstance(event, EndStepEvent):
diff --git a/python/paddle/fluid/tests/test_detection.py b/python/paddle/fluid/tests/test_detection.py
index 7d1b869cf5991dc5ef960ff4d72289979aae158a..e1c4c2eca08d4652ecda8e2579d342818c803f4a 100644
--- a/python/paddle/fluid/tests/test_detection.py
+++ b/python/paddle/fluid/tests/test_detection.py
@@ -474,17 +474,17 @@ class TestYoloDetection(unittest.TestCase):
         program = Program()
         with program_guard(program):
             x = layers.data(name='x', shape=[30, 7, 7], dtype='float32')
-            gtbox = layers.data(name='gtbox', shape=[10, 4], dtype='float32')
-            gtlabel = layers.data(name='gtlabel', shape=[10], dtype='int32')
-            gtscore = layers.data(name='gtscore', shape=[10], dtype='float32')
+            gt_box = layers.data(name='gt_box', shape=[10, 4], dtype='float32')
+            gt_label = layers.data(name='gt_label', shape=[10], dtype='int32')
+            gt_score = layers.data(name='gt_score', shape=[10], dtype='float32')
             loss = layers.yolov3_loss(
                 x,
-                gtbox,
-                gtlabel, [10, 13, 30, 13], [0, 1],
+                gt_box,
+                gt_label, [10, 13, 30, 13], [0, 1],
                 10,
                 0.7,
                 32,
-                gtscore=gtscore,
+                gt_score=gt_score,
                 use_label_smooth=False)
 
             self.assertIsNotNone(loss)
diff --git a/python/paddle/fluid/tests/unittests/test_activation_op.py b/python/paddle/fluid/tests/unittests/test_activation_op.py
index d587715d607c6da16da5c009db16322e8cd7d176..4d66b7a989732e37c48c73b9617943874ad07bba 100644
--- a/python/paddle/fluid/tests/unittests/test_activation_op.py
+++ b/python/paddle/fluid/tests/unittests/test_activation_op.py
@@ -192,6 +192,23 @@ class TestSqrt(TestActivation):
         self.check_grad(['X'], 'Out', max_relative_error=0.007)
 
 
+class TestRsqrt(TestActivation):
+    def setUp(self):
+        self.op_type = "rsqrt"
+        self.init_dtype()
+
+        x = np.random.uniform(0.1, 1, [2, 3]).astype(self.dtype)
+        out = 1.0 / np.sqrt(x)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.outputs = {'Out': out}
+
+    def test_check_grad(self):
+        if self.dtype == np.float16:
+            return
+        self.check_grad(['X'], 'Out', max_relative_error=0.0005)
+
+
 class TestAbs(TestActivation):
     def setUp(self):
         self.op_type = "abs"
diff --git a/python/paddle/fluid/tests/unittests/test_dygraph_multi_forward.py b/python/paddle/fluid/tests/unittests/test_dygraph_multi_forward.py
new file mode 100644
index 0000000000000000000000000000000000000000..8b8fdcc887beb4879b2ce1101184dabe6f819acf
--- /dev/null
+++ b/python/paddle/fluid/tests/unittests/test_dygraph_multi_forward.py
@@ -0,0 +1,201 @@
+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+
+from __future__ import print_function
+
+import contextlib
+import unittest
+import numpy as np
+import six
+
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid import core
+from paddle.fluid.optimizer import SGDOptimizer
+from paddle.fluid.dygraph.nn import Conv2D, Pool2D, FC
+from paddle.fluid.dygraph.base import to_variable
+from test_imperative_base import new_program_scope
+
+
+class SimpleImgConvPool(fluid.dygraph.Layer):
+    def __init__(self,
+                 name_scope,
+                 num_channels,
+                 num_filters,
+                 filter_size,
+                 pool_size,
+                 pool_stride,
+                 pool_padding=0,
+                 pool_type='max',
+                 global_pooling=False,
+                 conv_stride=1,
+                 conv_padding=0,
+                 conv_dilation=1,
+                 conv_groups=1,
+                 act=None,
+                 use_cudnn=False,
+                 param_attr=None,
+                 bias_attr=None):
+        super(SimpleImgConvPool, self).__init__(name_scope)
+
+        self._conv2d = Conv2D(
+            self.full_name(),
+            num_channels=num_channels,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            stride=conv_stride,
+            padding=conv_padding,
+            dilation=conv_dilation,
+            groups=conv_groups,
+            param_attr=None,
+            bias_attr=None,
+            use_cudnn=use_cudnn)
+
+        self._pool2d = Pool2D(
+            self.full_name(),
+            pool_size=pool_size,
+            pool_type=pool_type,
+            pool_stride=pool_stride,
+            pool_padding=pool_padding,
+            global_pooling=global_pooling,
+            use_cudnn=use_cudnn)
+
+    def forward(self, inputs):
+        x = self._conv2d(inputs)
+        x = self._pool2d(x)
+        return x
+
+
+class MNIST(fluid.dygraph.Layer):
+    def __init__(self, name_scope):
+        super(MNIST, self).__init__(name_scope)
+
+        self._simple_img_conv_pool_1 = SimpleImgConvPool(
