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

CMakeLists.txt

@@ -19,7 +19,7 @@ set(PADDLE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})
 
 include(system)
 
-project(paddle CXX C Go)
+project(paddle CXX C)
 message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
         "${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION}")
 message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
@@ -60,7 +60,7 @@ option(USE_NNPACK       "Compile PaddlePaddle with NNPACK library"      OFF)
 option(WITH_DISTRIBUTE  "Compile with grpc distributed support"         OFF)
 option(USE_EIGEN_FOR_BLAS   "Use matrix multiplication in Eigen"        OFF)
 option(WITH_ARM_FP16    "Use half precision support on armv8.2-a cpu"   OFF)
-option(WITH_FAST_BUNDLE_TEST    "Bundle tests that can be run in a single process together to reduce launch overhead"   ON)
+option(WITH_FAST_BUNDLE_TEST    "Bundle tests that can be run in a single process together to reduce launch overhead"   OFF)
 
 # CMAKE_BUILD_TYPE
 if(NOT CMAKE_BUILD_TYPE)
@@ -146,6 +146,7 @@ include(external/cares)
 include(external/grpc)
 
 include(cudnn)              # set cudnn libraries, must before configure
+include(cupti)
 include(configure)          # add paddle env configuration
 include(generic)            # simplify cmake module
 include(package)            # set paddle packages
@@ -201,13 +202,14 @@ endif()
 # "add_subdirectory(paddle)" and "add_subdirectory(python)" should be
 # placed after this block, because they depends on it.
 if(WITH_GOLANG)
+    enable_language(Go)
     add_subdirectory(go)
 endif(WITH_GOLANG)
 
 set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
 
-SET(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
-SET(CMAKE_C_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
+set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
+set(CMAKE_C_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
 
 add_subdirectory(paddle)
 if(WITH_PYTHON)

Dockerfile.android

@@ -21,16 +21,6 @@ RUN apt-get update && \
     wget curl tar unzip gcc g++ locales clang-format-3.8 swig cmake && \
     apt-get clean -y
 
-# Install Go and glide
-RUN wget -qO- go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \
-    tar -xz -C /usr/local && \
-    mkdir /root/gopath && \
-    mkdir /root/gopath/bin && \
-    mkdir /root/gopath/src
-ENV GOROOT=/usr/local/go GOPATH=/root/gopath
-# should not be in the same line with GOROOT definition, otherwise docker build could not find GOROOT.
-ENV PATH=${PATH}:${GOROOT}/bin:${GOPATH}/bin
-
 # git credential to skip password typing
 RUN git config --global credential.helper store
 

benchmark/cluster/vgg16/Dockerfile

-#FROM python:2.7.14
 FROM nvidia/cuda:8.0-cudnn5-runtime-ubuntu16.04
-RUN apt-get update && apt-get install -y python
-RUN pip install -U kubernetes opencv-python &&   apt-get update -y &&   apt-get install -y iputils-ping libgtk2.0-dev
-# NOTE: By default CI built wheel packages turn WITH_DISTRIBUTE=OFF,
-#       so we must build one with distribute support to install in this image.
+
+# you can get mirror list here:
+# https://launchpad.net/ubuntu/+archivemirrors
+ARG UBUNTU_MIRROR
+RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi'
+
+RUN apt-get update && apt-get install -y python python-dev python-pip iputils-ping libgtk2.0-dev
+RUN pip install -U kubernetes opencv-python
+
 RUN pip install paddlepaddle
+# if network is slowly, you may need to add proxy here.
+# ENV https_proxy=
 RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.cifar.train10()" | python'
 RUN pip uninstall -y paddlepaddle
+# unset proxy if it is setted.
+# ENV https_proxy=""
+
+# NOTE: By default CI built wheel packages turn WITH_DISTRIBUTE=OFF,
+#       so we must build one with distribute support to install in this image.
+ADD *.whl /
+RUN pip install /*.whl && rm -f /*.whl
+ENV LD_LIBRARY_PATH=/usr/local/lib
+
+# tf k8s
+RUN pip install tensorflow==1.4.0
+ADD tf_k8s /usr/bin
+RUN chmod +x /usr/bin/tf_k8s
+ADD vgg16_tf.py /workspace/
 
 # below lines may change a lot for debugging
 ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/paddle_k8s /usr/bin
 ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/k8s_tools.py /root
-ADD *.whl /
-RUN pip install /*.whl && rm -f /*.whl && \
-chmod +x /usr/bin/paddle_k8s
-ENV LD_LIBRARY_PATH=/usr/local/lib
+RUN chmod +x /usr/bin/paddle_k8s
 ADD vgg16_fluid.py vgg16_v2.py /workspace/

benchmark/cluster/vgg16/tf_k8s

+#!/bin/bash
+check_trainer_ret() {
+  ret=$1
+  stdbuf -oL echo "job returned $ret...setting pod return message..."
+  stdbuf -oL echo "==============================="
+
+  if [ $ret -eq 136 ] ; then
+    echo "Error Arithmetic Operation(Floating Point Exception)" > /dev/termination-log
+  elif [ $ret -eq 139 ] ; then
+    echo "Segmentation Fault" > /dev/termination-log
+  elif [ $ret -eq 1 ] ; then
+    echo "General Error" > /dev/termination-log
+  elif [ $ret -eq 134 ] ; then
+    echo "Program Abort" > /dev/termination-log
+  fi
+  stdbuf -oL echo "termination log wroted..."
+  exit $ret
+}
+
+g_pservers=""
+g_trainers=""
+
+wait_running_pods(){
+  pserver_label="tf-job-pserver=${JOB_NAME}"
+  trainer_label="tf-job-trainer=${JOB_NAME}"
+
+  stdbuf -oL python /root/k8s_tools.py wait_pods_running ${pserver_label} ${PSERVERS_NUM}
+  stdbuf -oL python /root/k8s_tools.py wait_pods_running ${trainer_label} ${TRAINERS_NUM}
+
+  g_pservers=$(python /root/k8s_tools.py fetch_endpoints ${pserver_label} ${PORT})
+  g_trainers=$(python /root/k8s_tools.py fetch_endpoints ${trainer_label} ${PORT})
+}
+
+start_tf_pserver(){
+  wait_running_pods
+
+  label="tf-job-pserver=${JOB_NAME}"
+  pserver_id=$(python /root/k8s_tools.py fetch_id ${label})
+
+  cmd="${ENTRY} --ps_hosts=${g_pservers} --worker_hosts=${g_trainers} \
+  --job_name=${TF_JOB_NAME} --task_index=${pserver_id}"
+
+  stdbuf -oL sh -c "cd ${TRAINER_PACKAGE} && ${cmd}"
+}
+
+start_tf_trainer(){
+  wait_running_pods
+
+  label="tf-job-trainer=${JOB_NAME}"
+  trainer_id=$(python /root/k8s_tools.py fetch_id ${label})
+
+  cmd="${ENTRY} --ps_hosts=${g_pservers} --worker_hosts=${g_trainers} \
+  --job_name=${TF_JOB_NAME} --task_index=${trainer_id} --batch_size=${BATCH_SIZE}"
+
+  stdbuf -oL sh -c "cd ${TRAINER_PACKAGE} && ${cmd}"
+  check_trainer_ret $?
+}
+
+start_tf(){
+    if [[ "${TF_JOB_NAME}" == "worker" ]]; then
+        start_tf_trainer
+    else
+        start_tf_pserver
+    fi
+}
+
+usage() {
+    echo "usage: tf_k8s [<args>]:"
+    echo "  start_tf         Start tensorflow jobs"
+}
+
+case "$1" in
+    start_tf)
+        start_tf
+        ;;
+    --help)
+        usage
+        ;;
+    *)
+        usage
+        ;;
+esac

benchmark/cluster/vgg16/tf_pserver.yaml

+apiVersion: extensions/v1beta1
+kind: ReplicaSet
+metadata:
+  name: vgg16job-tf-pserver
+spec:
+  replicas: 10
+  template:
+    metadata:
+      labels:
+        tf-job-pserver: vgg16job-tf
+    spec:
+      hostNetwork: true
+      imagePullSecrets:
+      - name: job-registry-secret
+      containers:
+      - name: pserver
+        image: "registry.baidu.com/paddlepaddle/fluid_benchmark_tf:vgg16"
+        imagePullPolicy: Always
+        command: ["tf_k8s", "start_tf"]
+        ports:
+        - name: jobport-30236
+          containerPort: 30236
+        env:
+        - name: PORT
+          value: "32036"
+        - name: ENTRY
+          value: "python vgg16_tf.py"
+        - name: JOB_NAME
+          value: vgg16job-tf
+        - name: PSERVERS_NUM
+          value: "10"
+        - name: TF_JOB_NAME 
+          value: "ps"
+        - name: TRAINERS_NUM
+          value: "20"
+        - name: BATCH_SIZE
+          value: "128"
+        - name: TRAINER_PACKAGE
+          value: "/workspace"
+        - name: NUM_PASSES
+          value: "1"
+        - name: NAMESPACE
+          valueFrom:
+            fieldRef:
+              fieldPath: "metadata.namespace"
+        - name: POD_IP
+          valueFrom:
+            fieldRef:
+              fieldPath: "status.podIP"
+        resources:
+          requests:
+            memory: 10Gi
+            cpu: 4
+          limits:
+            memory: 10Gi
+            cpu: 4

benchmark/cluster/vgg16/tf_trainer.yaml

+apiVersion: batch/v1
+kind: Job
+metadata:
+  name: vgg16job-tf-trainer
+spec:
+  parallelism: 20
+  completions: 20
+  template:
+    metadata:
+      labels:
+        tf-job-trainer: vgg16job-tf
+    spec:
+      imagePullSecrets:
+      - name: job-registry-secret
+      hostNetwork: true
+      containers:
+      - name: trainer
+        image: "registry.baidu.com/paddlepaddle/fluid_benchmark_tf:vgg16"
+        imagePullPolicy: Always
+        command: ["tf_k8s", "start_tf"]
+        ports:
+        - name: jobport-30236
+          containerPort: 30236
+        env:
+        - name: PORT
+          value: "32036"
+        - name: JOB_NAME
+          value: vgg16job-tf
+        - name: TF_JOB_NAME 
+          value: "worker"
+        - name: ENTRY
+          value: "python vgg16_tf.py"
+        - name: PSERVERS_NUM
+          value: "10"
+        - name: BATCH_SIZE
+          value: "128"
+        - name: TRAINERS_NUM
+          value: "20"
+        - name: TRAINER_PACKAGE
+          value: "/workspace"
+        - name: NUM_PASSES
+          value: "1"
+        - name: NAMESPACE
+          valueFrom:
+            fieldRef:
+              fieldPath: "metadata.namespace"
+        - name: POD_IP
+          valueFrom:
+            fieldRef:
+              fieldPath: "status.podIP"
+        resources:
+          requests:
+            memory: 40Gi
+            cpu: 2
+          limits:
+            memory: 40Gi
+            cpu: 2
+      restartPolicy: Never

benchmark/cluster/vgg16/vgg16_fluid.py

@@ -68,6 +68,21 @@ parser.add_argument(
     type=str2bool,
     default=True,
     help='Whether to run as local mode.')
+
+parser.add_argument(
+    "--ps_hosts",
+    type=str,
+    default="",
+    help="Comma-separated list of hostname:port pairs")
+parser.add_argument(
+    "--trainer_hosts",
+    type=str,
+    default="",
+    help="Comma-separated list of hostname:port pairs")
+
+# Flags for defining the tf.train.Server
+parser.add_argument(
+    "--task_index", type=int, default=0, help="Index of task within the job")
 args = parser.parse_args()
 
 
@@ -180,8 +195,9 @@ def main():
                     iters += 1
                     num_samples += len(data)
                     print(
-                        "Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, spent %f"
-                        % (pass_id, iters, loss, acc, time.time() - ts)
+                        "Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, Speed = %.2f img/s"
+                        % (pass_id, iters, loss, acc,
+                           len(data) / (time.time() - ts))
                     )  # The accuracy is the accumulation of batches, but not the current batch.
 
             pass_elapsed = time.time() - start_time
@@ -209,27 +225,24 @@ def main():
             batch_size=args.batch_size)
         train_loop(exe, fluid.default_main_program())
     else:
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # all pserver endpoints
-        eplist = []
-        for ip in pserver_ips.split(","):
-            eplist.append(':'.join([ip, "6174"]))
-        pserver_endpoints = ",".join(eplist)
-        print("pserver endpoints: ", pserver_endpoints)
         trainers = int(os.getenv("TRAINERS"))  # total trainer count
         print("trainers total: ", trainers)
-        current_endpoint = os.getenv(
-            "POD_IP") + ":6174"  # current pserver endpoint
+
         training_role = os.getenv(
             "TRAINING_ROLE",
             "TRAINER")  # get the training role: trainer/pserver
+
         t = fluid.DistributeTranspiler()
         t.transpile(
             optimize_ops,
             params_grads,
-            pservers=pserver_endpoints,
+            trainer_id=args.task_index,
+            pservers=args.ps_hosts,
             trainers=trainers)
 
         if training_role == "PSERVER":
+            current_endpoint = os.getenv("POD_IP") + ":" + os.getenv(
+                "PADDLE_INIT_PORT")
             if not current_endpoint:
                 print("need env SERVER_ENDPOINT")
                 exit(1)

