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

CMakeLists.txt

@@ -16,6 +16,7 @@ cmake_minimum_required(VERSION 3.0)
 
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 set(PROJ_ROOT ${CMAKE_CURRENT_SOURCE_DIR})
+set(PROJ_BINARY_ROOT ${CMAKE_CURRENT_BINARY_DIR})
 
 include(system)
 

cmake/generic.cmake

@@ -88,7 +88,7 @@
 #
 
 # including binary directory for generated headers.
-include_directories(${CMAKE_BINARY_DIR})
+include_directories(${CMAKE_CURRENT_BINARY_DIR})
 
 if(NOT APPLE)
     find_package(Threads REQUIRED)

paddle/optimizer/adadelta_optimizer.cc

@@ -27,22 +27,24 @@ void AdadeltaOptimizer::Update(const Tensor* gradient) {
 
 const char* AdadeltaOptimizer::SerializeState(int* state_len) {
   AdadeltaOptimizerState state;
-  // TODO(zhihong) : add lr_policy serialization
   state.set_num_sample_passed(num_sample_passed_);
+  std::string lr_str = this->lr_policy_->SerializeState(state_len);
+  state.mutable_lr_state()->ParseFromString(lr_str);
 
   TensorToProto(*parameter_, state.mutable_parameter());
   TensorToProto(*accum_gradient_, state.mutable_accum_gradient());
   TensorToProto(*accum_delta_, state.mutable_accum_delta());
   TensorToProto(*update_delta_, state.mutable_update_delta());
   auto str = state.SerializeAsString();
-  *state_len = str.size();
+  *state_len += str.size();
   return str.c_str();
 }
 
 void AdadeltaOptimizer::DeserializeState(const std::string& str) {
   AdadeltaOptimizerState state;
   state.ParseFromString(str);
-  // TODO(zhihong) : add lr_policy DeserializeState
+  auto lr_state = state.lr_state();
+  this->lr_policy_->DeserializeState(lr_state.SerializeAsString());
   num_sample_passed_ = state.num_sample_passed();
 
   ProtoToTensor(state.parameter(), parameter_);

paddle/optimizer/adagrad_optimizer.cc

@@ -19,20 +19,23 @@ void AdagradOptimizer::Update(const Tensor* gradient) {
 }
 const char* AdagradOptimizer::SerializeState(int* state_len) {
   AdagradOptimizerState state;
-  // TODO(zhihong) : add lr_policy serialization
   state.set_num_sample_passed(num_sample_passed_);
+  std::string lr_str = this->lr_policy_->SerializeState(state_len);
+  state.mutable_lr_state()->ParseFromString(lr_str);
 
   TensorToProto(*parameter_, state.mutable_parameter());
   TensorToProto(*accum_gradient_, state.mutable_accum_gradient());
   auto str = state.SerializeAsString();
-  *state_len = str.size();
+  *state_len += str.size();
   return str.c_str();
 }
 
 void AdagradOptimizer::DeserializeState(const std::string& str) {
   AdagradOptimizerState state;
   state.ParseFromString(str);
-  // TODO(zhihong) : add lr_policy DeserializeState
+  auto lr_state = state.lr_state();
+  this->lr_policy_->DeserializeState(lr_state.SerializeAsString());
+
   num_sample_passed_ = state.num_sample_passed();
   ProtoToTensor(state.parameter(), parameter_);
   ProtoToTensor(state.accum_gradient(), accum_gradient_);

paddle/optimizer/adam_optimizer.cc

@@ -24,20 +24,23 @@ void AdamOptimizer::Update(const Tensor *gradient) {
 
 const char *AdamOptimizer::SerializeState(int *state_len) {
   AdamOptimizerState state;
-  // TODO(zhihong) : add lr_policy serialization
+  std::string lr_str = this->lr_policy_->SerializeState(state_len);
+  state.mutable_lr_state()->ParseFromString(lr_str);
   state.set_num_sample_passed(num_sample_passed_);
+
   TensorToProto(*parameter_, state.mutable_parameter());
   TensorToProto(*momentums_, state.mutable_momentums());
   TensorToProto(*velocitys_, state.mutable_velocitys());
   auto str = state.SerializeAsString();
-  *state_len = str.size();
+  *state_len += str.size();
   return str.c_str();
 }
 
 void AdamOptimizer::DeserializeState(const std::string &str) {
   AdamOptimizerState state;
   state.ParseFromString(str);
-  // TODO(zhihong) : add lr_policy DeserializeState
+  auto lr_state = state.lr_state();
+  this->lr_policy_->DeserializeState(lr_state.SerializeAsString());
   num_sample_passed_ = state.num_sample_passed();
 
