Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into imperative_fix_growing_dict

test=develop

paddle/fluid/API.spec

@@ -13,6 +13,7 @@ paddle.fluid.name_scope (ArgSpec(args=['prefix'], varargs=None, keywords=None, d
 paddle.fluid.cuda_places (ArgSpec(args=['device_ids'], varargs=None, keywords=None, defaults=(None,)), ('document', '7d9a51fc9cf3c5245b5227080a8064c3'))
 paddle.fluid.cpu_places (ArgSpec(args=['device_count'], varargs=None, keywords=None, defaults=(None,)), ('document', '4c0cd83f0b401fc2ff84c70974e5d210'))
 paddle.fluid.cuda_pinned_places (ArgSpec(args=['device_count'], varargs=None, keywords=None, defaults=(None,)), ('document', 'd0c3ebd813c39958c92b78e3eef7e912'))
+paddle.fluid.in_dygraph_mode (ArgSpec(args=[], varargs=None, keywords=None, defaults=None), ('document', 'f06314a1cb30c96b5808dde2219c2dae'))
 paddle.fluid.Executor.__init__ (ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.Executor.close (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', 'f5369953dd0c443961cf79f7a00e1a03'))
 paddle.fluid.Executor.infer_from_dataset (ArgSpec(args=['self', 'program', 'dataset', 'scope', 'thread', 'debug', 'fetch_list', 'fetch_info', 'print_period'], varargs=None, keywords=None, defaults=(None, None, None, 0, False, None, None, 100)), ('document', '9c7decb955b9c4f718114179c8985581'))
@@ -155,10 +156,10 @@ paddle.fluid.layers.label_smooth (ArgSpec(args=['label', 'prior_dist', 'epsilon'
 paddle.fluid.layers.roi_pool (ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0)), ('document', 'c317aa595deb31649083c8faa91cdb97'))
 paddle.fluid.layers.roi_align (ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None)), ('document', '12c5bbb8b38c42e623fbc47611d766e1'))
 paddle.fluid.layers.dice_loss (ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,)), ('document', '1ba0508d573f65feecf3564dce22aa1d'))
-paddle.fluid.layers.image_resize (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None, True, 1)), ('document', '7a1966d7c3a48f1fc0881cdaf5d83b0b'))
+paddle.fluid.layers.image_resize (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None, True, 1)), ('document', 'd1b08c11bb9277386fcf6ae70b6622d1'))
 paddle.fluid.layers.image_resize_short (ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',)), ('document', '06211aefc50c5a3e940d7204d859cdf7'))
-paddle.fluid.layers.resize_bilinear (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, None, True, 1)), ('document', 'e4fb4ed511b2293b8f04f7e872afbfd7'))
-paddle.fluid.layers.resize_nearest (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners'], varargs=None, keywords=None, defaults=(None, None, None, None, True)), ('document', '735fa9758a6d7ff3b47d7b827f961c1d'))
+paddle.fluid.layers.resize_bilinear (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, None, True, 1)), ('document', 'c45591fbc4f64a178fbca219e1546a58'))
+paddle.fluid.layers.resize_nearest (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners'], varargs=None, keywords=None, defaults=(None, None, None, None, True)), ('document', 'ae6d73cdc7f3a138d8a338ecdb33c1ae'))
 paddle.fluid.layers.gather (ArgSpec(args=['input', 'index'], varargs=None, keywords=None, defaults=None), ('document', '98f1c86716b9b7f4dda83f20e2adeee2'))
 paddle.fluid.layers.scatter (ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '65f8e9d8ddfd0b412f940579c4faa342'))
 paddle.fluid.layers.sequence_scatter (ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '15b522457dfef103f0c20ca9d397678b'))

paddle/fluid/framework/details/inplace_op_pass.cc

@@ -305,6 +305,12 @@ void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
 
     VLOG(4) << "Try to inplace " << in_var_name << " with " << out_var_name;
 
+    if (var_nodes_[in_var_name].back() != in_node) {
+      VLOG(4) << "SKIP since " << in_var_name
+              << " is also used as output by other ops";
+      continue;
+    }
+
     bool can_replace = true;
     if (in_var_name == out_var_name) {
       can_replace = false;
@@ -527,6 +533,9 @@ void GraphView::Build(ir::Graph* g) {
   };
   for (auto& node : g->Nodes()) {
     if (!node->IsOp()) continue;
+    // avoid optimize the variable used in sub-blocks
+    if (OpHasSubBlock(node->Op())) update_skip_set(node);
+
     if (node->Name() == "send") update_skip_set(node);
     if (node->Name() == "recv") update_skip_set(node);
     if (node->Name() == "prefetch") update_skip_set(node);

paddle/fluid/framework/details/op_registry.h

@@ -233,6 +233,12 @@ struct OpInfoFiller<T, kNoNeedBufferVarsInference> {
   }
 };
 
+// A fake OpInfoFiller of void
+template <>
+struct OpInfoFiller<void, kUnknown> {
+  void operator()(const char* op_type, OpInfo* info) const {}
+};
+
 }  // namespace details
 
 }  // namespace framework

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -123,8 +123,8 @@ CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
       // will enhance this pass later.
       "runtime_context_cache_pass",     //
       "attention_lstm_fuse_pass",       //
-      "seqpool_concat_fuse_pass",       //
       "seqconv_eltadd_relu_fuse_pass",  //
+      // "seqpool_concat_fuse_pass",    //
       // "embedding_fc_lstm_fuse_pass", //
       "fc_lstm_fuse_pass",             //
       "mul_lstm_fuse_pass",            //

paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc

@@ -150,6 +150,9 @@ void SetConfig(AnalysisConfig *cfg, bool use_mkldnn = false) {
   if (use_mkldnn) {
     cfg->EnableMKLDNN();
   }
+  // Enable seqpool_concat_fuse_pass, disabled by default since it takes much
+  // time
+  cfg->pass_builder()->InsertPass(2, "seqpool_concat_fuse_pass");
 }
 
 void profile(bool use_mkldnn = false) {

paddle/fluid/op_use_default_grad_op_maker.spec

-abs
-acos
-asin
-atan
 attention_lstm
-brelu
 conv_shift
-cos
 cos_sim
 dequantize
-elu
 fc
 flatten
 fsp
@@ -21,13 +14,8 @@ fusion_seqconv_eltadd_relu
 fusion_seqexpand_concat_fc
 fusion_seqpool_concat
 fusion_squared_mat_sub
-gelu
 gru
-hard_shrink
 hierarchical_sigmoid
-leaky_relu
-log
-logsigmoid
 lrn
 lstm_unit
 lstmp
@@ -38,7 +26,6 @@ modified_huber_loss
 nce
 pool2d
 pool3d
-pow
 prelu
 quantize
 rank_loss
@@ -50,20 +37,10 @@ reduce_sum
 requantize
 reshape
 rnn_memory_helper
-round
 sequence_softmax
-sin
-softplus
-softshrink
-softsign
 spp
-square
 squeeze
-stanh
-swish
-tanh_shrink
 tensor_array_to_tensor
-thresholded_relu
 transpose
 unpool
 unsqueeze

paddle/fluid/operators/activation_cudnn_op.cu.cc

@@ -12,6 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <memory>
+#include <string>
+#include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/activation_op.h"
 #include "paddle/fluid/platform/cudnn_desc.h"
@@ -82,6 +85,8 @@ template <typename T>
 struct CudnnReluGradFunctor : public CudnnActivationGradFunctor<T> {
   explicit CudnnReluGradFunctor(const CUDADeviceContext& ctx)
       : CudnnActivationGradFunctor<T>(ctx, 0.0, CUDNN_ACTIVATION_RELU) {}
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 template <typename T>
@@ -94,6 +99,8 @@ struct CudnnRelu6GradFunctor : public CudnnActivationGradFunctor<T> {
   explicit CudnnRelu6GradFunctor(const CUDADeviceContext& ctx)
       : CudnnActivationGradFunctor<T>(ctx, 6.0, CUDNN_ACTIVATION_CLIPPED_RELU) {
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 template <typename T>
@@ -105,6 +112,8 @@ template <typename T>
 struct CudnnSigmoidGradFunctor : public CudnnActivationGradFunctor<T> {
   explicit CudnnSigmoidGradFunctor(const CUDADeviceContext& ctx)
       : CudnnActivationGradFunctor<T>(ctx, 0.0, CUDNN_ACTIVATION_SIGMOID) {}
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 template <typename T>
@@ -116,6 +125,8 @@ template <typename T>
 struct CudnnTanhGradFunctor : public CudnnActivationGradFunctor<T> {
   explicit CudnnTanhGradFunctor(const CUDADeviceContext& ctx)
       : CudnnActivationGradFunctor<T>(ctx, 0.0, CUDNN_ACTIVATION_TANH) {}
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 template <typename Functor>
@@ -140,10 +151,13 @@ class CudnnActivationGradKernel
  public:
   using T = typename Functor::ELEMENT_TYPE;
   void Compute(const framework::ExecutionContext& context) const override {
+    static_assert(Functor::FwdDeps() == kDepOut, "Forward deps must be Out.");
+
     const framework::Tensor *X, *Out, *dOut;
     X = Out = dOut = nullptr;
     framework::Tensor* dX = nullptr;
-    ExtractActivationGradTensor(context, &X, &Out, &dOut, &dX);
+    ExtractActivationGradTensor<Functor::FwdDeps()>(context, &X, &Out, &dOut,
+                                                    &dX);
     dX->mutable_data<T>(context.GetPlace());
     auto& dev_ctx = context.template device_context<CUDADeviceContext>();
     Functor functor(dev_ctx);

paddle/fluid/operators/activation_op.cc

@@ -15,7 +15,9 @@ limitations under the License. */
 #include "paddle/fluid/operators/activation_op.h"
 #include <memory>
 #include <string>
+#include <type_traits>
 #include <unordered_map>
+#include <vector>
 #include "paddle/fluid/operators/mkldnn/mkldnn_activation_op.h"
 #include "paddle/fluid/platform/port.h"
 #ifdef PADDLE_WITH_CUDA
@@ -27,6 +29,25 @@ namespace operators {
 
 using paddle::framework::Tensor;
 
+template <typename GradFunctor>
+static constexpr bool CanInplaceAct() {
+  return GradFunctor::FwdDeps() == kDepOut || GradFunctor::FwdDeps() == kNoDeps;
+}
+
+std::unique_ptr<std::unordered_set<std::string>> GetInplaceOpSet() {
+  std::unique_ptr<std::unordered_set<std::string>> ret(
+      new std::unordered_set<std::string>());
+#define INSERT_INTO_INPLACE_OP_SET(op_type, __omitted, fwd_functor, \
+                                   bwd_functor)                     \
+  if (CanInplaceAct<bwd_functor<float>>()) {                        \
+    ret->insert(#op_type);                                          \
+  }
+
+  FOR_EACH_ACTIVATION_OP(INSERT_INTO_INPLACE_OP_SET);
+#undef INSERT_INTO_INPLACE_OP_SET
+  return ret;
+}
+
 #define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT)                    \
   class OP_NAME##OpMaker                                                     \
       : public ::paddle::framework::OpProtoAndCheckerMaker {                 \
@@ -50,26 +71,32 @@ using paddle::framework::Tensor;
     }                                                                        \
   }
 
-#define REGISTER_ACTIVATION_OP_GRAD_MAKER(OP_NAME, KERNEL_TYPE)              \
-  class OP_NAME##GradMaker                                                   \
-      : public ::paddle::framework::SingleGradOpDescMaker {                  \
-   public:                                                                   \
-    using ::paddle::framework::SingleGradOpDescMaker::SingleGradOpDescMaker; \
-                                                                             \
-   protected:                                                                \
-    std::unique_ptr<::paddle::framework::OpDesc> Apply() const override {    \
-      auto* op = new ::paddle::framework::OpDesc();                          \
-      op->SetType(#KERNEL_TYPE "_grad");                                     \
-      op->SetInput("Out", Output("Out"));                                    \
-      op->SetInput(::paddle::framework::GradVarName("Out"),                  \
-                   OutputGrad("Out"));                                       \
-                                                                             \
-      op->SetAttrMap(Attrs());                                               \
-                                                                             \
-      op->SetOutput(::paddle::framework::GradVarName("X"), InputGrad("X"));  \
-      return std::unique_ptr<::paddle::framework::OpDesc>(op);               \
-    }                                                                        \
+template <ActBwdOpFwdDeps kDepValue>
+class ActivationGradOpDescMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    std::unique_ptr<framework::OpDesc> op(new framework::OpDesc());
+    op->SetType(ForwardOpType() + "_grad");
+    op->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
+    op->SetAttrMap(Attrs());
+
+    if (static_cast<int>(kDepValue) &
+        static_cast<int>(ActBwdOpFwdDeps::kDepX)) {
+      op->SetInput("X", Input("X"));
+    }
+
+    if (static_cast<int>(kDepValue) &
+        static_cast<int>(ActBwdOpFwdDeps::kDepOut)) {
+      op->SetInput("Out", Output("Out"));
+    }
+
+    return op;
   }
+};
 
 framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx,
                                       const framework::OperatorWithKernel& oper,
@@ -129,14 +156,15 @@ class ActivationOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
   void InferShape(framework::InferShapeContext* ctx) const override {
-    ctx->ShareDim("Out", framework::GradVarName("X"));
-    ctx->ShareLoD("Out", framework::GradVarName("X"));
+    auto out_grad_name = framework::GradVarName("Out");
+    ctx->ShareDim(out_grad_name, framework::GradVarName("X"));
+    ctx->ShareLoD(out_grad_name, framework::GradVarName("X"));
   }
 
  protected:
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
-    return GetKernelType(ctx, *this, "Out");
+    return GetKernelType(ctx, *this, framework::GradVarName("Out"));
   }
 };
 
@@ -558,79 +586,27 @@ REGISTER_ACTIVATION_OP_MAKER(Log, LogDoc);
 REGISTER_ACTIVATION_OP_MAKER(Square, SquareDoc);
 REGISTER_ACTIVATION_OP_MAKER(Softplus, SoftplusDoc);
 REGISTER_ACTIVATION_OP_MAKER(Softsign, SoftsignDoc);
-
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Sigmoid, sigmoid);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Relu, relu);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Gelu, gelu);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Exp, exp);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Tanh, tanh);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Ceil, ceil);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Floor, floor);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Sqrt, sqrt);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(SoftRelu, soft_relu);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Relu6, relu6);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Reciprocal, reciprocal);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(HardSigmoid, hard_sigmoid);
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 
-#define FOR_EACH_INPLACE_OP_FUNCTOR(__macro) \
-  __macro(Sigmoid, sigmoid);                 \
-  __macro(Relu, relu);                       \
-  __macro(Exp, exp);                         \
-  __macro(Tanh, tanh);                       \
-  __macro(Ceil, ceil);                       \
-  __macro(Floor, floor);                     \
-  __macro(Sqrt, sqrt);                       \
-  __macro(SoftRelu, soft_relu);              \
-  __macro(Relu6, relu6);                     \
-  __macro(Reciprocal, reciprocal);           \
-  __macro(HardSigmoid, hard_sigmoid);
-
-#define FOR_EACH_OP_FUNCTOR(__macro) \
-  __macro(LogSigmoid, logsigmoid);   \
-  __macro(SoftShrink, softshrink);   \
-  __macro(Abs, abs);                 \
-  __macro(Cos, cos);                 \
-  __macro(Acos, acos);               \
-  __macro(Sin, sin);                 \
-  __macro(Asin, asin);               \
-  __macro(Atan, atan);               \
-  __macro(Round, round);             \
-  __macro(Log, log);                 \
-  __macro(Square, square);           \
-  __macro(Gelu, gelu);               \
-  __macro(BRelu, brelu);             \
-  __macro(Pow, pow);                 \
-  __macro(STanh, stanh);             \
-  __macro(Softplus, softplus);       \
-  __macro(Softsign, softsign);       \
-  __macro(LeakyRelu, leaky_relu);    \
-  __macro(TanhShrink, tanh_shrink);  \
-  __macro(ELU, elu);                 \
-  __macro(HardShrink, hard_shrink);  \
-  __macro(Swish, swish);             \
-  __macro(ThresholdedRelu, thresholded_relu);
-
-#define REGISTER_INPLACE_ACTIVATION_OP(OP_NAME, KERNEL_TYPE)                   \
-  REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp,            \
-                    ::paddle::operators::OP_NAME##OpMaker,                     \
-                    ::paddle::operators::ActivationOpInferVarType,             \
-                    ::paddle::operators::OP_NAME##GradMaker,                   \
-                    ::paddle::framework::SingleOpInplaceInToOut);              \
-  REGISTER_OPERATOR(KERNEL_TYPE##_grad, ::paddle::operators::ActivationOpGrad, \
-                    ::paddle::framework::SingleOpInplaceInToOut)
-
-#define REGISTER_ACTIVATION_OP(OP_NAME, KERNEL_TYPE)                    \
-  REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp,     \
-                    ::paddle::operators::OP_NAME##OpMaker,              \
-                    ::paddle::operators::ActivationOpInferVarType,      \
-                    ::paddle::framework::DefaultGradOpDescMaker<true>); \
-  REGISTER_OPERATOR(KERNEL_TYPE##_grad, ::paddle::operators::ActivationOpGrad)
-
-#define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor)   \
+#define REGISTER_ACTIVATION_OP(KERNEL_TYPE, OP_NAME, functor, grad_functor) \
+  REGISTER_OPERATOR(                                                        \
+      KERNEL_TYPE, ops::ActivationOp, ops::OP_NAME##OpMaker,                \
+      ops::ActivationOpInferVarType,                                        \
+      ops::ActivationGradOpDescMaker<ops::grad_functor<float>::FwdDeps()>,  \
+      std::conditional<ops::CanInplaceAct<ops::grad_functor<float>>(),      \
+                       ::paddle::framework::SingleOpInplaceInToOut,         \
+                       void>::type);                                        \
+  REGISTER_OPERATOR(                                                        \
+      KERNEL_TYPE##_grad, ops::ActivationOpGrad,                            \
+      std::conditional<ops::CanInplaceAct<ops::grad_functor<float>>(),      \
+                       ::paddle::framework::SingleOpInplaceInToOut,         \
+                       void>::type)
+
+#define REGISTER_ACTIVATION_CPU_KERNEL(act_type, op_name, functor,        \
+                                       grad_functor)                      \
   REGISTER_OP_CPU_KERNEL(                                                 \
       act_type, ops::ActivationKernel<paddle::platform::CPUDeviceContext, \
                                       ops::functor<float>>,               \
@@ -643,6 +619,5 @@ namespace ops = paddle::operators;
       ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,       \
                                 ops::grad_functor<double>>);
 
-FOR_EACH_OP_FUNCTOR(REGISTER_ACTIVATION_OP);
-FOR_EACH_INPLACE_OP_FUNCTOR(REGISTER_INPLACE_ACTIVATION_OP);
-FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL);
+FOR_EACH_ACTIVATION_OP(REGISTER_ACTIVATION_OP);
+FOR_EACH_ACTIVATION_OP(REGISTER_ACTIVATION_CPU_KERNEL);

paddle/fluid/operators/activation_op.cu

@@ -15,7 +15,8 @@ limitations under the License. */
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 
-#define REGISTER_ACTIVATION_CUDA_KERNEL(act_type, functor, grad_functor)    \
+#define REGISTER_ACTIVATION_CUDA_KERNEL(act_type, op_name, functor,         \
+                                        grad_functor)                       \
   REGISTER_OP_CUDA_KERNEL(                                                  \
       act_type,                                                             \
       ops::ActivationKernel<plat::CUDADeviceContext, ops::functor<float>>,  \
@@ -30,4 +31,4 @@ namespace plat = paddle::platform;
       ops::ActivationGradKernel<plat::CUDADeviceContext,                    \
                                 ops::grad_functor<plat::float16>>);
 
-FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL);
+FOR_EACH_ACTIVATION_OP(REGISTER_ACTIVATION_CUDA_KERNEL);

paddle/fluid/operators/activation_op.h

@@ -12,6 +12,7 @@ limitations under the License. */
 #pragma once
 #include <glog/logging.h>
 #include <algorithm>
+#include <memory>
 #include <string>
 #include <unordered_set>
 #include <utility>
@@ -35,21 +36,29 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-/* Use ugly global variable, for the using in python layer side
-   Please refer to the layer_helper.py and get the details.
- */
-static std::unordered_set<std::string> InplaceOpSet = {
-    "sigmoid", "exp",        "relu",  "tanh",      "sqrt",        "ceil",
-    "floor",   "reciprocal", "relu6", "soft_relu", "hard_sigmoid"};
+enum ActBwdOpFwdDeps {
+  kNoDeps = 0x00,  // Do not need any forward input/output
+  kDepX = 0x01,    // Only need forward input X
+  kDepOut = 0x02,  // Only need forward output Out
+
+  // Never add kDepXOut, because Out can be always calculated
+  // by forward input X in backward part.
+  // FIXME(zjl): but in MKLDNN abs, X and Out are all needed...
+  // Developers should not rely on this enum value!
+  kDepXOut = 0x03
+};
+
+std::unique_ptr<std::unordered_set<std::string>> GetInplaceOpSet();
 
 static bool IsInplace(const std::string& op) {
-  bool inplace = InplaceOpSet.count(op);
+  static auto InplaceOpSet = GetInplaceOpSet();
+  bool inplace = InplaceOpSet->count(op);
   // for op_grad
   const int kGradSuffixLen = 4;
   if (op.size() > kGradSuffixLen &&
       op.compare(op.size() - kGradSuffixLen - 1, kGradSuffixLen, "grad")) {
     inplace =
-        InplaceOpSet.count(op.substr(0, op.size() - (kGradSuffixLen + 1)));
+        InplaceOpSet->count(op.substr(0, op.size() - (kGradSuffixLen + 1)));
   }
   return inplace;
 }
@@ -85,16 +94,21 @@ inline void ExtractActivationTensor(const framework::ExecutionContext& context,
                  context.op().Output("Out"));
 }
 