benchmark/cluster/vgg16/vgg16_tf.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""VGG16 benchmark in TensorFlow
+You can get distribution example template structure here:
+https://medium.com/clusterone/how-to-write-distributed-tensorflow-code-with-an-example-on-tensorport-70bf3306adcb
+https://www.tensorflow.org/deploy/distributed
+"""
+
+import tensorflow as tf
+import paddle.v2 as paddle
+import numpy as np
+import argparse
+import time
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument(
+    '--batch_size', type=int, default=128, help="Batch size for training.")
+parser.add_argument(
+    '--learning_rate',
+    type=float,
+    default=1e-3,
+    help="Learning rate for training.")
+parser.add_argument('--num_passes', type=int, default=50, help="No. of passes.")
+parser.add_argument(
+    '--device',
+    type=str,
+    default='CPU',
+    choices=['CPU', 'GPU'],
+    help="The device type.")
+parser.add_argument(
+    '--data_format',
+    type=str,
+    default='NHWC',
+    choices=['NCHW', 'NHWC'],
+    help='The data order, NCHW=[batch, channels, height, width].'
+    'Only support NHWC right now.')
+parser.add_argument(
+    '--data_set',
+    type=str,
+    default='cifar10',
+    choices=['cifar10', 'flowers'],
+    help='Optional dataset for benchmark.')
+
+parser.add_argument(
+    "--ps_hosts",
+    type=str,
+    default="",
+    help="Comma-separated list of hostname:port pairs")
+parser.add_argument(
+    "--worker_hosts",
+    type=str,
+    default="",
+    help="Comma-separated list of hostname:port pairs")
+parser.add_argument(
+    "--job_name", type=str, default="", help="One of 'worker', 'ps'")
+# Flags for defining the tf.train.Server
+parser.add_argument(
+    "--task_index", type=int, default=0, help="Index of task within the job")
+
+args = parser.parse_args()
+
+
+class VGG16Model(object):
+    def __init__(self):
+        self.parameters = []
+
+    def batch_norm_relu(self, inputs, is_training):
+        """Performs a batch normalization followed by a ReLU."""
+        # We set fused=True for a significant speed boost. See
+        # https://www.tensorflow.org/speed/speed_guide#common_fused_ops
+        inputs = tf.layers.batch_normalization(
+            inputs=inputs,
+            axis=1 if args.data_format == 'NCHW' else -1,
+            momentum=0.9,
+            epsilon=1e-05,
+            center=True,
+            scale=True,
+            training=is_training,
+            fused=True)
+        inputs = tf.nn.relu(inputs)
+        return inputs
+
+    def conv_bn_layer(self,
+                      name,
+                      images,
+                      kernel_shape,
+                      is_training,
+                      drop_rate=0.0):
+        with tf.name_scope(name) as scope:
+            kernel = tf.Variable(
+                tf.truncated_normal(
+                    kernel_shape, dtype=tf.float32, stddev=1e-1),
+                name='weights')
+            conv = tf.nn.conv2d(
+                images,
+                kernel, [1, 1, 1, 1],
+                data_format=args.data_format,
+                padding='SAME')
+            biases = tf.Variable(
+                tf.constant(
+                    0.0, shape=[kernel_shape[-1]], dtype=tf.float32),
+                trainable=True,
+                name='biases')
+            out = tf.nn.bias_add(conv, biases)
+            out = self.batch_norm_relu(out, is_training)
+            out = tf.layers.dropout(out, rate=drop_rate, training=is_training)
+            return out
+
+    def fc_layer(self, name, inputs, shape):
+        with tf.name_scope(name) as scope:
+            fc_w = tf.Variable(
+                tf.truncated_normal(
+                    shape, dtype=tf.float32, stddev=1e-1),
+                name='weights')
+            fc_b = tf.Variable(
+                tf.constant(
+                    0.0, shape=[shape[-1]], dtype=tf.float32),
+                trainable=True,
+                name='biases')
+            out = tf.nn.bias_add(tf.matmul(inputs, fc_w), fc_b)
+            return out
+
+    def network(self, images, class_dim, is_training):
+        """ VGG16 model structure.
+
+            TODO(kuke): enable this network to support the 'NCHW' data format
+        """
+
+        # conv1
+        conv1_1 = self.conv_bn_layer(
+            'conv1_1', images, [3, 3, 3, 64], is_training, drop_rate=0.3)
+        conv1_2 = self.conv_bn_layer(
+            'conv1_2', conv1_1, [3, 3, 64, 64], is_training, drop_rate=0.0)
+        # pool1
+        pool1 = tf.nn.max_pool(
+            conv1_2,
+            ksize=[1, 2, 2, 1],
+            strides=[1, 2, 2, 1],
+            padding='SAME',
+            name='pool1')
+        # conv2
+        conv2_1 = self.conv_bn_layer(
+            'conv2_1', pool1, [3, 3, 64, 128], is_training, drop_rate=0.4)
+        conv2_2 = self.conv_bn_layer(
+            'conv2_2', conv2_1, [3, 3, 128, 128], is_training, drop_rate=0.0)
+        # pool2
+        pool2 = tf.nn.max_pool(
+            conv2_2,
+            ksize=[1, 2, 2, 1],
+            strides=[1, 2, 2, 1],
+            padding='SAME',
+            name='pool2')
+        # conv3
+        conv3_1 = self.conv_bn_layer(
+            'conv3_1', pool2, [3, 3, 128, 256], is_training, drop_rate=0.4)
+        conv3_2 = self.conv_bn_layer(
+            'conv3_2', conv3_1, [3, 3, 256, 256], is_training, drop_rate=0.4)
+        conv3_3 = self.conv_bn_layer(
+            'conv3_3', conv3_2, [3, 3, 256, 256], is_training, drop_rate=0.0)
+        # pool3
+        pool3 = tf.nn.max_pool(
+            conv3_3,
+            ksize=[1, 2, 2, 1],
+            strides=[1, 2, 2, 1],
+            padding='SAME',
+            name='pool3')
+        # conv4
+        conv4_1 = self.conv_bn_layer(
+            'conv4_1', pool3, [3, 3, 256, 512], is_training, drop_rate=0.4)
+        conv4_2 = self.conv_bn_layer(
+            'conv4_2', conv4_1, [3, 3, 512, 512], is_training, drop_rate=0.4)
+        conv4_3 = self.conv_bn_layer(
+            'conv4_3', conv4_2, [3, 3, 512, 512], is_training, drop_rate=0.0)
+        # pool4
+        pool4 = tf.nn.max_pool(
+            conv4_3,
+            ksize=[1, 2, 2, 1],
+            strides=[1, 2, 2, 1],
+            padding='SAME',
+            name='pool4')
+        # conv5
+        conv5_1 = self.conv_bn_layer(
+            'conv5_1', pool4, [3, 3, 512, 512], is_training, drop_rate=0.4)
+        conv5_2 = self.conv_bn_layer(
+            'conv5_2', conv5_1, [3, 3, 512, 512], is_training, drop_rate=0.4)
+        conv5_3 = self.conv_bn_layer(
+            'conv5_3', conv5_2, [3, 3, 512, 512], is_training, drop_rate=0.0)
+        # pool5
+        pool5 = tf.nn.max_pool(
+            conv5_3,
+            ksize=[1, 2, 2, 1],
+            strides=[1, 2, 2, 1],
+            padding='SAME',
+            name='pool4')
+        # flatten
+        shape = int(np.prod(pool5.get_shape()[1:]))
+        pool5_flat = tf.reshape(pool5, [-1, shape])
+        # fc1
+        drop = tf.layers.dropout(pool5_flat, rate=0.5, training=is_training)
+        fc1 = self.fc_layer('fc1', drop, [shape, 512])
+        # fc2
+        bn = self.batch_norm_relu(fc1, is_training)
+        drop = tf.layers.dropout(bn, rate=0.5, training=is_training)
+        fc2 = self.fc_layer('fc2', drop, [512, 512])
+
+        fc3 = self.fc_layer('fc3', fc2, [512, class_dim])
+
+        return fc3
+
+
+def run_benchmark(cluster_spec, server):
+    """Run benchmark on cifar10 or flowers."""
+
+    if args.data_set == "cifar10":
+        class_dim = 10
+        raw_shape = (3, 32, 32)
+        dat_shape = (None, 32, 32, 3) if args.data_format == 'NHWC' else (
+            None, 3, 32, 32)
+    else:
+        class_dim = 102
+        raw_shape = (3, 224, 224)
+        dat_shape = (None, 224, 224, 3) if args.data_format == 'NHWC' else (
+            None, 3, 224, 224)
+
+    device = tf.train.replica_device_setter(
+        worker_device="/job:worker/task:{}".format(args.task_index),
+        cluster=cluster_spec)
+
+    with tf.device(device):
+        images = tf.placeholder(tf.float32, shape=dat_shape)
+        labels = tf.placeholder(tf.int64, shape=(None, ))
+        is_training = tf.placeholder('bool')
+        onehot_labels = tf.one_hot(labels, depth=class_dim)
+
+        vgg16 = VGG16Model()
+        logits = vgg16.network(images, class_dim, is_training)
+        loss = tf.losses.softmax_cross_entropy(
+            onehot_labels=onehot_labels, logits=logits)
+        avg_loss = tf.reduce_mean(loss)
+
+        correct = tf.equal(tf.argmax(logits, 1), labels)
+        accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
+
+        optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
+        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
+        global_step = tf.Variable(0, name='global_step', trainable=False)
+        with tf.control_dependencies(update_ops):
+            train_op = optimizer.minimize(avg_loss, global_step=global_step)
+
+        summary_op = tf.summary.merge_all()
+        init_op = tf.global_variables_initializer()
+
+    # data reader
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.cifar.train10()
+            if args.data_set == 'cifar10' else paddle.dataset.flowers.train(),
+            buf_size=5120),
+        batch_size=args.batch_size)
+    test_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.cifar.test10()
+            if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
+            buf_size=5120),
+        batch_size=args.batch_size)
+
+    # test
+    def test():
+        test_accs = []
+        for batch_id, data in enumerate(test_reader()):
+            test_images = np.array(
+         map(lambda x: np.transpose(x[0].reshape(raw_shape),
+         axes=[1, 2, 0]) if args.data_format == 'NHWC' else x[0], data)).astype("float32")
+            test_labels = np.array(map(lambda x: x[1], data)).astype('int64')
+            test_accs.append(
+                accuracy.eval(feed_dict={
+                    images: test_images,
+                    labels: test_labels,
+                    is_training: False
+                }))
+        return np.mean(test_accs)
+
+    config = tf.ConfigProto(
+        intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)
+    config.gpu_options.allow_growth = True
+
+    hooks = [tf.train.StopAtStepHook(last_step=1000000)]
+
+    with tf.train.MonitoredTrainingSession(
+            master=server.target, is_chief=(args.task_index == 0),
+            hooks=hooks) as sess:
+        iters, num_samples, start_time = 0, 0, 0.0
+        for pass_id in range(args.num_passes):
+            # train
+            num_samples = 0
+            start_time = time.time()
+            for batch_id, data in enumerate(train_reader()):
+                train_images = np.array(
+                    map(lambda x: np.transpose(x[0].reshape(raw_shape),
+                    axes=[1, 2, 0]) if args.data_format == 'NHWC' else x[0], data)).astype("float32")
+                train_labels = np.array(map(lambda x: x[1], data)).astype(
+                    'int64')
+                iter_begin_time = time.time()
+                _, loss, acc = sess.run([train_op, avg_loss, accuracy],
+                                        feed_dict={
+                                            images: train_images,
+                                            labels: train_labels,
+                                            is_training: True
+                                        })
+                iters += 1
+                print(
+                    "Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, Speed=%.2f imgs/sec"
+                    % (pass_id, iters, loss, acc,
+                       len(data) / (time.time() - iter_begin_time)))
+                num_samples += len(data)
+            train_elapsed = time.time() - start_time
+            # test
+            pass_test_acc = test()
+            print("Pass = %d, Train speed = %f imgs/s, Test accuracy = %f\n" %
+                  (pass_id, num_samples / train_elapsed, pass_test_acc))
+
+
+def print_arguments():
+    print('-----------  Configuration Arguments -----------')
+    for arg, value in sorted(vars(args).iteritems()):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------------')
+
+
+if __name__ == '__main__':
+    print_arguments()
+
+    ps_hosts = args.ps_hosts.split(",")
+    worker_hosts = args.worker_hosts.split(",")
+
+    # Create a cluster from the parameter server and worker hosts.
+    cluster_spec = tf.train.ClusterSpec({
+        "ps": ps_hosts,
+        "worker": worker_hosts
+    })
+
+    # Create and start a server for the local task.
+    server = tf.train.Server(
+        cluster_spec, job_name=args.job_name, task_index=args.task_index)
+
+    if args.job_name == "ps":
+        print("start pserver")
+        server.join()
+    elif args.job_name == "worker":
+        print("start worker")
+        run_benchmark(cluster_spec, server)

cmake/configure.cmake

@@ -59,6 +59,7 @@ endif(NOT WITH_GOLANG)
 
 if(NOT WITH_GPU)
     add_definitions(-DHPPL_STUB_FUNC)
+    add_definitions("-DCUPTI_LIB_PATH=\"\"")
 
     list(APPEND CMAKE_CXX_SOURCE_FILE_EXTENSIONS cu)
 else()
@@ -73,7 +74,14 @@ else()
     if(NOT CUDNN_FOUND)
         message(FATAL_ERROR "Paddle needs cudnn to compile")
     endif()
-
+    if(CUPTI_FOUND)
+        include_directories(${CUPTI_INCLUDE_DIR})
+        add_definitions(-DPADDLE_WITH_CUPTI)
+        add_definitions("-DCUPTI_LIB_PATH=\"${CUPTI_LIBRARY_PATH}\"")
+    else()
+        add_definitions("-DCUPTI_LIB_PATH=\"\"")
+        message(STATUS "Cannot find CUPTI, GPU Profiling is incorrect.")
+    endif()
     set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-Xcompiler ${SIMD_FLAG}")
 
     # Include cuda and cudnn

cmake/cuda.cmake

@@ -155,7 +155,8 @@ endif()
 include_directories(${CUDA_INCLUDE_DIRS})
 list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
 if(NOT WITH_DSO)
-    list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY})
+    # TODO(panyx0718): CUPTI only allows DSO?
+    list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUPTI_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY})
 endif(NOT WITH_DSO)
 
 # setting nvcc arch flags

cmake/cupti.cmake

+if(NOT WITH_GPU)
+    return()
+endif()
+
+
+set(CUPTI_ROOT "/usr" CACHE PATH "CUPTI ROOT")
+find_path(CUPTI_INCLUDE_DIR cupti.h
+        PATHS ${CUPTI_ROOT} ${CUPTI_ROOT}/include
+        $ENV{CUPTI_ROOT} $ENV{CUPTI_ROOT}/include
+        ${CUDA_TOOLKIT_ROOT_DIR}/extras/CUPTI/include
+        NO_DEFAULT_PATH
+        )
+
+get_filename_component(__libpath_hist ${CUDA_CUDART_LIBRARY} PATH)
+
+set(TARGET_ARCH "x86_64")
+if(NOT ${CMAKE_SYSTEM_PROCESSOR})
+    set(TARGET_ARCH ${CMAKE_SYSTEM_PROCESSOR})
+endif()
+
+list(APPEND CUPTI_CHECK_LIBRARY_DIRS
+        ${CUPTI_ROOT}
+        ${CUPTI_ROOT}/lib64
+        ${CUPTI_ROOT}/lib
+        ${CUPTI_ROOT}/lib/${TARGET_ARCH}-linux-gnu
+        $ENV{CUPTI_ROOT}
+        $ENV{CUPTI_ROOT}/lib64
+        $ENV{CUPTI_ROOT}/lib
+        /usr/lib
+        ${CUDA_TOOLKIT_ROOT_DIR}/extras/CUPTI/lib64)
+find_library(CUPTI_LIBRARY NAMES libcupti.so libcupti.dylib # libcupti_static.a
+       PATHS ${CUPTI_CHECK_LIBRARY_DIRS} ${CUPTI_INCLUDE_DIR} ${__libpath_hist}
+       NO_DEFAULT_PATH
+       DOC "Path to cuPTI library.")
+
+get_filename_component(CUPTI_LIBRARY_PATH ${CUPTI_LIBRARY} DIRECTORY)
+if(CUPTI_INCLUDE_DIR AND CUPTI_LIBRARY)
+    set(CUPTI_FOUND ON)
+else()
+    set(CUPTI_FOUND OFF)
+endif()

doc/build_and_install/build_from_source_cn.rst

@@ -189,7 +189,7 @@ PaddlePaddle的编译选项，包括生成CPU/GPU二进制文件、链接何种B
     "WITH_TESTING", "是否开启单元测试", "OFF"
     "WITH_DOC", "是否编译中英文文档", "OFF"
     "WITH_SWIG_PY", "是否编译PYTHON的SWIG接口，该接口可用于预测和定制化训练", "Auto"
-    "WITH_GOLANG", "是否编译go语言的可容错parameter server", "ON"
+    "WITH_GOLANG", "是否编译go语言的可容错parameter server", "OFF"
     "WITH_MKL", "是否使用MKL数学库，如果为否则是用OpenBLAS", "ON"
 
 BLAS

doc/build_and_install/build_from_source_en.rst

@@ -191,7 +191,7 @@ You can add :code:`-D` argument to pass such options, like:
     "WITH_TESTING", "Build unit tests", "OFF"
     "WITH_DOC", "Build documentations", "OFF"
     "WITH_SWIG_PY", "Build Python SWIG interface for V2 API", "Auto"
-    "WITH_GOLANG", "Build fault-tolerant parameter server written in go", "ON"
+    "WITH_GOLANG", "Build fault-tolerant parameter server written in go", "OFF"
     "WITH_MKL", "Use MKL as BLAS library, else use OpenBLAS", "ON"
 
 

doc/howto/capi/workflow_of_capi_cn.md

@@ -65,6 +65,7 @@
     output_file = "output.paddle.model"
     merge_v2_model(net, param_file, output_file)
     ```
+
     对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)这个示例，可直接运行 `python` [merge_v2_model.py](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense/merge_v2_model.py)。序列化结果会写入当前运行目录下的`output.paddle.model`文件中。使用这种方式，运行时C-API可以通过指定`output.paddle.model`文件的路径来加载预测模型。
 
 #### 注意事项

doc/howto/cluster/multi_cluster/index_cn.rst

 在不同集群中运行
 ================
+用户的集群环境不尽相同，为了方便大家的部署，我们提供了多种的集群部署方式，方便提交集群训练任务，以下将一一介绍:
 
-PaddlePaddle可以使用多种分布式计算平台构建分布式计算任务，包括：
-- `Kubernetes <http://kubernetes.io>`_ Google开源的容器集群的调度框架，支持大规模集群生产环境的完整集群方案。
-- `OpenMPI <https://www.open-mpi.org>`_ 成熟的高性能并行计算框架。
-- `Fabric <http://www.fabfile.org>`_ 集群管理工具。可以使用`Fabric`编写集群任务提交和管理脚本。
+`Kubernetes <http://kubernetes.io>`_ 是Google开源的容器集群的调度框架，支持大规模集群生产环境的完整集群方案。以下指南展示了PaddlePaddle对Kubernetes的支持：
 
-对于不同的集群平台，会分别介绍集群作业的启动和停止方法。这些例子都可以在 `cluster_train_v2 <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/scripts/cluster_train_v2>`_ 找到。
+..  toctree::
+  :maxdepth: 1
+
+  k8s_cn.md
+  k8s_distributed_cn.md
 
-在使用分布式计算平台进行训练时，任务被调度在集群中时，分布式计算平台通常会通过API或者环境变量提供任务运行需要的参数，比如节点的ID、IP和任务节点个数等。
+`OpenMPI <https://www.open-mpi.org>`_  是成熟的高性能并行计算框架，在HPC领域使用非常的广泛。以下指南介绍了如何使用OpenMPI来搭建PaddlePaddle的集群训练任务:
 
