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

reyoung · reyoung · commit 96b703c007b8 · 2018-05-08T15:07:22.000+08:00
diff --git a/contrib/float16/README.md b/contrib/float16/README.md
diff --git a/doc/v2/api/config/layer.rst b/doc/v2/api/config/layer.rst
@@ -142,7 +142,7 @@ gated_unit
 -----------
 ..  autoclass:: paddle.v2.layer.gated_unit
     :noindex:
-    
+
 Recurrent Layer Group
 =====================
 
@@ -354,7 +354,7 @@ dropout
 --------
 ..  autoclass:: paddle.v2.layer.dropout
     :noindex:
-    
+
 dot_prod
 ---------
 .. autoclass:: paddle.v2.layer.dot_prod
@@ -460,6 +460,11 @@ multi_binary_label_cross_entropy_cost
 ..  autoclass:: paddle.v2.layer.multi_binary_label_cross_entropy_cost
     :noindex:
 
+classification_cost
+-------------------
+.. autoclass:: paddle.v2.layer.classification_cost
+   :noindex:
+
 huber_regression_cost
 -------------------------
 ..  autoclass:: paddle.v2.layer.huber_regression_cost
@@ -534,7 +539,7 @@ detection_output
 ----------------
 ..  autoclass:: paddle.v2.layer.detection_output
     :noindex:
-    
+
 Check Layer
 ============
 
diff --git a/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md b/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md
@@ -41,7 +41,7 @@ Training docker image needs to package the paddle pserver and paddle trainer run
 - Generating the initialization arguments for `Paddle PServer` and `Paddle Training` processes.
 
 Since the paddlepaddle official docker image already has the runtimes we need, we'll take it as the base image and pack some additional scripts for the processes mentioned above to build our training image. for more detail, please find from the following link:
-- https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/usage/cluster/src/k8s_train/Dockerfile
+- https://github.com/PaddlePaddle/Paddle/tree/develop/doc/v2/howto/cluster/multi_cluster/src/k8s_train/Dockerfile
 