+template <ActBwdOpFwdDeps kDepValue>
 inline void ExtractActivationGradTensor(
     const framework::ExecutionContext& context, const framework::Tensor** X,
     const framework::Tensor** Out, const framework::Tensor** dOut,
     framework::Tensor** dX) {
-  auto out_var = context.InputVar("Out");
   auto out_grad_var = context.InputVar(framework::GradVarName("Out"));
   auto x_grad_var = context.OutputVar(framework::GradVarName("X"));
-  PADDLE_ENFORCE(out_var != nullptr,
-                 "Cannot get input Variable Out, variable name = %s",
-                 context.op().Input("Out"));
+  const framework::Variable* out_var = nullptr;
+
+  if (static_cast<int>(kDepValue) & static_cast<int>(kDepOut)) {
+    out_var = context.InputVar("Out");
+    PADDLE_ENFORCE(out_var != nullptr,
+                   "Cannot get input Variable Out, variable name = %s",
+                   context.op().Input("Out"));
+  }
   PADDLE_ENFORCE(out_grad_var != nullptr,
                  "Cannot get input Variable %s, variable name = %s",
                  framework::GradVarName("Out"),
@@ -105,23 +119,36 @@ inline void ExtractActivationGradTensor(
                  context.op().Output(framework::GradVarName("X")));
 
   if (CanBeUsedBySelectedRows.count(context.op().Type())) {
-    *Out = paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*out_var);
     *dOut = paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(
         *out_grad_var);
     *dX = paddle::framework::GetMutableLoDTensorOrSelectedRowsValueFromVar(
         x_grad_var);
+
+    if (out_var) {
+      *Out =
+          paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*out_var);
+    } else {
+      *Out = *dOut;  // fake out
+    }
+
   } else {
     *Out = context.Input<framework::Tensor>("Out");
     *dOut = context.Input<framework::Tensor>(framework::GradVarName("Out"));
     *dX = context.Output<framework::Tensor>(framework::GradVarName("X"));
+
+    if (out_var) {
+      *Out = &(out_var->Get<framework::LoDTensor>());
+    } else {
+      *Out = *dOut;  // fake out
+    }
   }
+
   PADDLE_ENFORCE(*dX != nullptr,
                  "Cannot get output tensor %s, variable name = %s",
                  framework::GradVarName("X"),
                  context.op().Output(framework::GradVarName("X")));
 
-  bool inplace = IsInplace(context.op().Type());
-  if (!inplace) {
+  if (static_cast<int>(kDepValue) & static_cast<int>(kDepX)) {
     auto x_var = context.InputVar("X");
     PADDLE_ENFORCE(x_var != nullptr,
                    "Cannot get input tensor X, variable name = %s",
@@ -172,7 +199,8 @@ class ActivationGradKernel
     const framework::Tensor *X, *Out, *dOut;
     framework::Tensor* dX = nullptr;
     X = Out = dOut = nullptr;
-    ExtractActivationGradTensor(context, &X, &Out, &dOut, &dX);
+    ExtractActivationGradTensor<Functor::FwdDeps()>(context, &X, &Out, &dOut,
+                                                    &dX);
     dX->mutable_data<T>(context.GetPlace());
     auto dout = framework::EigenVector<T>::Flatten(detail::Ref(dOut));
     auto out = framework::EigenVector<T>::Flatten(detail::Ref(Out));
@@ -222,6 +250,8 @@ struct SigmoidGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = dout * out * (static_cast<T>(1) - out);
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 // Originally: logsigmoid(x) = -log (1 + exp(-x))
@@ -258,6 +288,8 @@ struct LogSigmoidGradFunctor : public BaseActivationFunctor<T> {
     dx.device(d) =
         dout * ((-x - temp).exp() / ((-temp).exp() + (-x - temp).exp()));
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 // exp(x) = e^x
@@ -276,6 +308,8 @@ struct ExpGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = dout * out;
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 // relu(x) = max(x, 0)
@@ -294,6 +328,8 @@ struct ReluGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = dout * (out > static_cast<T>(0)).template cast<T>();
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 // gelu(x) = 0.5 * x *  (1 + erf(x / sqrt(2)))
@@ -338,6 +374,8 @@ struct GeluGradFunctor : BaseActivationFunctor<T> {
                   (-static_cast<T>(0.5) * x.square()).exp();
     dx.device(d) = dout * (first + second);
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 // tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x))
@@ -356,6 +394,8 @@ struct TanhGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = dout * (static_cast<T>(1) - out * out);
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 // tanhshrink(x) = x - tanh(x)
@@ -375,6 +415,8 @@ struct TanhShrinkGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = dout * (x.tanh() * x.tanh());
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 // tanhshrink(x) = x - tanh(x)
@@ -409,6 +451,8 @@ struct HardShrinkGradFunctor : public BaseActivationFunctor<T> {
     auto temp2 = (x > static_cast<T>(threshold)).template cast<T>().eval();
     dx.device(d) = dout * (temp1 + temp2).template cast<T>();
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 // softshrink(x) = x - lambda, if x > lambda; x + lambda, if x < -lambda; 0
@@ -443,6 +487,8 @@ struct SoftShrinkGradFunctor : public BaseActivationFunctor<T> {
     auto temp2 = (x < -lambdaT).template cast<T>().eval();
     dx.device(d) = dout * (temp1 + temp2).template cast<T>();
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 // sqrt(x) = x^(1/2)
@@ -461,6 +507,8 @@ struct SqrtGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = static_cast<T>(0.5) * dout / out;
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 // ceil(x) = ceiling(x)
@@ -479,6 +527,8 @@ struct ZeroGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = static_cast<T>(0) / out;
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kNoDeps; }
 };
 
 // floor(x) = flooring(x)
@@ -522,6 +572,8 @@ struct CosGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = -dout * x.unaryExpr(Sine<T>());
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 // cosine(x) = cos(x)
@@ -541,6 +593,8 @@ struct SinGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = dout * x.unaryExpr(Cosine<T>());
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 // sine(x) = sin(x)
@@ -582,6 +636,8 @@ struct AcosGradFunctor : public BaseActivationFunctor<T> {
     dx.device(d) =
         -dout * static_cast<T>(1) / (static_cast<T>(1) - x.square()).sqrt();
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 template <typename T>
@@ -614,6 +670,8 @@ struct AsinGradFunctor : public BaseActivationFunctor<T> {
     dx.device(d) =
         dout * static_cast<T>(1) / (static_cast<T>(1) - x.square()).sqrt();
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 template <typename T>
@@ -645,6 +703,8 @@ struct AtanGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = dout * static_cast<T>(1) / (static_cast<T>(1) + x.square());
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 // round(x) = [x]
@@ -672,6 +732,8 @@ struct AbsGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = dout * x.sign();
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepXOut; }
 };
 
 // reciprocal(x) = 1 / x
@@ -690,6 +752,8 @@ struct ReciprocalGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = dout * static_cast<T>(-1) * out * out;
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 // log(x) = natural logarithm of x
@@ -708,6 +772,8 @@ struct LogGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = dout * (static_cast<T>(1) / x);
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 // square(x) = x^2
@@ -726,6 +792,8 @@ struct SquareGradFunctor : public BaseActivationFunctor<T> {
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = dout * static_cast<T>(2) * x;
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 template <typename T>
@@ -760,6 +828,8 @@ struct BReluGradFunctor : public BaseActivationFunctor<T> {
                    ((x > static_cast<T>(t_min)) * (x < static_cast<T>(t_max)))
                        .template cast<T>();
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 // relu6(x) = min(max(0, x), 6)
@@ -792,6 +862,8 @@ struct Relu6GradFunctor : public BaseActivationFunctor<T> {
         ((out > static_cast<T>(0)) * (out < static_cast<T>(threshold)))
             .template cast<T>();
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 // softplus(x) = log(1 + exp(x))
@@ -821,6 +893,8 @@ struct SoftplusGradFunctor : public BaseActivationFunctor<T> {
     dx.device(d) =
         dout * ((x - temp).exp() / ((-temp).exp() + (x - temp).exp()));
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 // softsign(x) = x / (1 + |x|)
@@ -842,6 +916,8 @@ struct SoftsignGradFunctor : public BaseActivationFunctor<T> {
     dx.device(d) =
         dout * (static_cast<T>(1) / (static_cast<T>(1) + x.abs()).square());
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 template <typename T>
@@ -872,6 +948,8 @@ struct SoftReluGradFunctor : public BaseActivationFunctor<T> {
     auto temp = ((out > -tmp) * (out < tmp)).template cast<T>().eval();
     dx.device(d) = dout * (static_cast<T>(1) - (-out).exp()) * temp;
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 template <typename T>
@@ -901,6 +979,8 @@ struct LeakyReluGradFunctor : public BaseActivationFunctor<T> {
     auto temp2 = (x >= static_cast<T>(0)).template cast<T>().eval();
     dx.device(d) = dout * (temp1 + temp2).template cast<T>();
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 template <typename T>
@@ -928,9 +1008,11 @@ struct ELUGradFunctor : public BaseActivationFunctor<T> {
             typename dX>
   void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
     dx.device(d) = dout * (x > static_cast<T>(0)).template cast<T>() +
-                   dout * (out + static_cast<T>(alpha)) *
+                   dout * static_cast<T>(alpha) * x.exp() *
                        (x < static_cast<T>(0)).template cast<T>();
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 // FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5198
@@ -958,6 +1040,8 @@ struct PowGradFunctor : public BaseActivationFunctor<T> {
     dx.device(d) = dout * static_cast<T>(factor) *
                    x.pow(static_cast<T>(factor) - static_cast<T>(1));
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 template <typename T>
@@ -991,6 +1075,8 @@ struct STanhGradFunctor : public BaseActivationFunctor<T> {
     auto temp = (a * x).tanh() * (a * x).tanh();
     dx.device(d) = dout * a * b * (static_cast<T>(1) - temp);
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 template <typename T>
@@ -1020,6 +1106,8 @@ struct ThresholdedReluGradFunctor : public BaseActivationFunctor<T> {
     auto th = static_cast<T>(threshold);
     dx.device(d) = dout * (x > th).template cast<T>();
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 template <typename T>
@@ -1053,6 +1141,8 @@ struct HardSigmoidGradFunctor : public BaseActivationFunctor<T> {
                        .template cast<T>() *
                    static_cast<T>(slope);
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 template <typename T>
@@ -1077,49 +1167,54 @@ struct SwishGradFunctor : 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 {
+  void operator()(Device d, X x, Out fake_out, dOut dout, dX dx) const {
     auto temp1 = static_cast<T>(1) /
                  (static_cast<T>(1) + (static_cast<T>(-beta) * x).exp());
+    auto out = x * temp1;
     auto temp2 = temp1 * (static_cast<T>(1) - (static_cast<T>(beta) * out));
     dx.device(d) = dout * ((static_cast<T>(beta) * out) + temp2);
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepX; }
 };
 
 }  // namespace operators
 }  // namespace paddle
 
-#define FOR_EACH_KERNEL_FUNCTOR(__macro)                             \
-  __macro(sigmoid, SigmoidFunctor, SigmoidGradFunctor);              \
-  __macro(logsigmoid, LogSigmoidFunctor, LogSigmoidGradFunctor);     \
-  __macro(exp, ExpFunctor, ExpGradFunctor);                          \
-  __macro(relu, ReluFunctor, ReluGradFunctor);                       \
-  __macro(gelu, GeluFunctor, GeluGradFunctor);                       \
-  __macro(tanh, TanhFunctor, TanhGradFunctor);                       \
-  __macro(atan, AtanFunctor, AtanGradFunctor);                       \
-  __macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor);     \
-  __macro(sqrt, SqrtFunctor, SqrtGradFunctor);                       \
-  __macro(abs, AbsFunctor, AbsGradFunctor);                          \
-  __macro(ceil, CeilFunctor, ZeroGradFunctor);                       \
-  __macro(floor, FloorFunctor, ZeroGradFunctor);                     \
-  __macro(cos, CosFunctor, CosGradFunctor);                          \
-  __macro(acos, AcosFunctor, AcosGradFunctor);                       \
-  __macro(sin, SinFunctor, SinGradFunctor);                          \
-  __macro(asin, AsinFunctor, AsinGradFunctor);                       \
-  __macro(round, RoundFunctor, ZeroGradFunctor);                     \
-  __macro(reciprocal, ReciprocalFunctor, ReciprocalGradFunctor);     \
-  __macro(log, LogFunctor, LogGradFunctor);                          \
-  __macro(square, SquareFunctor, SquareGradFunctor);                 \
-  __macro(brelu, BReluFunctor, BReluGradFunctor);                    \
-  __macro(soft_relu, SoftReluFunctor, SoftReluGradFunctor);          \
-  __macro(pow, PowFunctor, PowGradFunctor);                          \
-  __macro(stanh, STanhFunctor, STanhGradFunctor);                    \
-  __macro(softplus, SoftplusFunctor, SoftplusGradFunctor);           \
-  __macro(softsign, SoftsignFunctor, SoftsignGradFunctor);           \
-  __macro(relu6, Relu6Functor, Relu6GradFunctor);                    \
-  __macro(leaky_relu, LeakyReluFunctor, LeakyReluGradFunctor);       \
-  __macro(tanh_shrink, TanhShrinkFunctor, TanhShrinkGradFunctor);    \
-  __macro(elu, ELUFunctor, ELUGradFunctor);                          \
-  __macro(hard_shrink, HardShrinkFunctor, HardShrinkGradFunctor);    \
-  __macro(hard_sigmoid, HardSigmoidFunctor, HardSigmoidGradFunctor); \
-  __macro(swish, SwishFunctor, SwishGradFunctor);                    \
-  __macro(thresholded_relu, ThresholdedReluFunctor, ThresholdedReluGradFunctor);
+#define FOR_EACH_ACTIVATION_OP(__macro)                                       \
+  __macro(sigmoid, Sigmoid, SigmoidFunctor, SigmoidGradFunctor);              \
+  __macro(logsigmoid, LogSigmoid, LogSigmoidFunctor, LogSigmoidGradFunctor);  \
+  __macro(exp, Exp, ExpFunctor, ExpGradFunctor);                              \
+  __macro(relu, Relu, ReluFunctor, ReluGradFunctor);                          \
+  __macro(gelu, Gelu, GeluFunctor, GeluGradFunctor);                          \
+  __macro(tanh, Tanh, TanhFunctor, TanhGradFunctor);                          \
+  __macro(atan, Atan, AtanFunctor, AtanGradFunctor);                          \
+  __macro(softshrink, SoftShrink, SoftShrinkFunctor, SoftShrinkGradFunctor);  \
+  __macro(sqrt, Sqrt, SqrtFunctor, SqrtGradFunctor);                          \
+  __macro(abs, Abs, AbsFunctor, AbsGradFunctor);                              \
+  __macro(ceil, Ceil, CeilFunctor, ZeroGradFunctor);                          \
+  __macro(floor, Floor, FloorFunctor, ZeroGradFunctor);                       \
+  __macro(cos, Cos, CosFunctor, CosGradFunctor);                              \
+  __macro(acos, Acos, AcosFunctor, AcosGradFunctor);                          \
+  __macro(sin, Sin, SinFunctor, SinGradFunctor);                              \
+  __macro(asin, Asin, AsinFunctor, AsinGradFunctor);                          \
+  __macro(round, Round, RoundFunctor, ZeroGradFunctor);                       \
+  __macro(reciprocal, Reciprocal, ReciprocalFunctor, ReciprocalGradFunctor);  \
+  __macro(log, Log, LogFunctor, LogGradFunctor);                              \
+  __macro(square, Square, SquareFunctor, SquareGradFunctor);                  \
+  __macro(brelu, BRelu, BReluFunctor, BReluGradFunctor);                      \
+  __macro(soft_relu, SoftRelu, SoftReluFunctor, SoftReluGradFunctor);         \
+  __macro(pow, Pow, PowFunctor, PowGradFunctor);                              \
+  __macro(stanh, STanh, STanhFunctor, STanhGradFunctor);                      \
+  __macro(softplus, Softplus, SoftplusFunctor, SoftplusGradFunctor);          \
+  __macro(softsign, Softsign, SoftsignFunctor, SoftsignGradFunctor);          \
+  __macro(relu6, Relu6, Relu6Functor, Relu6GradFunctor);                      \
+  __macro(leaky_relu, LeakyRelu, LeakyReluFunctor, LeakyReluGradFunctor);     \
+  __macro(tanh_shrink, TanhShrink, TanhShrinkFunctor, TanhShrinkGradFunctor); \
+  __macro(elu, ELU, ELUFunctor, ELUGradFunctor);                              \
+  __macro(hard_shrink, HardShrink, HardShrinkFunctor, HardShrinkGradFunctor); \
+  __macro(hard_sigmoid, HardSigmoid, HardSigmoidFunctor,                      \
+          HardSigmoidGradFunctor);                                            \
+  __macro(swish, Swish, SwishFunctor, SwishGradFunctor);                      \
+  __macro(thresholded_relu, ThresholdedRelu, ThresholdedReluFunctor,          \
+          ThresholdedReluGradFunctor);

paddle/fluid/operators/distributed/grpc/grpc_serde.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 #include <nccl.h>
 #endif
 #include <limits>
+#include <memory>
 #include <thread>  // NOLINT
 
 #include "google/protobuf/io/coded_stream.h"
@@ -104,8 +105,10 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
   e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber,
                             payload->memory_size());
   if (payload->memory_size() >= std::numeric_limits<int>::max()) {
-    LOG(FATAL) << "AppendZeroCopy varname:" << name
-               << ", vlen:" << payload->memory_size();
+    LOG(FATAL) << "FATAL error: varname:" << name
+               << ", vlen:" << payload->memory_size()
+               << " >= std::numeric_limits<int>::max():"
+               << std::numeric_limits<int>::max() << ", so exit!";
   }
   // steal reference of tensor data
   ::grpc::Slice slices[4];  // metadata, tensor, rows meta, rows

paddle/fluid/operators/interpolate_op.cc

@@ -37,10 +37,19 @@ class InterpolateOp : public framework::OperatorWithKernel {
         "Interpolation method can only be \"bilinear\" or \"nearest\".");
 
     auto dim_x = ctx->GetInputDim("X");  // NCHW format
-    int out_h = ctx->Attrs().Get<int>("out_h");
-    int out_w = ctx->Attrs().Get<int>("out_w");
     PADDLE_ENFORCE_EQ(dim_x.size(), 4, "X's dimension must be 4");
 
+    int out_h, out_w;
+    float scale = ctx->Attrs().Get<float>("scale");
+    if (scale > 0) {
+      // round down
+      out_h = static_cast<int>(dim_x[2] * scale);
+      out_w = static_cast<int>(dim_x[3] * scale);
+    } else {
+      out_h = ctx->Attrs().Get<int>("out_h");
+      out_w = ctx->Attrs().Get<int>("out_w");
+    }
+
     if (ctx->HasInput("OutSize") && ctx->IsRuntime()) {
       auto out_size_dim = ctx->GetInputDim("OutSize");
       PADDLE_ENFORCE_EQ(out_size_dim.size(), 1,
@@ -77,6 +86,7 @@ class InterpolateOpMaker : public framework::OpProtoAndCheckerMaker {
 
     AddAttr<int>("out_h", "output height of interpolate op.");
     AddAttr<int>("out_w", "output width of interpolate op.");
+    AddAttr<float>("scale", "scale factor of interpolate op.").SetDefault(0.);
     AddAttr<std::string>("interp_method",
                          "(string, default \"bilinear\"), interpolation "
                          "method, can be \"bilinear\" for "

paddle/fluid/operators/interpolate_op.cu

@@ -192,9 +192,21 @@ class InterpolateOpCUDAKernel : public framework::OpKernel<T> {
     auto* output = ctx.Output<Tensor>("Out");
     auto* input_data = input->data<T>();
 
+    int n = input->dims()[0];
+    int c = input->dims()[1];
+    int in_h = input->dims()[2];
+    int in_w = input->dims()[3];
+
     auto interp_method = ctx.Attr<std::string>("interp_method");
     int out_h = ctx.Attr<int>("out_h");
     int out_w = ctx.Attr<int>("out_w");
+
+    float scale = ctx.Attr<float>("scale");
+    if (scale > 0) {
+      out_h = in_h * scale;
+      out_w = in_w * scale;
+    }
+
     auto out_size = ctx.Input<Tensor>("OutSize");
     if (out_size != nullptr) {
       Tensor sizes;
@@ -207,11 +219,6 @@ class InterpolateOpCUDAKernel : public framework::OpKernel<T> {
     bool align_corners = ctx.Attr<bool>("align_corners");
     int align_mode = ctx.Attr<int>("align_mode");
 
-    int n = input->dims()[0];
-    int c = input->dims()[1];
-    int in_h = input->dims()[2];
-    int in_w = input->dims()[3];
-
     auto* output_data =
         output->mutable_data<T>({n, c, out_h, out_w}, ctx.GetPlace());
 
@@ -268,14 +275,20 @@ class InterpolateGradOpCUDAKernel : public framework::OpKernel<T> {
     math::SetConstant<platform::CUDADeviceContext, T> zero;
     zero(device_ctx, input_grad, static_cast<T>(0.0));
 