   ProtoToTensor(state.parameter(), parameter_);

paddle/optimizer/lr_policy.h

@@ -17,36 +17,56 @@ public:
 // constant learning rate policy
 class ConstLr final : public LrPolicy {
 public:
-  ConstLr(double lr) : learning_rate(lr){};
+  ConstLr(double lr) : learning_rate_(lr){};
   double LearningRate(const uint64_t num_sample_passed) {
-    return learning_rate;
+    return learning_rate_;
+  }
+  const char *SerializeState(int *state_len) {
+    LrPolicyState state;
+    state.set_learning_rate(learning_rate_);
+    auto str = state.SerializeAsString();
+    *state_len = str.size();
+    return str.c_str();
+  }
+  void DeserializeState(const std::string &str) {
+    LrPolicyState state;
+    state.ParseFromString(str);
+    learning_rate_ = state.learning_rate();
   }
-  const char *SerializeState(int *state_len) { return nullptr; }
-  void DeserializeState(const std::string &state) {}
 
 private:
-  double learning_rate;
+  double learning_rate_;
 };
 
 class LinearLr final : public LrPolicy {
 public:
   LinearLr(double lr, double lr_decay_a, double lr_decay_b)
-      : learning_rate(lr), lr_decay_a(lr_decay_a), lr_decay_b(lr_decay_b) {}
+      : learning_rate_(lr), lr_decay_a_(lr_decay_a), lr_decay_b_(lr_decay_b) {}
   double LearningRate(const uint64_t num_sample_passed) {
-    return std::max(learning_rate - lr_decay_a * num_sample_passed, lr_decay_b);
+    return std::max(learning_rate_ - lr_decay_a_ * num_sample_passed,
+                    lr_decay_b_);
   }
   const char *SerializeState(int *state_len) {
-    // TODO(zhihong) : add lr_policy serialization
-    return nullptr;
+    LrPolicyState state;
+    state.set_learning_rate(learning_rate_);
+    state.set_lr_decay_a(lr_decay_a_);
+    state.set_lr_decay_b(lr_decay_b_);
+    auto str = state.SerializeAsString();
+    *state_len = str.size();
+    return str.c_str();
   }
-  void DeserializeState(const std::string &state) {
-    // TODO(zhihong) : add lr_policy serialization
+  void DeserializeState(const std::string &str) {
+    LrPolicyState state;
+    state.ParseFromString(str);
+    learning_rate_ = state.learning_rate();
+    lr_decay_a_ = state.lr_decay_a();
+    lr_decay_b_ = state.lr_decay_b();
   }
 
 private:
-  double learning_rate;
-  double lr_decay_a;
-  double lr_decay_b;
+  double learning_rate_;
+  double lr_decay_a_;
+  double lr_decay_b_;
 };
 
 }  // namespace optimizer

paddle/optimizer/sgd_optimizer.cc

@@ -30,16 +30,20 @@ void SGDOptimizer::Update(const Tensor *gradient) {
 const char *SGDOptimizer::SerializeState(int *state_len) {
   SGDOptimizerState state;
   state.set_num_sample_passed(num_sample_passed_);
+  std::string lr_str = this->lr_policy_->SerializeState(state_len);
+  state.mutable_lr_state()->ParseFromString(lr_str);
   TensorToProto(*parameter_, state.mutable_parameter());
   if (momentum_ != 0.0) TensorToProto(*momentums_, state.mutable_momentums());
   auto str = state.SerializeAsString();
-  *state_len = str.size();
+  *state_len += str.size();
   return str.c_str();
 }
 
 void SGDOptimizer::DeserializeState(const std::string &str) {
   SGDOptimizerState state;
   state.ParseFromString(str);
+  auto lr_state = state.lr_state();
+  this->lr_policy_->DeserializeState(lr_state.SerializeAsString());
   num_sample_passed_ = state.num_sample_passed();
   ProtoToTensor(state.parameter(), parameter_);
   if (momentum_ != 0.0) ProtoToTensor(state.parameter(), momentums_);

paddle/platform/CMakeLists.txt

@@ -4,3 +4,5 @@ nv_test(cuda_test SRCS cuda_test.cu)
 
 cc_library(place SRCS place.cc)
 cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
+
+nv_test(device_context_test SRCS device_context_test.cc DEPS dynamic_loader place eigen3 glog gflags)