 ..  toctree::
   :maxdepth: 1
 
-  fabric_cn.md
   openmpi_cn.md
-  k8s_cn.md
-  k8s_distributed_cn.md
+
+`Fabric <http://www.fabfile.org>`_ 是一个方便的程序部署和管理工具。我们提供了使用Fabric 进行部署、管理的方法，如果想详细了解，请阅读以下指南:
+
+..  toctree::
+  :maxdepth: 1
+
+  fabric_cn.md
+
+我们也支持在AWS上部署PaddlePaddle，详细请了解:
+
+..  toctree::
+  :maxdepth: 1
+
   k8s_aws_cn.md
+
+您可以在 `cluster_train_v2 <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/scripts/cluster_train_v2>`_ 找到以上相关的例子。
+

doc/mobile/cross_compiling_for_android_cn.md

@@ -23,6 +23,12 @@ $ docker build -t username/paddle-android:dev . -f Dockerfile.android
 $ docker pull paddlepaddle/paddle:latest-dev-android
 ```
 
+对于国内用户，我们提供了加速访问的镜像源：
+
+```bash
+$ docker pull docker.paddlepaddlehub.com/paddle:latest-dev-android
+```
+
 ### 编译PaddlePaddle C-API库
 构建好开发镜像后，即可使用开发镜像来编译Android版PaddlePaddle C-API库。
 Android的Docker开发镜像向用户提供两个可配置的参数：
@@ -56,15 +62,15 @@ Android的Docker开发镜像向用户提供两个可配置的参数：
 
 - 编译`armeabi-v7a`，`Android API 21`的PaddlePaddle库
 
-  ```bash
-  $ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" username/paddle-android:dev
-  ```
+```bash
+$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" username/paddle-android:dev
+```
 
 - 编译`arm64-v8a`，`Android API 21`的PaddlePaddle库
 
-  ```bash
-  $ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=arm64-v8a" -e "ANDROID_API=21" username/paddle-android:dev
-  ```
+```bash
+$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=arm64-v8a" -e "ANDROID_API=21" username/paddle-android:dev
+```
 
 执行上述`docker run`命令时，容器默认执行[paddle/scripts/docker/build_android.sh](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build_android.sh)脚本。该脚本中记录了交叉编译Android版PaddlePaddle库常用的CMake配置，并且会根据`ANDROID_ABI`和`ANDROID_API`自动构建独立工具链、进行编译和安装。由于arm64架构要求Android API不小于21。因此当`ANDROID_ABI=arm64-v8a`，`ANDROID_API<21`时，Docker容器中将默认使用`Android API 21`的编译工具链。用户可以参考下文[配置交叉编译参数](#配置交叉编译参数)章节，根据个人的需求修改定制Docker容器所执行的脚本。编译安装结束之后，PaddlePaddle的C-API库将被安装到`$PWD/install_android`目录，所依赖的第三方库同时也被安装到`$PWD/install_android/third_party`目录。
 
@@ -155,7 +161,11 @@ cmake -DCMAKE_SYSTEM_NAME=Android \
       ..
 ```
 
-用户还可根据自己的需求设置其他编译参数。比如希望最小化生成的库的大小，可以设置`CMAKE_BUILD_TYPE`为`MinSizeRel`；若希望最快的执行速度，则可设置`CMAKE_BUILD_TYPE`为`Release`。亦可以通过手动设置`CMAKE_C/CXX_FLAGS`来影响PaddlePaddle的编译过程。
+用户还可根据自己的需求设置其他编译参数。
+
+- 设置`CMAKE_BUILD_TYPE`为`MinSizeRel`，最小化生成的库的大小。
+- 设置`CMAKE_BUILD_TYPE`为`Release`，获得最快的执行速度，
+- 用户亦可以通过手动设置`CMAKE_C/CXX_FLAGS`来影响PaddlePaddle的编译过程。
 
 **性能TIPS**，为了达到最快的计算速度，在CMake参数配置上，有以下建议：
 

doc/mobile/cross_compiling_for_android_en.md

@@ -25,6 +25,12 @@ Users can directly use the published Docker image.
 $ docker pull paddlepaddle/paddle:latest-dev-android
 ```
 
+For users in China, we provide a faster mirror.
+
+```bash
+$ docker pull docker.paddlepaddlehub.com/paddle:latest-dev-android
+```
+
 ### Build the Inference Library
 
 We can run the Docker image we just created to build the inference library of PaddlePaddle for Android using the command below:
@@ -86,19 +92,19 @@ Android NDK includes everything we need to build the [*standalone toolchain*](ht
 
 - To build the standalone toolchain for `armeabi-v7a` and Android API level 21:
 
-  ```bash
-  your/path/to/android-ndk-r14b-linux-x86_64/build/tools/make-standalone-toolchain.sh \
+```bash
+your/path/to/android-ndk-r14b-linux-x86_64/build/tools/make-standalone-toolchain.sh \
         --arch=arm --platform=android-21 --install-dir=your/path/to/arm_standalone_toolchain
-  ```
+```
   
   The generated standalone toolchain will be in `your/path/to/arm_standalone_toolchain`.
 
 - To build the standalone toolchain for `arm64-v8a` and Android API level 21:
 