+    int n = input_grad->dims()[0];
+    int c = input_grad->dims()[1];
+    int in_h = input_grad->dims()[2];
+    int in_w = input_grad->dims()[3];
+
     auto interp_method = ctx.Attr<std::string>("interp_method");
     int out_h = ctx.Attr<int>("out_h");
     int out_w = ctx.Attr<int>("out_w");
+    float scale = ctx.Attr<float>("scale");
+    if (scale > 0) {
+      out_h = in_h * scale;
+      out_w - in_w* scale;
+    }
     auto out_size = ctx.Input<Tensor>("OutSize");
-
-    bool align_corners = ctx.Attr<bool>("align_corners");
-    int align_mode = ctx.Attr<int>("align_mode");
-
     if (out_size != nullptr) {
       Tensor sizes;
       framework::TensorCopy(*out_size, platform::CPUPlace(), &sizes);
@@ -284,10 +297,8 @@ class InterpolateGradOpCUDAKernel : public framework::OpKernel<T> {
       out_w = size_data[1];
     }
 
-    int n = input_grad->dims()[0];
-    int c = input_grad->dims()[1];
-    int in_h = input_grad->dims()[2];
-    int in_w = input_grad->dims()[3];
+    bool align_corners = ctx.Attr<bool>("align_corners");
+    int align_mode = ctx.Attr<int>("align_mode");
 
     int in_hw = in_h * in_w;
     int out_hw = out_h * out_w;

paddle/fluid/operators/interpolate_op.h

@@ -163,9 +163,21 @@ class InterpolateKernel : public framework::OpKernel<T> {
     auto* input = ctx.Input<Tensor>("X");
     auto* output = ctx.Output<Tensor>("Out");
 
+    const int n = input->dims()[0];
+    const int c = input->dims()[1];
+    const int in_h = input->dims()[2];
+    const int in_w = input->dims()[3];
+
     std::string interp_method = ctx.Attr<std::string>("interp_method");
     int out_h = ctx.Attr<int>("out_h");
     int out_w = ctx.Attr<int>("out_w");
+
+    float scale = ctx.Attr<float>("scale");
+    if (scale > 0) {
+      out_h = static_cast<int>(in_h * scale);
+      out_w = static_cast<int>(in_w * scale);
+    }
+
     auto out_size = ctx.Input<Tensor>("OutSize");
     if (out_size != nullptr) {
       auto out_size_data = out_size->data<int>();
@@ -175,11 +187,6 @@ class InterpolateKernel : public framework::OpKernel<T> {
     bool align_corners = ctx.Attr<bool>("align_corners");
     int align_mode = ctx.Attr<int>("align_mode");
 
-    const int n = input->dims()[0];
-    const int c = input->dims()[1];
-    const int in_h = input->dims()[2];
-    const int in_w = input->dims()[3];
-
     output->mutable_data<T>({n, c, out_h, out_w}, ctx.GetPlace());
     auto& device_ctx =
         ctx.template device_context<platform::CPUDeviceContext>();
@@ -221,23 +228,31 @@ class InterpolateGradKernel : public framework::OpKernel<T> {
     auto* input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto* output_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
 
+    const int n = input->dims()[0];
+    const int c = input->dims()[1];
+    const int in_h = input->dims()[2];
+    const int in_w = input->dims()[3];
+
     std::string interp_method = ctx.Attr<std::string>("interp_method");
     int out_h = ctx.Attr<int>("out_h");
     int out_w = ctx.Attr<int>("out_w");
+
+    float scale = ctx.Attr<float>("scale");
+    if (scale > 0) {
+      out_h = static_cast<int>(in_h * scale);
+      out_w = static_cast<int>(in_w * scale);
+    }
+
     auto out_size = ctx.Input<Tensor>("OutSize");
     if (out_size != nullptr) {
       auto out_size_data = out_size->data<int>();
       out_h = out_size_data[0];
       out_w = out_size_data[1];
     }
+
     bool align_corners = ctx.Attr<bool>("align_corners");
     int align_mode = ctx.Attr<int>("align_mode");
 
-    const int n = input->dims()[0];
-    const int c = input->dims()[1];
-    const int in_h = input->dims()[2];
-    const int in_w = input->dims()[3];
-
     input_grad->mutable_data<T>({n, c, in_h, in_w}, ctx.GetPlace());
     auto& device_ctx =
         ctx.template device_context<platform::CPUDeviceContext>();

paddle/scripts/paddle_build.sh

@@ -291,8 +291,12 @@ function build() {
     Building in /paddle/build ...
     ============================================
 EOF
+    parallel_number=`nproc`
+    if [[ "$1" != "" ]]; then
+      parallel_number=$1
+    fi
     make clean
-    make -j `nproc`
+    make -j ${parallel_number}
     make install -j `nproc`
 }
 
@@ -737,9 +741,13 @@ function gen_fluid_lib() {
     Generating fluid library for train and inference ...
     ========================================
 EOF
+    parallel_number=`nproc`
+    if [[ "$1" != "" ]]; then
+      parallel_number=$1
+    fi
     cmake .. -DWITH_DISTRIBUTE=OFF -DON_INFER=ON
-    make -j `nproc` fluid_lib_dist
-    make -j `nproc` inference_lib_dist
+    make -j ${parallel_number} fluid_lib_dist
+    make -j ${parallel_number} inference_lib_dist
 }
 
 function tar_fluid_lib() {
@@ -770,11 +778,22 @@ EOF
 
 function main() {
     local CMD=$1
+    local parallel_number=$2
     init
     case $CMD in
+      build_only)
+        cmake_gen ${PYTHON_ABI:-""}
+        build ${parallel_number}
+        ;;
+      build_and_check)
+        cmake_gen ${PYTHON_ABI:-""}
+        build ${parallel_number}
+        assert_api_not_changed ${PYTHON_ABI:-""}
+        assert_api_spec_approvals
+        ;;
       build)
         cmake_gen ${PYTHON_ABI:-""}
-        build
+        build ${parallel_number}
         gen_dockerfile ${PYTHON_ABI:-""}
         ;;
       test)
@@ -797,7 +816,7 @@ function main() {
         ;;
       fluid_inference_lib)
         cmake_gen ${PYTHON_ABI:-""}
-        gen_fluid_lib
+        gen_fluid_lib ${parallel_number}
         tar_fluid_lib
         test_fluid_lib
         ;;
@@ -806,16 +825,16 @@ function main() {
         ;;
       cicheck)
         cmake_gen ${PYTHON_ABI:-""}
-        build
+        build ${parallel_number}
         assert_api_not_changed ${PYTHON_ABI:-""}
         run_test
-        gen_fluid_lib
+        gen_fluid_lib ${parallel_number}
         test_fluid_lib
         assert_api_spec_approvals
         ;;
       cicheck_brpc)
         cmake_gen ${PYTHON_ABI:-""}
-        build
+        build ${parallel_number}
         run_brpc_test
         ;;
       assert_api)
@@ -823,7 +842,7 @@ function main() {
         assert_api_spec_approvals
         ;;
       test_inference)
-        gen_fluid_lib
+        gen_fluid_lib ${parallel_number}
         test_fluid_lib
         ;;
       assert_api_approvals)
@@ -840,7 +859,7 @@ function main() {
         ;;
       cicheck_py35)
         cmake_gen ${PYTHON_ABI:-""}
-        build
+        build ${parallel_number}
         run_test
         assert_api_not_changed ${PYTHON_ABI:-""}
         ;;
@@ -848,7 +867,7 @@ function main() {
         cmake_gen ${PYTHON_ABI:-""}
         ;;
       gen_fluid_lib)
-        gen_fluid_lib
+        gen_fluid_lib ${parallel_number}
         ;;
       test_fluid_lib)
         test_fluid_lib

python/paddle/fluid/__init__.py

@@ -66,6 +66,8 @@ from . import compiler
 from .compiler import *
 from paddle.fluid.layers.math_op_patch import monkey_patch_variable
 from . import install_check
+from .dygraph.nn import *
+from .dygraph.layers import *
 
 Tensor = LoDTensor
 

python/paddle/fluid/dygraph/base.py

@@ -22,7 +22,7 @@ __all__ = ['enabled', 'guard', 'to_variable']
 
 
 def enabled():
-    return framework._in_dygraph_mode()
+    return framework.in_dygraph_mode()
 
 
 @signature_safe_contextmanager

python/paddle/fluid/dygraph/checkpoint.py

@@ -97,20 +97,12 @@ def load_persistables(vardict, dirname, filename=None):
 
     Examples:
         .. code-block:: python
-            my_layer = layer(fluid.dygraph.Layer)
+            my_layer = layer(fluid.Layer)
             param_path = "./my_paddle_model"
 
             param_dict = fluid.dygraph.load_persistables(my_layer.parameters(), param_path)
             param_1 = param_dict['PtbModel_0.w_1']
 
-            or:
-            my_layer = layer(fluid.dygraph.Layer)
-            param_path = "./my_paddle_model"
-            filename = "model.file"
-            param_dict = fluid.dygraph.load_persistables(my_layer.state_dict(), param_path,
-                                                                       filename=filename)
-            param_1 = param_dict['PtbModel_0.w_1']
-
         """
     if isinstance(vardict, collections.OrderedDict):
         return _load_var_from_file(vardict, dirname, filename)

python/paddle/fluid/dygraph/layer_object_helper.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import copy
 import six
-from ..framework import Parameter, _in_dygraph_mode
+from ..framework import Parameter, in_dygraph_mode
 from ..param_attr import ParamAttr
 from .. import core
 from six.moves import zip

python/paddle/fluid/dygraph/layers.py

@@ -139,14 +139,14 @@ class Layer(core.Layer):
 
     def clear_gradients(self):
         for p in self.parameters():
-            p._clear_gradient()
+            p.clear_gradient()
 
-    def _build_once(self, *args):
+    def build_once(self, *args):
         pass
 
     def __call__(self, *inputs):
         if not self._built:
-            self._build_once(*inputs)
+            self.build_once(*inputs)
 
         outputs = self.forward(*inputs)
         self._built = True

python/paddle/fluid/dygraph/nn.py

@@ -19,7 +19,7 @@ from six.moves import reduce
 from .. import core
 from ..layers import utils
 from . import layers
-from ..framework import Variable, _in_dygraph_mode, OpProtoHolder, Parameter
+from ..framework import Variable, in_dygraph_mode, OpProtoHolder, Parameter
 from ..param_attr import ParamAttr
 from ..initializer import Normal, Constant, NumpyArrayInitializer
 import numpy as np
@@ -33,6 +33,109 @@ __all__ = [
 
 
 class Conv2D(layers.Layer):
+    """
+    The convolution2D layer calculates the output based on the input, filter
+    and strides, paddings, dilations, groups parameters. Input and
+    Output are in NCHW format, where N is batch size, C is the number of
+    channels, H is the height of the feature, and W is the width of the feature.
+    Filter is in MCHW format, where M is the number of output image channels,
+    C is the number of input image channels, H is the height of the filter,
+    and W is the width of the filter. If the groups is greater than 1,
+    C will equal the number of input image channels divided by the groups.
+    Please refer to UFLDL's `convolution
+    <http://ufldl.stanford.edu/tutorial/supervised/FeatureExtractionUsingConvolution/>`_
+    for more detials.
+    If bias attribution and activation type are provided, bias is added to the
+    output of the convolution, and the corresponding activation function is
+    applied to the final result.
+
+    For each input :math:`X`, the equation is:
+
+    .. math::
+
+        Out = \sigma (W \\ast X + b)
+
+    Where:
+
+    * :math:`X`: Input value, a tensor with NCHW format.
+    * :math:`W`: Filter value, a tensor with MCHW format.
+    * :math:`\\ast`: Convolution operation.
+    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
+    * :math:`\\sigma`: Activation function.
+    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
+
+    Example:
+
+        - Input:
+
+          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`
+
+          Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)`
+
+        - Output:
+
+          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`
+
+        Where
+
+        .. math::
+
+            H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\
+            W_{out}&= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]} + 1
+
+    Args:
+        input (Variable): The input image with [N, C, H, W] format.
+        num_filters(int): The number of filter. It is as same as the output
+            image channel.
+        filter_size (int|tuple|None): The filter size. If filter_size is a tuple,
+            it must contain two integers, (filter_size_H, filter_size_W).
+            Otherwise, the filter will be a square.
+        stride (int|tuple): The stride size. If stride is a tuple, it must
+            contain two integers, (stride_H, stride_W). Otherwise, the
+            stride_H = stride_W = stride. Default: stride = 1.
+        padding (int|tuple): The padding size. If padding is a tuple, it must
+            contain two integers, (padding_H, padding_W). Otherwise, the
+            padding_H = padding_W = padding. Default: padding = 0.
+        dilation (int|tuple): The dilation size. If dilation is a tuple, it must
+            contain two integers, (dilation_H, dilation_W). Otherwise, the
+            dilation_H = dilation_W = dilation. Default: dilation = 1.
+        groups (int): The groups number of the Conv2d Layer. According to grouped
+            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
+            the first half of the filters is only connected to the first half
+            of the input channels, while the second half of the filters is only
+            connected to the second half of the input channels. Default: groups=1.
+        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
+            of conv2d. If it is set to None or one attribute of ParamAttr, conv2d
+            will create ParamAttr as param_attr. If the Initializer of the param_attr
+            is not set, the parameter is initialized with :math:`Normal(0.0, std)`,
+            and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None.
+        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d.
+            If it is set to False, no bias will be added to the output units.
+            If it is set to None or one attribute of ParamAttr, conv2d
+            will create ParamAttr as bias_attr. If the Initializer of the bias_attr
+            is not set, the bias is initialized zero. Default: None.
+        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
+            library is installed. Default: True
+        act (str): Activation type, if it is set to None, activation is not appended.
+            Default: None
+        name (str|None): A name for this layer(optional). If set None, the layer
+            will be named automatically. Default: None
+
+    Returns:
+        Variable: The tensor variable storing the convolution and \
+                  non-linearity activation result.
+
+    Raises:
+        ValueError: If the shapes of input, filter_size, stride, padding and
+                    groups mismatch.
+
+    Examples:
+        .. code-block:: python
+
+          data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
+          conv2d = fluid.layers.conv2d(input=data, num_filters=2, filter_size=3, act="relu")
+    """
+
     def __init__(self,
                  name_scope,
                  num_channels,
@@ -265,7 +368,7 @@ class Conv3D(layers.Layer):
         self._param_attr = param_attr
         self._bias_attr = bias_attr
 
-    def _build_once(self, input):
+    def build_once(self, input):
         num_channels = input.shape[1]
         self._dtype = self._helper.input_dtype(input)
 
@@ -332,6 +435,116 @@ class Conv3D(layers.Layer):
 
 
 class Conv3DTranspose(layers.Layer):
+    """
+    **Convlution3D transpose layer**
+
+    The convolution3D transpose layer calculates the output based on the input,
+    filter, and dilations, strides, paddings. Input(Input) and output(Output)
+    are in NCDHW format. Where N is batch size, C is the number of channels,
+    D is the depth of the feature, H is the height of the feature, and W
+    is the width of the feature. Parameters(dilations, strides, paddings) are
+    two elements. These two elements represent height and width, respectively.
+    The details of convolution transpose layer, please refer to the following
+    explanation and references `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
+    If bias attribution and activation type are provided, bias is added to
+    the output of the convolution, and the corresponding activation function
+    is applied to the final result.
+
+    For each input :math:`X`, the equation is:
+
+    .. math::
+
+        Out = \sigma (W \\ast X + b)
+
+    In the above equation:
+
+    * :math:`X`: Input value, a tensor with NCDHW format.
+    * :math:`W`: Filter value, a tensor with MCDHW format.
+    * :math:`\\ast`: Convolution operation.
+    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
+    * :math:`\\sigma`: Activation function.
+    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
+
+    Example:
+
+        - Input:
+
+          Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`
+
+          Filter shape: :math:`(C_{in}, C_{out}, D_f, H_f, W_f)`
+
+        - Output:
+
+          Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})`
+
+        Where
+
+        .. math::
+
+           D_{out} &= (D_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (D_f - 1) + 1 \\\\
+           H_{out} &= (H_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (H_f - 1) + 1 \\\\
+           W_{out} &= (W_{in} - 1) * strides[2] - 2 * paddings[2] + dilations[2] * (W_f - 1) + 1
+
+    Args:
+        input(Variable): The input image with [N, C, D, H, W] format.
+        num_filters(int): The number of the filter. It is as same as the output
+            image channel.
+        output_size(int|tuple|None): The output image size. If output size is a
+            tuple, it must contain three integers, (image_D, image_H, image_W). This
+            parameter only works when filter_size is None.
+        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
+            it must contain three integers, (filter_size_D, filter_size_H, filter_size_W).
+            Otherwise, the filter will be a square. None if use output size to
+            calculate filter_size.
+        padding(int|tuple): The padding size. If padding is a tuple, it must
+            contain three integers, (padding_D, padding_H, padding_W). Otherwise, the
+            padding_D = padding_H = padding_W = padding. Default: padding = 0.
+        stride(int|tuple): The stride size. If stride is a tuple, it must
+            contain three integers, (stride_D, stride_H, stride_W). Otherwise, the
+            stride_D = stride_H = stride_W = stride. Default: stride = 1.
+        dilation(int|tuple): The dilation size. If dilation is a tuple, it must
+            contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the
+            dilation_D = dilation_H = dilation_W = dilation. Default: dilation = 1.
+        groups(int): The groups number of the Conv3d transpose layer. Inspired by
+            grouped convolution in Alex Krizhevsky's Deep CNN paper, in which
+            when group=2, the first half of the filters is only connected to the
+            first half of the input channels, while the second half of the
+            filters is only connected to the second half of the input channels.
+            Default: groups=1
+        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
+            of conv3d_transpose. If it is set to None or one attribute of ParamAttr, conv3d_transpose
+            will create ParamAttr as param_attr. If the Initializer of the param_attr
+            is not set, the parameter is initialized with Xavier. Default: None.
+        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv3d_transpose.
+            If it is set to False, no bias will be added to the output units.
+            If it is set to None or one attribute of ParamAttr, conv3d_transpose
+            will create ParamAttr as bias_attr. If the Initializer of the bias_attr
+            is not set, the bias is initialized zero. Default: None.
+        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
+            library is installed. Default: True
+        act (str): Activation type, if it is set to None, activation is not appended.
+            Default: None.
+        name(str|None): A name for this layer(optional). If set None, the layer
+            will be named automatically.
+
+    Returns:
+        Variable: The tensor variable storing the convolution transpose result.
+
+    Raises:
+        ValueError: If the shapes of input, filter_size, stride, padding and
+                    groups mismatch.
+
+    Examples:
+       .. code-block:: python
+
+          conv3d_transpose = nn.Conv3DTranspose(
+                'Conv3DTranspose',
+                num_filters=12,
+                filter_size=12,
+                use_cudnn=False)
+          transpose_res = conv3d_transpose(base.to_variable(input_array))
+    """
+
     def __init__(self,
                  name_scope,
                  num_filters,
@@ -362,7 +575,7 @@ class Conv3DTranspose(layers.Layer):
         self._bias_attr = bias_attr
         self._act = act
 
-    def _build_once(self, input):
+    def build_once(self, input):
         self._dtype = self._helper.input_dtype(input)
         self._input_channel = input.shape[1]
 