paddle/platform/device_context.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "paddle/framework/enforce.h"
+#ifndef PADDLE_ONLY_CPU
+#include "paddle/platform/cuda.h"
+#include "paddle/platform/dynload/cublas.h"
+#include "paddle/platform/dynload/cudnn.h"
+#include "paddle/platform/dynload/curand.h"
+#define EIGEN_USE_GPU
+#endif
+#include "paddle/platform/place.h"
+#include "unsupported/Eigen/CXX11/Tensor"
+
+namespace paddle {
+namespace platform {
+
+class DeviceContext {
+ public:
+  virtual ~DeviceContext() {}
+};
+
+class CPUDeviceContext : public DeviceContext {};
+
+#ifndef PADDLE_ONLY_CPU
+class GPUPlaceGuard {
+ public:
+  explicit GPUPlaceGuard(GPUPlace new_place) : previous_(GetCurrentDeviceId()) {
+    if (previous_ != new_place) {
+      paddle::platform::SetDeviceId(new_place.device);
+    }
+  }
+
+  ~GPUPlaceGuard() { paddle::platform::SetDeviceId(previous_.device); }
+
+ private:
+  GPUPlace previous_;
+};
+
+class CUDADeviceContext : public DeviceContext {
+ public:
+  explicit CUDADeviceContext(const GPUPlace gpu_place) : gpu_place_(gpu_place) {
+    GPUPlaceGuard guard(gpu_place_);
+    paddle::platform::throw_on_error(cudaStreamCreate(&stream_),
+                                     "cudaStreamCreate failed");
+    eigen_stream_ = new Eigen::CudaStreamDevice(&stream_);
+    eigen_device_ = new Eigen::GpuDevice(eigen_stream_);
+  }
+
+  void Wait() {
+    paddle::platform::throw_on_error(cudaStreamSynchronize(stream_),
+                                     "cudaStreamSynchronize failed");
+  }
+
+  cudaStream_t stream() { return stream_; }
+
+  Eigen::GpuDevice eigen_device() { return *eigen_device_; }
+
+  cublasHandle_t cublas_handle() {
+    if (!blas_handle_) {
+      GPUPlaceGuard guard(gpu_place_);
+      PADDLE_ENFORCE(paddle::platform::dynload::cublasCreate(&blas_handle_) ==
+                         CUBLAS_STATUS_SUCCESS,
+                     "cublasCreate failed");
+      PADDLE_ENFORCE(paddle::platform::dynload::cublasSetStream(
+                         blas_handle_, stream_) == CUBLAS_STATUS_SUCCESS,
+                     "cublasSetStream failed");
+    }
+    return blas_handle_;
+  }
+
+  cudnnHandle_t cudnn_handle() {
+    if (!dnn_handle_) {
+      GPUPlaceGuard guard(gpu_place_);
+      PADDLE_ENFORCE(paddle::platform::dynload::cudnnCreate(&dnn_handle_) ==
+                         CUDNN_STATUS_SUCCESS,
+                     "cudnnCreate failed");
+      PADDLE_ENFORCE(paddle::platform::dynload::cudnnSetStream(
+                         dnn_handle_, stream_) == CUDNN_STATUS_SUCCESS,
+                     "cudnnSetStream failed");
+    }
+    return dnn_handle_;
+  }
+
+  curandGenerator_t curand_generator() {
+    if (!rand_generator_) {
+      GPUPlaceGuard guard(gpu_place_);
+      PADDLE_ENFORCE(paddle::platform::dynload::curandCreateGenerator(
+                         &rand_generator_, CURAND_RNG_PSEUDO_DEFAULT) ==
+                         CURAND_STATUS_SUCCESS,
+                     "curandCreateGenerator failed");
+      PADDLE_ENFORCE(
+          paddle::platform::dynload::curandSetPseudoRandomGeneratorSeed(
+              rand_generator_, random_seed_) == CURAND_STATUS_SUCCESS,
+          "curandSetPseudoRandomGeneratorSeed failed");
+      PADDLE_ENFORCE(paddle::platform::dynload::curandSetStream(
+                         rand_generator_, stream_) == CURAND_STATUS_SUCCESS,
+                     "curandSetStream failed");
+    }
+    return rand_generator_;
+  }
+
+  ~CUDADeviceContext() {
+    Wait();
+    if (blas_handle_) {
+      PADDLE_ENFORCE(paddle::platform::dynload::cublasDestroy(blas_handle_) ==
+                         CUBLAS_STATUS_SUCCESS,
+                     "cublasDestroy failed");
+    }
+
+    if (dnn_handle_) {
+      PADDLE_ENFORCE(paddle::platform::dynload::cudnnDestroy(dnn_handle_) ==
+                         CUDNN_STATUS_SUCCESS,
+                     "cudnnDestroy failed");
+    }
+
+    if (rand_generator_) {
+      PADDLE_ENFORCE(paddle::platform::dynload::curandDestroyGenerator(
+                         rand_generator_) == CURAND_STATUS_SUCCESS,
+                     "curandDestroyGenerator failed");
+    }
+
+    delete eigen_stream_;
+    delete eigen_device_;
+
+    paddle::platform::throw_on_error(cudaStreamDestroy(stream_),
+                                     "cudaStreamDestroy failed");
+  }
+
+ private:
+  GPUPlace gpu_place_;
+  cudaStream_t stream_;
+
+  Eigen::CudaStreamDevice* eigen_stream_;
+  Eigen::GpuDevice* eigen_device_;
+
+  cublasHandle_t blas_handle_{nullptr};
+
+  cudnnHandle_t dnn_handle_{nullptr};
+
+  int random_seed_;
+  curandGenerator_t rand_generator_{nullptr};
+};
+#endif
+}  // namespace platform
+}  // namespace paddle