-  ```bash
-  your/path/to/android-ndk-r14b-linux-x86_64/build/tools/make-standalone-toolchain.sh \
+```bash
+your/path/to/android-ndk-r14b-linux-x86_64/build/tools/make-standalone-toolchain.sh \
         --arch=arm64 --platform=android-21 --install-dir=your/path/to/arm64_standalone_toolchain
-  ```
+```
 
   The generated standalone toolchain will be in `your/path/to/arm64_standalone_toolchain`.
 

doc/templates/conf.py.cn.in

@@ -18,7 +18,6 @@ import shlex
 from recommonmark import parser, transform
 import paddle
 import paddle.v2
-import paddle.fluid
 
 MarkdownParser = parser.CommonMarkParser
 AutoStructify = transform.AutoStructify

doc/templates/conf.py.en.in

@@ -18,7 +18,6 @@ import shlex
 from recommonmark import parser, transform
 import paddle
 import paddle.v2
-import paddle.fluid
 
 
 MarkdownParser = parser.CommonMarkParser

paddle/fluid/framework/executor.cc

@@ -58,13 +58,13 @@ static void CreateTensor(Variable* var, proto::VarType::Type var_type) {
     var->GetMutable<ReaderHolder>();
   } else if (var_type == proto::VarType::CHANNEL) {
     var->GetMutable<ChannelHolder>();
-  } else if (var_type == proto::VarType::NCCL_COM) {
-    // GetMutable will be called in ncclInit
+  } else if (var_type == proto::VarType::RAW) {
+    // GetMutable will be called in operator
   } else {
     PADDLE_THROW(
         "Variable type %d is not in "
         "[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, "
-        "LOD_RANK_TABLE, PLACE_LIST, READER, CHANNEL, NCCL_COM]",
+        "LOD_RANK_TABLE, PLACE_LIST, READER, CHANNEL, RAW]",
         var_type);
   }
 }
@@ -127,7 +127,9 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
 
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    platform::RecordEvent record_event(op->Type(), pool.Get(place_));
+    // TODO(panyx0718): Need a program id to distinguish programs.
+    platform::RecordEvent record_event(op->Type(), pool.Get(place_),
+                                       op_desc->Block()->ID());
 
     VLOG(3) << place_ << " " << op->DebugStringEx(local_scope);
     op->Run(*local_scope, place_);

paddle/fluid/framework/framework.proto

@@ -113,7 +113,10 @@ message VarType {
     PLACE_LIST = 14;
     READER = 15;
     CHANNEL = 16;
-    NCCL_COM = 17;
+    // Any runtime decided variable type is raw
+    // raw variables should manage their own allocations
+    // in operators like nccl_op
+    RAW = 17;
   }
 
   required Type type = 1;
@@ -164,4 +167,6 @@ message BlockDesc {
 // Please refer to
 // https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md
 // for more details.
+// TODO(panyx0718): A model can have multiple programs. Need a
+// way to distinguish them. Maybe ID or name?
 message ProgramDesc { repeated BlockDesc blocks = 1; }

paddle/fluid/framework/lod_tensor.cc

@@ -31,8 +31,14 @@ std::ostream &operator<<(std::ostream &os, const LoD &lod) {
   os << "{";
   for (auto &v : lod) {
     os << "{";
+    bool is_first = true;
     for (auto &i : v) {
-      os << i << ",";
+      if (is_first) {
+        os << i;
+        is_first = false;
+      } else {
+        os << ", " << i;
+      }
     }
     os << "}";
   }

paddle/fluid/framework/op_desc.h

@@ -125,6 +125,8 @@ class OpDesc {
 
   BlockDesc *Block() { return this->block_; }
 
+  const BlockDesc &BlockRef() const { return *this->block_; }
+
   void SetBlock(BlockDesc *block) { this->block_ = block; }
 
  private:

paddle/fluid/inference/io.cc

@@ -32,24 +32,12 @@ void ReadBinaryFile(const std::string& filename, std::string& contents) {
   inputfs.close();
 }
 
-bool IsParameter(const framework::VarDesc* var,
-                 const framework::ProgramDesc& main_program) {
-  if (var->Persistable()) {
-    // There are many unreachable variables in the program
-    for (size_t i = 0; i < main_program.Size(); ++i) {
-      const framework::BlockDesc& block = main_program.Block(i);
-      for (auto* op : block.AllOps()) {
-        if (op->Type() == framework::kFeedOpType) {
-          continue;
-        }
-        for (auto input_argument_name : op->InputArgumentNames()) {
-          if (input_argument_name == var->Name()) {
+bool IsPersistable(const framework::VarDesc* var) {
+  if (var->Persistable() &&
+      var->GetType() != framework::proto::VarType::FEED_MINIBATCH &&
+      var->GetType() != framework::proto::VarType::FETCH_LIST) {
     return true;
   }
-        }
-      }
-    }
-  }
   return false;
 }
 
@@ -65,8 +53,8 @@ void LoadPersistables(framework::Executor& executor,
   std::vector<std::string> paramlist;
 
   for (auto* var : global_block.AllVars()) {
-    if (IsParameter(var, main_program)) {
-      VLOG(3) << "parameter's name: " << var->Name();
+    if (IsPersistable(var)) {
+      VLOG(3) << "persistable variable's name: " << var->Name();
 
       framework::VarDesc* new_var = load_block->Var(var->Name());
       new_var->SetShape(var->GetShape());
@@ -101,7 +89,6 @@ void LoadPersistables(framework::Executor& executor,
 
   executor.Run(*load_program, &scope, 0, true, true);
 
-  VLOG(3) << "Ran loading successfully";
   delete load_program;
 }
 

paddle/fluid/inference/tests/book/CMakeLists.txt

@@ -30,5 +30,5 @@ inference_test(label_semantic_roles)
 inference_test(recognize_digits ARGS mlp conv)
 inference_test(recommender_system)
 #inference_test(rnn_encoder_decoder)
-inference_test(understand_sentiment)
+inference_test(understand_sentiment ARGS conv)
 inference_test(word2vec)

paddle/fluid/inference/tests/book/test_inference_label_semantic_roles.cc

@@ -32,16 +32,42 @@ TEST(inference, label_semantic_roles) {
   paddle::framework::LoDTensor word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1,
       ctx_p2, mark;
   paddle::framework::LoD lod{{0, 4, 10}};
-
-  SetupLoDTensor(word, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(
-      predicate, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_n2, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_n1, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_0, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_p1, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(ctx_p2, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
-  SetupLoDTensor(mark, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
+  int64_t word_dict_len = 44068;
+  int64_t predicate_dict_len = 3162;
+  int64_t mark_dict_len = 2;
+
+  SetupLoDTensor(word,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(predicate,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(predicate_dict_len - 1));
+  SetupLoDTensor(ctx_n2,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(ctx_n1,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(ctx_0,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(ctx_p1,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(ctx_p2,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
+  SetupLoDTensor(mark,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(mark_dict_len - 1));
 
   std::vector<paddle::framework::LoDTensor*> cpu_feeds;
   cpu_feeds.push_back(&word);

paddle/fluid/inference/tests/book/test_inference_understand_sentiment.cc

@@ -31,7 +31,12 @@ TEST(inference, understand_sentiment) {
 
   paddle::framework::LoDTensor words;
   paddle::framework::LoD lod{{0, 4, 10}};
-  SetupLoDTensor(words, lod, static_cast<int64_t>(0), static_cast<int64_t>(10));
+  int64_t word_dict_len = 5147;
+
+  SetupLoDTensor(words,
+                 lod,
+                 static_cast<int64_t>(0),
+                 static_cast<int64_t>(word_dict_len - 1));
 
   std::vector<paddle::framework::LoDTensor*> cpu_feeds;
   cpu_feeds.push_back(&words);

paddle/fluid/inference/tests/book/test_inference_word2vec.cc

@@ -31,12 +31,12 @@ TEST(inference, word2vec) {
 
   paddle::framework::LoDTensor first_word, second_word, third_word, fourth_word;
   paddle::framework::LoD lod{{0, 1}};
-  int64_t dict_size = 2072;  // Hard-coding the size of dictionary
+  int64_t dict_size = 2073;  // The size of dictionary
 
-  SetupLoDTensor(first_word, lod, static_cast<int64_t>(0), dict_size);
-  SetupLoDTensor(second_word, lod, static_cast<int64_t>(0), dict_size);
-  SetupLoDTensor(third_word, lod, static_cast<int64_t>(0), dict_size);
-  SetupLoDTensor(fourth_word, lod, static_cast<int64_t>(0), dict_size);
+  SetupLoDTensor(first_word, lod, static_cast<int64_t>(0), dict_size - 1);
+  SetupLoDTensor(second_word, lod, static_cast<int64_t>(0), dict_size - 1);
+  SetupLoDTensor(third_word, lod, static_cast<int64_t>(0), dict_size - 1);
+  SetupLoDTensor(fourth_word, lod, static_cast<int64_t>(0), dict_size - 1);
 
   std::vector<paddle::framework::LoDTensor*> cpu_feeds;
   cpu_feeds.push_back(&first_word);

paddle/fluid/inference/tests/test_helper.h

@@ -101,8 +101,8 @@ void TestInference(const std::string& dirname,
   if (IsCombined) {
     // All parameters are saved in a single file.
     // Hard-coding the file names of program and parameters in unittest.
-    // Users are free to specify different filename
-    // (provided: the filenames are changed in the python api as well: io.py)
+    // The file names should be consistent with that used in Python API
+    //  `fluid.io.save_inference_model`.
     std::string prog_filename = "__model_combined__";
     std::string param_filename = "__params_combined__";
     inference_program = paddle::inference::Load(executor,

paddle/fluid/operators/CMakeLists.txt

@@ -11,6 +11,8 @@ function(op_library TARGET)
     set(cc_srcs)
     set(cu_srcs)
     set(cu_cc_srcs)
+    set(cudnn_cu_cc_srcs)
+    set(CUDNN_FILE)
     set(op_common_deps operator op_registry math_function)
     set(options "")
     set(oneValueArgs "")
@@ -30,10 +32,16 @@ function(op_library TARGET)
         if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
             list(APPEND cu_srcs ${TARGET}.cu)
         endif()
+        string(REPLACE "_op" "_cudnn_op" CUDNN_FILE "${TARGET}")
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${CUDNN_FILE}.cu.cc)
+            list(APPEND cudnn_cu_cc_srcs ${CUDNN_FILE}.cu.cc)
+        endif()
     else()
         foreach(src ${op_library_SRCS})
             if (${src} MATCHES ".*\\.cu$")
                 list(APPEND cu_srcs ${src})
+            elseif(${src} MATCHES ".*_cudnn_op.cu.cc$")
+                list(APPEND cudnn_cu_cc_srcs ${src})
             elseif(${src} MATCHES ".*\\.cu.cc$")
                 list(APPEND cu_cc_srcs ${src})
             elseif(${src} MATCHES ".*\\.cc$")
@@ -54,7 +62,7 @@ function(op_library TARGET)
         set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE)
     endif()
     if (WITH_GPU)
-        nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
+        nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
                 ${op_common_deps})
     else()
         cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${op_library_DEPS}
@@ -98,6 +106,12 @@ function(op_library TARGET)
         set(pybind_flag 1)
     endif()
 
+    # pybind USE_OP_DEVICE_KERNEL for CUDNN
+    list(LENGTH cudnn_cu_cc_srcs cudnn_cu_cc_srcs_len)
+    if (WITH_GPU AND ${cudnn_cu_cc_srcs_len} GREATER 0)
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
+    endif()
+
     # pybind USE_OP
     if (${pybind_flag} EQUAL 0)
         file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
@@ -141,6 +155,7 @@ op_library(print_op DEPS lod_tensor)
 op_library(adagrad_op DEPS selected_rows_functor)
 op_library(maxout_op DEPS maxouting)
 op_library(unpool_op DEPS unpooling)
+op_library(pool_op DEPS pooling)
 op_library(pool_with_index_op DEPS pooling)
 op_library(lod_rank_table_op DEPS lod_rank_table)
 op_library(lod_tensor_to_array_op DEPS lod_rank_table_op)
@@ -152,43 +167,17 @@ op_library(lstm_op DEPS sequence2batch lstm_compute)
 op_library(lstmp_op DEPS sequence2batch lstm_compute)
 op_library(gru_op DEPS sequence2batch gru_compute)
 op_library(recurrent_op DEPS executor)
-op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale math_function)
+op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
 op_library(cos_sim_op DEPS cos_sim_functor)
 op_library(parallel_do_op DEPS executor)
 op_library(create_reader_op DEPS reader)
 
-# Regist multiple Kernel to pybind
 if (WITH_GPU)
-
-op_library(conv_op SRCS conv_op.cc conv_op.cu.cc conv_cudnn_op.cu.cc DEPS
-    vol2col depthwise_conv)
-
-op_library(edit_distance_op SRCS edit_distance_op.cc edit_distance_op.cu DEPS math_function)
-op_library(pool_op SRCS pool_op.cc pool_op.cu.cc pool_cudnn_op.cu.cc DEPS pooling)
-op_library(conv_transpose_op SRCS conv_transpose_op.cc conv_transpose_op.cu.cc
-  conv_transpose_cudnn_op.cu.cc DEPS vol2col)
-file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(conv2d, CUDNN);\n")
-file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(pool2d, CUDNN);\n")
-file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(conv2d_transpose, CUDNN);\n")
+    op_library(conv_op DEPS vol2col depthwise_conv)
 else()
-op_library(conv_op SRCS conv_op.cc DEPS vol2col)
-op_library(pool_op SRCS pool_op.cc DEPS pooling)
-op_library(conv_transpose_op SRCS conv_transpose_op.cc DEPS vol2col)
+    op_library(conv_op DEPS vol2col)
 endif()
-
-cc_library(batch_size_like SRCS batch_size_like.cc DEPS op_registry)
-
-op_library(fill_constant_batch_size_like_op
-  SRCS fill_constant_batch_size_like_op.cc fill_constant_batch_size_like_op.cu.cc
-  DEPS batch_size_like)
-
-op_library(uniform_random_batch_size_like_op
-  SRCS uniform_random_batch_size_like_op.cc
-  DEPS batch_size_like uniform_random_op)
-
-op_library(gaussian_random_batch_size_like_op
-  SRCS gaussian_random_batch_size_like_op.cc
-  DEPS batch_size_like gaussian_random_op)
+op_library(conv_transpose_op DEPS vol2col)
 
 # FIXME(typhoonzero): save/load depends lodtensor serialization functions
 op_library(save_op DEPS lod_tensor)

paddle/fluid/operators/batch_size_like.cc

-/* Copyright (c) 2018 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/fluid/operators/batch_size_like.h"
-
-namespace paddle {
-namespace operators {
-
-void BatchSizeLikeOp::InferShape(framework::InferShapeContext *ctx) const {
-  PADDLE_ENFORCE(ctx->HasInput("Input"),
-                 "Input(Input) of %s should not be null.", Type());
-  PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of %s should not be null.",
-                 Type());
-
-  auto &shape = ctx->Attrs().Get<std::vector<int>>("shape");
-  PADDLE_ENFORCE_GT(shape.size(), 0);
-  std::vector<int64_t> shape_int64(shape.size(), 0);
-  std::transform(shape.begin(), shape.end(), shape_int64.begin(),
-                 [](int a) { return static_cast<int64_t>(a); });
-  auto output_dim = framework::make_ddim(shape_int64);
-
-  int input_dim_idx = ctx->Attrs().Get<int>("input_dim_idx");
-  PADDLE_ENFORCE_GE(input_dim_idx, 0);
-  PADDLE_ENFORCE_GT(ctx->GetInputDim("Input").size(), input_dim_idx);
-
-  int output_dim_idx = ctx->Attrs().Get<int>("output_dim_idx");
-  PADDLE_ENFORCE_GE(output_dim_idx, 0);
-  PADDLE_ENFORCE_GT(static_cast<int>(shape.size()), output_dim_idx);
-
-  output_dim[output_dim_idx] = ctx->GetInputDim("Input")[input_dim_idx];
-  ctx->SetOutputDim("Out", output_dim);
-}
-
-BatchSizeLikeOpMaker::BatchSizeLikeOpMaker(OpProto *proto,
-                                           OpAttrChecker *op_checker)
-    : framework::OpProtoAndCheckerMaker(proto, op_checker) {
-  AddInput("Input",
-           "(Tensor) Tensor "
-           "whose input_dim_idx'th dimension specifies the batch_size");
-  AddOutput("Out",
-            "(Tensor) Tensor of specified shape will be filled "
-            "with the specified value");
-  AddAttr<std::vector<int>>("shape", "(vector<int>) The shape of the output");
-  AddAttr<int>("input_dim_idx",
-               "(int, default 0) The index of input's batch size dimension")
-      .SetDefault(0);
-  AddAttr<int>("output_dim_idx",
-               "(int, default 0) The index of output's batch size dimension")
-      .SetDefault(0);
-}
-
-}  // namespace operators
-}  // namespace paddle

paddle/fluid/operators/batch_size_like.h

@@ -24,12 +24,50 @@ class BatchSizeLikeOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
-  void InferShape(framework::InferShapeContext *ctx) const override;
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"),
+                   "Input(Input) of %s should not be null.", Type());
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of %s should not be null.", Type());
+
+    auto &shape = ctx->Attrs().Get<std::vector<int>>("shape");
+    PADDLE_ENFORCE_GT(shape.size(), 0);
+    std::vector<int64_t> shape_int64(shape.size(), 0);
+    std::transform(shape.begin(), shape.end(), shape_int64.begin(),
+                   [](int a) { return static_cast<int64_t>(a); });
+    auto output_dim = framework::make_ddim(shape_int64);
+
+    int input_dim_idx = ctx->Attrs().Get<int>("input_dim_idx");
+    PADDLE_ENFORCE_GE(input_dim_idx, 0);
+    PADDLE_ENFORCE_GT(ctx->GetInputDim("Input").size(), input_dim_idx);
+
+    int output_dim_idx = ctx->Attrs().Get<int>("output_dim_idx");
+    PADDLE_ENFORCE_GE(output_dim_idx, 0);
+    PADDLE_ENFORCE_GT(static_cast<int>(shape.size()), output_dim_idx);
+
+    output_dim[output_dim_idx] = ctx->GetInputDim("Input")[input_dim_idx];
+    ctx->SetOutputDim("Out", output_dim);
+  }
 };
 
 class BatchSizeLikeOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker);
+  BatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Input",
+             "(Tensor) Tensor "
+             "whose input_dim_idx'th dimension specifies the batch_size");
+    AddOutput("Out",
+              "(Tensor) Tensor of specified shape will be filled "
+              "with the specified value");
+    AddAttr<std::vector<int>>("shape", "(vector<int>) The shape of the output");
+    AddAttr<int>("input_dim_idx",
+                 "(int, default 0) The index of input's batch size dimension")
+        .SetDefault(0);
+    AddAttr<int>("output_dim_idx",
+                 "(int, default 0) The index of output's batch size dimension")
+        .SetDefault(0);
+  }
 };
 
 }  // namespace operators

paddle/fluid/operators/bipartite_match_op.cc

@@ -94,6 +94,38 @@ class BipartiteMatchKernel : public framework::OpKernel<T> {
     }
   }
 
+  void ArgMaxMatch(const Tensor& dist, int* match_indices, T* match_dist,
+                   T overlap_threshold) const {
+    constexpr T kEPS = static_cast<T>(1e-6);
+    int64_t row = dist.dims()[0];
+    int64_t col = dist.dims()[1];
+    auto* dist_data = dist.data<T>();
+    for (int64_t j = 0; j < col; ++j) {
+      if (match_indices[j] != -1) {
+        // the j-th column has been matched to one entity.
+        continue;
+      }
+      int max_row_idx = -1;
+      T max_dist = -1;
+      for (int i = 0; i < row; ++i) {
+        T dist = dist_data[i * col + j];
+        if (dist < kEPS) {
+          // distance is 0 between m-th row and j-th column