@@ -436,6 +649,54 @@ class Conv3DTranspose(layers.Layer):
 
 
 class Pool2D(layers.Layer):
+    """
+    ${comment}
+
+    Args:
+        input (Variable): The input tensor of pooling operator. The format of
+                          input tensor is NCHW, where N is batch size, C is
+                          the number of channels, H is the height of the
+                          feature, and W is the width of the feature.
+        pool_size (int|list|tuple): The pool kernel size. If pool kernel size is a tuple or list,
+            it must contain two integers, (pool_size_Height, pool_size_Width).
+            Otherwise, the pool kernel size will be a square of an int.
+        pool_type: ${pooling_type_comment}
+        pool_stride (int|list|tuple): The pool stride size. If pool stride size is a tuple or list,
+            it must contain two integers, (pool_stride_Height, pool_stride_Width).
+            Otherwise, the pool stride size will be a square of an int.
+        pool_padding (int|list|tuple): The pool padding size. If pool padding size is a tuple,
+            it must contain two integers, (pool_padding_on_Height, pool_padding_on_Width).
+            Otherwise, the pool padding size will be a square of an int.
+        global_pooling (bool): ${global_pooling_comment}
+        use_cudnn (bool): ${use_cudnn_comment}
+        ceil_mode (bool): ${ceil_mode_comment}
+        name (str|None): A name for this layer(optional). If set None, the
+                        layer will be named automatically.
+        exclusive (bool): Whether to exclude padding points in average pooling
+                          mode, default is true
+
+    Returns:
+        Variable: The pooling result.
+
+    Raises:
+        ValueError: If 'pool_type' is not "max" nor "avg"
+        ValueError: If 'global_pooling' is False and 'pool_size' is -1
+        ValueError: If 'use_cudnn' is not a bool value.
+
+    Examples:
+
+        .. code-block:: python
+
+          data = fluid.layers.data(
+              name='data', shape=[3, 32, 32], dtype='float32')
+          pool2d = fluid.Pool2D("pool2d",pool_size=2,
+                            pool_type='max',
+                            pool_stride=1,
+                            global_pooling=False)
+
+          pool2d_res = pool2d(data)
+    """
+
     def __init__(self,
                  name_scope,
                  pool_size=-1,
@@ -495,6 +756,102 @@ class Pool2D(layers.Layer):
 
 
 class FC(layers.Layer):
+    """
+    **Fully Connected Layer**
+
+    This function creates a fully connected layer in the network. It can take
+    one or multiple tensors as its inputs(input can be a list of Variable, see
+    Args in detail). It creates a variable called weights for each input tensor,
+    which represents a fully connected weight matrix from each input unit to
+    each output unit. The fully connected layer multiplies each input tensor
+    with its corresponding weight to produce an output Tensor with shape [M, `size`],
+    where M is batch size. If multiple input tensors are given, the results of
+    multiple output tensors with shape [M, `size`] will be summed up. If bias_attr
+    is not None, a bias variable will be created and added to the output.
+    Finally, if activation is not None, it will be applied to the output as well.
+
+    When the input is single tensor:
+
+    .. math::
+
+        Out = Act({XW + b})
+
+    When the input are multiple tensors:
+
+    .. math::
+
+        Out = Act({\sum_{i=0}^{N-1}X_iW_i + b})
+
+    In the above equation:
+
+    * :math:`N`: Number of the input. N equals to len(input) if input is list of Variable.
+    * :math:`X_i`: The i-th input tensor.
+    * :math:`W_i`: The i-th weights matrix corresponding i-th input tensor.
+    * :math:`b`: The bias parameter created by this layer (if needed).
+    * :math:`Act`: The activation function.
+    * :math:`Out`: The output tensor.
+
+    See below for an example.
+
+    .. code-block:: text
+
+        Given:
+            data_1.data = [[[0.1, 0.2],
+                           [0.3, 0.4]]]
+            data_1.shape = (1, 2, 2) # 1 is batch_size
+
+            data_2 = [[[0.1, 0.2, 0.3]]]
+            data_2.shape = (1, 1, 3)
+
+            out = fluid.layers.fc(input=[data_1, data_2], size=2)
+
+        Then:
+            out.data = [[0.18669507, 0.1893476]]
+            out.shape = (1, 2)
+
+    Args:
+        input (Variable|list of Variable): The input tensor(s) of this layer, and the dimension of
+            the input tensor(s) is at least 2.
+        size(int): The number of output units in this layer.
+        num_flatten_dims (int, default 1): The fc layer can accept an input tensor with more than
+            two dimensions. If this happens, the multidimensional tensor will first be flattened
+            into a 2-dimensional matrix. The parameter `num_flatten_dims` determines how the input
+            tensor is flattened: the first `num_flatten_dims` (inclusive, index starts from 1)
+            dimensions will be flatten to form the first dimension of the final matrix (height of
+            the matrix), and the rest `rank(X) - num_flatten_dims` dimensions are flattened to
+            form the second dimension of the final matrix (width of the matrix). For example, suppose
+            `X` is a 5-dimensional tensor with a shape [2, 3, 4, 5, 6], and `num_flatten_dims` = 3.
+            Then, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = [24, 30].
+        param_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for learnable
+            parameters/weights of this layer.
+        bias_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for the bias
+            of this layer. If it is set to False, no bias will be added to the output units.
+            If it is set to None, the bias is initialized zero. Default: None.
+        act (str, default None): Activation to be applied to the output of this layer.
+        is_test(bool): A flag indicating whether execution is in test phase.
+        name (str, default None): The name of this layer.
+
+    Returns:
+        Variable: The transformation result.
+
+    Raises:
+        ValueError: If rank of the input tensor is less than 2.
+
+    Examples:
+        .. code-block:: python
+
+          # when input is single tensor
+          data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32")
+          fc = fluid.FC("fc", size=1000, act="tanh")
+          fc_res = fc(data)
+
+          # when input are multiple tensors
+          data_1 = fluid.layers.data(name="data_1", shape=[32, 32], dtype="float32")
+          data_2 = fluid.layers.data(name="data_2", shape=[24, 36], dtype="float32")
+          fc = fluid.FC("fc", size=1000, act="tanh")
+          fc_res = fc([data_1, data_2])
+    """
+
     def __init__(self,
                  name_scope,
                  size,
@@ -522,7 +879,7 @@ class FC(layers.Layer):
         assert isinstance(value, Parameter)
         self.__w[i] = value
 
-    def _build_once(self, input):
+    def build_once(self, input):
         i = 0
         for inp, param in self._helper.iter_inputs_and_params(input,
                                                               self._param_attr):
@@ -591,6 +948,91 @@ class FC(layers.Layer):
 
 
 class BatchNorm(layers.Layer):
+    """
+    **Batch Normalization Layer**
+
+    Can be used as a normalizer function for conv2d and fully_connected operations.
+    The required data format for this layer is one of the following:
+
+    1. NHWC `[batch, in_height, in_width, in_channels]`
+
+    2. NCHW `[batch, in_channels, in_height, in_width]`
+
+    Refer to `Batch Normalization: Accelerating Deep Network Training by Reducing
+    Internal Covariate Shift <https://arxiv.org/pdf/1502.03167.pdf>`_
+    for more details.
+
+    :math:`input` is the input features over a mini-batch.
+
+    ..  math::
+
+        \\mu_{\\beta} &\\gets \\frac{1}{m} \\sum_{i=1}^{m} x_i \\qquad &//\\
+        \ mini-batch\ mean \\\\
+        \\sigma_{\\beta}^{2} &\\gets \\frac{1}{m} \\sum_{i=1}^{m}(x_i - \\
+        \\mu_{\\beta})^2 \\qquad &//\ mini-batch\ variance \\\\
+        \\hat{x_i} &\\gets \\frac{x_i - \\mu_\\beta} {\\sqrt{\\
+        \\sigma_{\\beta}^{2} + \\epsilon}} \\qquad &//\ normalize \\\\
+        y_i &\\gets \\gamma \\hat{x_i} + \\beta \\qquad &//\ scale\ and\ shift
+
+
+    When use_global_stats = True, the :math:`\\mu_{\\beta}`
+    and :math:`\\sigma_{\\beta}^{2}` are not the statistics of one mini-batch.
+    They are global (or running) statistics. (It usually got from the
+    pre-trained model.)
+    The training and testing (or inference) have the same behavior:
+
+    ..  math::
+
+        \\hat{x_i} &\\gets \\frac{x_i - \\mu_\\beta} {\\sqrt{\\
+        \\sigma_{\\beta}^{2} + \\epsilon}}  \\\\
+        y_i &\\gets \\gamma \\hat{x_i} + \\beta
+
+    Args:
+        input(variable): The rank of input variable can be 2, 3, 4, 5.
+        act(string, Default None): Activation type, linear|relu|prelu|...
+        is_test (bool, Default False): A flag indicating whether it is in
+            test phrase or not.
+        momentum(float, Default 0.9): The value used for the moving_mean and
+            moving_var computation. The updated formula is:
+            :math:`moving\_mean = moving\_mean * momentum + new\_mean * (1. - momentum)`
+            :math:`moving\_var = moving\_var * momentum + new\_var * (1. - momentum)`
+            Default is 0.9.
+        epsilon(float, Default 1e-05): A value added to the denominator for
+            numerical stability. Default is 1e-5.
+        param_attr(ParamAttr|None): The parameter attribute for Parameter `scale`
+             of batch_norm. If it is set to None or one attribute of ParamAttr, batch_norm
+             will create ParamAttr as param_attr. If the Initializer of the param_attr
+             is not set, the parameter is initialized with Xavier. Default: None.
+        bias_attr(ParamAttr|None): The parameter attribute for the bias of batch_norm.
+             If it is set to None or one attribute of ParamAttr, batch_norm
+             will create ParamAttr as bias_attr. If the Initializer of the bias_attr
+             is not set, the bias is initialized zero. Default: None.
+        data_layout(string, default NCHW): NCHW|NHWC
+        in_place(bool, Default False): Make the input and output of batch norm reuse memory.
+        name(string, Default None): A name for this layer(optional). If set None, the layer
+            will be named automatically.
+        moving_mean_name(string, Default None): The name of moving_mean which store the global Mean.
+        moving_variance_name(string, Default None): The name of the moving_variance which store the global Variance.
+        do_model_average_for_mean_and_var(bool, Default False): Do model average for mean and variance or not.
+        fuse_with_relu (bool): if True, this OP performs relu after batch norm.
+        use_global_stats(bool, Default False): Whether to use global mean and
+            variance. In inference or test mode, set use_global_stats to true
+            or is_test to true, and the behavior is equivalent.
+            In train mode, when setting use_global_stats True, the global mean
+            and variance are also used during train period.
+
+    Returns:
+        Variable: A tensor variable which is the result after applying batch normalization on the input.
+
+    Examples:
+
+        .. code-block:: python
+            fc = fluid.FC('fc', size=200, param_attr='fc1.w')
+            hidden1 = fc(x)
+            batch_norm = fluid.BatchNorm("batch_norm", 10)
+            hidden2 = batch_norm(hidden1)
+    """
+
     def __init__(self,
                  name_scope,
                  num_channels,
@@ -629,7 +1071,7 @@ class BatchNorm(layers.Layer):
             dtype=self._dtype,
             default_initializer=Constant(1.0))
         if use_global_stats and self._param_attr.learning_rate == 0.:
-            self._scale._stop_gradient = True
+            self._scale.stop_gradient = True
 
         self._bias = self.create_parameter(
             attr=self._param_attr,
@@ -637,7 +1079,7 @@ class BatchNorm(layers.Layer):
             dtype=self._dtype,
             is_bias=True)
         if use_global_stats and self._param_attr.learning_rate == 0.:
-            self._bias._stop_gradient = True
+            self._bias.stop_gradient = True
 
         self._mean = self.create_parameter(
             attr=ParamAttr(
@@ -647,7 +1089,7 @@ class BatchNorm(layers.Layer):
                 do_model_average=do_model_average_for_mean_and_var),
             shape=param_shape,
             dtype=self._dtype)
-        self._mean._stop_gradient = True
+        self._mean.stop_gradient = True
 
         self._variance = self.create_parameter(
             attr=ParamAttr(
@@ -657,7 +1099,7 @@ class BatchNorm(layers.Layer):
                 do_model_average=do_model_average_for_mean_and_var),
             shape=param_shape,
             dtype=self._dtype)
-        self._variance._stop_gradient = True
+        self._variance.stop_gradient = True
 
         self._in_place = in_place
         self._momentum = momentum
@@ -666,7 +1108,7 @@ class BatchNorm(layers.Layer):
         self._fuse_with_relu = fuse_with_relu
         self._use_global_stats = use_global_stats
 
-    def _build_once(self, input):
+    def build_once(self, input):
         pass
 
     def forward(self, input):
@@ -747,7 +1189,7 @@ class Embedding(layers.Layer):
 
           dict_size = len(dataset.ids)
           input = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
-          embedding = fluid.dygraph.Embedding(size=[dict_size, 16])
+          embedding = fluid.Embedding(size=[dict_size, 16])
           fc = embedding(input)
     """
 
@@ -797,70 +1239,70 @@ class Embedding(layers.Layer):
 
 
 class LayerNorm(layers.Layer):
-    def __init__(self,
-                 name_scope,
-                 scale=True,
-                 shift=True,
-                 begin_norm_axis=1,
-                 epsilon=1e-05,
-                 param_attr=None,
-                 bias_attr=None,
-                 act=None):
-        """
-        ${comment}
+    """
+    ${comment}
 
-        The formula is as follows:
+    The formula is as follows:
 
-        ..  math::
+    ..  math::
 
-            \\mu & = \\frac{1}{H}\\sum_{i=1}^{H} a_i
+        \\mu & = \\frac{1}{H}\\sum_{i=1}^{H} a_i
 
-            \\sigma & = \\sqrt{\\frac{1}{H}\sum_{i=1}^{H}(a_i - \\mu)^2}
+        \\sigma & = \\sqrt{\\frac{1}{H}\sum_{i=1}^{H}(a_i - \\mu)^2}
 
-            h & = f(\\frac{g}{\\sigma}(a - \\mu) + b)
+        h & = f(\\frac{g}{\\sigma}(a - \\mu) + b)
 
-        * :math:`a`: the vector representation of the summed inputs to the neurons
-        in that layer.
+    * :math:`a`: the vector representation of the summed inputs to the neurons
+    in that layer.
 
-        * :math:`H`: the number of hidden units in a layers
+    * :math:`H`: the number of hidden units in a layers
 
-        * :math:`g`: the trainable scale parameter.
+    * :math:`g`: the trainable scale parameter.
 
-        * :math:`b`: the trainable bias parameter.
+    * :math:`b`: the trainable bias parameter.
 
-        Args:
-            input(Variable): The input tensor variable.
-            scale(bool): Whether to learn the adaptive gain :math:`g` after
-                normalization. Default True.
-            shift(bool): Whether to learn the adaptive bias :math:`b` after
-                normalization. Default True.
-            begin_norm_axis(int): The normalization will be performed along
-                dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
-                Default 1.
-            epsilon(float): The small value added to the variance to prevent
-                division by zero. Default 1e-05.
-            param_attr(ParamAttr|None): The parameter attribute for the learnable
-                gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is
-                omitted. If :attr:`scale` is True and :attr:`param_attr` is None,
-                a default :code:`ParamAttr` would be added as scale. The
-                :attr:`param_attr` is initialized as 1 if it is added. Default None.
-            bias_attr(ParamAttr|None): The parameter attribute for the learnable
-                bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is
-                omitted. If :attr:`shift` is True and :attr:`param_attr` is None,
-                a default :code:`ParamAttr` would be added as bias. The
-                :attr:`bias_attr` is initialized as 0 if it is added. Default None.
-            act(str): Activation to be applied to the output of layer normalizaiton.
-                      Default None.
-        Returns:
-            ${y_comment}
+    Args:
+        input(Variable): The input tensor variable.
+        scale(bool): Whether to learn the adaptive gain :math:`g` after
+            normalization. Default True.
+        shift(bool): Whether to learn the adaptive bias :math:`b` after
+            normalization. Default True.
+        begin_norm_axis(int): The normalization will be performed along
+            dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
+            Default 1.
+        epsilon(float): The small value added to the variance to prevent
+            division by zero. Default 1e-05.
+        param_attr(ParamAttr|None): The parameter attribute for the learnable
+            gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is
+            omitted. If :attr:`scale` is True and :attr:`param_attr` is None,
+            a default :code:`ParamAttr` would be added as scale. The
+            :attr:`param_attr` is initialized as 1 if it is added. Default None.
+        bias_attr(ParamAttr|None): The parameter attribute for the learnable
+            bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is
+            omitted. If :attr:`shift` is True and :attr:`param_attr` is None,
+            a default :code:`ParamAttr` would be added as bias. The
+            :attr:`bias_attr` is initialized as 0 if it is added. Default None.
+        act(str): Activation to be applied to the output of layer normalizaiton.
+                  Default None.
+    Returns:
+        ${y_comment}
 
-        Examples:
+    Examples:
 
-            >>> data = fluid.layers.data(name='data', shape=[3, 32, 32],
-            >>>                          dtype='float32')
-            >>> x = fluid.layers.layer_norm(input=data, begin_norm_axis=1)
-        """
+        >>> data = fluid.layers.data(name='data', shape=[3, 32, 32],
+        >>>                          dtype='float32')
+        >>> x = fluid.layers.layer_norm(input=data, begin_norm_axis=1)
+    """
 
+    def __init__(self,
+                 name_scope,
+                 scale=True,
+                 shift=True,
+                 begin_norm_axis=1,
+                 epsilon=1e-05,
+                 param_attr=None,
+                 bias_attr=None,
+                 act=None):
         super(LayerNorm, self).__init__(name_scope)
         self._scale = scale
         self._shift = shift
@@ -870,7 +1312,7 @@ class LayerNorm(layers.Layer):
         self._bias_attr = bias_attr
         self._act = act
 
-    def _build_once(self, input):
+    def build_once(self, input):
         self._dtype = self._helper.input_dtype(input)
         input_shape = input.shape
         param_shape = [
@@ -1232,7 +1674,7 @@ class NCE(layers.Layer):
             'remote_prefetch': remote_prefetch
         }
 
-    def _build_once(self, input, label, sample_weight=None):
+    def build_once(self, input, label, sample_weight=None):
         assert isinstance(input, Variable)
         assert isinstance(label, Variable)
 
@@ -1318,7 +1760,7 @@ class PRelu(layers.Layer):
             raise ValueError('mode should be one of all, channel, element.')
         self._alpha_shape = [1]
 
-    def _build_once(self, input):
+    def build_once(self, input):
         if self._mode == 'channel':
             self._alpha_shape = [1, input.shape[1], 1, 1]
         elif self._mode == 'element':
@@ -1396,7 +1838,7 @@ class BilinearTensorProduct(layers.Layer):
         self._name = name
         self._inputs = dict()
 
-    def _build_once(self, x, y):
+    def build_once(self, x, y):
         self._dtype = self._helper.input_dtype(x)
 
         param_shape = [self._size, x.shape[1], y.shape[1]]
@@ -1572,7 +2014,7 @@ class Conv2DTranspose(layers.Layer):
         self._output_size = output_size
         self._op_type = 'conv2d_transpose'
 
-    def _build_once(self, input):
+    def build_once(self, input):
         input_channel = input.shape[1]
         if (input_channel == self._groups and
                 self._num_filters == input_channel and not self._use_cudnn):
@@ -1686,7 +2128,7 @@ class SequenceConv(layers.Layer):
                  bias_attr=None,
                  param_attr=None,
                  act=None):
-        assert not _in_dygraph_mode(
+        assert not in_dygraph_mode(
         ), "SequenceConv is not supported by dynamic graph mode yet!"
         super(SequenceConv, self).__init__(name_scope)
         self._num_filters = num_filters
@@ -1696,7 +2138,7 @@ class SequenceConv(layers.Layer):
         self._bias_attr = bias_attr
         self._param_attr = param_attr
 
-    def _build_once(self, input):
+    def build_once(self, input):
         self._dtype = self._helper.input_dtype(input)
         filter_shape = [self._filter_size * input.shape[1], self._num_filters]
         self._filter_param = self.create_parameter(
@@ -1726,14 +2168,14 @@ class RowConv(layers.Layer):
                  future_context_size,
                  param_attr=None,
                  act=None):
-        assert not _in_dygraph_mode(
+        assert not in_dygraph_mode(
         ), "RowConv is not supported by dynamic graph mode yet!"
         super(RowConv, self).__init__(name_scope)
         self._act = act
         self._param_attr = param_attr
         self._future_context_size = future_context_size
 
-    def _build_once(self, input):
+    def build_once(self, input):
         self._dtype = self._helper.input_dtype(input)
         filter_shape = [self._future_context_size + 1, input.shape[1]]
         self._filter_param = self.create_parameter(
@@ -1796,7 +2238,7 @@ class GroupNorm(layers.Layer):
         if data_layout != 'NCHW':
             raise ValueError("unsupported data layout:" + data_layout)
 
-    def _build_once(self, input):
+    def build_once(self, input):
         self._dtype = self._helper.input_dtype(input)
         param_shape = [input.shape[1]]
         if self._bias_attr:
@@ -1849,7 +2291,7 @@ class SpectralNorm(layers.Layer):
         self._eps = eps
         self._dim = dim
 
-    def _build_once(self, weight):
+    def build_once(self, weight):
         self._dtype = self._helper.input_dtype(weight)
         input_shape = weight.shape
         h = input_shape[self._dim]
@@ -1904,7 +2346,7 @@ class TreeConv(layers.Layer):
         self._bias_attr = bias_attr
         self._param_attr = param_attr
 
-    def _build_once(self, nodes_vector, edge_set):
+    def build_once(self, nodes_vector, edge_set):
         assert isinstance(nodes_vector, Variable)
         assert isinstance(edge_set, Variable)
         self._dtype = self._helper.input_dtype(nodes_vector)

python/paddle/fluid/framework.py

@@ -67,6 +67,7 @@ __all__ = [
     'cuda_places',
     'cpu_places',
     'cuda_pinned_places',
+    'in_dygraph_mode',
 ]
 
 EMPTY_VAR_NAME = core.kEmptyVarName()
@@ -79,7 +80,10 @@ _dygraph_tracer_ = None
 _dygraph_current_expected_place_ = None
 
 
-def _in_dygraph_mode():
+def in_dygraph_mode():
+    '''
+    Returns(bool): True if the program is running in dynamic graph mode
+    '''
     return _dygraph_tracer_ is not None
 
 
@@ -396,7 +400,7 @@ class Variable(object):
             if not isinstance(dtype, core.VarDesc.VarType):
                 dtype = convert_np_dtype_to_dtype_(dtype)
 
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             # record vars in tracer rather than blocks
             self._ivar = kwargs.get("ivar", None)
             if not self._ivar:
@@ -482,21 +486,21 @@ class Variable(object):
 
             self.block.vars[name] = self
             self.op = None
-            self.stop_gradient = stop_gradient
+            self._stop_gradient = stop_gradient
             self.is_data = is_data
 
-    def _numpy(self):
+    def numpy(self):
         new_ivar = self._ivar._copy_to(core.CPUPlace(), True)
         return np.array(new_ivar.value().get_tensor())
 
-    def _backward(self):
+    def backward(self):
         self._ivar._run_backward()
 
-    def _gradient(self):
+    def gradient(self):
         new_ivar = self._ivar._grad_ivar()._copy_to(core.CPUPlace(), True)
         return np.array(new_ivar.value().get_tensor())
 
-    def _clear_gradient(self):
+    def clear_gradient(self):
         self._ivar._clear_gradient()
 
     def __str__(self):
@@ -516,7 +520,7 @@ class Variable(object):
         Returns:
             str: The debug string.
         """
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             # TODO(panyx0718): add more dygraph debug info.
             return 'name %s, dtype: %s shape: %s' % (self.name, self.dtype,
                                                      self.shape)
@@ -535,7 +539,7 @@ class Variable(object):
 
     __repr__ = __str__
 
-    def _set_desc(self, input):
+    def set_desc(self, input):
         """
         Set the variable description.
 