paddle/platform/device_context_test.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/platform/device_context.h"
+#include "gtest/gtest.h"
+
+TEST(CUDADeviceContext, Init) {
+  int count = paddle::platform::GetDeviceCount();
+  for (int i = 0; i < count; i++) {
+    paddle::platform::CUDADeviceContext* device_context =
+        new paddle::platform::CUDADeviceContext(i);
+    Eigen::GpuDevice gpu_device = device_context->eigen_device();
+    ASSERT_NE(nullptr, gpu_device.stream());
+    cudnnHandle_t cudnn_handle = device_context->cudnn_handle();
+    ASSERT_NE(nullptr, cudnn_handle);
+    cublasHandle_t cublas_handle = device_context->cublas_handle();
+    ASSERT_NE(nullptr, cublas_handle);
+    curandGenerator_t curand_handle = device_context->curand_generator();
+    ASSERT_NE(nullptr, curand_handle);
+    delete device_context;
+  }
+}

proto/OptimizerConfig.proto

@@ -78,11 +78,15 @@ enum DataType {
   repeated bytes content = 2;
 }
 
+message LrPolicyState {
+  // learninRate Policy
+  optional double learning_rate = 1 [default = 1.0];
+  optional double lr_decay_a = 2;
+  optional double lr_decay_b = 3;
+}
+
 message SGDOptimizerState {
-  // learning rate policy
-  optional double learning_rate = 101;
-  optional double lr_decay_a = 102;
-  optional double lr_decay_b = 103;
+  optional LrPolicyState lr_state = 101;
   optional double num_sample_passed = 104;
   // state
   optional TensorProto parameter = 1;
@@ -91,9 +95,7 @@ message SGDOptimizerState {
 
 message AdadeltaOptimizerState {
   // learning rate policy
-  optional double learning_rate = 101;
-  optional double lr_decay_a = 102;
-  optional double lr_decay_b = 103;
+  optional LrPolicyState lr_state = 101;
   optional double num_sample_passed = 104;
   // state
   optional TensorProto parameter = 1;
@@ -102,11 +104,9 @@ message AdadeltaOptimizerState {
   optional TensorProto update_delta = 4;
 }
 
+
 message AdagradOptimizerState {
-  // learning rate policy
-  optional double learning_rate = 101;
-  optional double lr_decay_a = 102;
-  optional double lr_decay_b = 103;
+  optional LrPolicyState lr_state = 101;
   optional double num_sample_passed = 104;
   // state
   optional TensorProto parameter = 1;
@@ -114,10 +114,7 @@ message AdagradOptimizerState {
 }
 
 message AdamOptimizerState {
-  // learning rate policy
-  optional double learning_rate = 101;
-  optional double lr_decay_a = 102;
-  optional double lr_decay_b = 103;
+  optional LrPolicyState lr_state = 101;
   optional double num_sample_passed = 104;
   // state
   optional TensorProto parameter = 1;

python/setup.py.in

@@ -34,6 +34,6 @@ setup(name='paddle',
           '': '${CMAKE_CURRENT_SOURCE_DIR}',
           # The paddle.v2.framework.proto will be generated while compiling.
           # So that package points to other directory.
-          'paddle.v2.framework.proto': '${CMAKE_BINARY_DIR}/paddle/framework'
+          'paddle.v2.framework.proto': '${PROJ_BINARY_ROOT}/paddle/framework'
       },
 )