+          continue;
+        }
+        if (dist >= overlap_threshold && dist > max_dist) {
+          max_row_idx = i;
+          max_dist = dist;
+        }
+      }
+      if (max_row_idx != -1) {
+        PADDLE_ENFORCE_EQ(match_indices[j], -1);
+        match_indices[j] = max_row_idx;
+        match_dist[j] = max_dist;
+      }
+    }
+  }
+
   void Compute(const framework::ExecutionContext& context) const override {
     auto* dist_mat = context.Input<LoDTensor>("DistMat");
     auto* match_indices = context.Output<Tensor>("ColToRowMatchIndices");
@@ -120,13 +152,21 @@ class BipartiteMatchKernel : public framework::OpKernel<T> {
 
     int* indices = match_indices->data<int>();
     T* dist = match_dist->data<T>();
+    auto type = context.Attr<std::string>("match_type");
+    auto threshold = context.Attr<float>("dist_threshold");
     if (n == 1) {
       BipartiteMatch(*dist_mat, indices, dist);
+      if (type == "per_prediction") {
+        ArgMaxMatch(*dist_mat, indices, dist, threshold);
+      }
     } else {
       auto lod = dist_mat->lod().back();
       for (size_t i = 0; i < lod.size() - 1; ++i) {
         Tensor one_ins = dist_mat->Slice(lod[i], lod[i + 1]);
         BipartiteMatch(one_ins, indices + i * col, dist + i * col);
+        if (type == "per_prediction") {
+          ArgMaxMatch(one_ins, indices + i * col, dist + i * col, threshold);
+        }
       }
     }
   }
@@ -147,6 +187,19 @@ class BipartiteMatchOpMaker : public framework::OpProtoAndCheckerMaker {
         "This tensor can contain LoD information to represent a batch of "
         "inputs. One instance of this batch can contain different numbers of "
         "entities.");
+    AddAttr<std::string>(
+        "match_type",
+        "(string, defalut: per_prediction) "
+        "The type of matching method, should be 'bipartite' or "
+        "'per_prediction', 'bipartite' by defalut.")
+        .SetDefault("bipartite")
+        .InEnum({"bipartite", "per_prediction"});
+    AddAttr<float>(
+        "dist_threshold",
+        "(float, defalut: 0.5) "
+        "If `match_type` is 'per_prediction', this threshold is to determine "
+        "the extra matching bboxes based on the maximum distance.")
+        .SetDefault(0.5);
     AddOutput("ColToRowMatchIndices",
               "(Tensor) A 2-D Tensor with shape [N, M] in int type. "
               "N is the batch size. If ColToRowMatchIndices[i][j] is -1, it "
@@ -168,10 +221,10 @@ distance matrix. For input 2D matrix, the bipartite matching algorithm can
 find the matched column for each row, also can find the matched row for
 each column. And this operator only calculate matched indices from column
 to row. For each instance, the number of matched indices is the number of
-of columns of the input ditance matrix.
+of columns of the input distance matrix.
 
 There are two outputs to save matched indices and distance.
-A simple description, this algothrim matched the best (maximum distance)
+A simple description, this algorithm matched the best (maximum distance)
 row entity to the column entity and the matched indices are not duplicated
 in each row of ColToRowMatchIndices. If the column entity is not matched
 any row entity, set -1 in ColToRowMatchIndices.

paddle/fluid/operators/concat_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <utility>
 #include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/strided_memcpy.h"
@@ -34,6 +35,39 @@ class ConcatKernel : public framework::OpKernel<T> {
     auto out_stride = framework::stride_numel(out->dims());
 
     size_t output_offset = 0;
+
+    // If axis >=1, copy to out immediately need to call many times
+    // of cuda memcpy. Copy the input to cpu and do the stride copy,
+    // then copy to gpu output.
+
+    if (platform::is_gpu_place(place) && axis >= 1) {
+      platform::CPUPlace copy_place;
+      auto& cpu_ctx = *platform::DeviceContextPool::Instance().Get(copy_place);
+      framework::Tensor cpu_out;
+      cpu_out.Resize(out->dims());
+      cpu_out.mutable_data<T>(copy_place);
+      auto& dev_ctx = ctx.device_context();
+      std::vector<std::unique_ptr<framework::Tensor>> cpu_ins;
+      for (auto* in : ins) {
+        std::unique_ptr<framework::Tensor> cpu_in(new framework::Tensor);
+        framework::TensorCopy(*in, copy_place, dev_ctx, cpu_in.get());
+        cpu_ins.emplace_back(std::move(cpu_in));
+      }
+      // TODO(dzhwinter): overlap copy and compute stream
+      // https://devblogs.nvidia.com/how-overlap-data-transfers-cuda-cc/
+      dev_ctx.Wait();
+
+      for (auto& in : cpu_ins) {
+        auto& cpu_in = *in.get();
+        auto in_stride = framework::stride_numel(cpu_in.dims());
+
+        StridedNumelCopyWithAxis<T>(
+            cpu_ctx, axis, cpu_out.data<T>() + output_offset, out_stride,
+            cpu_in.data<T>(), in_stride, in_stride[axis]);
+        output_offset += in_stride[axis];
+      }
+      framework::TensorCopy(cpu_out, place, dev_ctx, out);
+    } else {
       for (auto* in : ins) {
         auto in_stride = framework::stride_numel(in->dims());
         StridedNumelCopyWithAxis<T>(ctx.device_context(), axis,
@@ -42,6 +76,7 @@ class ConcatKernel : public framework::OpKernel<T> {
         output_offset += in_stride[axis];
       }
     }
+  }
 };
 
 template <typename DeviceContext, typename T>

paddle/fluid/operators/conv_op.cc

@@ -54,12 +54,6 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
 
   std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
   for (size_t i = 0; i < strides.size(); ++i) {
-    PADDLE_ENFORCE(in_dims[i + 2] + 2 * paddings[i] -
-                           (dilations[i] * (filter_dims[i + 2] - 1) + 1) >
-                       0,
-                   "Due to the settings of paddings, filter_dims and "
-                   "dilations, the output size is less than 0, please check "
-                   "again.");
     output_shape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
                                           dilations[i], paddings[i],
                                           strides[i]));

paddle/fluid/operators/conv_op.h

@@ -31,7 +31,14 @@ using Tensor = framework::Tensor;
 inline int ConvOutputSize(int input_size, int filter_size, int dilation,
                           int padding, int stride) {
   const int dkernel = dilation * (filter_size - 1) + 1;
-  const int output_size = (input_size + 2 * padding - dkernel) / stride + 1;
+  int output_size = (input_size + 2 * padding - dkernel) / stride + 1;
+  PADDLE_ENFORCE(
+      output_size > 0,
+      "Due to the settings of padding(%d), filter_size(%d), dilation(%d) and "
+      "stride(%d), the output size is less than 0, please check "
+      "again. Input_size:%d",
+      padding, filter_size, dilation, stride, input_size);
+
   return output_size;
 }
 inline bool IsExpand(std::vector<int64_t>& filter_dim,

paddle/fluid/operators/detail/grpc_client.cc

@@ -177,8 +177,8 @@ std::shared_ptr<grpc::Channel> RPCClient::GetChannel(const std::string& ep) {
   args.SetMaxSendMessageSize(std::numeric_limits<int>::max());
   args.SetMaxReceiveMessageSize(std::numeric_limits<int>::max());
 
-  auto ch = std::shared_ptr<grpc::Channel>(
-      grpc::CreateCustomChannel(ep, grpc::InsecureChannelCredentials(), args));
+  auto ch =
+      grpc::CreateCustomChannel(ep, grpc::InsecureChannelCredentials(), args);
 
   channels_[ep] = ch;
   return ch;

paddle/fluid/operators/listen_and_serv_op.cc

@@ -129,6 +129,8 @@ class ListenAndServOp : public framework::OperatorBase {
       }
       if (exit_flag) {
         rpc_service_->ShutDown();
+        rpc_service_->SetCond(1);
+        break;
       }
       try {
         executor.Run(*program, &recv_scope, block->ID(), /*global_block*/

paddle/fluid/operators/nccl_op.cc

@@ -65,7 +65,7 @@ class NCCLInitOpVarTypeInference : public framework::VarTypeInference {
                   framework::BlockDesc *block) const override {
     auto out_var_name = op_desc.Output("Communicator").front();
     auto &out_var = block->FindRecursiveOrCreateVar(out_var_name);
-    auto var_type = framework::proto::VarType::NCCL_COM;
+    auto var_type = framework::proto::VarType::RAW;
     out_var.SetType(var_type);
   }
 };

paddle/fluid/operators/pool_op.cc

@@ -19,6 +19,11 @@ namespace operators {
 
 int PoolOutputSize(int input_size, int filter_size, int padding, int stride) {
   int output_size = (input_size - filter_size + 2 * padding) / stride + 1;
+  PADDLE_ENFORCE(output_size > 0,
+                 "Due to the settings of padding(%d), filter_size(%d) and "
+                 "stride(%d), the output size is less than 0, please check "
+                 "again. Input_size:%d",
+                 padding, filter_size, stride, input_size);
   return output_size;
 }
 

paddle/fluid/operators/reshape_op.cc

@@ -121,10 +121,15 @@ class ReshapeGradOp : public framework::OperatorWithKernel {
 }  // namespace operators
 }  // namespace paddle
 namespace ops = paddle::operators;
+using CPU = paddle::platform::CPUDeviceContext;
 
 REGISTER_OP(reshape, ops::ReshapeOp, ops::ReshapeOpMaker, reshape_grad,
             ops::ReshapeGradOp);
-REGISTER_OP_CPU_KERNEL(reshape,
-                       ops::ReshapeKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(
-    reshape_grad, ops::ReshapeGradKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(reshape, ops::ReshapeKernel<CPU, float>,
+                       ops::ReshapeKernel<CPU, double>,
+                       ops::ReshapeKernel<CPU, int>,
+                       ops::ReshapeKernel<CPU, int64_t>);
+REGISTER_OP_CPU_KERNEL(reshape_grad, ops::ReshapeGradKernel<CPU, float>,
+                       ops::ReshapeGradKernel<CPU, double>,
+                       ops::ReshapeGradKernel<CPU, int>,
+                       ops::ReshapeGradKernel<CPU, int64_t>);

paddle/fluid/operators/reshape_op.cu

@@ -13,10 +13,14 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/reshape_op.h"
+using CUDA = paddle::platform::CUDADeviceContext;
 
-REGISTER_OP_CUDA_KERNEL(
-    reshape,
-    paddle::operators::ReshapeKernel<paddle::platform::CUDAPlace, float>);
-REGISTER_OP_CUDA_KERNEL(
-    reshape_grad,
-    paddle::operators::ReshapeGradKernel<paddle::platform::CUDAPlace, float>);
+REGISTER_OP_CUDA_KERNEL(reshape, paddle::operators::ReshapeKernel<CUDA, float>,
+                        paddle::operators::ReshapeKernel<CUDA, double>,
+                        paddle::operators::ReshapeKernel<CUDA, int>,
+                        paddle::operators::ReshapeKernel<CUDA, int64_t>);
+REGISTER_OP_CUDA_KERNEL(reshape_grad,
+                        paddle::operators::ReshapeGradKernel<CUDA, float>,
+                        paddle::operators::ReshapeGradKernel<CUDA, double>,
+                        paddle::operators::ReshapeGradKernel<CUDA, int>,
+                        paddle::operators::ReshapeGradKernel<CUDA, int64_t>);

paddle/fluid/operators/send_op.cc

@@ -121,9 +121,27 @@ This operator will send tensor to recv_op at the parameter server.
   }
 };
 
+class SendOpVarTypeInference : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc& op_desc,
+                  framework::BlockDesc* block) const override {
+    auto out_var_name = op_desc.Output("RPCClient").front();
+    auto& out_var = block->FindRecursiveOrCreateVar(out_var_name);
+    auto var_type = framework::proto::VarType::RAW;
+    out_var.SetType(var_type);
+  }
+};
+
+class SendOpShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* ctx) const override {}
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 
-REGISTER_OPERATOR(send, ops::SendOp, ops::SendOpMaker);
+REGISTER_OPERATOR(send, ops::SendOp, paddle::framework::EmptyGradOpMaker,
+                  ops::SendOpMaker, ops::SendOpVarTypeInference,
+                  ops::SendOpShapeInference);

paddle/fluid/operators/send_recv_op_test.cc

@@ -95,7 +95,7 @@ void AddOp(const std::string &type, const f::VariableNameMap &inputs,
   for (auto kv : outputs) {
     for (auto v : kv.second) {
       auto var = block->Var(v);
-      var->SetDataType(f::proto::DataType::FP32);
+      var->SetDataType(f::proto::VarType::FP32);
     }
   }
 
@@ -122,33 +122,37 @@ void StartServerNet(bool is_sparse) {
 
   // sub program run in listen_and_serv_op, for simple test we use sum
   f::ProgramDesc program;
-  f::BlockDesc *block = program.MutableBlock(0);
+  f::BlockDesc *optimize_block = program.MutableBlock(0);
   // X for server side tensors, RX for received tensers, must be of same shape.
-  AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, block);
+  AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, optimize_block);
 
   f::AttributeMap attrs;
   attrs.insert({"endpoint", std::string("127.0.0.1:6174")});
+  attrs.insert({"Fanin", 1});
   attrs.insert({"ParamList", std::vector<std::string>({"Out"})});
   attrs.insert({"GradList", std::vector<std::string>({"x1"})});
-  attrs.insert({"OptimizeBlock", block});
+  attrs.insert({"OptimizeBlock", optimize_block});
   listen_and_serv_op =
-      f::OpRegistry::CreateOp("listen_and_serv", {}, {}, attrs);
+      f::OpRegistry::CreateOp("listen_and_serv", {{"X", {"x1"}}}, {}, attrs);
   listen_and_serv_op->Run(scope, place);
 }
 
 TEST(SendRecvOp, CPUDense) {
   std::thread server_thread(StartServerNet, false);
-  sleep(10);  // wait server to start
+  sleep(5);  // wait server to start
   // local net
   f::Scope scope;
   p::CPUPlace place;
   InitTensorsInScope(scope, place);
+  // create rpc client var
+  scope.Var("RPC_CLIENT_VAR");
 
   f::AttributeMap attrs;
   attrs.insert({"endpoints", std::vector<std::string>({"127.0.0.1:6174"})});
   attrs.insert({"epmap", std::vector<std::string>({"127.0.0.1:6174"})});
-  auto send_op = f::OpRegistry::CreateOp("send", {{"X", {"x1"}}},
-                                         {{"Out", {"Out"}}}, attrs);
+  auto send_op = f::OpRegistry::CreateOp(
+      "send", {{"X", {"x1"}}},
+      {{"Out", {"Out"}}, {"RPCClient", {"RPC_CLIENT_VAR"}}}, attrs);
   send_op->Run(scope, place);
 
   auto in_var = scope.Var("x1");
@@ -175,11 +179,13 @@ TEST(SendRecvOp, CPUSparse) {
   p::CPUPlace place;
   p::CPUDeviceContext ctx(place);
   InitSelectedRowsInScope(scope, place);
+  scope.Var("RPC_CLIENT_VAR");
   f::AttributeMap attrs;
   attrs.insert({"endpoints", std::vector<std::string>({"127.0.0.1:6174"})});
   attrs.insert({"epmap", std::vector<std::string>({"127.0.0.1:6174"})});
-  auto send_op = f::OpRegistry::CreateOp("send", {{"X", {"x1"}}},
-                                         {{"Out", {"Out"}}}, attrs);
+  auto send_op = f::OpRegistry::CreateOp(
+      "send", {{"X", {"x1"}}},
+      {{"Out", {"Out"}}, {"RPCClient", {"RPC_CLIENT_VAR"}}}, attrs);
   send_op->Run(scope, place);
 
   auto x0 = scope.Var("x0")->GetMutable<f::SelectedRows>();

paddle/fluid/platform/CMakeLists.txt

+proto_library(profiler_proto SRCS profiler.proto)
+
 if(WITH_GPU)
   cc_library(enforce SRCS enforce.cc DEPS)
 else()
@@ -37,7 +39,8 @@ nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
 nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place device_context)
 nv_test(nccl_test SRCS nccl_test.cu DEPS dynload_cuda gpu_info device_context)
 
-cc_library(profiler SRCS profiler.cc DEPS device_context)
+cc_library(device_tracer SRCS device_tracer.cc DEPS profiler_proto ${GPU_CTX_DEPS})
+cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer)
 cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
 
 nv_test(float16_gpu_test SRCS float16_test.cu)

paddle/fluid/platform/device_tracer.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/platform/device_tracer.h"
+#include <map>
+#include <mutex>
+#include "glog/logging.h"
+#include "paddle/fluid/framework/block_desc.h"
+#include "paddle/fluid/string/printf.h"
+
+namespace paddle {
+namespace platform {
+namespace {
+
+thread_local const char *cur_annotation = nullptr;
+std::once_flag tracer_once_flag;
+DeviceTracer *tracer = nullptr;
+}  // namespace
+#ifdef PADDLE_WITH_CUPTI
+
+namespace {
+// TODO(panyx0718): Revisit the buffer size here.
+uint64_t kBufSize = 32 * 1024;
+uint64_t kAlignSize = 8;
+
+#define ALIGN_BUFFER(buffer, align)                                 \
+  (((uintptr_t)(buffer) & ((align)-1))                              \
+       ? ((buffer) + (align) - ((uintptr_t)(buffer) & ((align)-1))) \
+       : (buffer))
+
+#define CUPTI_CALL(call)                                                   \
+  do {                                                                     \
+    CUptiResult _status = call;                                            \
+    if (_status != CUPTI_SUCCESS) {                                        \
+      const char *errstr;                                                  \
+      dynload::cuptiGetResultString(_status, &errstr);                     \
+      fprintf(stderr, "%s:%d: error: function %s failed with error %s.\n", \
+              __FILE__, __LINE__, #call, errstr);                          \
+      exit(-1);                                                            \
+    }                                                                      \
+  } while (0)
+
+void EnableActivity() {
+  // Device activity record is created when CUDA initializes, so we
+  // want to enable it before cuInit() or any CUDA runtime call.
+  CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMCPY));
+  CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_KERNEL));
+  CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DEVICE));
+  CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMSET));
+  CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_OVERHEAD));
+  // We don't track these activities for now.
+  // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_CONTEXT));
+  // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DRIVER));
+  // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_RUNTIME));
+  // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_NAME));
+  // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MARKER));
+}
+
+void DisableActivity() {
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MEMCPY));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_KERNEL));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_DEVICE));
+  // Disable all other activity record kinds.
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_CONTEXT));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_DRIVER));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_RUNTIME));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MEMSET));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_NAME));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MARKER));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_OVERHEAD));
+}
+
+void CUPTIAPI bufferRequested(uint8_t **buffer, size_t *size,