@@ -548,43 +552,43 @@ class Variable(object):
         self.desc = input
 
     @property
-    def _stop_gradient(self):
-        if _in_dygraph_mode():
+    def stop_gradient(self):
+        if in_dygraph_mode():
             return self._ivar.stop_gradient
         else:
-            return self.stop_gradient
+            return self._stop_gradient
 
-    @_stop_gradient.setter
-    def _stop_gradient(self, s):
-        if _in_dygraph_mode():
+    @stop_gradient.setter
+    def stop_gradient(self, s):
+        if in_dygraph_mode():
             self._ivar.stop_gradient = s
         else:
-            self.stop_gradient = s
+            self._stop_gradient = s
 
     @property
     def persistable(self):
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             return self._ivar.persistable
         else:
             return self.desc.persistable()
 
     @persistable.setter
     def persistable(self, p):
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             return self._ivar.persistable
         else:
             self.desc.set_persistable(p)
 
     @property
     def name(self):
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             return self._ivar.name
         else:
             return cpt.to_text(self.desc.name())
 
     @name.setter
     def name(self, new_name):
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             self._ivar.name = new_name
         else:
             self.desc.set_name(new_name)
@@ -592,14 +596,14 @@ class Variable(object):
     @property
     def shape(self):
         # convert to tuple, make it as same as numpy API.
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             return self._ivar.shape
         else:
             return tuple(self.desc.shape())
 
     @property
     def dtype(self):
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             return self._ivar.dtype
         else:
             return self.desc.dtype()
@@ -611,7 +615,7 @@ class Variable(object):
 
     @property
     def type(self):
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             return self._ivar.dtype
         else:
             return self.desc.type()
@@ -721,7 +725,7 @@ class Variable(object):
                 name=unique_name.generate(".".join(self.name)),
                 dtype=self.dtype,
                 persistable=self.persistable,
-                stop_gradient=self._stop_gradient, )
+                stop_gradient=self.stop_gradient, )
         else:
             return self
 
@@ -930,7 +934,7 @@ class Operator(object):
                  inputs=None,
                  outputs=None,
                  attrs=None):
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             if type is None:
                 raise ValueError(
                     "`type` to initialized an Operator can not be None.")
@@ -1049,7 +1053,7 @@ class Operator(object):
                     for arg in out_args:
                         out_arg_names.append(cpt.to_text(arg.name))
                         # TODO(minqiyang): could we remove variable's op in static mode?
-                        if not _in_dygraph_mode():
+                        if not in_dygraph_mode():
                             arg.op = self
                     self.desc.set_output(out_proto.name, out_arg_names)
 
@@ -1095,7 +1099,7 @@ class Operator(object):
 
     @property
     def type(self):
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             return self.iop.type
         else:
             return self.desc.type()
@@ -1638,7 +1642,7 @@ class Block(object):
         Returns:
             Operator: the append Operator.
         """
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             op = Operator(
                 block=self,
                 desc=None,
@@ -1710,7 +1714,7 @@ class Block(object):
         return self.ops[start:end]
 
     def _prepend_op(self, *args, **kwargs):
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             op = Operator(
                 self,
                 None,

python/paddle/fluid/initializer.py

@@ -165,7 +165,7 @@ class ConstantInitializer(Initializer):
                 'force_cpu': self._force_cpu or force_init_on_cpu()
             },
             stop_gradient=True)
-        if not framework._in_dygraph_mode():
+        if not framework.in_dygraph_mode():
             var.op = op
         return op
 
@@ -245,7 +245,7 @@ class UniformInitializer(Initializer):
                 attrs={"in_dtype": out_var.dtype,
                        "out_dtype": var.dtype})
 
-        if not framework._in_dygraph_mode():
+        if not framework.in_dygraph_mode():
             var.op = op
         return op
 
@@ -324,7 +324,7 @@ class NormalInitializer(Initializer):
                 outputs={"Out": var},
                 attrs={"in_dtype": out_var.dtype,
                        "out_dtype": var.dtype})
-        if not framework._in_dygraph_mode():
+        if not framework.in_dygraph_mode():
             var.op = op
         return op
 
@@ -403,7 +403,7 @@ class TruncatedNormalInitializer(Initializer):
                 outputs={"Out": var},
                 attrs={"in_dtype": out_var.dtype,
                        "out_dtype": var.dtype})
-        if not framework._in_dygraph_mode():
+        if not framework.in_dygraph_mode():
             var.op = op
         return op
 
@@ -509,7 +509,7 @@ class XavierInitializer(Initializer):
                     "seed": self._seed
                 },
                 stop_gradient=True)
-        if not framework._in_dygraph_mode():
+        if not framework.in_dygraph_mode():
             var.op = op
         return op
 
@@ -610,7 +610,7 @@ class MSRAInitializer(Initializer):
                     "seed": self._seed
                 },
                 stop_gradient=True)
-        if not framework._in_dygraph_mode():
+        if not framework.in_dygraph_mode():
             var.op = op
         return op
 
@@ -709,7 +709,7 @@ class BilinearInitializer(Initializer):
                 'shape': list(shape),
                 value_name: values
             })
-        if not framework._in_dygraph_mode():
+        if not framework.in_dygraph_mode():
             var.op = op
         return op
 
@@ -768,7 +768,7 @@ class NumpyArrayInitializer(Initializer):
                 value_name: values
             },
             stop_gradient=True)
-        if not framework._in_dygraph_mode():
+        if not framework.in_dygraph_mode():
             var.op = op
         return op
 

python/paddle/fluid/layer_helper.py

@@ -17,7 +17,7 @@ from __future__ import print_function
 import copy
 import six
 
-from .framework import Parameter, dtype_is_floating, _in_dygraph_mode
+from .framework import Parameter, dtype_is_floating, in_dygraph_mode
 from . import unique_name
 from paddle.fluid.initializer import Constant, Xavier
 from .param_attr import ParamAttr

python/paddle/fluid/layer_helper_base.py

@@ -17,7 +17,7 @@ from __future__ import print_function
 import copy
 import numpy as np
 
-from .framework import Variable, default_main_program, default_startup_program, _in_dygraph_mode, _current_expected_place
+from .framework import Variable, default_main_program, default_startup_program, in_dygraph_mode, _current_expected_place
 from . import unique_name
 from .param_attr import ParamAttr, WeightNormParamAttr
 from . import core
@@ -54,7 +54,7 @@ class LayerHelperBase(object):
         Return Variable construct from value
         """
         if isinstance(value, np.ndarray):
-            assert _in_dygraph_mode(
+            assert in_dygraph_mode(
             ), "to_variable could only be called in dygraph mode"
 
             if not block:
@@ -302,7 +302,7 @@ class LayerHelperBase(object):
             param = self._create_weight_normalize(attr, shape, dtype)
             WeightNormParamAttr.params_with_weight_norm.append(param)
             return param
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             # In dygraph mode, we want the returned parameter to be
             # initialized so that it can be used imperatively.
             return self.main_program.global_block().create_parameter(
@@ -370,7 +370,7 @@ class LayerHelperBase(object):
                initializer: initializer to use
         """
         assert isinstance(var, Variable)
-        if _in_dygraph_mode():
+        if in_dygraph_mode():
             initializer(var, var.block)
         else:
             self.startup_program.global_block().create_var(

python/paddle/fluid/layers/nn.py

@@ -23,7 +23,7 @@ import os
 import inspect
 from ..layer_helper import LayerHelper
 from ..initializer import Normal, Constant, NumpyArrayInitializer
-from ..framework import Variable, OpProtoHolder, _in_dygraph_mode
+from ..framework import Variable, OpProtoHolder, in_dygraph_mode
 from ..dygraph import base
 from ..param_attr import ParamAttr
 from .layer_function_generator import autodoc, templatedoc, _generate_doc_string_
@@ -481,7 +481,7 @@ def dynamic_lstm(input,
             forward, _ = fluid.layers.dynamic_lstm(
                 input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
     """
-    assert _in_dygraph_mode(
+    assert in_dygraph_mode(
     ) is not True, "please use lstm instead of dynamic_lstm in dygraph mode!"
     assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp."
     helper = LayerHelper('lstm', **locals())
@@ -867,7 +867,7 @@ def dynamic_lstmp(input,
                                                      proj_activation="tanh")
     """
 
-    assert _in_dygraph_mode(
+    assert in_dygraph_mode(
     ) is not True, "please use lstm instead of dynamic_lstmp in dygraph mode!"
 
     assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp."
@@ -1041,7 +1041,7 @@ def dynamic_gru(input,
             hidden = fluid.layers.dynamic_gru(input=x, size=hidden_dim)
     """
 
-    assert _in_dygraph_mode(
+    assert in_dygraph_mode(
     ) is not True, "please use gru instead of dynamic_gru in dygraph mode!"
 
     helper = LayerHelper('gru', **locals())
@@ -1760,7 +1760,7 @@ def sequence_conv(input,
         Variable: output of sequence_conv
     """
 
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper('sequence_conv', **locals())
     dtype = helper.input_dtype()
@@ -1821,7 +1821,7 @@ def sequence_softmax(input, use_cudnn=False, name=None):
                               dtype='float32', lod_level=1)
              x_sequence_softmax = fluid.layers.sequence_softmax(input=x)
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper('sequence_softmax', **locals())
     dtype = helper.input_dtype()
@@ -2315,7 +2315,7 @@ def sequence_pool(input, pool_type, is_test=False):
              last_x = fluid.layers.sequence_pool(input=x, pool_type='last')
              first_x = fluid.layers.sequence_pool(input=x, pool_type='first')
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper('sequence_pool', **locals())
     dtype = helper.input_dtype()
@@ -2356,7 +2356,7 @@ def sequence_concat(input, name=None):
 
            out = fluid.layers.sequence_concat(input=[seq1, seq2, seq3])
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper('sequence_concat', **locals())
     out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
@@ -2485,7 +2485,7 @@ def sequence_slice(input, offset, length, name=None):
              subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset,
                                                    length=length)
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper("sequence_slice", **locals())
     dtype = helper.input_dtype()
@@ -3307,7 +3307,7 @@ def layer_norm(input,
         >>>                          dtype='float32')
         >>> x = fluid.layers.layer_norm(input=data, begin_norm_axis=1)
     """
-    assert _in_dygraph_mode(
+    assert in_dygraph_mode(
     ) is not True, "please use FC instead of fc in dygraph mode!"
     helper = LayerHelper('layer_norm', **locals())
     dtype = helper.input_dtype()
@@ -3946,7 +3946,7 @@ def sequence_expand(x, y, ref_level=-1, name=None):
                              dtype='float32', lod_level=1)
             out = layers.sequence_expand(x=x, y=y, ref_level=0)
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper('sequence_expand', input=x, **locals())
     dtype = helper.input_dtype()
@@ -4014,7 +4014,7 @@ def sequence_expand_as(x, y, name=None):
                              dtype='float32', lod_level=1)
             out = layers.sequence_expand_as(x=x, y=y)
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper('sequence_expand_as', input=x, **locals())
     dtype = helper.input_dtype()
@@ -4062,7 +4062,7 @@ def sequence_pad(x, pad_value, maxlen=None, name=None):
             out = fluid.layers.sequence_pad(x=x, pad_value=pad_value)
     """
 
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper('sequence_pad', input=x, **locals())
     dtype = helper.input_dtype()
@@ -4130,7 +4130,7 @@ def sequence_unpad(x, length, name=None):
             out = fluid.layers.sequence_unpad(x=x, length=len)
     """
 
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper('sequence_unpad', input=x, **locals())
     dtype = helper.input_dtype()
@@ -5305,7 +5305,7 @@ def sequence_reshape(input, new_dim):
             x = fluid.layers.data(shape=[5, 20], dtype='float32', lod_level=1)
             x_reshaped = fluid.layers.sequence_reshape(input=x, new_dim=10)
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper('sequence_reshape', **locals())
     out = helper.create_variable_for_type_inference(helper.input_dtype())
@@ -5841,7 +5841,7 @@ def im2sequence(input,
                 input=layer, stride=[1, 1], filter_size=[2, 2])
 
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
 
     if isinstance(filter_size, int):
@@ -6485,7 +6485,7 @@ def squeeze(input, axes, name=None):
             x = layers.data(name='x', shape=[5, 1, 10])
             y = layers.sequeeze(input=x, axes=[1])
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "squeeze layer is not supported in dygraph mode yet.")
     helper = LayerHelper("squeeze", **locals())
     out = helper.create_variable_for_type_inference(dtype=input.dtype)
@@ -7138,10 +7138,10 @@ def image_resize(input,
         out_shape(list|tuple|Variable|None): Output shape of image resize
                                     layer, the shape is (out_h, out_w).
                                     Default: None
-        scale(float|None): The multiplier for the input height or width.
-                         At least one of out_shape or scale must be set.
-                         And out_shape has a higher priority than scale.
-                         Default: None
+        scale(float|None): The multiplier for the input height or width. At
+             least one of :attr:`out_shape` or :attr:`scale` must be set. 
+             And :attr:`out_shape` has a higher priority than :attr:`scale`. 
+             Default: None.
         name(str|None): A name for this layer(optional). If set None, the layer
                         will be named automatically.
         resample(str): The resample method. It supports 'BILINEAR' and 'NEAREST'
@@ -7179,6 +7179,7 @@ def image_resize(input,
                     or 'NEAREST' currently.
         ValueError: One of out_shape and scale must not be None.
         ValueError: out_shape length should be 2.
+        ValueError: scale should be greater than zero.
         TypeError: align_corners shoule be a bool value
         ValueError: align_mode can only be '0' or '1'
 
@@ -7210,26 +7211,36 @@ def image_resize(input,
     def _is_list_or_turple_(data):
         return (isinstance(data, list) or isinstance(data, tuple))
 
-    out_h = 0
-    out_w = 0
     inputs = {"X": input}
+    attrs = {
+        "out_h": 0,
+        "out_w": 0,
+        "interp_method": resample_type,
+        "align_corners": align_corners,
+        "align_mode": align_mode
+    }
+
     if out_shape is not None:
         if isinstance(out_shape, Variable):
             warnings.warn("out_shape as Variable type is deprecated, \
                     it is recommended to use actual_shape instead of \
                     out_shape to specify output shape dynamically.")
             inputs['OutSize'] = out_shape
-        elif not (_is_list_or_turple_(out_shape)):
-            raise TypeError("out_shape should be a list or tuple or Variable.")
-        elif len(out_shape) != 2:
-            raise ValueError("out_shape length should be 2.")
-
-        out_shape = list(map(int, out_shape))
-        out_h = out_shape[0]
-        out_w = out_shape[1]
+        else:
+            if not (_is_list_or_turple_(out_shape)):
+                raise TypeError(
+                    "out_shape should be a list or tuple or Variable.")
+            if len(out_shape) != 2:
+                raise ValueError("out_shape length should be 2.")
+
+            out_shape = list(map(int, out_shape))
+            attrs['out_h'] = out_shape[0]
+            attrs['out_w'] = out_shape[1]
+
     else:
-        out_h = int(input.shape[2] * scale)
-        out_w = int(input.shape[3] * scale)
+        if scale <= 0:
+            raise ValueError("scale should be greater than zero.")
+        attrs['scale'] = float(scale)
 
     if isinstance(actual_shape, Variable):
         inputs["OutSize"] = actual_shape
@@ -7241,13 +7252,7 @@ def image_resize(input,
         type='{}_interp'.format(resample_type),
         inputs=inputs,
         outputs={"Out": out},
-        attrs={
-            "out_h": out_h,
-            "out_w": out_w,
-            "interp_method": resample_type,
-            "align_corners": align_corners,
-            "align_mode": align_mode
-        })
+        attrs=attrs)
     return out
 
 
@@ -7315,11 +7320,14 @@ def resize_bilinear(input,
     Args:
         input(${x_type}): ${x_comment}.
 
-        out_shape(${out_size_type}): ${out_size_comment}.
+        out_shape(list|tuple|Variable|None): Output shape of resize bilinear
+                                    layer, the shape is (out_h, out_w).
+                                    Default: None
 
         scale(float|None): The multiplier for the input height or width. At
-             least one of out_shape or scale must be set. And out_shape has
-             a higher priority than scale. Default: None.
+             least one of :attr:`out_shape` or :attr:`scale` must be set. 
+             And :attr:`out_shape` has a higher priority than :attr:`scale`. 
+             Default: None.
 
         name(str|None): The output variable name.
         actual_shape(Variable): An optional input to specify output shape
@@ -7406,11 +7414,14 @@ def resize_nearest(input,
     Args:
         input(${x_type}): ${x_comment}.
 
-        out_shape(${out_size_type}): ${out_size_comment}.
+        out_shape(list|tuple|Variable|None): Output shape of resize nearest
+                                    layer, the shape is (out_h, out_w).
+                                    Default: None
 
         scale(float|None): The multiplier for the input height or width. At
-             least one of out_shape or scale must be set. And out_shape has
-             a higher priority than scale. Default: None.
+             least one of :attr:`out_shape` or :attr:`scale` must be set. 
+             And :attr:`out_shape` has a higher priority than :attr:`scale`. 
+             Default: None.
 
         name(str|None): The output variable name.
         actual_shape(Variable): An optional input to specify output shape
@@ -7620,7 +7631,7 @@ def sequence_scatter(input, index, updates, name=None):
             output = fluid.layers.sequence_scatter(input, index, updates)
 
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper('sequence_scatter', **locals())
     dtype = helper.input_dtype()
@@ -8710,7 +8721,7 @@ def sequence_enumerate(input, win_size, pad_value=0, name=None):
             x = fluid.layers.data(shape[30, 1], dtype='int32', lod_level=1)
             out = fluid.layers.sequence_enumerate(input=x, win_size=3, pad_value=0)
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper('sequence_enumerate', **locals())
     out = helper.create_variable_for_type_inference(
@@ -8751,7 +8762,7 @@ def sequence_mask(x, maxlen=None, dtype='int64', name=None):
         Variable: The output sequence mask.
 
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
 
     helper = LayerHelper('sequence_mask', **locals())
@@ -9230,7 +9241,7 @@ def _elementwise_op(helper):
     op_type = helper.layer_type
     x = helper.kwargs.get('x', None)
     y = helper.kwargs.get('y', None)
-    if _in_dygraph_mode():
+    if in_dygraph_mode():
         x = base.to_variable(x)
         y = base.to_variable(y)
 
@@ -9803,7 +9814,7 @@ def sequence_reverse(x, name=None):
     Returns:
         out(${y_type}): ${y_comment}
     """
-    assert not _in_dygraph_mode(), (
+    assert not in_dygraph_mode(), (
         "sequence layer is not supported in dygraph mode yet.")
     helper = LayerHelper("sequence_reverse", **locals())
     if name is None:

python/paddle/fluid/optimizer.py

@@ -55,7 +55,7 @@ class Optimizer(object):
     """
 
     def __init__(self, learning_rate, regularization=None, name=None):
-        if framework._in_dygraph_mode():
+        if framework.in_dygraph_mode():
             if not isinstance(learning_rate, float) and \
                     not isinstance(learning_rate, LearningRateDecay):
                 raise TypeError(
@@ -205,7 +205,7 @@ class Optimizer(object):
             name = self._name + "_" + name
         if (name in self._accumulators and
                 param.name in self._accumulators[name]):
-            if framework._in_dygraph_mode():
+            if framework.in_dygraph_mode():
                 return self._accumulators[name][param.name]
             raise Exception("Accumulator {} already exists for parameter {}".
                             format(name, param.name))
@@ -374,7 +374,7 @@ class Optimizer(object):
             See examples in `apply_gradients`.
         """
         self._dtype = loss.dtype
-        if framework._in_dygraph_mode():
+        if framework.in_dygraph_mode():
             if parameter_list is not None:
                 parameters = parameter_list
             else:
@@ -459,7 +459,7 @@ class Optimizer(object):
         Returns:
             list: A list of operators appended to the current program.
         """
-        if framework._in_dygraph_mode():
+        if framework.in_dygraph_mode():
             with program_guard(framework.default_main_program(),
                                framework.default_startup_program()):
                 optimize_ops = self._create_optimization_pass(params_grads)
@@ -639,16 +639,16 @@ class DGCMomentumOptimizer(MomentumOptimizer):
 
     Original paper is https://arxiv.org/abs/1712.01887
 
-    DGC reduce the communication bandwidth by sending only the important gradients (sparse update):\
+    DGC reduces the communication bandwidth by sending only the important gradients (sparse update):\
         only gradients larger than a threshold are transmitted.
 
-    To avoid losing information, DGC accumulate the rest of the gradients locally.
+    To avoid losing information, DGC accumulates the rest of the gradients locally.
 
     Eventually, these gradients become large enough to be transmitted.
 
-    Thus, DGC send the large gradients immediately but eventually send all of the gradients over time.
+    Thus, DGC sends the large gradients immediately but eventually send all of the gradients over time.
 
-    To ensure no loss of accuracy, DGC employs momentum correc-tionandlocal gradient clipping on top of the gradient sparsification to maintain model performance.
+    To ensure no loss of accuracy, DGC employs momentum correction and local gradient clipping on top of the gradient sparsification to maintain model performance.
 
     DGC also uses momentum factor masking and warmup training to overcome the staleness problem caused by reduced communication.
 