 
 ```bash
@@ -62,7 +62,7 @@ represent the Docker Image which built in this step.
 ### Prepare Training Data
 
 We can download and split the training job by creating a Kubernetes Job, or custom your image
-by editing [k8s_train](./src/k8s_train/).
+by editing [k8s_train](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/v2/howto/cluster/multi_cluster/src/k8s_train).
 
 Before creating a Job, we need to bind a [persistenVolumeClaim](https://kubernetes.io/docs/user-guide/persistent-volumes) by the different type of
 the different file system, the generated dataset would be saved on this volume.
diff --git a/paddle/fluid/inference/tensorrt/CMakeLists.txt b/paddle/fluid/inference/tensorrt/CMakeLists.txt
@@ -1,5 +1,4 @@
 nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader)
 nv_test(test_tensorrt_engine SRCS test_engine.cc engine.cc DEPS dynload_cuda)
-nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor)
 set(ENGINE_FILE ${CMAKE_CURRENT_SOURCE_DIR}/engine.cc)
 add_subdirectory(convert)
diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt
@@ -1,3 +1,4 @@
-nv_test(test_tensorrt_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES})
-nv_test(test_tensorrt_activation_op SRCS test_activation_op.cc ${ENGINE_FILE} activation_op.cc 
+nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES})
+nv_test(test_trt_activation_op SRCS test_activation_op.cc ${ENGINE_FILE} activation_op.cc 
   DEPS ${FLUID_CORE_MODULES} activation_op)
+nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor)
diff --git a/paddle/fluid/inference/tensorrt/convert/io_converter.cc b/paddle/fluid/inference/tensorrt/convert/io_converter.cc
@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/inference/tensorrt/io_converter.h"
+#include "paddle/fluid/inference/tensorrt/convert/io_converter.h"
 #include <cuda.h>
 #include "paddle/fluid/platform/enforce.h"
 
@@ -50,7 +50,7 @@ class DefaultInputConverter : public EngineInputConverter {
   }
 };
 
-REGISTER_TENSORRT_INPUT_CONVERTER(mul, DefaultInputConverter);
+REGISTER_TENSORRT_INPUT_CONVERTER(default, DefaultInputConverter);
 
 }  // namespace tensorrt
 }  // namespace inference
diff --git a/paddle/fluid/inference/tensorrt/convert/io_converter.h b/paddle/fluid/inference/tensorrt/convert/io_converter.h
@@ -40,7 +40,8 @@ class EngineInputConverter {
   static void Run(const std::string& in_op_type, const LoDTensor& in, void* out,
                   size_t max_size, cudaStream_t* stream) {
     PADDLE_ENFORCE(stream != nullptr);
-    auto* converter = Registry<EngineInputConverter>::Lookup(in_op_type);
+    auto* converter = Registry<EngineInputConverter>::Lookup(
+        in_op_type, "default" /* default_type */);
     PADDLE_ENFORCE_NOT_NULL(converter);
     converter->SetStream(stream);
     (*converter)(in, out, max_size);
diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h
@@ -19,6 +19,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/tensorrt/engine.h"
+#include "paddle/fluid/inference/utils/singleton.h"
 
 namespace paddle {
 namespace inference {
@@ -32,34 +33,23 @@ class OpConverter {
   OpConverter() {}
   virtual void operator()(const framework::OpDesc& op) {}
 
-  void Execute(const framework::OpDesc& op, TensorRTEngine* engine) {
+  void Run(const framework::OpDesc& op, TensorRTEngine* engine) {
     std::string type = op.Type();
-    auto it = converters_.find(type);
-    PADDLE_ENFORCE(it != converters_.end(), "no OpConverter for optype [%s]",
-                   type);
-    it->second->SetEngine(engine);
-    (*it->second)(op);
-  }
-
-  static OpConverter& Global() {
-    static auto* x = new OpConverter;
-    return *x;
-  }
-
-  template <typename T>
-  void Register(const std::string& key) {
-    converters_[key] = new T;
+    auto* it = Registry<OpConverter>::Lookup(type);
+    PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", type);
+    it->SetEngine(engine);
+    (*it)(op);
   }
 
   // convert fluid op to tensorrt layer
   void ConvertOp(const framework::OpDesc& op, TensorRTEngine* engine) {
-    OpConverter::Global().Execute(op, engine);
+    OpConverter::Run(op, engine);
   }
 
   // convert fluid block to tensorrt network
   void ConvertBlock(const framework::BlockDesc& block, TensorRTEngine* engine) {
     for (auto op : block.AllOps()) {
-      OpConverter::Global().Execute(*op, engine);
+      OpConverter::Run(*op, engine);
     }
   }
 
@@ -78,12 +68,12 @@ class OpConverter {
   framework::Scope* scope_{nullptr};
 };
 
-#define REGISTER_TRT_OP_CONVERTER(op_type__, Converter__)      \
-  struct trt_##op_type__##_converter {                         \
-    trt_##op_type__##_converter() {                            \
-      OpConverter::Global().Register<Converter__>(#op_type__); \
-    }                                                          \
-  };                                                           \
+#define REGISTER_TRT_OP_CONVERTER(op_type__, Converter__)       \
+  struct trt_##op_type__##_converter {                          \
+    trt_##op_type__##_converter() {                             \
+      Registry<OpConverter>::Register<Converter__>(#op_type__); \
+    }                                                           \
+  };                                                            \
   trt_##op_type__##_converter trt_##op_type__##_converter__;
 
 }  // namespace tensorrt
