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

Author: JiayiFeng

Dockerfile

@@ -24,7 +24,7 @@ COPY ./paddle/scripts/docker/root/ /root/
 
 RUN apt-get update && \
     apt-get install -y --allow-downgrades \
-    git python-pip python-dev openssh-server bison \
+    git python-pip python-dev python-opencv openssh-server bison \
     libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 \
     wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
     curl sed grep graphviz libjpeg-dev zlib1g-dev  \
@@ -76,8 +76,7 @@ RUN easy_install -U pip && \
     pip install sphinx-rtd-theme==0.1.9 recommonmark
 
 RUN pip install pre-commit 'ipython==5.3.0' && \
-    pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
-    pip install opencv-python
+    pip install 'ipykernel==4.6.0' 'jupyter==1.0.0'
 
 #For docstring checker
 RUN pip install pylint pytest astroid isort

benchmark/fluid/Dockerfile

@@ -1,8 +1,8 @@
 FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04
-RUN apt-get update && apt-get install -y python python-pip iputils-ping libgtk2.0-dev wget vim net-tools iftop
+RUN apt-get update && apt-get install -y python python-pip iputils-ping libgtk2.0-dev wget vim net-tools iftop python-opencv
 RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/lib/libcudnn.so && ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/lib/libnccl.so
 RUN pip install -U pip
-RUN pip install -U kubernetes opencv-python paddlepaddle
+RUN pip install -U kubernetes paddlepaddle
 
 # IMPORTANT:
 # Add "ENV http_proxy=http://ip:port" if your download is slow, and don't forget to unset it at runtime.

benchmark/fluid/fluid_benchmark.py

@@ -69,6 +69,11 @@ def parse_args():
         type=int,
         default=1,
         help='If gpus > 1, will use ParallelExecutor to run, else use Executor.')
+    parser.add_argument(
+        '--cpus',
+        type=int,
+        default=1,
+        help='If cpus > 1, will use ParallelDo to run, else use Executor.')
     parser.add_argument(
         '--data_set',
         type=str,
@@ -85,8 +90,8 @@ def parse_args():
         help='If set, use nvprof for CUDA.')
     parser.add_argument(
         '--no_test',
-        action='store_false',
-        help='If set, test the testset during training.')
+        action='store_true',
+        help='If set, do not test the testset during training.')
     parser.add_argument(
         '--memory_optimize',
         action='store_true',
@@ -229,9 +234,9 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
             print("Pass: %d, Iter: %d, Loss: %f\n" %
                   (pass_id, iters, np.mean(train_losses)))
         print_train_time(start_time, time.time(), num_samples)
-        print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses)))
+        print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses))),
         # evaluation
-        if not args.no_test and batch_acc != None:
+        if not args.no_test and batch_acc:
             pass_test_acc = test(exe, infer_prog, test_reader, feeder,
                                  batch_acc)
             print(", Test Accuracy: %f" % pass_test_acc)
@@ -310,7 +315,7 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
                 print("Pass %d, batch %d, loss %s" %
                       (pass_id, batch_id, np.array(loss)))
         print_train_time(start_time, time.time(), num_samples)
-        if not args.no_test and batch_acc != None:
+        if not args.no_test and batch_acc:
             test_acc = test(startup_exe, infer_prog, test_reader, feeder,
                             batch_acc)
             print("Pass: %d, Test Accuracy: %f\n" % (pass_id, test_acc))

benchmark/fluid/models/mnist.py

@@ -69,15 +69,30 @@ def get_model(args):
     images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE)
     label = fluid.layers.data(name='label', shape=[1], dtype='int64')
 
-    # Train program
-    predict = cnn_model(images)
-    cost = fluid.layers.cross_entropy(input=predict, label=label)
-    avg_cost = fluid.layers.mean(x=cost)
-
-    # Evaluator
-    batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
-    batch_acc = fluid.layers.accuracy(
-        input=predict, label=label, total=batch_size_tensor)
+    if args.device == 'CPU' and args.cpus > 1:
+        places = fluid.layers.get_places(args.cpus)
+        pd = fluid.layers.ParallelDo(places)
+        with pd.do():
+            predict = cnn_model(pd.read_input(images))
+            label = pd.read_input(label)
+            cost = fluid.layers.cross_entropy(input=predict, label=label)
+            avg_cost = fluid.layers.mean(x=cost)
+            batch_acc = fluid.layers.accuracy(input=predict, label=label)
+
+            pd.write_output(avg_cost)
+            pd.write_output(batch_acc)
+
+        avg_cost, batch_acc = pd()
+        avg_cost = fluid.layers.mean(avg_cost)
+        batch_acc = fluid.layers.mean(batch_acc)
+    else:
+        # Train program
+        predict = cnn_model(images)
+        cost = fluid.layers.cross_entropy(input=predict, label=label)
+        avg_cost = fluid.layers.mean(x=cost)
+
+        # Evaluator
+        batch_acc = fluid.layers.accuracy(input=predict, label=label)
 