@@ -663,7 +663,7 @@ class DGCMomentumOptimizer(MomentumOptimizer):
         learning_rate (float|Variable): the learning rate used to update parameters. \
             Can be a float value or a Variable with one float value as data element.
         momentum (float): Momentum factor.
-        rampup_begin_step (int): The begining step from which gradient compression is implemented.
+        rampup_begin_step (int): The beginning step from which gradient compression is implemented.
         rampup_step (int): How long it use the sparsity periods. Default is 1.
             for example: If the sparsity is [0.75, 0.9375, 0.984375, 0.996, 0.999], and the rampup_step is 5, \
                 it will use 0.75 at 0 step, and 0.9375 at 1 step, and so on. And when reach sparsity array ends, \
@@ -671,9 +671,9 @@ class DGCMomentumOptimizer(MomentumOptimizer):
         sparsity (list[float]): Get top important element from gradient tensor, the ratio is (1 - current sparsity).
         use_nesterov (bool): Enables Nesterov momentum. True means use nesterov.
         local_grad_clip_norm (float): Clip norm value if needed.
-        num_trainers: The number of training node.
+        num_trainers: The number of training nodes.
         regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer.
-        name: A optional name prefix.
+        name: An optional name prefix.
 
     Examples:
         .. code-block:: python

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

+# Copyright (c) 2019 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.
+
+import numpy as np
+import six
+
+
+def fake_imdb_reader(word_dict_size,
+                     sample_num,
+                     lower_seq_len=100,
+                     upper_seq_len=200,
+                     class_dim=2):
+    def __reader__():
+        for _ in six.moves.range(sample_num):
+            length = np.random.random_integers(
+                low=lower_seq_len, high=upper_seq_len, size=[1])[0]
+            ids = np.random.random_integers(
+                low=0, high=word_dict_size - 1, size=[length]).astype('int64')
+            label = np.random.random_integers(
+                low=0, high=class_dim - 1, size=[1]).astype('int64')[0]
+            yield ids, label
+
+    return __reader__

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

@@ -18,7 +18,7 @@ import numpy as np
 import paddle.fluid as fluid
 
 
-class L1(fluid.dygraph.Layer):
+class L1(fluid.Layer):
     def __init__(self, prefix):
         super(L1, self).__init__(prefix)
         self._param_attr = fluid.ParamAttr(
@@ -32,7 +32,7 @@ class L1(fluid.dygraph.Layer):
         return self.w1 + self.w2
 
 
-class L2(fluid.dygraph.Layer):
+class L2(fluid.Layer):
     def __init__(self, prefix):
         super(L2, self).__init__(prefix)
         self.layer1 = L1(self.full_name())
@@ -42,7 +42,7 @@ class L2(fluid.dygraph.Layer):
         return self.layer1() + self.layer2()
 
 
-class L3(fluid.dygraph.Layer):
+class L3(fluid.Layer):
     def __init__(self, prefix):
         super(L3, self).__init__(prefix)
         self.layer1 = L2(self.full_name())
@@ -59,7 +59,7 @@ class TestBaseLayer(unittest.TestCase):
             ret = l()
             self.assertEqual(l.w1.name, "test_one_level/L1_0.w_0")
             self.assertEqual(l.w2.name, "test_one_level/L1_0.w_1")
-            self.assertTrue(np.allclose(ret._numpy(), 0.2 * np.ones([2, 2])))
+            self.assertTrue(np.allclose(ret.numpy(), 0.2 * np.ones([2, 2])))
 
     def test_three_level(self):
         with fluid.dygraph.guard():
@@ -72,7 +72,7 @@ class TestBaseLayer(unittest.TestCase):
             self.assertEqual(names[3], "test_three_level/L3_0/L2_0/L1_1.w_1")
             self.assertEqual(names[4], "test_three_level/L3_0/L2_1/L1_0.w_0")
             self.assertEqual(names[5], "test_three_level/L3_0/L2_1/L1_0.w_1")
-            self.assertTrue(np.allclose(ret._numpy(), 0.8 * np.ones([2, 2])))
+            self.assertTrue(np.allclose(ret.numpy(), 0.8 * np.ones([2, 2])))
 
 
 if __name__ == '__main__':

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

@@ -91,17 +91,26 @@ class TestBilinearInterpOp(OpTest):
         self.op_type = "bilinear_interp"
         input_np = np.random.random(self.input_shape).astype("float32")
 
-        output_np = bilinear_interp_np(input_np, self.out_h, self.out_w,
-                                       self.out_size, self.actual_shape,
-                                       self.align_corners, self.align_mode)
+        if self.scale > 0:
+            out_h = int(self.input_shape[2] * self.scale)
+            out_w = int(self.input_shape[3] * self.scale)
+        else:
+            out_h = self.out_h
+            out_w = self.out_w
+
+        output_np = bilinear_interp_np(input_np, out_h, out_w, self.out_size,
+                                       self.actual_shape, self.align_corners,
+                                       self.align_mode)
         self.inputs = {'X': input_np}
         if self.out_size is not None:
             self.inputs['OutSize'] = self.out_size
         if self.actual_shape is not None:
             self.inputs['OutSize'] = self.actual_shape
+
         self.attrs = {
             'out_h': self.out_h,
             'out_w': self.out_w,
+            'scale': self.scale,
             'interp_method': self.interp_method,
             'align_corners': self.align_corners,
             'align_mode': self.align_mode
@@ -119,6 +128,7 @@ class TestBilinearInterpOp(OpTest):
         self.input_shape = [2, 3, 4, 4]
         self.out_h = 2
         self.out_w = 2
+        self.scale = 0.
         self.out_size = np.array([3, 3]).astype("int32")
         self.align_corners = True
         self.align_mode = 1
@@ -130,6 +140,7 @@ class TestBilinearInterpCase1(TestBilinearInterpOp):
         self.input_shape = [4, 1, 7, 8]
         self.out_h = 1
         self.out_w = 1
+        self.scale = 0.
         self.align_corners = True
         self.align_mode = 1
 
@@ -140,6 +151,7 @@ class TestBilinearInterpCase2(TestBilinearInterpOp):
         self.input_shape = [3, 3, 9, 6]
         self.out_h = 12
         self.out_w = 12
+        self.scale = 0.
         self.align_corners = True
         self.align_mode = 1
 
@@ -150,6 +162,7 @@ class TestBilinearInterpCase3(TestBilinearInterpOp):
         self.input_shape = [1, 1, 128, 64]
         self.out_h = 64
         self.out_w = 128
+        self.scale = 0.
         self.align_corners = True
         self.align_mode = 1
 
@@ -160,6 +173,7 @@ class TestBilinearInterpCase4(TestBilinearInterpOp):
         self.input_shape = [4, 1, 7, 8]
         self.out_h = 1
         self.out_w = 1
+        self.scale = 0.
         self.out_size = np.array([2, 2]).astype("int32")
         self.align_corners = True
         self.align_mode = 1
@@ -171,6 +185,7 @@ class TestBilinearInterpCase5(TestBilinearInterpOp):
         self.input_shape = [3, 3, 9, 6]
         self.out_h = 12
         self.out_w = 12
+        self.scale = 0.
         self.out_size = np.array([11, 11]).astype("int32")
         self.align_corners = True
         self.align_mode = 1
@@ -182,6 +197,7 @@ class TestBilinearInterpCase6(TestBilinearInterpOp):
         self.input_shape = [1, 1, 128, 64]
         self.out_h = 64
         self.out_w = 128
+        self.scale = 0.
         self.out_size = np.array([65, 129]).astype("int32")
         self.align_corners = True
         self.align_mode = 1
@@ -193,6 +209,7 @@ class TestBilinearInterpActualShape(TestBilinearInterpOp):
         self.input_shape = [3, 2, 32, 16]
         self.out_h = 64
         self.out_w = 32
+        self.scale = 0.
         self.out_size = np.array([66, 40]).astype("int32")
         self.align_corners = True
         self.align_mode = 1
@@ -206,15 +223,25 @@ class TestBilinearInterpOpUint8(OpTest):
         self.op_type = "bilinear_interp"
         input_np = np.random.randint(
             low=0, high=256, size=self.input_shape).astype("uint8")
-        output_np = bilinear_interp_np(input_np, self.out_h, self.out_w,
-                                       self.out_size, self.actual_shape,
-                                       self.align_corners, self.align_mode)
+
+        if self.scale > 0:
+            out_h = int(self.input_shape[2] * self.scale)
+            out_w = int(self.input_shape[3] * self.scale)
+        else:
+            out_h = self.out_h
+            out_w = self.out_w
+
+        output_np = bilinear_interp_np(input_np, out_h, out_w, self.out_size,
+                                       self.actual_shape, self.align_corners,
+                                       self.align_mode)
         self.inputs = {'X': input_np}
         if self.out_size is not None:
             self.inputs['OutSize'] = self.out_size
+
         self.attrs = {
             'out_h': self.out_h,
             'out_w': self.out_w,
+            'scale': self.scale,
             'interp_method': self.interp_method,
             'align_corners': self.align_corners,
             'align_mode': self.align_mode
@@ -229,6 +256,7 @@ class TestBilinearInterpOpUint8(OpTest):
         self.input_shape = [1, 3, 9, 6]
         self.out_h = 10
         self.out_w = 9
+        self.scale = 0.
         self.align_corners = True
         self.align_mode = 1
 
@@ -239,6 +267,7 @@ class TestBilinearInterpCase1Uint8(TestBilinearInterpOpUint8):
         self.input_shape = [2, 3, 128, 64]
         self.out_h = 120
         self.out_w = 50
+        self.scale = 0.
         self.align_corners = True
         self.align_mode = 1
 
@@ -249,6 +278,7 @@ class TestBilinearInterpCase2Uint8(TestBilinearInterpOpUint8):
         self.input_shape = [4, 1, 7, 8]
         self.out_h = 5
         self.out_w = 13
+        self.scale = 0.
         self.out_size = np.array([6, 15]).astype("int32")
         self.align_corners = True
         self.align_mode = 1
@@ -272,5 +302,38 @@ class TestBilinearInterpWithMethod3(TestBilinearInterpOp):
         self.align_mode = 0
 
 
+class TestBilinearInterpScale1(TestBilinearInterpOp):
+    def init_test_case(self):
+        self.interp_method = 'bilinear'
+        self.input_shape = [2, 3, 16, 32]
+        self.out_h = 60
+        self.out_w = 25
+        self.scale = 2.
+        self.align_corners = True
+        self.align_mode = 1
+
+
+class TestBilinearInterpScale2(TestBilinearInterpOp):
+    def init_test_case(self):
+        self.interp_method = 'bilinear'
+        self.input_shape = [2, 3, 16, 32]
+        self.out_h = 60
+        self.out_w = 25
+        self.scale = 1.
+        self.align_corners = True
+        self.align_mode = 1
+
+
+class TestBilinearInterpScale3(TestBilinearInterpOp):
+    def init_test_case(self):
+        self.interp_method = 'bilinear'
+        self.input_shape = [2, 3, 16, 32]
+        self.out_h = 60
+        self.out_w = 25
+        self.scale = 1.5
+        self.align_corners = True
+        self.align_mode = 1
+
+
 if __name__ == "__main__":
     unittest.main()

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

@@ -19,7 +19,7 @@ import time
 import six
 import unittest
 
-EPOCH_NUM = 60
+EPOCH_NUM = 20
 BATCH_SIZE = 32
 CLASS_NUM = 10
 

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

@@ -22,6 +22,8 @@ import paddle
 import paddle.fluid.core as core
 import paddle.fluid as fluid
 from paddle.fluid import compiler
+import numpy as np
+from fake_reader import fake_imdb_reader
 
 
 def train(network, use_cuda, use_parallel_executor, batch_size=32, pass_num=2):
@@ -35,16 +37,16 @@ def train(network, use_cuda, use_parallel_executor, batch_size=32, pass_num=2):
         )
         return
 
-    word_dict = paddle.dataset.imdb.word_dict()
-    train_reader = paddle.batch(
-        paddle.dataset.imdb.train(word_dict), batch_size=batch_size)
+    word_dict_size = 5147
+    reader = fake_imdb_reader(word_dict_size, batch_size * 40)
+    train_reader = paddle.batch(reader, batch_size=batch_size)
 
     data = fluid.layers.data(
         name="words", shape=[1], dtype="int64", lod_level=1)
 
     label = fluid.layers.data(name="label", shape=[1], dtype="int64")
 
-    cost = network(data, label, len(word_dict))
+    cost = network(data, label, word_dict_size)
     cost.persistable = True
     optimizer = fluid.optimizer.Adagrad(learning_rate=0.2)
     optimizer.minimize(cost)

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

@@ -19,6 +19,8 @@ import numpy as np
 import paddle
 import paddle.fluid.core as core
 import paddle.fluid as fluid
+import six
+from fake_reader import fake_imdb_reader
 
 
 def bow_net(data,
@@ -48,11 +50,10 @@ def bow_net(data,
 
 class TestGradientClip(unittest.TestCase):
     def setUp(self):
-        self.word_dict = paddle.dataset.imdb.word_dict()
+        self.word_dict_len = 5147
         self.BATCH_SIZE = 2
-        self.train_data = paddle.batch(
-            paddle.dataset.imdb.train(self.word_dict),
-            batch_size=self.BATCH_SIZE)
+        reader = fake_imdb_reader(self.word_dict_len, self.BATCH_SIZE * 100)
+        self.train_data = paddle.batch(reader, batch_size=self.BATCH_SIZE)
 
     def get_places(self):
         places = [core.CPUPlace()]
@@ -131,7 +132,7 @@ class TestGradientClip(unittest.TestCase):
             data = fluid.layers.data(
                 name="words", shape=[1], dtype="int64", lod_level=1)
             label = fluid.layers.data(name="label", shape=[1], dtype="int64")
-            cost = bow_net(data, label, len(self.word_dict))
+            cost = bow_net(data, label, self.word_dict_len)
 
             fluid.clip.set_gradient_clip(
                 clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=5.0))

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

@@ -18,11 +18,11 @@ import numpy as np
 
 import paddle.fluid as fluid
 from paddle.fluid import core
-from paddle.fluid.dygraph.nn import FC
+from paddle.fluid import FC
 from test_imperative_base import new_program_scope
 
 
-class MyLayer(fluid.dygraph.Layer):
+class MyLayer(fluid.Layer):
     def __init__(self, name_scope):
         super(MyLayer, self).__init__(name_scope)
 
@@ -34,7 +34,7 @@ class MyLayer(fluid.dygraph.Layer):
         return [x]
 
 
-class MyPyLayer(fluid.dygraph.PyLayer):
+class MyPyLayer(fluid.PyLayer):
     def __init__(self):
         super(MyPyLayer, self).__init__()
 
@@ -48,7 +48,7 @@ class MyPyLayer(fluid.dygraph.PyLayer):
         return np.array(dout) * (1 - np.square(np.array(out)))
 
 
-class MLP(fluid.dygraph.Layer):
+class MLP(fluid.Layer):
     def __init__(self, name_scope):
         super(MLP, self).__init__(name_scope)
         self._fc1 = FC(self.full_name(),
@@ -71,7 +71,7 @@ class MLP(fluid.dygraph.Layer):
         return x
 
 
-class SimpleRNNCell(fluid.dygraph.Layer):
+class SimpleRNNCell(fluid.Layer):
     def __init__(self, name_scope, step_input_size, hidden_size, output_size,
                  param_attr):
         super(SimpleRNNCell, self).__init__(name_scope)
@@ -81,7 +81,7 @@ class SimpleRNNCell(fluid.dygraph.Layer):
         self._dtype = core.VarDesc.VarType.FP32
         self.param_attr = param_attr
 
-    def _build_once(self, inputs, pre_hidden):
+    def build_once(self, inputs, pre_hidden):
         i2h_param_shape = [self.step_input_size, self.hidden_size]
         h2h_param_shape = [self.hidden_size, self.hidden_size]
         h2o_param_shape = [self.output_size, self.hidden_size]
@@ -159,7 +159,7 @@ class SimpleRNNCell(fluid.dygraph.Layer):
         return reduce_out, hidden
 
 
-class SimpleRNN(fluid.dygraph.Layer):
+class SimpleRNN(fluid.Layer):
     def __init__(self, name_scope):
         super(SimpleRNN, self).__init__(name_scope)
         self.seq_len = 4
@@ -200,22 +200,22 @@ class TestImperative(unittest.TestCase):
                 inputs.append(fluid.dygraph.base.to_variable(x))
             ret = fluid.layers.sums(inputs)
             loss = fluid.layers.reduce_sum(ret)
-            loss._backward()
-            self.assertTrue(np.allclose(ret._numpy(), x * 10))
-            self.assertTrue(np.allclose(inputs[0]._gradient(), x))
+            loss.backward()
+            self.assertTrue(np.allclose(ret.numpy(), x * 10))
+            self.assertTrue(np.allclose(inputs[0].gradient(), x))
 
     def test_layer(self):
         with fluid.dygraph.guard():
             cl = core.Layer()
             cl.forward([])
-            l = fluid.dygraph.Layer("l")
+            l = fluid.Layer("l")
             self.assertRaises(NotImplementedError, l.forward, [])
 
     def test_pylayer_func_id(self):
 
         with fluid.dygraph.guard():
 
-            class PyLayer1(fluid.dygraph.PyLayer):
+            class PyLayer1(fluid.PyLayer):
                 def __init__(self):
                     super(PyLayer1, self).__init__()
 
@@ -227,7 +227,7 @@ class TestImperative(unittest.TestCase):
                 def backward(input):
                     return input
 
-            class PyLayer2(fluid.dygraph.PyLayer):
+            class PyLayer2(fluid.PyLayer):
                 def __init__(self):
                     super(PyLayer2, self).__init__()
 
@@ -257,9 +257,9 @@ class TestImperative(unittest.TestCase):
             my_py_layer = MyPyLayer()
             var_inp = fluid.dygraph.base.to_variable(np_inp)
             outs = my_py_layer(var_inp)
-            dy_out = np.sum(outs[0]._numpy())
-            outs[0]._backward()
-            dy_grad = var_inp._gradient()
+            dy_out = np.sum(outs[0].numpy())
+            outs[0].backward()
+            dy_grad = var_inp.gradient()
 
         with new_program_scope():
             inp = fluid.layers.data(
@@ -287,9 +287,9 @@ class TestImperative(unittest.TestCase):
             l = MyLayer("my_layer")
             x = l(var_inp)[0]
             self.assertIsNotNone(x)
-            dy_out = x._numpy()
-            x._backward()
-            dy_grad = l._x_for_debug._gradient()
+            dy_out = x.numpy()
+            x.backward()
+            dy_grad = l._x_for_debug.gradient()
 
         with new_program_scope():
             inp = fluid.layers.data(
@@ -314,9 +314,9 @@ class TestImperative(unittest.TestCase):
             var_inp = fluid.dygraph.base.to_variable(np_inp)
             mlp = MLP("mlp")
             out = mlp(var_inp)
-            dy_out = out._numpy()
-            out._backward()
-            dy_grad = mlp._fc1._w._gradient()
+            dy_out = out.numpy()
+            out.backward()
+            dy_grad = mlp._fc1._w.gradient()
 
         with new_program_scope():
             inp = fluid.layers.data(
@@ -358,7 +358,7 @@ class TestImperative(unittest.TestCase):
                 x = fluid.layers.elementwise_add(inp1, inp2)
             else:
                 x = fluid.layers.elementwise_sub(inp1, inp2)
-            dygraph_result = x._numpy()
+            dygraph_result = x.numpy()
 
         # static graph
         with new_program_scope():
@@ -407,11 +407,11 @@ class TestImperative(unittest.TestCase):
             var_inp = fluid.layers.reshape(var_inp, shape=[1, 4, 3])
             simple_rnn = SimpleRNN("simple_rnn")
             outs, pre_hiddens = simple_rnn.forward(var_inp)
-            dy_out = outs[3]._numpy()
-            outs[3]._backward()
-            dy_grad_h2o = simple_rnn._cell._h2o_w._gradient()
-            dy_grad_h2h = simple_rnn._cell._h2h_w._gradient()
-            dy_grad_i2h = simple_rnn._cell._i2h_w._gradient()
+            dy_out = outs[3].numpy()
+            outs[3].backward()
+            dy_grad_h2o = simple_rnn._cell._h2o_w.gradient()
+            dy_grad_h2h = simple_rnn._cell._h2h_w.gradient()
+            dy_grad_i2h = simple_rnn._cell._i2h_w.gradient()
 
         with new_program_scope():
             inp = fluid.layers.data(

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

@@ -18,11 +18,11 @@ import numpy as np
 import paddle
 import paddle.fluid as fluid
 from paddle.fluid.optimizer import SGDOptimizer
-from paddle.fluid.dygraph.nn import Conv2D, Pool2D, FC
+from paddle.fluid import Conv2D, Pool2D, FC
 from paddle.fluid.dygraph.base import to_variable
 
 
-class SimpleImgConvPool(fluid.dygraph.Layer):
+class SimpleImgConvPool(fluid.Layer):
     def __init__(self,
                  name_scope,
                  num_channels,
@@ -71,7 +71,7 @@ class SimpleImgConvPool(fluid.dygraph.Layer):
         return x
 
 
-class MNIST(fluid.dygraph.Layer):
+class MNIST(fluid.Layer):
     def __init__(self, name_scope):
         super(MNIST, self).__init__(name_scope)
 
@@ -125,21 +125,21 @@ class TestDygraphCheckpoint(unittest.TestCase):
 
                     img = to_variable(dy_x_data)
                     label = to_variable(y_data)
-                    label._stop_gradient = True
+                    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()
+                    dy_out = avg_loss.numpy()
 