diff --git a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc
@@ -26,7 +26,7 @@ namespace paddle {
 namespace inference {
 namespace tensorrt {
 
-void compare(float input, float expect) {
+void Compare(float input, float expect) {
   framework::Scope scope;
   platform::CUDAPlace place;
   platform::CUDADeviceContext ctx(place);
@@ -85,8 +85,8 @@ void compare(float input, float expect) {
 }
 
 TEST(OpConverter, ConvertRelu) {
-  compare(1, 1);   // relu(1) = 1
-  compare(-5, 0);  // relu(-5) = 0
+  Compare(1, 1);   // relu(1) = 1
+  Compare(-5, 0);  // relu(-5) = 0
 }
 
 }  // namespace tensorrt
diff --git a/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc
@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/lod_tensor.h"
-#include "paddle/fluid/inference/tensorrt/io_converter.h"
+#include "paddle/fluid/inference/tensorrt/convert/io_converter.h"
 
 #include <gtest/gtest.h>
 
@@ -34,7 +34,7 @@ TEST_F(EngineInputConverterTester, DefaultCPU) {
   ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0);
 
   cudaStream_t stream;
-  EngineInputConverter::Run("mul", tensor, buffer, tensor.memory_size(),
+  EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(),
                             &stream);
 }
 
@@ -44,7 +44,7 @@ TEST_F(EngineInputConverterTester, DefaultGPU) {
   ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0);
 
   cudaStream_t stream;
-  EngineInputConverter::Run("mul", tensor, buffer, tensor.memory_size(),
+  EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(),
                             &stream);
 }
 
diff --git a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc
@@ -20,7 +20,7 @@ namespace paddle {
 namespace inference {
 namespace tensorrt {
 
-TEST(BlockConverter, ConvertBlock) {
+TEST(OpConverter, ConvertBlock) {
   framework::ProgramDesc prog;
   auto* block = prog.MutableBlock(0);
   auto* mul_op = block->AppendOp();
diff --git a/paddle/fluid/inference/utils/singleton.h b/paddle/fluid/inference/utils/singleton.h
@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <string>
 #include <unordered_map>
 #include "paddle/fluid/platform/enforce.h"
 
@@ -49,9 +50,15 @@ struct Registry {
     items_[name] = new ItemChild;
   }
 
-  static ItemParent* Lookup(const std::string& name) {
+  static ItemParent* Lookup(const std::string& name,
+                            const std::string& default_name = "") {
     auto it = items_.find(name);
-    if (it == items_.end()) return nullptr;
+    if (it == items_.end()) {
+      if (default_name == "")
+        return nullptr;
+      else
+        return items_.find(default_name)->second;
+    }
     return it->second;
   }
 
diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py
@@ -60,6 +60,7 @@
     'io',
     'initializer',
     'layers',
+    'transpiler'
     'nets',
     'optimizer',
     'learning_rate_decay',
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
@@ -1042,13 +1042,14 @@ def clone(self, for_test=False):
         Returns(Program):
             The cloned Program object.
         """
-        p = Program()
         if for_test:
-            p.desc = core.inference_optimize(self.desc)
+            p = self.inference_optimize()
         else:
+            p = Program()
             p.desc = core.ProgramDesc(self.desc)
-        p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())]
-        p.sync_with_cpp()
+            p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())]
+            p.sync_with_cpp()
+
         p.copy_param_info_from(self)
         return p
 
@@ -1061,7 +1062,7 @@ def prune(self, targets):
                 if isinstance(t, Variable):
                     # After transpiler processing, the op that output this
                     # variable maybe has been changed, so t.op is not reliable
-                    # and we need to find the current op that generate this 
+                    # and we need to find the current op that generate this
                     # variable here.
                     t.op = None
                     global_block = self.global_block()
@@ -1087,8 +1088,16 @@ def prune(self, targets):
         return res
 
     def inference_optimize(self):
+        # this is an alternative implement before
+        # core.inference_optimize being fixed.
         res = Program()
-        res.desc = core.inference_optimize(self.desc)
+        res.desc = core.ProgramDesc(self.desc)
+        for i in xrange(res.desc.num_blocks()):
+            block = res.desc.block(i)
+            for j in xrange(block.op_size()):
+                op = block.op(j)
+                if op.has_attr('is_test'):
+                    op.set_attr('is_test', True)
         res.blocks = [Block(res, i) for i in xrange(res.desc.num_blocks())]
         res.sync_with_cpp()
         return res
diff --git a/python/paddle/fluid/tests/book/label_semantic_roles/no_test_label_semantic_roles.py b/python/paddle/fluid/tests/book/label_semantic_roles/no_test_label_semantic_roles.py
diff --git a/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_conv.py b/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_conv.py
diff --git a/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_mlp.py b/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_mlp.py
diff --git a/python/paddle/fluid/trainer.py b/python/paddle/fluid/trainer.py
diff --git a/python/setup.py.in b/python/setup.py.in