+                              size_t *maxNumRecords) {
+  uint8_t *buf = (uint8_t *)malloc(kBufSize + kAlignSize);
+  *size = kBufSize;
+  *buffer = ALIGN_BUFFER(buf, kAlignSize);
+  *maxNumRecords = 0;
+}
+
+void CUPTIAPI bufferCompleted(CUcontext ctx, uint32_t streamId, uint8_t *buffer,
+                              size_t size, size_t validSize) {
+  CUptiResult status;
+  CUpti_Activity *record = NULL;
+  if (validSize > 0) {
+    do {
+      status = dynload::cuptiActivityGetNextRecord(buffer, validSize, &record);
+      if (status == CUPTI_SUCCESS) {
+        switch (record->kind) {
+          case CUPTI_ACTIVITY_KIND_KERNEL:
+          case CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL: {
+            auto *kernel =
+                reinterpret_cast<const CUpti_ActivityKernel3 *>(record);
+            tracer->AddKernelRecords(kernel->start, kernel->end,
+                                     kernel->deviceId, kernel->streamId,
+                                     kernel->correlationId);
+            break;
+          }
+          default: { break; }
+        }
+      } else if (status == CUPTI_ERROR_MAX_LIMIT_REACHED) {
+        // Seems not an error in this case.
+        break;
+      } else {
+        CUPTI_CALL(status);
+      }
+    } while (1);
+
+    size_t dropped;
+    CUPTI_CALL(
+        dynload::cuptiActivityGetNumDroppedRecords(ctx, streamId, &dropped));
+    if (dropped != 0) {
+      fprintf(stderr, "Dropped %u activity records\n", (unsigned int)dropped);
+    }
+  }
+  free(buffer);
+}
+}  // namespace
+
+class DeviceTracerImpl : public DeviceTracer {
+ public:
+  DeviceTracerImpl() : enabled_(false) {}
+
+  void AddAnnotation(uint64_t id, const std::string &anno) {
+    std::lock_guard<std::mutex> l(trace_mu_);
+    correlations_[id] = anno;
+  }
+
+  void AddKernelRecords(uint64_t start, uint64_t end, uint32_t device_id,
+                        uint32_t stream_id, uint32_t correlation_id) {
+    std::lock_guard<std::mutex> l(trace_mu_);
+    kernel_records_.push_back(
+        KernelRecord{start, end, device_id, stream_id, correlation_id});
+  }
+
+  bool IsEnabled() {
+    std::lock_guard<std::mutex> l(trace_mu_);
+    return enabled_;
+  }
+
+  void Enable() {
+    std::lock_guard<std::mutex> l(trace_mu_);
+    if (enabled_) {
+      fprintf(stderr, "DeviceTracer already enabled\n");
+      return;
+    }
+    EnableActivity();
+
+    // Register callbacks for buffer requests and completed by CUPTI.
+    CUPTI_CALL(dynload::cuptiActivityRegisterCallbacks(bufferRequested,
+                                                       bufferCompleted));
+
+    CUptiResult ret;
+    ret = dynload::cuptiSubscribe(
+        &subscriber_, static_cast<CUpti_CallbackFunc>(ApiCallback), this);
+    if (ret == CUPTI_ERROR_MAX_LIMIT_REACHED) {
+      fprintf(stderr, "CUPTI subcriber limit reached.\n");
+    } else if (ret != CUPTI_SUCCESS) {
+      fprintf(stderr, "Failed to create CUPTI subscriber.\n");
+    }
+    CUPTI_CALL(
+        dynload::cuptiEnableCallback(1, subscriber_, CUPTI_CB_DOMAIN_DRIVER_API,
+                                     CUPTI_DRIVER_TRACE_CBID_cuLaunchKernel));
+
+    CUPTI_CALL(dynload::cuptiGetTimestamp(&start_ns_));
+    enabled_ = true;
+  }
+
+  proto::Profile GenProfile() {
+    std::lock_guard<std::mutex> l(trace_mu_);
+    proto::Profile profile_pb;
+    profile_pb.set_start_ns(start_ns_);
+    profile_pb.set_end_ns(end_ns_);
+    std::map<std::string, std::vector<uint64_t>> event_times;
+    for (const KernelRecord &r : kernel_records_) {
+      if (correlations_.find(r.correlation_id) == correlations_.end()) {
+        fprintf(stderr, "cannot relate a kernel activity\n");
+        continue;
+      }
+      auto *event = profile_pb.add_events();
+      event->set_name(correlations_.at(r.correlation_id));
+      event->set_start_ns(r.start_ns);
+      event->set_end_ns(r.end_ns);
+      event->set_stream_id(r.stream_id);
+      event->set_device_id(r.device_id);
+      event_times[event->name()].push_back(r.end_ns - r.start_ns);
+    }
+    for (const auto &et : event_times) {
+      fprintf(
+          stderr, "%s: total: %fms invoked cuda kernels: %lu\n",
+          et.first.c_str(),
+          std::accumulate(et.second.begin(), et.second.end(), 0) / 1000000.0,
+          et.second.size());
+    }
+    return profile_pb;
+  }
+
+  void Disable() {
+    // flush might cause additional calls to DeviceTracker.
+    dynload::cuptiActivityFlushAll(CUPTI_ACTIVITY_FLAG_FLUSH_FORCED);
+    std::lock_guard<std::mutex> l(trace_mu_);
+    DisableActivity();
+    dynload::cuptiUnsubscribe(subscriber_);
+    CUPTI_CALL(dynload::cuptiGetTimestamp(&end_ns_));
+    PADDLE_ENFORCE(dynload::cuptiFinalize());
+    enabled_ = false;
+  }
+
+ private:
+  static void CUPTIAPI ApiCallback(void *userdata, CUpti_CallbackDomain domain,
+                                   CUpti_CallbackId cbid, const void *cbdata) {
+    auto *cbInfo = reinterpret_cast<const CUpti_CallbackData *>(cbdata);
+    DeviceTracer *tracer = reinterpret_cast<DeviceTracer *>(userdata);
+
+    if ((domain == CUPTI_CB_DOMAIN_DRIVER_API) &&
+        (cbid == CUPTI_DRIVER_TRACE_CBID_cuLaunchKernel)) {
+      if (cbInfo->callbackSite == CUPTI_API_ENTER) {
+        const std::string anno =
+            cur_annotation ? cur_annotation : cbInfo->symbolName;
+        tracer->AddAnnotation(cbInfo->correlationId, anno);
+      }
+    } else {
+      VLOG(1) << "Unhandled API Callback for " << domain << " " << cbid;
+    }
+  }
+
+  std::mutex trace_mu_;
+  bool enabled_;
+  uint64_t start_ns_;
+  uint64_t end_ns_;
+  std::vector<KernelRecord> kernel_records_;
+  std::unordered_map<uint32_t, std::string> correlations_;
+  CUpti_SubscriberHandle subscriber_;
+};
+
+#endif  // PADDLE_WITH_CUPTI
+
+class DeviceTracerDummy : public DeviceTracer {
+ public:
+  DeviceTracerDummy() {}
+
+  void AddAnnotation(uint64_t id, const std::string &anno) {}
+
+  void AddKernelRecords(uint64_t start, uint64_t end, uint32_t device_id,
+                        uint32_t stream_id, uint32_t correlation_id) {}
+
+  bool IsEnabled() { return false; }
+
+  void Enable() {}
+
+  proto::Profile GenProfile() { return proto::Profile(); }
+
+  void Disable() {}
+};
+
+void CreateTracer(DeviceTracer **t) {
+#ifdef PADDLE_WITH_CUPTI
+  *t = new DeviceTracerImpl();
+#else
+  *t = new DeviceTracerDummy();
+#endif  // PADDLE_WITH_CUPTI
+}
+
+DeviceTracer *GetDeviceTracer() {
+  std::call_once(tracer_once_flag, CreateTracer, &tracer);
+  return tracer;
+}
+
+void SetCurAnnotation(const char *anno) { cur_annotation = anno; }
+
+void ClearCurAnnotation() { cur_annotation = nullptr; }
+
+}  // namespace platform
+}  // namespace paddle

paddle/fluid/platform/device_tracer.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/fluid/platform/dynload/cupti.h"
+#include "paddle/fluid/platform/profiler.pb.h"
+
+namespace paddle {
+namespace platform {
+
+///////////////////////
+// WARN: Under Development. Don't depend on it yet.
+//////////////////////
+
+// DeviceTracer performs the following tasks:
+// 1. Register cuda callbacks for various events: kernel, memcpy, etc.
+// 2. Collect cuda statistics: start/end ts, memory, etc.
+// 3. Generate a protobuf for further analysis.
+class DeviceTracer {
+ public:
+  struct KernelRecord {
+    uint64_t start_ns;
+    uint64_t end_ns;
+    uint32_t device_id;
+    uint32_t stream_id;
+    uint32_t correlation_id;
+  };
+
+  virtual ~DeviceTracer() {}
+  // Needs to be called once before use.
+  virtual void Enable() = 0;
+  // Needs to be called once after use.
+  virtual void Disable() = 0;
+
+  // Add a pair to correlate internal cuda id with high level
+  // annotation (string). So cuda statistics can be represented by
+  // human-readable annotations.
+  virtual void AddAnnotation(uint64_t id, const std::string& anno) = 0;
+
+  // Add a cuda kernel stats. `correlation_id` will be mapped to annotation
+  // added before for human readability.
+  virtual void AddKernelRecords(uint64_t start, uint64_t end,
+                                uint32_t device_id, uint32_t stream_id,
+                                uint32_t correlation_id) = 0;
+
+  // Generate a proto after done (Disabled).
+  virtual proto::Profile GenProfile() = 0;
+
+  virtual bool IsEnabled() = 0;
+};
+
+// Get a DeviceTracer.
+DeviceTracer* GetDeviceTracer();
+
+// Set a name for the cuda kernel operation being launched by the thread.
+void SetCurAnnotation(const char* anno);
+// Clear the name after the operation is done.
+void ClearCurAnnotation();
+
+}  // namespace platform
+}  // namespace paddle

paddle/fluid/platform/dynload/CMakeLists.txt

 cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags enforce)
-nv_library(dynload_cuda SRCS cublas.cc cudnn.cc curand.cc nccl.cc
-        DEPS dynamic_loader)
+
+list(APPEND CUDA_SRCS cublas.cc cudnn.cc curand.cc nccl.cc)
+if (CUPTI_FOUND)
+    list(APPEND CUDA_SRCS cupti.cc)
+endif(CUPTI_FOUND)
+nv_library(dynload_cuda SRCS ${CUDA_SRCS} DEPS dynamic_loader)
 cc_library(dynload_warpctc SRCS warpctc.cc DEPS dynamic_loader warpctc)

paddle/fluid/platform/dynload/cupti.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#ifdef PADDLE_WITH_CUPTI
+
+#include "paddle/fluid/platform/dynload/cupti.h"
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace platform {
+namespace dynload {
+
+std::once_flag cupti_dso_flag;
+void *cupti_dso_handle = nullptr;
+
+#define DEFINE_WRAP(__name) DynLoad__##__name __name
+
+CUPTI_ROUTINE_EACH(DEFINE_WRAP);
+
+}  // namespace dynload
+}  // namespace platform
+}  // namespace paddle
+
+#endif  // PADDLE_WITH_CUPTI

paddle/fluid/platform/dynload/cupti.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#ifdef PADDLE_WITH_CUPTI
+#include <cuda.h>
+#include <cupti.h>
+#include <dlfcn.h>
+#include <mutex>
+#include "paddle/fluid/platform/dynload/dynamic_loader.h"
+
+namespace paddle {
+namespace platform {
+namespace dynload {
+
+extern std::once_flag cupti_dso_flag;
+extern void *cupti_dso_handle;
+
+/**
+ * The following macro definition can generate structs
+ * (for each function) to dynamic load cupti routine
+ * via operator overloading.
+ *
+ * note: default dynamic linked libs
+ */
+#ifdef PADDLE_USE_DSO
+#define DECLARE_DYNAMIC_LOAD_CUPTI_WRAP(__name)                    \
+  struct DynLoad__##__name {                                       \
+    template <typename... Args>                                    \
+    inline CUptiResult CUPTIAPI operator()(Args... args) {         \
+      typedef CUptiResult CUPTIAPI (*cuptiFunc)(Args...);          \
+      std::call_once(cupti_dso_flag,                               \
+                     paddle::platform::dynload::GetCUPTIDsoHandle, \
+                     &cupti_dso_handle);                           \
+      void *p_##__name = dlsym(cupti_dso_handle, #__name);         \
+      return reinterpret_cast<cuptiFunc>(p_##__name)(args...);     \
+    }                                                              \
+  };                                                               \
+  extern DynLoad__##__name __name
+#else
+#define DECLARE_DYNAMIC_LOAD_CUPTI_WRAP(__name)            \
+  struct DynLoad__##__name {                               \
+    template <typename... Args>                            \
+    inline CUptiResult CUPTIAPI operator()(Args... args) { \
+      return __name(args...);                              \
+    }                                                      \
+  };                                                       \
+  extern DynLoad__##__name __name
+#endif
+
+#define CUPTI_ROUTINE_EACH(__macro)           \
+  __macro(cuptiActivityEnable);               \
+  __macro(cuptiActivityDisable);              \
+  __macro(cuptiActivityRegisterCallbacks);    \
+  __macro(cuptiActivityGetAttribute);         \
+  __macro(cuptiActivitySetAttribute);         \
+  __macro(cuptiGetTimestamp);                 \
+  __macro(cuptiActivityGetNextRecord);        \
+  __macro(cuptiGetResultString);              \
+  __macro(cuptiActivityGetNumDroppedRecords); \
+  __macro(cuptiActivityFlushAll);             \
+  __macro(cuptiFinalize);                     \
+  __macro(cuptiSubscribe);                    \
+  __macro(cuptiUnsubscribe);                  \
+  __macro(cuptiEnableCallback);
+
+CUPTI_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUPTI_WRAP);
+
+#undef DECLARE_DYNAMIC_LOAD_CUPTI_WRAP
+}  // namespace dynload
+}  // namespace platform
+}  // namespace paddle
+
+#endif  // PADDLE_WITH_CUPTI

paddle/fluid/platform/dynload/dynamic_loader.cc

@@ -40,10 +40,14 @@ DEFINE_string(nccl_dir, "",
               "libcurand. For instance, /usr/local/cuda/lib64. If default, "
               "dlopen will search cuda from LD_LIBRARY_PATH");
 
+DEFINE_string(cupti_dir, "", "Specify path for loading cupti.so.");
+
 namespace paddle {
 namespace platform {
 namespace dynload {
 
+static const char* cupti_lib_path = CUPTI_LIB_PATH;
+
 static inline std::string join(const std::string& part1,
                                const std::string& part2) {
   // directory separator
@@ -143,6 +147,18 @@ void GetCUDNNDsoHandle(void** dso_handle) {
 #endif
 }
 
+void GetCUPTIDsoHandle(void** dso_handle) {
+  std::string cupti_path = cupti_lib_path;
+  if (!FLAGS_cupti_dir.empty()) {
+    cupti_path = FLAGS_cupti_dir;
+  }
+#if defined(__APPLE__) || defined(__OSX__)
+  GetDsoHandleFromSearchPath(cupti_path, "libcupti.dylib", dso_handle, false);
+#else
+  GetDsoHandleFromSearchPath(cupti_path, "libcupti.so", dso_handle, false);
+#endif
+}
+
 void GetCurandDsoHandle(void** dso_handle) {
 #if defined(__APPLE__) || defined(__OSX__)
   GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcurand.dylib", dso_handle);

paddle/fluid/platform/dynload/dynamic_loader.h

@@ -34,6 +34,8 @@ void GetCublasDsoHandle(void** dso_handle);
  */
 void GetCUDNNDsoHandle(void** dso_handle);
 
+void GetCUPTIDsoHandle(void** dso_handle);
+
 /**
  * @brief    load the DSO of CURAND
  *

paddle/fluid/platform/profiler.cc

@@ -15,7 +15,13 @@ limitations under the License. */
 #include "paddle/fluid/platform/profiler.h"
 #include <iomanip>
 #include <map>
+#ifdef PADDLE_WITH_CUDA
+#include <cuda.h>
+#endif  // PADDLE_WITH_CUDA
 #include "glog/logging.h"
+#include "paddle/fluid/framework/block_desc.h"
+#include "paddle/fluid/platform/device_tracer.h"
+#include "paddle/fluid/string/printf.h"
 
 namespace paddle {
 namespace platform {
@@ -126,15 +132,20 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx) {
   GetEventList().Record(EventKind::kPopRange, name, g_thread_id, dev_ctx);
 }
 
-RecordEvent::RecordEvent(const std::string& name,
-                         const DeviceContext* dev_ctx) {
+RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx,
+                         int32_t block_id) {
   if (g_state == ProfilerState::kDisabled) return;
   dev_ctx_ = dev_ctx;
   name_ = name;
   PushEvent(name_, dev_ctx_);
+
+  full_name_ = string::Sprintf("%s_b%d", name, block_id);
+  // Maybe need the same push/pop behavior.
+  SetCurAnnotation(full_name_.c_str());
 }
 
 RecordEvent::~RecordEvent() {
+  ClearCurAnnotation();
   if (g_state == ProfilerState::kDisabled) return;
   PopEvent(name_, dev_ctx_);
 }
@@ -147,7 +158,14 @@ void EnableProfiler(ProfilerState state) {
                  "The profiling state should be disabled when calling ",
                  "EnableProfiler.");
   g_state = state;
-  g_profiler_place = (g_state == ProfilerState::kCUDA) ? "CUDA" : "CPU";
+  if (g_state == ProfilerState::kCUDA) {
+    g_profiler_place = "CUDA";
+  } else if (g_state == ProfilerState::kCPU) {
+    g_profiler_place = "CPU";
+  } else {
+    g_profiler_place = "All";
+    GetDeviceTracer()->Enable();
+  }
 #ifdef PADDLE_WITH_CUDA
   if (g_state == ProfilerState::kCUDA) {
     // Generate some dummy evenets first to reduce the startup overhead.
@@ -190,6 +208,12 @@ void DisableProfiler(EventSortingKey sorted_key) {
   Mark("_stop_profiler_", nullptr);
   g_state = ProfilerState::kDisabled;
 
+  DeviceTracer* tracer = GetDeviceTracer();
+  if (g_profiler_place == "All" && tracer && tracer->IsEnabled()) {
+    tracer->Disable();
+    tracer->GenProfile();
+  }
+
   std::vector<std::vector<Event>> all_events = GetAllEvents();
   ParseEvents(all_events, sorted_key);
   ResetProfiler();
@@ -254,9 +278,11 @@ void ParseEvents(std::vector<std::vector<Event>>& events,
         }
 
         if (rit != pushed_events.rend()) {
-          double event_time = (g_profiler_place == "CUDA")
+          double event_time =
+              (g_profiler_place == "CUDA" || g_profiler_place == "All")
                   ? rit->CudaElapsedMs(events[i][j])
                   : rit->CpuElapsedMs(events[i][j]);
+
           std::string event_name =
               "thread" + std::to_string(rit->thread_id()) + "::" + rit->name();
           max_name_width = std::max(max_name_width, event_name.size());

paddle/fluid/platform/profiler.h

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <mutex>
 #include <vector>
 #include "paddle/fluid/platform/device_context.h"
+#include "paddle/fluid/platform/profiler.pb.h"
 
 namespace paddle {
 namespace platform {
@@ -93,6 +94,7 @@ enum ProfilerState {
   kDisabled,  // disabled state
   kCPU,       // CPU profiling state
   kCUDA,      // GPU profiling state
+  kAll,       // Profile both CPU and GPU. (Currently experimental).
 };
 
 void Mark(const std::string& name, const DeviceContext* dev_ctx);
@@ -102,7 +104,8 @@ void PushEvent(const std::string& name, const DeviceContext* dev_ctx);
 void PopEvent(const std::string& name, const DeviceContext* dev_ctx);
 
 struct RecordEvent {
-  explicit RecordEvent(const std::string& name, const DeviceContext* dev_ctx);
+  RecordEvent(const std::string& name, const DeviceContext* dev_ctx,
+              int32_t block_id);
 
   ~RecordEvent();
 
@@ -110,9 +113,12 @@ struct RecordEvent {
   const DeviceContext* dev_ctx_;
   // Event name
   std::string name_;
+  // Need to distinguish name by op type, block_id, program_id and perhaps
+  // different kernel invocations within an op.
+  std::string full_name_;
 };
 
-// Return the event list of all threads. Asummed the returned value calls
+// Return the event list of all threads. Assumed the returned value calls
 // event_lists, event_lists[i][j] represents the j-th Event of i-th thread.
 std::vector<std::vector<Event>> GetAllEvents();
 

paddle/fluid/platform/profiler.proto

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+syntax = "proto2";
+package paddle.platform.proto;
+
+message Event {
+  optional string name = 1;
+  optional uint64 start_ns = 2;
+  optional uint64 end_ns = 3;
+  optional uint32 device_id = 5;
+  optional uint32 stream_id = 6;
+}
+
+message Profile {
+  repeated Event events = 1;