+            self.full_name(), 1, 20, 5, 2, 2, act="relu")
+
+        self._simple_img_conv_pool_2 = SimpleImgConvPool(
+            self.full_name(), 20, 50, 5, 2, 2, act="relu")
+
+        pool_2_shape = 50 * 4 * 4
+        SIZE = 10
+        scale = (2.0 / (pool_2_shape**2 * SIZE))**0.5
+        self._fc = FC(self.full_name(),
+                      10,
+                      param_attr=fluid.param_attr.ParamAttr(
+                          initializer=fluid.initializer.NormalInitializer(
+                              loc=0.0, scale=scale)),
+                      act="softmax")
+
+    def forward(self, inputs):
+        x = self._simple_img_conv_pool_1(inputs)
+        x = self._simple_img_conv_pool_2(x)
+        x = self._fc(x)
+        return x
+
+
+class TestDygraphMultiForward(unittest.TestCase):
+    def test_mnist_forward_float32(self):
+        seed = 90
+        epoch_num = 1
+        with fluid.dygraph.guard():
+            fluid.default_startup_program().random_seed = seed
+            fluid.default_main_program().random_seed = seed
+
+            mnist = MNIST("mnist")
+            sgd = SGDOptimizer(learning_rate=1e-3)
+            train_reader = paddle.batch(
+                paddle.dataset.mnist.train(), batch_size=128, drop_last=True)
+
+            dy_param_init_value = {}
+            mnist.eval()
+            for epoch in range(epoch_num):
+                for batch_id, data in enumerate(train_reader()):
+                    dy_x_data = np.array(
+                        [x[0].reshape(1, 28, 28)
+                         for x in data]).astype('float32')
+                    y_data = np.array(
+                        [x[1] for x in data]).astype('int64').reshape(128, 1)
+
+                    img = to_variable(dy_x_data)
+                    label = to_variable(y_data)
+                    label.stop_gradient = True
+
+                    cost = mnist(img)
+                    loss = fluid.layers.cross_entropy(cost, label)
+                    avg_loss = fluid.layers.mean(loss)
+
+                    dy_out = avg_loss.numpy()
+
+                    if epoch == 0 and batch_id == 0:
+                        for param in mnist.parameters():
+                            dy_param_init_value[param.name] = param.numpy()
+
+        with new_program_scope():
+            fluid.default_startup_program().random_seed = seed
+            fluid.default_main_program().random_seed = seed
+
+            exe = fluid.Executor(fluid.CPUPlace(
+            ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
+
+            mnist = MNIST("mnist")
+            sgd = SGDOptimizer(learning_rate=1e-3)
+            train_reader = paddle.batch(
+                paddle.dataset.mnist.train(), batch_size=128, drop_last=True)
+
+            img = fluid.layers.data(
+                name='pixel', shape=[1, 28, 28], dtype='float32')
+            label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+            cost = mnist(img)
+            loss = fluid.layers.cross_entropy(cost, label)
+            avg_loss = fluid.layers.mean(loss)
+
+            # initialize params and fetch them
+            static_param_init_value = {}
+            static_param_name_list = []
+            for param in mnist.parameters():
+                static_param_name_list.append(param.name)
+
+            out = exe.run(fluid.default_startup_program(),
+                          fetch_list=static_param_name_list)
+
+            for i in range(len(static_param_name_list)):
+                static_param_init_value[static_param_name_list[i]] = out[i]
+
+            for epoch in range(epoch_num):
+                for batch_id, data in enumerate(train_reader()):
+                    static_x_data = np.array(
+                        [x[0].reshape(1, 28, 28)
+                         for x in data]).astype('float32')
+                    y_data = np.array(
+                        [x[1] for x in data]).astype('int64').reshape([128, 1])
+
+                    fetch_list = [avg_loss.name]
+                    out = exe.run(
+                        fluid.default_main_program(),
+                        feed={"pixel": static_x_data,
+                              "label": y_data},
+                        fetch_list=fetch_list)
+
+                    static_out = out[0]
+
+        self.assertTrue(np.allclose(dy_x_data.all(), static_x_data.all()))
+
+        for key, value in six.iteritems(static_param_init_value):
+            self.assertTrue(np.allclose(value, dy_param_init_value[key]))
+
+        self.assertTrue(np.allclose(static_out, dy_out))
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py
index 8960cbcdd2f574a647229894c44c2b6ea188b7d4..b1851f4c78ddf984b06cf67f628099d5b60c771e 100644
--- a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py
+++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py
@@ -65,7 +65,9 @@ class ModelHyperParams(object):