-                    avg_loss._backward()
+                    avg_loss.backward()
                     sgd.minimize(avg_loss)
                     fluid.dygraph.save_persistables(mnist, "save_dir")
                     mnist.clear_gradients()
 
                     for param in mnist.parameters():
-                        dy_param_init_value[param.name] = param._numpy()
+                        dy_param_init_value[param.name] = param.numpy()
 
                     mnist.load_dict(
                         fluid.dygraph.load_persistables(mnist, "save_dir"))

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

@@ -32,11 +32,11 @@ NUM_BATCHES = int(os.environ.get('NUM_BATCHES', 5))
 NUM_EPOCHES = int(os.environ.get('NUM_EPOCHES', 1))
 
 
-class DMF(fluid.dygraph.Layer):
+class DMF(fluid.Layer):
     def __init__(self, name_scope):
         super(DMF, self).__init__(name_scope)
-        self._user_latent = fluid.dygraph.FC(self.full_name(), 256)
-        self._item_latent = fluid.dygraph.FC(self.full_name(), 256)
+        self._user_latent = fluid.FC(self.full_name(), 256)
+        self._item_latent = fluid.FC(self.full_name(), 256)
 
         self._user_layers = []
         self._item_layers = []
@@ -45,13 +45,11 @@ class DMF(fluid.dygraph.Layer):
             self._user_layers.append(
                 self.add_sublayer(
                     'user_layer_%d' % i,
-                    fluid.dygraph.FC(
-                        self.full_name(), self._hid_sizes[i], act='relu')))
+                    fluid.FC(self.full_name(), self._hid_sizes[i], act='relu')))
             self._item_layers.append(
                 self.add_sublayer(
                     'item_layer_%d' % i,
-                    fluid.dygraph.FC(
-                        self.full_name(), self._hid_sizes[i], act='relu')))
+                    fluid.FC(self.full_name(), self._hid_sizes[i], act='relu')))
 
     def forward(self, users, items):
         users = self._user_latent(users)
@@ -63,19 +61,18 @@ class DMF(fluid.dygraph.Layer):
         return fluid.layers.elementwise_mul(users, items)
 
 
-class MLP(fluid.dygraph.Layer):
+class MLP(fluid.Layer):
     def __init__(self, name_scope):
         super(MLP, self).__init__(name_scope)
-        self._user_latent = fluid.dygraph.FC(self.full_name(), 256)
-        self._item_latent = fluid.dygraph.FC(self.full_name(), 256)
+        self._user_latent = fluid.FC(self.full_name(), 256)
+        self._item_latent = fluid.FC(self.full_name(), 256)
         self._match_layers = []
         self._hid_sizes = [128, 64]
         for i in range(len(self._hid_sizes)):
             self._match_layers.append(
                 self.add_sublayer(
                     'match_layer_%d' % i,
-                    fluid.dygraph.FC(
-                        self.full_name(), self._hid_sizes[i], act='relu')))
+                    fluid.FC(self.full_name(), self._hid_sizes[i], act='relu')))
         self._mat
 
     def forward(self, users, items):
@@ -88,7 +85,7 @@ class MLP(fluid.dygraph.Layer):
         return match_vec
 
 
-class DeepCF(fluid.dygraph.Layer):
+class DeepCF(fluid.Layer):
     def __init__(self, name_scope, num_users, num_items, matrix):
         super(DeepCF, self).__init__(name_scope)
         self._num_users = num_users
@@ -99,11 +96,11 @@ class DeepCF(fluid.dygraph.Layer):
             matrix.dtype,
             is_bias=False,
             default_initializer=fluid.initializer.NumpyArrayInitializer(matrix))
-        self._rating_matrix._stop_gradient = True
+        self._rating_matrix.stop_gradient = True
 
         self._mlp = MLP(self.full_name())
         self._dmf = DMF(self.full_name())
-        self._match_fc = fluid.dygraph.FC(self.full_name(), 1, act='sigmoid')
+        self._match_fc = fluid.FC(self.full_name(), 1, act='sigmoid')
 
     def forward(self, users, items):
         # users_emb = self._user_emb(users)
@@ -255,10 +252,10 @@ class TestDygraphDeepCF(unittest.TestCase):
                         fluid.layers.log_loss(prediction,
                                               to_variable(labels_np[
                                                   slice:slice + BATCH_SIZE])))
-                    loss._backward()
+                    loss.backward()
                     adam.minimize(loss)
                     deepcf.clear_gradients()
-                    dy_loss = loss._numpy()
+                    dy_loss = loss.numpy()
                     sys.stderr.write('dynamic loss: %s %s\n' % (slice, dy_loss))
 
         self.assertEqual(static_loss, dy_loss)

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

@@ -22,12 +22,12 @@ import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 from paddle.fluid.optimizer import SGDOptimizer
-from paddle.fluid.dygraph.nn import Conv2D, Pool2D, FC
+from paddle.fluid import Conv2D, Pool2D, FC
 from test_imperative_base import new_program_scope
 from paddle.fluid.dygraph.base import to_variable
 
 
-class Discriminator(fluid.dygraph.Layer):
+class Discriminator(fluid.Layer):
     def __init__(self, name_scope):
         super(Discriminator, self).__init__(name_scope)
         self._fc1 = FC(self.full_name(), size=32, act='elu')
@@ -38,7 +38,7 @@ class Discriminator(fluid.dygraph.Layer):
         return self._fc2(x)
 
 
-class Generator(fluid.dygraph.Layer):
+class Generator(fluid.Layer):
     def __init__(self, name_scope):
         super(Generator, self).__init__(name_scope)
         self._fc1 = FC(self.full_name(), size=64, act='elu')
@@ -150,7 +150,7 @@ class TestDygraphGAN(unittest.TestCase):
                     x=d_fake, label=to_variable(np.zeros([2, 1], np.float32))))
 
             d_loss = d_loss_real + d_loss_fake
-            d_loss._backward()
+            d_loss.backward()
             sgd.minimize(d_loss)
             discriminator.clear_gradients()
             generator.clear_gradients()
@@ -160,15 +160,15 @@ class TestDygraphGAN(unittest.TestCase):
             g_loss = fluid.layers.reduce_mean(
                 fluid.layers.sigmoid_cross_entropy_with_logits(
                     x=d_fake, label=to_variable(np.ones([2, 1], np.float32))))
-            g_loss._backward()
+            g_loss.backward()
             sgd.minimize(g_loss)
             for p in discriminator.parameters():
-                dy_params[p.name] = p._numpy()
+                dy_params[p.name] = p.numpy()
             for p in generator.parameters():
-                dy_params[p.name] = p._numpy()
+                dy_params[p.name] = p.numpy()
 
-            dy_g_loss = g_loss._numpy()
-            dy_d_loss = d_loss._numpy()
+            dy_g_loss = g_loss.numpy()
+            dy_d_loss = d_loss.numpy()
 
         self.assertEqual(dy_g_loss, static_g_loss)
         self.assertEqual(dy_d_loss, static_d_loss)

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

@@ -15,14 +15,12 @@
 import contextlib
 import unittest
 import numpy as np
-import six
 import sys
 
 import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 from paddle.fluid.optimizer import AdamOptimizer
-from paddle.fluid.dygraph.nn import Conv2D, Pool2D, FC
 from test_imperative_base import new_program_scope
 from paddle.fluid.dygraph.base import to_variable
 
@@ -31,7 +29,7 @@ def gen_data():
     pass
 
 
-class GraphConv(fluid.dygraph.Layer):
+class GraphConv(fluid.Layer):
     def __init__(self, name_scope, in_features, out_features):
         super(GraphConv, self).__init__(name_scope)
 
@@ -50,7 +48,7 @@ class GraphConv(fluid.dygraph.Layer):
         return fluid.layers.matmul(adj, support) + self.bias
 
 
-class GCN(fluid.dygraph.Layer):
+class GCN(fluid.Layer):
     def __init__(self, name_scope, num_hidden):
         super(GCN, self).__init__(name_scope)
         self.gc = GraphConv(self.full_name(), num_hidden, 32)
@@ -134,10 +132,9 @@ class TestDygraphGNN(unittest.TestCase):
             loss = fluid.layers.reduce_sum(loss)
             adam = AdamOptimizer(learning_rate=1e-3)
             adam.minimize(loss)
-            self.assertEqual(static_loss, loss._numpy())
-            self.assertTrue(
-                np.allclose(static_weight, model.gc.weight._numpy()))
-            sys.stderr.write('%s %s\n' % (static_loss, loss._numpy()))
+            self.assertEqual(static_loss, loss.numpy())
+            self.assertTrue(np.allclose(static_weight, model.gc.weight.numpy()))
+            sys.stderr.write('%s %s\n' % (static_loss, loss.numpy()))
 
 
 if __name__ == '__main__':

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

@@ -128,25 +128,25 @@ class TestImperativeMnist(unittest.TestCase):
 
                     img = to_variable(dy_x_data)
                     label = to_variable(y_data)
-                    label._stop_gradient = True
+                    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()
+                    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()
+                            dy_param_init_value[param.name] = param.numpy()
 
-                    avg_loss._backward()
+                    avg_loss.backward()
                     sgd.minimize(avg_loss)
                     mnist.clear_gradients()
 
                     dy_param_value = {}
                     for param in mnist.parameters():
-                        dy_param_value[param.name] = param._numpy()
+                        dy_param_value[param.name] = param.numpy()
 
         with new_program_scope():
             fluid.default_startup_program().random_seed = seed

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

@@ -28,7 +28,7 @@ from paddle.fluid.dygraph.base import to_variable
 from test_imperative_base import new_program_scope
 
 
-class MLP(fluid.dygraph.Layer):
+class MLP(fluid.Layer):
     def __init__(self, name_scope, param_attr=None, bias_attr=None):
         super(MLP, self).__init__(name_scope)
 
@@ -75,18 +75,18 @@ class TestImperativeOptimizerBase(unittest.TestCase):
 
                 cost = mlp(img)
                 avg_loss = fluid.layers.reduce_mean(cost)
-                dy_out = avg_loss._numpy()
+                dy_out = avg_loss.numpy()
 
                 if batch_id == 0:
                     for param in mlp.parameters():
-                        dy_param_init_value[param.name] = param._numpy()
+                        dy_param_init_value[param.name] = param.numpy()
 
-                avg_loss._backward()
+                avg_loss.backward()
                 optimizer.minimize(avg_loss)
                 mlp.clear_gradients()
                 dy_param_value = {}
                 for param in mlp.parameters():
-                    dy_param_value[param.name] = param._numpy()
+                    dy_param_value[param.name] = param.numpy()
 
         with new_program_scope():
             fluid.default_startup_program().random_seed = seed

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

@@ -24,10 +24,9 @@ from paddle.fluid.dygraph.base import to_variable
 from test_imperative_base import new_program_scope
 import numpy as np
 import six
-from paddle.fluid.backward import append_backward
 
 
-class SimpleLSTMRNN(fluid.dygraph.Layer):
+class SimpleLSTMRNN(fluid.Layer):
     def __init__(self,
                  name_scope,
                  hidden_size,
@@ -45,7 +44,7 @@ class SimpleLSTMRNN(fluid.dygraph.Layer):
         self.cell_array = []
         self.hidden_array = []
 
-    def _build_once(self, input_embedding, init_hidden=None, init_cell=None):
+    def build_once(self, input_embedding, init_hidden=None, init_cell=None):
         self.weight_1_arr = []
         self.weight_2_arr = []
         self.bias_arr = []
@@ -132,7 +131,7 @@ class SimpleLSTMRNN(fluid.dygraph.Layer):
         return real_res, last_hidden, last_cell
 
 
-class PtbModel(fluid.dygraph.Layer):
+class PtbModel(fluid.Layer):
     def __init__(self,
                  name_scope,
                  hidden_size,
@@ -177,7 +176,7 @@ class PtbModel(fluid.dygraph.Layer):
             default_initializer=fluid.initializer.UniformInitializer(
                 low=-self.init_scale, high=self.init_scale))
 
-    def _build_once(self, input, label, init_hidden, init_cell):
+    def build_once(self, input, label, init_hidden, init_cell):
         pass
 
     def forward(self, input, label, init_hidden, init_cell):
@@ -260,13 +259,13 @@ class TestDygraphPtbRnn(unittest.TestCase):
                                                             init_cell)
                 if i == 0:
                     for param in ptb_model.parameters():
-                        dy_param_init[param.name] = param._numpy()
-                dy_loss._backward()
+                        dy_param_init[param.name] = param.numpy()
+                dy_loss.backward()
                 sgd.minimize(dy_loss)
                 ptb_model.clear_gradients()
                 if i == batch_num - 1:
                     for param in ptb_model.parameters():
-                        dy_param_updated[param.name] = param._numpy()
+                        dy_param_updated[param.name] = param.numpy()
 
         with new_program_scope():
             fluid.default_startup_program().random_seed = seed
@@ -334,11 +333,11 @@ class TestDygraphPtbRnn(unittest.TestCase):
                         static_param_updated[static_param_name_list[k -
                                                                     3]] = out[k]
 
-        self.assertTrue(np.array_equal(static_loss_value, dy_loss._numpy()))
+        self.assertTrue(np.array_equal(static_loss_value, dy_loss.numpy()))
         self.assertTrue(
-            np.array_equal(static_last_cell_value, last_cell._numpy()))
+            np.array_equal(static_last_cell_value, last_cell.numpy()))
         self.assertTrue(
-            np.array_equal(static_last_hidden_value, last_hidden._numpy()))
+            np.array_equal(static_last_hidden_value, last_hidden.numpy()))
         for key, value in six.iteritems(static_param_init):
             self.assertTrue(np.array_equal(value, dy_param_init[key]))
         for key, value in six.iteritems(static_param_updated):

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

@@ -21,7 +21,7 @@ import paddle
 import paddle.fluid as fluid
 from paddle.fluid import core
 from paddle.fluid.layer_helper import LayerHelper
-from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, FC
+from paddle.fluid import Conv2D, Pool2D, BatchNorm, FC
 from paddle.fluid.dygraph.base import to_variable
 from test_imperative_base import new_program_scope
 
@@ -68,7 +68,7 @@ def optimizer_setting(params):
     return optimizer
 
 
-class ConvBNLayer(fluid.dygraph.Layer):
+class ConvBNLayer(fluid.Layer):
     def __init__(self,
                  name_scope,
                  num_channels,
@@ -99,7 +99,7 @@ class ConvBNLayer(fluid.dygraph.Layer):
         return y
 
 
-class BottleneckBlock(fluid.dygraph.Layer):
+class BottleneckBlock(fluid.Layer):
     def __init__(self,
                  name_scope,
                  num_channels,
@@ -156,7 +156,7 @@ class BottleneckBlock(fluid.dygraph.Layer):
         return layer_helper.append_activation(y)
 
 
-class ResNet(fluid.dygraph.Layer):
+class ResNet(fluid.Layer):
     def __init__(self, name_scope, layers=50, class_dim=102):
         super(ResNet, self).__init__(name_scope)
 
@@ -247,7 +247,7 @@ class TestDygraphResnet(unittest.TestCase):
 
             dy_param_init_value = {}
             for param in resnet.parameters():
-                dy_param_init_value[param.name] = param._numpy()
+                dy_param_init_value[param.name] = param.numpy()
 
             for batch_id, data in enumerate(train_reader()):
                 if batch_id >= batch_num:
@@ -260,20 +260,20 @@ class TestDygraphResnet(unittest.TestCase):
 
                 img = to_variable(dy_x_data)
                 label = to_variable(y_data)
-                label._stop_gradient = True
+                label.stop_gradient = True
 
                 out = resnet(img)
                 loss = fluid.layers.cross_entropy(input=out, label=label)
                 avg_loss = fluid.layers.mean(x=loss)
 
-                dy_out = avg_loss._numpy()
+                dy_out = avg_loss.numpy()
 
                 if batch_id == 0:
                     for param in resnet.parameters():
                         if param.name not in dy_param_init_value:
-                            dy_param_init_value[param.name] = param._numpy()
+                            dy_param_init_value[param.name] = param.numpy()
 
-                avg_loss._backward()
+                avg_loss.backward()
 
                 dy_grad_value = {}
                 for param in resnet.parameters():
@@ -288,7 +288,7 @@ class TestDygraphResnet(unittest.TestCase):
 
                 dy_param_value = {}
                 for param in resnet.parameters():
-                    dy_param_value[param.name] = param._numpy()
+                    dy_param_value[param.name] = param.numpy()
 
         with new_program_scope():
             fluid.default_startup_program().random_seed = seed

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

@@ -333,7 +333,7 @@ class TestImperativeResneXt(unittest.TestCase):
 
             dy_param_init_value = {}
             for param in se_resnext.parameters():
-                dy_param_init_value[param.name] = param._numpy()
+                dy_param_init_value[param.name] = param.numpy()
             for epoch_id in range(epoch_num):
                 for batch_id, data in enumerate(train_reader()):
 
@@ -349,19 +349,19 @@ class TestImperativeResneXt(unittest.TestCase):
 
                     img = to_variable(dy_x_data)
                     label = to_variable(y_data)
-                    label._stop_gradient = True
+                    label.stop_gradient = True
 
                     out = se_resnext(img)
                     loss = fluid.layers.cross_entropy(input=out, label=label)
                     avg_loss = fluid.layers.mean(x=loss)
 
-                    dy_out = avg_loss._numpy()
+                    dy_out = avg_loss.numpy()
 
                     if batch_id == 0:
                         for param in se_resnext.parameters():
                             if param.name not in dy_param_init_value:
-                                dy_param_init_value[param.name] = param._numpy()
-                    avg_loss._backward()
+                                dy_param_init_value[param.name] = param.numpy()
+                    avg_loss.backward()
 
                     #dy_grad_value = {}
                     #for param in se_resnext.parameters():
@@ -375,7 +375,7 @@ class TestImperativeResneXt(unittest.TestCase):
 
                     dy_param_value = {}
                     for param in se_resnext.parameters():
-                        dy_param_value[param.name] = param._numpy()
+                        dy_param_value[param.name] = param.numpy()
 
         with new_program_scope():
             fluid.default_startup_program().random_seed = seed

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

@@ -16,7 +16,8 @@ from __future__ import print_function
 
 import unittest
 import paddle.fluid as fluid
-from paddle.fluid.dygraph import Embedding, LayerNorm, FC, to_variable, Layer, guard
+from paddle.fluid import Embedding, LayerNorm, FC, Layer
+from paddle.fluid.dygraph import to_variable, guard
 from test_imperative_base import new_program_scope
 from paddle.fluid import core
 import numpy as np
@@ -116,7 +117,7 @@ class ModelHyperParams(object):
     # to process after each sub-layer
     postprocess_cmd = "da"  # dropout + residual connection
     # random seed used in dropout for CE.
-    dropout_seed = 1
+    dropout_seed = None
     # the flag indicating whether to share embedding and softmax weights.
     # vocabularies in source and target should be same for weight sharing.
     weight_sharing = True
@@ -166,15 +167,21 @@ def create_data(is_static=False):
         ]
     else:
         enc_inputs = [
-            to_variable(src_word_np), to_variable(src_pos_np),
-            to_variable(src_slf_attn_bias_np)
+            to_variable(
+                src_word_np, name='src_word'), to_variable(
+                    src_pos_np, name='src_pos'), to_variable(
+                        src_slf_attn_bias_np, name='src_slf_attn_bias')
         ]
         dec_inputs = [
-            to_variable(trg_word_np), to_variable(trg_pos_np),
-            to_variable(trg_slf_attn_bias_np), to_variable(trg_src_attn_bias_np)
+            to_variable(
+                trg_word_np, name='trg_word'), to_variable(
+                    trg_pos_np, name='trg_pos'), to_variable(
+                        trg_slf_attn_bias_np, name='trg_slf_attn_bias'),
+            to_variable(
+                trg_src_attn_bias_np, name='trg_src_attn_bias')
         ]
-        label = to_variable(lbl_word_np)
-        weight = to_variable(lbl_weight_np)
+        label = to_variable(lbl_word_np, name='lbl_word')
+        weight = to_variable(lbl_weight_np, name='lbl_weight')
         return enc_inputs, dec_inputs, label, weight
 
 
@@ -211,7 +218,7 @@ def make_all_inputs(input_fields):
 # The placeholder for batch_size in compile time. Must be -1 currently to be
 # consistent with some ops' infer-shape output in compile time, such as the
 # sequence_expand op used in beamsearch decoder.
-batch_size = 32
+batch_size = -1
 # The placeholder for squence length in compile time.
 seq_len = ModelHyperParams.max_length
 # Here list the data shapes and data types of all inputs.
@@ -305,54 +312,40 @@ sync = False
 # how many batches we use
 batch_num = 5
 