+  optional uint64 start_ns = 2;
+  optional uint64 end_ns = 3;
+}
\ No newline at end of file

paddle/fluid/platform/profiler_test.cc

@@ -95,7 +95,7 @@ TEST(RecordEvent, RecordEvent) {
    */
   for (int i = 1; i < 5; ++i) {
     std::string name = "evs_op_" + std::to_string(i);
-    RecordEvent record_event(name, dev_ctx);
+    RecordEvent record_event(name, dev_ctx, 0);
     int counter = 1;
     while (counter != i * 1000) counter++;
   }

paddle/fluid/pybind/protobuf.cc

@@ -252,7 +252,7 @@ void BindVarDsec(py::module &m) {
       .value("CHANNEL", proto::VarType::CHANNEL)
       .value("PLACE_LIST", proto::VarType::PLACE_LIST)
       .value("READER", proto::VarType::READER)
-      .value("NCCL_COM", proto::VarType::NCCL_COM);
+      .value("RAW", proto::VarType::RAW);
 }
 
 void BindOpDesc(py::module &m) {

paddle/fluid/pybind/pybind.cc

@@ -459,6 +459,7 @@ All parameter, weight, gradient are variables in Paddle.
       .value("kDisabled", platform::ProfilerState::kDisabled)
       .value("kCPU", platform::ProfilerState::kCPU)
       .value("kCUDA", platform::ProfilerState::kCUDA)
+      .value("kAll", platform::ProfilerState::kAll)
       .export_values();
 
   py::enum_<platform::EventSortingKey>(m, "EventSortingKey", py::arithmetic())

paddle/scripts/docker/README.md

@@ -58,7 +58,7 @@ Users can specify the following Docker build arguments with either "ON" or "OFF"
 | `WITH_AVX` | OFF | Set to "ON" to enable AVX support. |
 | `WITH_TESTING` | OFF | Build unit tests binaries. |
 | `WITH_MKL` | ON | Build with [Intel® MKL](https://software.intel.com/en-us/mkl) and [Intel® MKL-DNN](https://github.com/01org/mkl-dnn) support. |
-| `WITH_GOLANG` | ON | Build fault-tolerant parameter server written in go. |
+| `WITH_GOLANG` | OFF | Build fault-tolerant parameter server written in go. |
 | `WITH_SWIG_PY` | ON | Build with SWIG python API support. |
 | `WITH_C_API` | OFF | Build capi libraries for inference. |
 | `WITH_PYTHON` | ON | Build with python support. Turn this off if build is only for capi. |

paddle/scripts/docker/build.sh

@@ -40,7 +40,7 @@ function cmake_gen() {
         -DWITH_DISTRIBUTE=${WITH_DISTRIBUTE:-OFF}
         -DWITH_MKL=${WITH_MKL:-ON}
         -DWITH_AVX=${WITH_AVX:-OFF}
-        -DWITH_GOLANG=${WITH_GOLANG:-ON}
+        -DWITH_GOLANG=${WITH_GOLANG:-OFF}
         -DCUDA_ARCH_NAME=${CUDA_ARCH_NAME:-All}
         -DWITH_SWIG_PY=ON
         -DWITH_C_API=${WITH_C_API:-OFF}
@@ -49,6 +49,7 @@ function cmake_gen() {
         -DCUDNN_ROOT=/usr/
         -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON}
         -DWITH_TESTING=${WITH_TESTING:-ON}
+        -DWITH_FAST_BUNDLE_TEST=ON
         -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
     ========================================
 EOF
@@ -64,7 +65,7 @@ EOF
         -DWITH_DISTRIBUTE=${WITH_DISTRIBUTE:-OFF} \
         -DWITH_MKL=${WITH_MKL:-ON} \
         -DWITH_AVX=${WITH_AVX:-OFF} \
-        -DWITH_GOLANG=${WITH_GOLANG:-ON} \
+        -DWITH_GOLANG=${WITH_GOLANG:-OFF} \
         -DCUDA_ARCH_NAME=${CUDA_ARCH_NAME:-All} \
         -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} \
         -DWITH_C_API=${WITH_C_API:-OFF} \
@@ -72,6 +73,7 @@ EOF
         -DCUDNN_ROOT=/usr/ \
         -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON} \
         -DWITH_TESTING=${WITH_TESTING:-ON} \
+        -DWITH_FAST_BUNDLE_TEST=ON \
         -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
 }
 

paddle/testing/paddle_gtest_main.cc

@@ -28,10 +28,9 @@ int main(int argc, char** argv) {
   }
 #ifdef PADDLE_WITH_CUDA
   new_argv.push_back(
-      strdup("--tryfromenv=fraction_of_gpu_memory_to_use,use_pinned_memory,"
-             "warpctc_dir"));
+      strdup("--tryfromenv=fraction_of_gpu_memory_to_use,use_pinned_memory"));
 #else
-  new_argv.push_back(strdup("--tryfromenv=use_pinned_memory,warpctc_dir"));
+  new_argv.push_back(strdup("--tryfromenv=use_pinned_memory"));
 #endif
   int new_argc = static_cast<int>(new_argv.size());
   char** new_argv_address = new_argv.data();

python/paddle/fluid/distribute_transpiler.py

@@ -226,8 +226,7 @@ class DistributeTranspiler:
         rpc_client_var = program.global_block().create_var(
             name="RPC_CLIENT_VAR",
             persistable=True,
-            dtype='float32',  # dtype and shape is not used in fact
-            shape=[0])
+            type=core.VarDesc.VarType.RAW)
 
         # create send_op
         program.global_block().append_op(

python/paddle/fluid/framework.py

@@ -784,6 +784,7 @@ class Block(object):
         elif type(v) == Variable:
             var = Variable(
                 self,
+                type=v.type,
                 name=new_name,
                 error_clip=error_clip,
                 stop_gradient=stop_gradient)

python/paddle/fluid/io.py

@@ -68,7 +68,7 @@ def save_vars(executor,
               main_program=None,
               vars=None,
               predicate=None,
-              save_file_name=None):
+              filename=None):
     """
     Save variables to directory by executor.
 
@@ -80,7 +80,7 @@ def save_vars(executor,
     as a bool. If it returns true, the corresponding input variable will be saved.
     :param vars: variables need to be saved. If vars is specified, program & predicate
     will be ignored
-    :param save_file_name: The name of a single file that all vars are saved to. 
+    :param filename: The name of a single file that all vars are saved to.
         If it is None, save variables to separate files.
 
     :return: None
@@ -95,7 +95,7 @@ def save_vars(executor,
             executor,
             dirname=dirname,
             vars=filter(predicate, main_program.list_vars()),
-            save_file_name=save_file_name)
+            filename=filename)
     else:
         save_program = Program()
         save_block = save_program.global_block()
@@ -103,7 +103,7 @@ def save_vars(executor,
         save_var_map = {}
         for each_var in vars:
             new_var = _clone_var_in_block_(save_block, each_var)
-            if save_file_name is None:
+            if filename is None:
                 save_block.append_op(
                     type='save',
                     inputs={'X': [new_var]},
@@ -112,7 +112,7 @@ def save_vars(executor,
             else:
                 save_var_map[new_var.name] = new_var
 
-        if save_file_name is not None:
+        if filename is not None:
             save_var_list = []
             for name in sorted(save_var_map.keys()):
                 save_var_list.append(save_var_map[name])
@@ -121,12 +121,12 @@ def save_vars(executor,
                 type='save_combine',
                 inputs={'X': save_var_list},
                 outputs={},
-                attrs={'file_path': os.path.join(dirname, save_file_name)})
+                attrs={'file_path': os.path.join(dirname, filename)})
 
         executor.run(save_program)
 
 
-def save_params(executor, dirname, main_program=None, save_file_name=None):
+def save_params(executor, dirname, main_program=None, filename=None):
     """
     Save all parameters to directory with executor.
     """
@@ -136,11 +136,10 @@ def save_params(executor, dirname, main_program=None, save_file_name=None):
         main_program=main_program,
         vars=None,
         predicate=is_parameter,
-        save_file_name=save_file_name)
+        filename=filename)
 
 
-def save_persistables(executor, dirname, main_program=None,
-                      save_file_name=None):
+def save_persistables(executor, dirname, main_program=None, filename=None):
     """
     Save all persistables to directory with executor.
     """
@@ -150,7 +149,7 @@ def save_persistables(executor, dirname, main_program=None,
         main_program=main_program,
         vars=None,
         predicate=is_persistable,
-        save_file_name=save_file_name)
+        filename=filename)
 
 
 def load_vars(executor,
@@ -158,7 +157,7 @@ def load_vars(executor,
               main_program=None,
               vars=None,
               predicate=None,
-              load_file_name=None):
+              filename=None):
     """
     Load variables from directory by executor.
 
@@ -170,7 +169,7 @@ def load_vars(executor,
     as a bool. If it returns true, the corresponding input variable will be loaded.
     :param vars: variables need to be loaded. If vars is specified, program &
     predicate will be ignored
-    :param load_file_name: The name of the single file that all vars are loaded from.   
+    :param filename: The name of the single file that all vars are loaded from.
         If it is None, load variables from separate files.
 
     :return: None
@@ -185,7 +184,7 @@ def load_vars(executor,
             executor,
             dirname=dirname,
             vars=filter(predicate, main_program.list_vars()),
-            load_file_name=load_file_name)
+            filename=filename)
     else:
         load_prog = Program()
         load_block = load_prog.global_block()
@@ -194,7 +193,7 @@ def load_vars(executor,
         for each_var in vars:
             assert isinstance(each_var, Variable)
             new_var = _clone_var_in_block_(load_block, each_var)
-            if load_file_name is None:
+            if filename is None:
                 load_block.append_op(
                     type='load',
                     inputs={},
@@ -203,7 +202,7 @@ def load_vars(executor,
             else:
                 load_var_map[new_var.name] = new_var
 
-        if load_file_name is not None:
+        if filename is not None:
             load_var_list = []
             for name in sorted(load_var_map.keys()):
                 load_var_list.append(load_var_map[name])
@@ -212,12 +211,12 @@ def load_vars(executor,
                 type='load_combine',
                 inputs={},
                 outputs={"Out": load_var_list},
-                attrs={'file_path': os.path.join(dirname, load_file_name)})
+                attrs={'file_path': os.path.join(dirname, filename)})
 
         executor.run(load_prog)
 
 
-def load_params(executor, dirname, main_program=None, load_file_name=None):
+def load_params(executor, dirname, main_program=None, filename=None):
     """
     load all parameters from directory by executor.
     """
@@ -226,11 +225,10 @@ def load_params(executor, dirname, main_program=None, load_file_name=None):
         dirname=dirname,
         main_program=main_program,
         predicate=is_parameter,
-        load_file_name=load_file_name)
+        filename=filename)
 
 
-def load_persistables(executor, dirname, main_program=None,
-                      load_file_name=None):
+def load_persistables(executor, dirname, main_program=None, filename=None):
     """
     load all persistables from directory by executor.
     """
@@ -239,7 +237,7 @@ def load_persistables(executor, dirname, main_program=None,
         dirname=dirname,
         main_program=main_program,
         predicate=is_persistable,
-        load_file_name=load_file_name)
+        filename=filename)
 
 
 def get_inference_program(target_vars, main_program=None):
@@ -299,7 +297,8 @@ def save_inference_model(dirname,
                          target_vars,
                          executor,
                          main_program=None,
-                         save_file_name=None):
+                         model_filename=None,
+                         params_filename=None):
     """
     Build a model especially for inference,
     and save it to directory by the executor.
@@ -310,8 +309,11 @@ def save_inference_model(dirname,
     :param executor: executor that save inference model
     :param main_program: original program, which will be pruned to build the inference model.
             Default default_main_program().
-    :param save_file_name: The name of a single file that all parameters are saved to. 
-    If it is None, save parameters to separate files.
+    :param model_filename: The name of file to save inference program.
+        If not specified, default filename `__model__` will be used.
+    :param params_filename: The name of file to save parameters.
+        It is used for the case that all parameters are saved in a single binary file.
+        If not specified, parameters are considered saved in separate files.
 
     :return: None
     """
@@ -342,15 +344,19 @@ def save_inference_model(dirname,
     prepend_feed_ops(inference_program, feeded_var_names)
     append_fetch_ops(inference_program, fetch_var_names)
 
-    if save_file_name == None:
-        model_file_name = dirname + "/__model__"
+    if model_filename is not None:
+        model_filename = os.path.basename(model_filename)
     else:
-        model_file_name = dirname + "/__model_combined__"
+        model_filename = "__model__"
+    model_filename = os.path.join(dirname, model_filename)
 
-    with open(model_file_name, "wb") as f:
+    if params_filename is not None:
+        params_filename = os.path.basename(params_filename)
+
+    with open(model_filename, "wb") as f:
         f.write(inference_program.desc.serialize_to_string())
 
-    save_persistables(executor, dirname, inference_program, save_file_name)
+    save_persistables(executor, dirname, inference_program, params_filename)
 
 
 def get_feed_targets_names(program):
@@ -371,14 +377,20 @@ def get_fetch_targets_names(program):
     return fetch_targets_names
 
 
-def load_inference_model(dirname, executor, load_file_name=None):
+def load_inference_model(dirname,
+                         executor,
+                         model_filename=None,
+                         params_filename=None):
     """
     Load inference model from a directory
 
     :param dirname: directory path
     :param executor: executor that load inference model
-    :param load_file_name: The name of the single file that all parameters are loaded from.   
-    If it is None, load parameters from separate files.
+    :param model_filename: The name of file to load inference program.
+        If not specified, default filename `__model__` will be used.
+    :param params_filename: The name of file to load parameters.
+        It is used for the case that all parameters are saved in a single binary file.
+        If not specified, parameters are considered saved in separate files.
 
     :return: [program, feed_target_names, fetch_targets]
              program: program especially for inference.
@@ -388,16 +400,20 @@ def load_inference_model(dirname, executor, load_file_name=None):
     if not os.path.isdir(dirname):
         raise ValueError("There is no directory named '%s'", dirname)
 
-    if load_file_name == None:
-        model_file_name = dirname + "/__model__"
+    if model_filename is not None:
+        model_filename = os.path.basename(model_filename)
     else:
-        model_file_name = dirname + "/__model_combined__"
+        model_filename = "__model__"
+    model_filename = os.path.join(dirname, model_filename)
+
+    if params_filename is not None:
+        params_filename = os.path.basename(params_filename)
 
-    with open(model_file_name, "rb") as f:
+    with open(model_filename, "rb") as f:
         program_desc_str = f.read()
 
     program = Program.parse_from_string(program_desc_str)
-    load_persistables(executor, dirname, program, load_file_name)
+    load_persistables(executor, dirname, program, params_filename)
 
     feed_target_names = get_feed_targets_names(program)
     fetch_target_names = get_fetch_targets_names(program)

python/paddle/fluid/layers/detection.py

@@ -16,6 +16,7 @@ All layers just related to the detection neural network.
 """
 
 from layer_function_generator import generate_layer_fn
+from layer_function_generator import autodoc
 from ..layer_helper import LayerHelper
 import tensor
 import ops
@@ -28,6 +29,7 @@ __all__ = [
     'target_assign',
     'detection_output',
     'ssd_loss',
+    'detection_map',
 ]
 
 __auto__ = [
@@ -132,7 +134,48 @@ def detection_output(scores,
     return nmsed_outs
 
 
-def bipartite_match(dist_matrix, name=None):
+@autodoc()
+def detection_map(detect_res,
+                  label,
+                  pos_count=None,
+                  true_pos=None,
+                  false_pos=None,
+                  overlap_threshold=0.3,
+                  evaluate_difficult=True,
+                  ap_type='integral'):
+    helper = LayerHelper("detection_map", **locals())
+
+    map_out = helper.create_tmp_variable(dtype='float32')
+    accum_pos_count_out = helper.create_tmp_variable(dtype='int32')
+    accum_true_pos_out = helper.create_tmp_variable(dtype='float32')
+    accum_false_pos_out = helper.create_tmp_variable(dtype='float32')
+    helper.append_op(
+        type="detection_map",
+        inputs={
+            'Label': label,
+            'DetectRes': detect_res,
+            'PosCount': pos_count,
+            'TruePos': true_pos,
+            'FalsePos': false_pos
+        },
+        outputs={
+            'MAP': map_out,
+            'AccumPosCount': accum_pos_count_out,
+            'AccumTruePos': accum_true_pos_out,
+            'AccumFalsePos': accum_false_pos_out
+        },
+        attrs={
+            'overlap_threshold': overlap_threshold,
+            'evaluate_difficult': evaluate_difficult,
+            'ap_type': ap_type
+        })
+    return map_out, accum_pos_count_out, accum_true_pos_out, accum_false_pos_out
+
+
+def bipartite_match(dist_matrix,
+                    match_type=None,
+                    dist_threshold=None,
+                    name=None):
     """
     **Bipartite matchint operator**
 
@@ -164,6 +207,11 @@ def bipartite_match(dist_matrix, name=None):
             This tensor can contain LoD information to represent a batch of
             inputs. One instance of this batch can contain different numbers of
             entities.
+        match_type(string|None): The type of matching method, should be
+           'bipartite' or 'per_prediction', 'bipartite' by defalut.
+        dist_threshold(float|None): If `match_type` is 'per_prediction',
+            this threshold is to determine the extra matching bboxes based
+            on the maximum distance, 0.5 by defalut.
     Returns:
         match_indices(Variable): A 2-D Tensor with shape [N, M] in int type.
             N is the batch size. If match_indices[i][j] is -1, it
@@ -183,6 +231,10 @@ def bipartite_match(dist_matrix, name=None):
     helper.append_op(
         type='bipartite_match',
         inputs={'DistMat': dist_matrix},
+        attrs={
+            'match_type': match_type,
+            'dist_threshold': dist_threshold,
+        },
         outputs={
             'ColToRowMatchIndices': match_indices,
             'ColToRowMatchDist': match_distance
@@ -333,7 +385,7 @@ def ssd_loss(location,
         loc_loss_weight (float): Weight for localization loss, 1.0 by default.