     # number of head used in multi-head attention.
     n_head = 8
     # number of sub-layers to be stacked in the encoder and decoder.
-    n_layer = 6
+    # NOTE(zcd): the origin number of layer is 6, to make this unit test faster,
+    # we should reduce the layer number to 4.
+    n_layer = 4
     # dropout rate used by all dropout layers.
     dropout = 0.1
 
diff --git a/python/paddle/fluid/tests/unittests/test_sigmoid_cross_entropy_with_logits_op.py b/python/paddle/fluid/tests/unittests/test_sigmoid_cross_entropy_with_logits_op.py
index ae1883f1f7e44e06e378ff6d16dbc3c5060027e4..ec10b634091fc521062457b780b0c4cafcbacec0 100644
--- a/python/paddle/fluid/tests/unittests/test_sigmoid_cross_entropy_with_logits_op.py
+++ b/python/paddle/fluid/tests/unittests/test_sigmoid_cross_entropy_with_logits_op.py
@@ -149,5 +149,98 @@ class TestSigmoidCrossEntropyWithNorm(OpTest):
         self.check_grad(['X'], 'Out')
 
 
+class TestSigmoidCrossEntropyWithLogitsOp5(OpTest):
+    """Test sigmoid_cross_entropy_with_logit_op with probabalistic label
+    """
+
+    def setUp(self):
+        self.op_type = "sigmoid_cross_entropy_with_logits"
+        batch_size = [10, 10]
+        num_classes = 20
+        self.inputs = {
+            'X': logit(
+                np.random.uniform(0, 1, tuple(batch_size + [num_classes]))
+                .astype("float32")),
+            'Label': np.random.uniform(0, 1, tuple(batch_size + [num_classes]))
+            .astype("float32")
+        }
+
+        # Fw Pass is implemented as elementwise sigmoid followed by
+        # elementwise logistic loss
+        # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X))
+        sigmoid_X = expit(self.inputs['X'])
+        term1 = self.inputs['Label'] * np.log(sigmoid_X)
+        term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
+        self.outputs = {'Out': -term1 - term2}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
+class TestSigmoidCrossEntropyWithNorm2(OpTest):
+    def setUp(self):
+        self.op_type = "sigmoid_cross_entropy_with_logits"
+        batch_size = [10, 10]
+        num_classes = 20
+        ignore_index = -1
+        self.inputs = {
+            'X': logit(
+                np.random.uniform(0, 1, tuple(batch_size + [num_classes]))
+                .astype("float32")),
+            'Label': np.random.randint(-1, 2, tuple(batch_size + [num_classes]))
+            .astype("float32")
+        }
+        self.attrs = {'ignore_index': ignore_index, 'normalize': True}
+        sigmoid_X = expit(self.inputs['X'])
+        term1 = self.inputs['Label'] * np.log(sigmoid_X)
+        term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
+        out = -term1 - term2
+        out[np.where(self.inputs['Label'] == ignore_index)] = 0
+        if self.attrs['normalize']:
+            out = out / float(
+                np.where(self.inputs['Label'] != ignore_index)[0].size)
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
+class TestSigmoidCrossEntropyWithLogitsOp6(OpTest):
+    """Test sigmoid_cross_entropy_with_logit_op with binary label
+    """
+
+    def setUp(self):
+        self.op_type = "sigmoid_cross_entropy_with_logits"
+        batch_size = [10, 10]
+        num_classes = 20
+        self.inputs = {
+            'X': logit(
+                np.random.uniform(0, 1, tuple(batch_size + [num_classes]))
+                .astype("float32")),
+            'Label': np.random.randint(0, 2, tuple(batch_size + [num_classes]))
+            .astype("float32")
+        }
+
+        # Fw Pass is implemented as elementwise sigmoid followed by
+        # elementwise logistic loss
+        # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X))
+        sigmoid_X = expit(self.inputs['X'])
+        term1 = self.inputs['Label'] * np.log(sigmoid_X)
+        term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
+        self.outputs = {'Out': -term1 - term2}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
 if __name__ == '__main__':
     unittest.main()