     # inference program
     inference_program = fluid.default_main_program().clone()

benchmark/fluid/models/resnet.py

@@ -132,18 +132,33 @@ def get_model(args):
 
     input = fluid.layers.data(name='data', shape=dshape, dtype='float32')
     label = fluid.layers.data(name='label', shape=[1], dtype='int64')
-    predict = model(input, class_dim)
-    cost = fluid.layers.cross_entropy(input=predict, label=label)
-    avg_cost = fluid.layers.mean(x=cost)
 
-    batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
-    batch_acc = fluid.layers.accuracy(
-        input=predict, label=label, total=batch_size_tensor)
+    if args.device == 'CPU' and args.cpus > 1:
+        places = fluid.layers.get_places(args.cpus)
+        pd = fluid.layers.ParallelDo(places)
+        with pd.do():
+            predict = model(pd.read_input(input), class_dim)
+            label = pd.read_input(label)
+            cost = fluid.layers.cross_entropy(input=predict, label=label)
+            avg_cost = fluid.layers.mean(x=cost)
+            batch_acc = fluid.layers.accuracy(input=predict, label=label)
+
+            pd.write_output(avg_cost)
+            pd.write_output(batch_acc)
+
+        avg_cost, batch_acc = pd()
+        avg_cost = fluid.layers.mean(avg_cost)
+        batch_acc = fluid.layers.mean(batch_acc)
+    else:
+        predict = model(input, class_dim)
+        cost = fluid.layers.cross_entropy(input=predict, label=label)
+        avg_cost = fluid.layers.mean(x=cost)
+        batch_acc = fluid.layers.accuracy(input=predict, label=label)
 
     inference_program = fluid.default_main_program().clone()
     with fluid.program_guard(inference_program):
         inference_program = fluid.io.get_inference_program(
-            target_vars=[batch_acc, batch_size_tensor])
+            target_vars=[batch_acc])
 
     optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
 

benchmark/fluid/models/stacked_dynamic_lstm.py

@@ -101,9 +101,8 @@ def gate_common(
     loss = fluid.layers.mean(x=loss)
 
     # add acc
-    batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
     batch_acc = fluid.layers.accuracy(input=logit, label=fluid.layers.data(name='label', \
-                shape=[1], dtype='int64'), total=batch_size_tensor)
+                shape=[1], dtype='int64'))
 
     inference_program = fluid.default_main_program().clone()
     with fluid.program_guard(inference_program):

cmake/external/grpc.cmake

@@ -45,6 +45,7 @@ ExternalProject_Add(
     #    checkout and clean other dirs under third_party
     # 4. remove .git, and package the directory.
     URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.8.x.tar.gz"
+    URL_MD5  "c9c58ee7d0e8929a63155af6a2ecdbd0"
     PREFIX          ${GRPC_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CONFIGURE_COMMAND ""