-np.random.seed = 1
+np.random.seed = 90
 src_word_np = np.random.randint(
     1,
     ModelHyperParams.src_vocab_size - 1,
-    size=(batch_size, seq_len, 1),
+    size=(TrainTaskConfig.batch_size, seq_len, 1),
     dtype='int64')
 src_pos_np = np.random.randint(
-    1, seq_len, size=(batch_size, seq_len, 1), dtype='int64')
-src_slf_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
-                                       seq_len, seq_len).astype('float32')
+    1, seq_len, size=(TrainTaskConfig.batch_size, seq_len, 1), dtype='int64')
+src_slf_attn_bias_np = np.random.randn(TrainTaskConfig.batch_size,
+                                       ModelHyperParams.n_head, seq_len,
+                                       seq_len).astype('float32')
 
 trg_word_np = np.random.randint(
     1,
     ModelHyperParams.src_vocab_size - 1,
-    size=(batch_size, seq_len, 1),
+    size=(TrainTaskConfig.batch_size, seq_len, 1),
     dtype='int64')
 trg_pos_np = np.random.randint(
-    1, seq_len, size=(batch_size, seq_len, 1), dtype='int64')
-trg_slf_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
-                                       seq_len, seq_len).astype('float32')
-trg_src_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
-                                       seq_len, seq_len).astype('float32')
+    1, seq_len, size=(TrainTaskConfig.batch_size, seq_len, 1), dtype='int64')
+trg_slf_attn_bias_np = np.random.randn(TrainTaskConfig.batch_size,
+                                       ModelHyperParams.n_head, seq_len,
+                                       seq_len).astype('float32')
+trg_src_attn_bias_np = np.random.randn(TrainTaskConfig.batch_size,
+                                       ModelHyperParams.n_head, seq_len,
+                                       seq_len).astype('float32')
 
 lbl_word_np = np.random.randint(
     1,
     ModelHyperParams.src_vocab_size - 1,
-    size=(batch_size * seq_len, 1),
+    size=(TrainTaskConfig.batch_size * seq_len, 1),
     dtype='int64')
-lbl_weight_np = np.random.randn(batch_size * seq_len, 1).astype('float32')
-
-# np.random.seed = 1
-# src_word_np = np.arange(0, 10).reshape([batch_size, seq_len, 1]).astype('int64')
-# src_pos_np = np.random.randint(
-#     1, seq_len, size=(batch_size, seq_len, 1), dtype='int64')
-# src_slf_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
-#                                        seq_len, seq_len).astype('float32')
-#
-# trg_word_np =  np.arange(0, 10).reshape([batch_size, seq_len, 1]).astype('int64')
-# trg_pos_np = np.random.randint(
-#     1, seq_len, size=(batch_size, seq_len, 1), dtype='int64')
-# trg_slf_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
-#                                        seq_len, seq_len).astype('float32')
-# trg_src_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
-#                                        seq_len, seq_len).astype('float32')
-#
-# lbl_word_np =  np.arange(0, 10).reshape([batch_size * seq_len, 1]).astype('int64')
-# lbl_weight_np = np.random.randn(batch_size * seq_len, 1).astype('float32')
-#
+lbl_weight_np = np.random.randn(TrainTaskConfig.batch_size * seq_len,
+                                1).astype('float32')
+
 pos_inp1 = position_encoding_init(ModelHyperParams.max_length,
                                   ModelHyperParams.d_model)
 pos_inp2 = position_encoding_init(ModelHyperParams.max_length,
@@ -466,7 +459,7 @@ class MultiHeadAttentionLayer(Layer):
             x=v, shape=[0, 0, self._n_head, self._d_value], inplace=False)
         transpose_v = fluid.layers.transpose(x=reshaped_v, perm=[0, 2, 1, 3])
 
-        #scale dot product attention
+        # scale dot product attention
         product = fluid.layers.matmul(
             x=transpose_q,
             y=transpose_k,
@@ -739,7 +732,7 @@ class DecoderSubLayer(Layer):
         enc_attn_output_pp = self._multihead_attention_layer2(
             pre_process_rlt2, enc_output, enc_output, dec_enc_attn_bias)
         enc_attn_output = self._post_process_layer2(
-            slf_attn_output, enc_attn_output_pp, self._postprocess_cmd,
+            slf_attn_output_pp, enc_attn_output_pp, self._postprocess_cmd,
             self._prepostprcess_dropout)
         pre_process_rlt3 = self._pre_process_layer3(None, enc_attn_output,
                                                     self._preprocess_cmd,
@@ -990,16 +983,18 @@ class TestDygraphTransformer(unittest.TestCase):
                 enc_inputs, dec_inputs, label, weights = create_data()
                 dy_sum_cost, dy_avg_cost, dy_predict, dy_token_num = transformer(
                     enc_inputs, dec_inputs, label, weights)
+
                 if i == 0:
                     for param in transformer.parameters():
-                        dy_param_init[param.name] = param._numpy()
+                        dy_param_init[param.name] = param.numpy()
 
-                dy_avg_cost._backward()
+                dy_avg_cost.backward()
                 optimizer.minimize(dy_avg_cost)
                 transformer.clear_gradients()
+
                 if i == batch_num - 1:
                     for param in transformer.parameters():
-                        dy_param_updated[param.name] = param._numpy()
+                        dy_param_updated[param.name] = param.numpy()
 
         with new_program_scope():
             fluid.default_startup_program().random_seed = seed
@@ -1043,7 +1038,6 @@ class TestDygraphTransformer(unittest.TestCase):
             static_param_name_list = list()
             static_sum_cost, static_avg_cost, static_predict, static_token_num = transformer(
                 enc_inputs, dec_inputs, label, weights)
-
             optimizer.minimize(static_avg_cost)
             for param in transformer.parameters():
                 static_param_name_list.append(param.name)
@@ -1061,8 +1055,8 @@ class TestDygraphTransformer(unittest.TestCase):
                     static_sum_cost, static_avg_cost, static_predict,
                     static_token_num
                 ]
-                fetch_list.extend(static_param_name_list)
 
+                fetch_list.extend(static_param_name_list)
                 out = exe.run(fluid.default_main_program(),
                               feed=feed_dict,
                               fetch_list=fetch_list)
@@ -1076,13 +1070,14 @@ class TestDygraphTransformer(unittest.TestCase):
                                                                     4]] = out[k]
 
         self.assertTrue(
-            np.array_equal(static_avg_cost_value, dy_avg_cost._numpy()))
+            np.array_equal(static_avg_cost_value, dy_avg_cost.numpy()))
         self.assertTrue(
-            np.array_equal(static_sum_cost_value, dy_sum_cost._numpy()))
+            np.array_equal(static_sum_cost_value, dy_sum_cost.numpy()))
         self.assertTrue(
-            np.array_equal(static_predict_value, dy_predict._numpy()))
+            np.array_equal(static_predict_value, dy_predict.numpy()))
         self.assertTrue(
-            np.array_equal(static_token_num_value, dy_token_num._numpy()))
+            np.array_equal(static_token_num_value, dy_token_num.numpy()))
+
         for key, value in six.iteritems(static_param_init):
             self.assertTrue(np.array_equal(value, dy_param_init[key]))
         for key, value in six.iteritems(static_param_updated):

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

@@ -114,7 +114,7 @@ class TestLayer(LayerTest):
             dy_ret = fc2(ret)
 
         self.assertTrue(np.array_equal(static_ret, static_ret2))
-        self.assertTrue(np.array_equal(static_ret, dy_ret._numpy()))
+        self.assertTrue(np.array_equal(static_ret, dy_ret.numpy()))
 
     def test_layer_norm(self):
         inp = np.ones([3, 32, 32], dtype='float32')
@@ -142,7 +142,7 @@ class TestLayer(LayerTest):
             dy_ret = lm(base.to_variable(inp))
 
         self.assertTrue(np.allclose(static_ret, static_ret2))
-        self.assertTrue(np.allclose(dy_ret._numpy(), static_ret2))
+        self.assertTrue(np.allclose(dy_ret.numpy(), static_ret2))
 
     def test_relu(self):
         with self.static_graph():
@@ -156,7 +156,7 @@ class TestLayer(LayerTest):
             t = np.ones([3, 3], dtype='float32')
             dy_ret = layers.relu(base.to_variable(t))
 
-        self.assertTrue(np.allclose(static_ret, dy_ret._numpy()))
+        self.assertTrue(np.allclose(static_ret, dy_ret.numpy()))
 
     def test_matmul(self):
         with self.static_graph():
@@ -177,7 +177,7 @@ class TestLayer(LayerTest):
             t2 = np.ones([3, 3], dtype='float32')
             dy_ret = layers.matmul(base.to_variable(t), base.to_variable(t2))
 
-        self.assertTrue(np.allclose(static_ret, dy_ret._numpy()))
+        self.assertTrue(np.allclose(static_ret, dy_ret.numpy()))
 
     def test_conv2d(self):
         with self.static_graph():
@@ -204,7 +204,7 @@ class TestLayer(LayerTest):
                 'conv2d', num_channels=3, num_filters=3, filter_size=[2, 2])
             dy_ret = conv2d(base.to_variable(images))
 
-        self.assertTrue(np.allclose(static_ret, dy_ret._numpy()))
+        self.assertTrue(np.allclose(static_ret, dy_ret.numpy()))
         self.assertTrue(np.allclose(static_ret, static_ret2))
 
     def test_gru_unit(self):
@@ -246,7 +246,7 @@ class TestLayer(LayerTest):
 
         for i in range(len(static_ret)):
             self.assertTrue(np.allclose(static_ret[i], static_ret2[i]))
-            self.assertTrue(np.allclose(static_ret[i], dy_ret[i]._numpy()))
+            self.assertTrue(np.allclose(static_ret[i], dy_ret[i].numpy()))
 
     def test_elementwise_math(self):
         n = np.ones([3, 3], dtype='float32')
@@ -288,8 +288,8 @@ class TestLayer(LayerTest):
             ret = layers.elementwise_sub(ret, n5)
             dy_ret = layers.elementwise_mul(ret, n6)
         self.assertTrue(
-            np.allclose(static_ret, dy_ret._numpy()),
-            '%s vs %s' % (static_ret, dy_ret._numpy()))
+            np.allclose(static_ret, dy_ret.numpy()),
+            '%s vs %s' % (static_ret, dy_ret.numpy()))
 
     def test_elementwise_minmax(self):
         n = np.ones([3, 3], dtype='float32')
@@ -299,8 +299,8 @@ class TestLayer(LayerTest):
             min_ret = layers.elementwise_min(n, n2)
             max_ret = layers.elementwise_max(n, n2)
 
-        self.assertTrue(np.allclose(n, min_ret._numpy()))
-        self.assertTrue(np.allclose(n2, max_ret._numpy()))
+        self.assertTrue(np.allclose(n, min_ret.numpy()))
+        self.assertTrue(np.allclose(n2, max_ret.numpy()))
 
     def test_sequence_conv(self):
         inp_np = np.arange(12).reshape([3, 4]).astype('float32')
@@ -367,7 +367,7 @@ class TestLayer(LayerTest):
                 'conv2d_transpose', num_filters=10, output_size=28)
             dy_rlt = conv2d_transpose(base.to_variable(inp_np))
         self.assertTrue(np.allclose(static_rlt2, static_rlt))
-        self.assertTrue(np.allclose(dy_rlt._numpy(), static_rlt))
+        self.assertTrue(np.allclose(dy_rlt.numpy(), static_rlt))
 
     def test_bilinear_tensor_product(self):
         inp_np_x = np.array([[1, 2, 3]]).astype('float32')
@@ -410,7 +410,7 @@ class TestLayer(LayerTest):
             dy_rlt = btp(base.to_variable(inp_np_x), base.to_variable(inp_np_y))
 
         self.assertTrue(np.allclose(static_rlt2, static_rlt))
-        self.assertTrue(np.allclose(dy_rlt._numpy(), static_rlt))
+        self.assertTrue(np.allclose(dy_rlt.numpy(), static_rlt))
 
     def test_prelu(self):
         inp_np = np.ones([5, 200, 100, 100]).astype('float32')
@@ -451,7 +451,7 @@ class TestLayer(LayerTest):
             dy_rlt = prelu(base.to_variable(inp_np))
 
         self.assertTrue(np.allclose(static_rlt2, static_rlt))
-        self.assertTrue(np.allclose(dy_rlt._numpy(), static_rlt))
+        self.assertTrue(np.allclose(dy_rlt.numpy(), static_rlt))
 
     def test_embeding(self):
         inp_word = np.array([[[1]]]).astype('int64')
@@ -484,7 +484,7 @@ class TestLayer(LayerTest):
             static_rlt3 = emb2(base.to_variable(inp_word))
 
         self.assertTrue(np.allclose(static_rlt2, static_rlt))
-        self.assertTrue(np.allclose(static_rlt3._numpy(), static_rlt))
+        self.assertTrue(np.allclose(static_rlt3.numpy(), static_rlt))
 
     def test_nce(self):
         window_size = 5
@@ -598,7 +598,7 @@ class TestLayer(LayerTest):
             nce_loss3 = nce(embs3, words[label_word])
 
         self.assertTrue(np.allclose(static_rlt2, static_rlt))
-        self.assertTrue(np.allclose(nce_loss3._numpy(), static_rlt))
+        self.assertTrue(np.allclose(nce_loss3.numpy(), static_rlt))
 
     def test_conv3d(self):
         with self.static_graph():
@@ -625,7 +625,7 @@ class TestLayer(LayerTest):
             conv3d = nn.Conv3D('conv3d', num_filters=3, filter_size=2)
             dy_ret = conv3d(base.to_variable(images))
 
-        self.assertTrue(np.allclose(static_ret, dy_ret._numpy()))
+        self.assertTrue(np.allclose(static_ret, dy_ret.numpy()))
         self.assertTrue(np.allclose(static_ret, static_ret2))
 
     def test_row_conv(self):
@@ -719,7 +719,7 @@ class TestLayer(LayerTest):
             groupNorm = nn.GroupNorm('GroupNorm', groups=2)
             dy_ret = groupNorm(base.to_variable(input))
 
-        self.assertTrue(np.allclose(static_ret, dy_ret._numpy()))
+        self.assertTrue(np.allclose(static_ret, dy_ret.numpy()))
         self.assertTrue(np.allclose(static_ret, static_ret2))
 
     def test_spectral_norm(self):
@@ -769,7 +769,7 @@ class TestLayer(LayerTest):
             spectralNorm = nn.SpectralNorm('SpectralNorm', dim=1, power_iters=2)
             dy_ret = spectralNorm(base.to_variable(input))
 
-        self.assertTrue(np.allclose(static_ret, dy_ret._numpy()))
+        self.assertTrue(np.allclose(static_ret, dy_ret.numpy()))
         self.assertTrue(np.allclose(static_ret, static_ret2))
 
     def test_tree_conv(self):
@@ -842,7 +842,7 @@ class TestLayer(LayerTest):
             dy_ret = treeConv(base.to_variable(vectors), base.to_variable(adj))
 
         self.assertTrue(np.allclose(static_ret, static_ret2))
-        self.assertTrue(np.allclose(static_ret, dy_ret._numpy()))
+        self.assertTrue(np.allclose(static_ret, dy_ret.numpy()))
 
     def test_conv3d_transpose(self):
         input_array = np.arange(0, 48).reshape(
@@ -872,7 +872,7 @@ class TestLayer(LayerTest):
                 use_cudnn=False)
             dy_rlt = conv3d_transpose(base.to_variable(input_array))
         self.assertTrue(np.allclose(static_rlt2, static_rlt))
-        self.assertTrue(np.allclose(dy_rlt._numpy(), static_rlt))
+        self.assertTrue(np.allclose(dy_rlt.numpy(), static_rlt))
 
 
 class TestBook(LayerTest):
@@ -907,7 +907,7 @@ class TestBook(LayerTest):
                 if isinstance(dy_result, tuple):
                     dy_result = dy_result[0]
 
-        self.assertTrue(np.array_equal(static_result[0], dy_result._numpy()))
+        self.assertTrue(np.array_equal(static_result[0], dy_result.numpy()))
 
     def _get_np_data(self, shape, dtype, append_batch_size=True):
         np.random.seed(self.seed)

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

@@ -73,7 +73,14 @@ class TestNearestInterpOp(OpTest):
         self.op_type = "nearest_interp"
         input_np = np.random.random(self.input_shape).astype("float32")
 
-        output_np = nearest_neighbor_interp_np(input_np, self.out_h, self.out_w,
+        if self.scale > 0:
+            out_h = int(self.input_shape[2] * self.scale)
+            out_w = int(self.input_shape[3] * self.scale)
+        else:
+            out_h = self.out_h
+            out_w = self.out_w
+
+        output_np = nearest_neighbor_interp_np(input_np, out_h, out_w,
                                                self.out_size, self.actual_shape,
                                                self.align_corners)
         self.inputs = {'X': input_np}
@@ -84,6 +91,7 @@ class TestNearestInterpOp(OpTest):
         self.attrs = {
             'out_h': self.out_h,
             'out_w': self.out_w,
+            'scale': self.scale,
             'interp_method': self.interp_method,
             'align_corners': self.align_corners,
         }
@@ -100,6 +108,7 @@ class TestNearestInterpOp(OpTest):
         self.input_shape = [2, 3, 4, 4]
         self.out_h = 2
         self.out_w = 2
+        self.scale = 0.
         self.out_size = np.array([3, 3]).astype("int32")
         self.align_corners = True
 
@@ -110,6 +119,7 @@ class TestNearestNeighborInterpCase1(TestNearestInterpOp):
         self.input_shape = [4, 1, 7, 8]
         self.out_h = 1
         self.out_w = 1
+        self.scale = 0.
         self.align_corners = True
 
 
@@ -119,6 +129,7 @@ class TestNearestNeighborInterpCase2(TestNearestInterpOp):
         self.input_shape = [3, 3, 9, 6]
         self.out_h = 12
         self.out_w = 12
+        self.scale = 0.
         self.align_corners = True
 
 
@@ -128,6 +139,7 @@ class TestNearestNeighborInterpCase3(TestNearestInterpOp):
         self.input_shape = [1, 1, 128, 64]
         self.out_h = 64
         self.out_w = 128
+        self.scale = 0.
         self.align_corners = True
 
 
@@ -137,6 +149,7 @@ class TestNearestNeighborInterpCase4(TestNearestInterpOp):
         self.input_shape = [4, 1, 7, 8]
         self.out_h = 1
         self.out_w = 1
+        self.scale = 0.
         self.out_size = np.array([2, 2]).astype("int32")
         self.align_corners = True
 
@@ -147,6 +160,7 @@ class TestNearestNeighborInterpCase5(TestNearestInterpOp):
         self.input_shape = [3, 3, 9, 6]
         self.out_h = 12
         self.out_w = 12
+        self.scale = 0.
         self.out_size = np.array([11, 11]).astype("int32")
         self.align_corners = True
 
@@ -157,6 +171,7 @@ class TestNearestNeighborInterpCase6(TestNearestInterpOp):
         self.input_shape = [1, 1, 128, 64]
         self.out_h = 64
         self.out_w = 128
+        self.scale = 0.
         self.out_size = np.array([65, 129]).astype("int32")
         self.align_corners = True
 
@@ -167,6 +182,7 @@ class TestNearestNeighborInterpActualShape(TestNearestInterpOp):
         self.input_shape = [3, 2, 32, 16]
         self.out_h = 64
         self.out_w = 32
+        self.scale = 0.
         self.out_size = np.array([66, 40]).astype("int32")
         self.align_corners = True
 
@@ -179,7 +195,15 @@ class TestNearestInterpOpUint8(OpTest):
         self.op_type = "nearest_interp"
         input_np = np.random.randint(
             low=0, high=256, size=self.input_shape).astype("uint8")
-        output_np = nearest_neighbor_interp_np(input_np, self.out_h, self.out_w,
+
+        if self.scale > 0:
+            out_h = int(self.input_shape[2] * self.scale)
+            out_w = int(self.input_shape[3] * self.scale)
+        else:
+            out_h = self.out_h
+            out_w = self.out_w
+
+        output_np = nearest_neighbor_interp_np(input_np, out_h, out_w,
                                                self.out_size, self.actual_shape,
                                                self.align_corners)
         self.inputs = {'X': input_np}
@@ -188,6 +212,7 @@ class TestNearestInterpOpUint8(OpTest):
         self.attrs = {
             'out_h': self.out_h,
             'out_w': self.out_w,
+            'scale': self.scale,
             'interp_method': self.interp_method,
             'align_corners': self.align_corners
         }
@@ -201,6 +226,7 @@ class TestNearestInterpOpUint8(OpTest):
         self.input_shape = [1, 3, 9, 6]
         self.out_h = 10
         self.out_w = 9
+        self.scale = 0.
         self.align_corners = True
 
 
@@ -210,6 +236,7 @@ class TestNearestNeighborInterpCase1Uint8(TestNearestInterpOpUint8):
         self.input_shape = [2, 3, 128, 64]
         self.out_h = 120
         self.out_w = 50
+        self.scale = 0.
         self.align_corners = True
 
 
@@ -219,6 +246,7 @@ class TestNearestNeighborInterpCase2Uint8(TestNearestInterpOpUint8):
         self.input_shape = [4, 1, 7, 8]
         self.out_h = 5
         self.out_w = 13
+        self.scale = 0.
         self.out_size = np.array([6, 15]).astype("int32")
         self.align_corners = True
 
@@ -228,5 +256,38 @@ class TestNearestInterpWithoutCorners(TestNearestInterpOp):
         self.align_corners = False
 
 
+class TestNearestNeighborInterpScale1(TestNearestInterpOp):
+    def init_test_case(self):
+        self.interp_method = 'nearest'
+        self.input_shape = [3, 2, 32, 16]
+        self.out_h = 64
+        self.out_w = 32
+        self.scale = 2.
+        self.out_size = np.array([66, 40]).astype("int32")
+        self.align_corners = True
+
+
+class TestNearestNeighborInterpScale2(TestNearestInterpOp):
+    def init_test_case(self):
+        self.interp_method = 'nearest'
+        self.input_shape = [3, 2, 32, 16]
+        self.out_h = 64
+        self.out_w = 32
+        self.scale = 1.5
+        self.out_size = np.array([66, 40]).astype("int32")
+        self.align_corners = True
+
+
+class TestNearestNeighborInterpScale3(TestNearestInterpOp):
+    def init_test_case(self):
+        self.interp_method = 'nearest'
+        self.input_shape = [3, 2, 32, 16]
+        self.out_h = 64
+        self.out_w = 32
+        self.scale = 1.
+        self.out_size = np.array([66, 40]).astype("int32")
+        self.align_corners = True
+
+
 if __name__ == "__main__":
     unittest.main()