doc/fluid/howto/optimization/host_memory_profiling_cn.md

@@ -0,0 +1,89 @@
+## 堆内存分析和优化
+
+计算机程序都可能有内存泄漏的风险。**内存泄漏**一般是由于程序在堆(heap)上分配了内存而没有释放，随着程序的运行占用的内存越来越大，一方面会影响程序的稳定性，可能让运行速度越来越慢，或者造成oom，甚至会影响运行程序的机器的稳定性，造成宕机。
+
+
+目前有很多内存泄漏分析工具，比较经典的有[valgrind](http://valgrind.org/docs/manual/quick-start.html#quick-start.intro), [gperftools](https://gperftools.github.io/gperftools/)。
+
+因为Fluid是用Python驱动C++ core来运行，valgrind直接分析非常困难，需要自己编译debug版本的、带valgrind支持的专用Python版本，而且输出的信息中大部分是Python自己的符号和调用信息，分析起来很困难，另外使用valgrind会让程序运行速度变得非常慢，所以不建议使用。
+
+本教程主要介绍[gperftools](https://gperftools.github.io/gperftools/)的使用。
+
+gperftool主要支持以下四个功能：
+
+- thread-caching malloc
+- heap-checking using tcmalloc
+- heap-profiling using tcmalloc
+- CPU profiler
+
+Paddle也提供了基于gperftool的[CPU性能分析教程](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/howto/optimization/cpu_profiling_cn.md)。
+
+对于堆内存的分析，主要用到thread-caching malloc和heap-profiling using tcmalloc。
+
+## 使用流程
+#### 环境
+本教程基于paddle提供的Docker开发环境paddlepaddle/paddle:latest-dev，基于Ubuntu 16.04.4 LTS环境。
+
+#### 使用流程
+
+- 安装google-perftools
+
+```
+apt-get install libunwind-dev 
+apt-get install google-perftools
+```
+
+- 安装pprof
+
+```
+go get -u github.com/google/pprof
+```
+
+- 设置运行环境
+
+```
+export PPROF_PATH=/root/gopath/bin/pprof
+export PPROF_BINARY_PATH=/root/gopath/bin/pprof
+export LD_PRELOAD=/usr/lib/libtcmalloc.so.4
+```
+
+- 使用heap profile来运行python程序。本质上是周期性的对堆的分配情况做一次快照。
+
+```
+# HEAPPROFILE 设置生成的堆分析文件的目录和文件前缀
+# HEAP_PROFILE_ALLOCATION_INTERVAL 设置每分配多少存储dump一次dump，默认1GB
+env HEAPPROFILE="./perf_log/test.log" HEAP_PROFILE_ALLOCATION_INTERVAL=209715200 python trainer.py
+```
+
+随着程序的运行，会在perf_log这个文件夹下生成很多文件，如下：
+
+```
+-rw-r--r-- 1 root root 1.0M Jun  1 15:00 test.log.0001.heap
+-rw-r--r-- 1 root root 1.0M Jun  1 15:00 test.log.0002.heap
+-rw-r--r-- 1 root root 1.0M Jun  1 15:00 test.log.0003.heap
+-rw-r--r-- 1 root root 1.0M Jun  1 15:00 test.log.0004.heap
+-rw-r--r-- 1 root root 1.0M Jun  1 15:00 test.log.0005.heap
+-rw-r--r-- 1 root root 1.0M Jun  1 15:00 test.log.0006.heap
+```
+
+- 使用pprof对heap文件进行分析。分析有两种模式：
+	- 完整模式。会对当前heap做一个分析，显示目前分配内存一些调用路径。
+
+	```
+	pprof --pdf python test.log.0012.heap
+	```
+	上述命令会生成一个profile00x.pdf的文件，可以直接打开，例如：[memory_cpu_allocator](https://github.com/jacquesqiao/Paddle/blob/bd2ea0e1f84bb6522a66d44a072598153634cade/doc/fluid/howto/optimization/memory_cpu_allocator.pdf)。从下图可以看出，在CPU版本fluid的运行过程中，分配存储最多的模块式CPUAllocator. 而别的模块相对而言分配内存较少，所以被忽略了，这对于分配内存泄漏是很不方便的，因为泄漏是一个缓慢的过程，在这种图中是无法看到的。
+	
+	![result](https://user-images.githubusercontent.com/3048612/40964027-a54033e4-68dc-11e8-836a-144910c4bb8c.png)
+	
+	- Diff模式。可以对两个时刻的heap做diff，把一些内存分配没有发生变化的模块去掉，而把增量部分显示出来。
+	```
+	pprof --pdf --base test.log.0010.heap python test.log.1045.heap
+	```
+	生成的结果为：[`memory_leak_protobuf`](https://github.com/jacquesqiao/Paddle/blob/bd2ea0e1f84bb6522a66d44a072598153634cade/doc/fluid/howto/optimization/memory_leak_protobuf.pdf)
+	
+	从图中可以看出：ProgramDesc这个结构，在两个版本之间增长了200MB+，所以这里有很大的内存泄漏的可能性，最终结果也确实证明是这里造成了泄漏。
+	
+	![result](https://user-images.githubusercontent.com/3048612/40964057-b434d5e4-68dc-11e8-894b-8ab62bcf26c2.png)
+	![result](https://user-images.githubusercontent.com/3048612/40964063-b7dbee44-68dc-11e8-9719-da279f86477f.png)
+	

paddle/contrib/inference/demo/simple_on_word2vec.cc

@@ -65,7 +65,10 @@ void Main(bool use_gpu) {
 }
 
 TEST(demo, word2vec_cpu) { Main(false /*use_gpu*/); }
+
+#ifdef PADDLE_WITH_CUDA
 TEST(demo, word2vec_gpu) { Main(true /*use_gpu*/); }
+#endif
 
 }  // namespace demo
 }  // namespace paddle

paddle/contrib/inference/paddle_inference_api.h

@@ -63,6 +63,7 @@ class PaddlePredictor {
   struct Config;
   PaddlePredictor() = default;
   PaddlePredictor(const PaddlePredictor&) = delete;
+  PaddlePredictor& operator=(const PaddlePredictor&) = delete;
 
   // Predict an record.
   // The caller should be responsible for allocating and releasing the memory of
@@ -76,7 +77,7 @@ class PaddlePredictor {
   virtual std::unique_ptr<PaddlePredictor> Clone() = 0;
 
   // Destroy the Predictor.
-  virtual ~PaddlePredictor() {}
+  virtual ~PaddlePredictor() = default;
 
   // The common configs for all the predictors.
   struct Config {