         conf_loss_weight (float): Weight for confidence loss, 1.0 by default.
         match_type (str): The type of matching method during training, should
-            be 'bipartite' or 'per_prediction'.
+            be 'bipartite' or 'per_prediction', 'per_prediction' by defalut.
         mining_type (str): The hard example mining type, should be 'hard_example'
             or 'max_negative', now only support `max_negative`.
 
@@ -381,7 +433,8 @@ def ssd_loss(location,
     #   1.1 Compute IOU similarity between ground-truth boxes and prior boxes.
     iou = iou_similarity(x=gt_box, y=prior_box)
     #   1.2 Compute matched boundding box by bipartite matching algorithm.
-    matched_indices, matched_dist = bipartite_match(iou)
+    matched_indices, matched_dist = bipartite_match(iou, match_type,
+                                                    overlap_threshold)
 
     # 2. Compute confidence for mining hard examples
     # 2.1. Get the target label based on matched indices

python/paddle/fluid/layers/nn.py

@@ -21,6 +21,7 @@ from ..framework import Variable
 from ..param_attr import ParamAttr
 from layer_function_generator import autodoc
 from tensor import concat
+import utils
 
 __all__ = [
     'fc',
@@ -1139,8 +1140,8 @@ def sequence_conv(input,
 def conv2d(input,
            num_filters,
            filter_size,
-           stride=None,
-           padding=None,
+           stride=1,
+           padding=0,
            groups=None,
            param_attr=None,
            bias_attr=None,
@@ -1253,12 +1254,10 @@ def conv2d(input,
             raise ValueError("num_channels must be divisible by groups.")
         num_filter_channels = num_channels / groups
 
-    if isinstance(filter_size, int):
-        filter_size = [filter_size, filter_size]
-    if isinstance(stride, int):
-        stride = [stride, stride]
-    if isinstance(padding, int):
-        padding = [padding, padding]
+    filter_size = utils.convert_to_list(filter_size, 2, 'filter_size')
+    stride = utils.convert_to_list(stride, 2, 'stride')
+    padding = utils.convert_to_list(padding, 2, 'padding')
+
     if not isinstance(use_cudnn, bool):
         raise ValueError("use_cudnn should be True or False")
 
@@ -1433,10 +1432,10 @@ def sequence_last_step(input):
 
 
 def pool2d(input,
-           pool_size,
-           pool_type,
-           pool_stride=None,
-           pool_padding=None,
+           pool_size=-1,
+           pool_type="max",
+           pool_stride=1,
+           pool_padding=0,
            global_pooling=False,
            use_cudnn=True,
            name=None):
@@ -1444,20 +1443,20 @@ def pool2d(input,
     This function adds the operator for pooling in 2 dimensions, using the
     pooling configurations mentioned in input parameters.
     """
-    if pool_padding is None:
-        pool_padding = [0, 0]
-    if pool_stride is None:
-        pool_stride = [1, 1]
     if pool_type not in ["max", "avg"]:
         raise ValueError(
             "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
             str(pool_type))
-    if isinstance(pool_size, int):
-        pool_size = [pool_size, pool_size]
-    if isinstance(pool_stride, int):
-        pool_stride = [pool_stride, pool_stride]
-    if isinstance(pool_padding, int):
-        pool_padding = [pool_padding, pool_padding]
+
+    if global_pooling is False and pool_size == -1:
+        raise ValueError(
+            "When the global_pooling is False, pool_size must be passed "
+            "and be a valid value. Received pool_size: " + str(pool_size))
+
+    pool_size = utils.convert_to_list(pool_size, 2, 'pool_size')
+    pool_padding = utils.convert_to_list(pool_padding, 2, 'pool_padding')
+    pool_stride = utils.convert_to_list(pool_stride, 2, 'pool_stride')
+
     if not isinstance(use_cudnn, bool):
         raise ValueError("use_cudnn should be True or False")
 
@@ -1686,9 +1685,9 @@ def conv2d_transpose(input,
                      num_filters,
                      output_size=None,
                      filter_size=None,
-                     padding=None,
-                     stride=None,
-                     dilation=None,
+                     padding=0,
+                     stride=1,
+                     dilation=1,
                      param_attr=None,
                      use_cudnn=True,
                      name=None):
@@ -1784,26 +1783,12 @@ def conv2d_transpose(input,
         raise TypeError("Input of conv2d_transpose must be Variable")
     input_channel = input.shape[1]
 
-    op_attr = dict()
-
-    if isinstance(padding, int):
-        op_attr['paddings'] = [padding, padding]
-    elif padding is not None:
-        op_attr['paddings'] = padding
-
-    if isinstance(stride, int):
-        op_attr['strides'] = [stride, stride]
-    elif stride is not None:
-        op_attr['strides'] = stride
-
-    if isinstance(dilation, int):
-        op_attr['dilations'] = [dilation, dilation]
-    elif dilation is not None:
-        op_attr['dilations'] = dilation
+    padding = utils.convert_to_list(padding, 2, 'padding')
+    stride = utils.convert_to_list(stride, 2, 'stride')
+    dilation = utils.convert_to_list(dilation, 2, 'dilation')
 
     if not isinstance(use_cudnn, bool):
         raise ValueError("use_cudnn should be True or False")
-    op_attr['use_cudnn'] = use_cudnn
 
     if filter_size is None:
         if output_size is None:
@@ -1811,10 +1796,6 @@ def conv2d_transpose(input,
         if isinstance(output_size, int):
             output_size = [output_size, output_size]
 
-        padding = op_attr.get('paddings', [0, 0])
-        stride = op_attr.get('strides', [1, 1])
-        dilation = op_attr.get('dilations', [1, 1])
-
         h_in = input.shape[2]
         w_in = input.shape[3]
 
@@ -1823,9 +1804,9 @@ def conv2d_transpose(input,
         filter_size_w = (output_size[1] - (w_in - 1) * stride[1] + 2 *
                          padding[1] - 1) / dilation[1] + 1
         filter_size = [filter_size_h, filter_size_w]
-
-    elif isinstance(filter_size, int):
-        filter_size = [filter_size, filter_size]
+    else:
+        filter_size = utils.convert_to_list(filter_size, 2,
+                                            'conv2d_transpose.filter_size')
 
     filter_shape = [input_channel, num_filters] + filter_size
     img_filter = helper.create_parameter(
@@ -1837,7 +1818,12 @@ def conv2d_transpose(input,
         inputs={'Input': [input],
                 'Filter': [img_filter]},
         outputs={'Output': out},
-        attrs=op_attr)
+        attrs={
+            'strides': stride,
+            'paddings': padding,
+            'dilations': dilation,
+            'use_cudnn': use_cudnn
+        })
 
     return out
 

python/paddle/fluid/layers/utils.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import numpy as np
+
+
+def convert_to_list(value, n, name, dtype=np.int):
+    """
+    Converts a single numerical type or iterable of numerical
+    types into an numerical type list.
+
+    Arguments:
+      value: The value to validate and convert. Could an int, or any iterable
+        of ints.
+      n: The size of the list to be returned.
+      name: The name of the argument being validated, e.g. "stride" or
+        "filter_size". This is only used to format error messages.
+      dtype: the numerical type of the element of the list to be returned.
+
+    Returns:
+      A list of n dtypes.
+
+    Raises:
+      ValueError: If something else than an int/long or iterable thereof was
+        passed.
+    """
+    if isinstance(value, dtype):
+        return [value, ] * n
+    else:
+        try:
+            value_list = list(value)
+        except TypeError:
+            raise ValueError("The " + name +
+                             "'s type must be list or tuple. Received: " + str(
+                                 value))
+        if len(value_list) != n:
+            raise ValueError("The " + name + "'s length must be " + str(n) +
+                             ". Received: " + str(value))
+        for single_value in value_list:
+            try:
+                dtype(single_value)
+            except (ValueError, TypeError):
+                raise ValueError(
+                    "The " + name + "'s type must be a list or tuple of " + str(
+                        n) + " " + str(dtype) + " . Received: " + str(
+                            value) + " "
+                    "including element " + str(single_value) + " of type" + " "
+                    + str(type(single_value)))
+        return value_list

python/paddle/fluid/learning_rate_decay.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import layers
-from framework import Variable
 from initializer import init_on_cpu
 
 __all__ = [

python/paddle/fluid/profiler.py

@@ -97,9 +97,14 @@ def profiler(state, sorted_key=None):
             The `ave` means sorting by the average execution time.
     """
 
-    if state not in ['CPU', 'GPU']:
-        raise ValueError("The state must be 'CPU' or 'GPU'.")
-    prof_state = core.ProfilerState.kCUDA if state == "GPU" else core.ProfilerState.kCPU
+    if state not in ['CPU', 'GPU', "All"]:
+        raise ValueError("The state must be 'CPU' or 'GPU' or 'All'.")
+    if state == "GPU":
+        prof_state = core.ProfilerState.kCUDA
+    elif state == "CPU":
+        prof_state = core.ProfilerState.kCPU
+    else:
+        prof_state = core.ProfilerState.kAll
     core.enable_profiler(prof_state)
     yield
 

python/paddle/fluid/tests/CMakeLists.txt

@@ -7,5 +7,4 @@ endforeach()
 
 add_subdirectory(unittests)
 add_subdirectory(book)
-add_subdirectory(book_distribute)
 add_subdirectory(book_memory_optimization)

python/paddle/fluid/tests/book/notest_rnn_encoder_decoer.py

@@ -228,6 +228,8 @@ def infer(use_cuda, save_dirname=None):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     exe = fluid.Executor(place)
 
+    inference_scope = fluid.core.Scope()
+    with fluid.scope_guard(inference_scope):
         # Use fluid.io.load_inference_model to obtain the inference program desc,
         # the feed_target_names (the names of variables that will be feeded
         # data using feed operators), and the fetch_targets (variables that

python/paddle/fluid/tests/book/test_fit_a_line.py

@@ -19,9 +19,10 @@ import numpy
 import unittest
 import math
 import sys
+import os
 
 
-def train(use_cuda, save_dirname):
+def train(use_cuda, save_dirname, is_local):
     x = fluid.layers.data(name='x', shape=[13], dtype='float32')
 
     y_predict = fluid.layers.fc(input=x, size=1, act=None)
@@ -32,7 +33,7 @@ def train(use_cuda, save_dirname):
     avg_cost = fluid.layers.mean(cost)
 
     sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
-    sgd_optimizer.minimize(avg_cost)
+    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
 
     BATCH_SIZE = 20
 
@@ -42,15 +43,16 @@ def train(use_cuda, save_dirname):
         batch_size=BATCH_SIZE)
 
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-    feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
     exe = fluid.Executor(place)
 
+    def train_loop(main_program):
+        feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
         exe.run(fluid.default_startup_program())
 
         PASS_NUM = 100
         for pass_id in range(PASS_NUM):
             for data in train_reader():
-            avg_loss_value, = exe.run(fluid.default_main_program(),
+                avg_loss_value, = exe.run(main_program,
                                           feed=feeder.feed(data),
                                           fetch_list=[avg_cost])
                 print(avg_loss_value)
@@ -64,6 +66,35 @@ def train(use_cuda, save_dirname):
         raise AssertionError("Fit a line cost is too large, {0:2.2}".format(
             avg_loss_value[0]))
 
+    if is_local:
+        train_loop(fluid.default_main_program())
+    else:
+        port = os.getenv("PADDLE_INIT_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        eplist = []
+        for ip in pserver_ips.split(","):
+            eplist.append(':'.join([ip, port]))
+        pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
+        trainers = int(os.getenv("TRAINERS"))
+        current_endpoint = os.getenv("POD_IP") + ":" + port
+        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
+        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        t = fluid.DistributeTranspiler()
+        t.transpile(
+            optimize_ops,
+            params_grads,
+            trainer_id,
+            pservers=pserver_endpoints,
+            trainers=trainers)
+        if training_role == "PSERVER":
+            pserver_prog = t.get_pserver_program(current_endpoint)
+            pserver_startup = t.get_startup_program(current_endpoint,
+                                                    pserver_prog)
+            exe.run(pserver_startup)
+            exe.run(pserver_prog)
+        elif training_role == "TRAINER":
+            train_loop(t.get_trainer_program())
+
 
 def infer(use_cuda, save_dirname=None):
     if save_dirname is None:
@@ -72,6 +103,8 @@ def infer(use_cuda, save_dirname=None):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     exe = fluid.Executor(place)
 
+    inference_scope = fluid.core.Scope()
+    with fluid.scope_guard(inference_scope):
         # Use fluid.io.load_inference_model to obtain the inference program desc,
         # the feed_target_names (the names of variables that will be feeded
         # data using feed operators), and the fetch_targets (variables that
@@ -82,7 +115,8 @@ def infer(use_cuda, save_dirname=None):
         # The input's dimension should be 2-D and the second dim is 13
         # The input data should be >= 0
         batch_size = 10
-    tensor_x = numpy.random.uniform(0, 10, [batch_size, 13]).astype("float32")
+        tensor_x = numpy.random.uniform(0, 10,
+                                        [batch_size, 13]).astype("float32")
         assert feed_target_names[0] == 'x'
         results = exe.run(inference_program,
                           feed={feed_target_names[0]: tensor_x},
@@ -91,14 +125,14 @@ def infer(use_cuda, save_dirname=None):
         print("infer results: ", results[0])
 
 
-def main(use_cuda):
+def main(use_cuda, is_local=True):
     if use_cuda and not fluid.core.is_compiled_with_cuda():
         return
 
     # Directory for saving the trained model
     save_dirname = "fit_a_line.inference.model"
 
-    train(use_cuda, save_dirname)
+    train(use_cuda, save_dirname, is_local)
     infer(use_cuda, save_dirname)
 
 

python/paddle/fluid/tests/book/test_image_classification.py

@@ -21,6 +21,7 @@ import math
 import sys
 import numpy
 import unittest
+import os
 
 
 def resnet_cifar10(input, depth=32):
@@ -92,7 +93,7 @@ def vgg16_bn_drop(input):
     return fc2
 
 
-def train(net_type, use_cuda, save_dirname):
+def train(net_type, use_cuda, save_dirname, is_local):
     classdim = 10
     data_shape = [3, 32, 32]
 
@@ -117,7 +118,7 @@ def train(net_type, use_cuda, save_dirname):
     test_program = fluid.default_main_program().clone()
 
     optimizer = fluid.optimizer.Adam(learning_rate=0.001)
-    optimizer.minimize(avg_cost)
+    optimize_ops, params_grads = optimizer.minimize(avg_cost)
 
     BATCH_SIZE = 128
     PASS_NUM = 1
@@ -133,12 +134,13 @@ def train(net_type, use_cuda, save_dirname):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     exe = fluid.Executor(place)
     feeder = fluid.DataFeeder(place=place, feed_list=[images, label])
-    exe.run(fluid.default_startup_program())
 
+    def train_loop(main_program):
+        exe.run(fluid.default_startup_program())
         loss = 0.0
         for pass_id in range(PASS_NUM):
             for batch_id, data in enumerate(train_reader()):
-            exe.run(feed=feeder.feed(data))
+                exe.run(main_program, feed=feeder.feed(data))
 
                 if (batch_id % 10) == 0:
                     acc_list = []
@@ -166,6 +168,35 @@ def train(net_type, use_cuda, save_dirname):
                                                       [predict], exe)
                         return
 
+    if is_local:
+        train_loop(fluid.default_main_program())
+    else:
+        port = os.getenv("PADDLE_INIT_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        eplist = []
+        for ip in pserver_ips.split(","):
+            eplist.append(':'.join([ip, port]))
+        pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
+        trainers = int(os.getenv("TRAINERS"))
+        current_endpoint = os.getenv("POD_IP") + ":" + port
+        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
+        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        t = fluid.DistributeTranspiler()
+        t.transpile(
+            optimize_ops,
+            params_grads,
+            trainer_id,
+            pservers=pserver_endpoints,
+            trainers=trainers)
+        if training_role == "PSERVER":
+            pserver_prog = t.get_pserver_program(current_endpoint)
+            pserver_startup = t.get_startup_program(current_endpoint,
+                                                    pserver_prog)
+            exe.run(pserver_startup)
+            exe.run(pserver_prog)
+        elif training_role == "TRAINER":
+            train_loop(t.get_trainer_program())
+
 
 def infer(use_cuda, save_dirname=None):
     if save_dirname is None:
@@ -174,6 +205,8 @@ def infer(use_cuda, save_dirname=None):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     exe = fluid.Executor(place)
 
+    inference_scope = fluid.core.Scope()
+    with fluid.scope_guard(inference_scope):
         # Use fluid.io.load_inference_model to obtain the inference program desc,
         # the feed_target_names (the names of variables that will be feeded
         # data using feed operators), and the fetch_targets (variables that
@@ -182,7 +215,9 @@ def infer(use_cuda, save_dirname=None):
          fetch_targets] = fluid.io.load_inference_model(save_dirname, exe)
 
         # The input's dimension of conv should be 4-D or 5-D.
-    tensor_img = numpy.random.rand(1, 3, 32, 32).astype("float32")
+        # Use normilized image pixels as input data, which should be in the range [0, 1.0].
+        batch_size = 1
+        tensor_img = numpy.random.rand(batch_size, 3, 32, 32).astype("float32")
 
         # Construct feed as a dictionary of {feed_target_name: feed_target_data}
         # and results will contain a list of data corresponding to fetch_targets.
@@ -192,14 +227,14 @@ def infer(use_cuda, save_dirname=None):
         print("infer results: ", results[0])
 
 
-def main(net_type, use_cuda):
+def main(net_type, use_cuda, is_local=True):
     if use_cuda and not fluid.core.is_compiled_with_cuda():
         return
 
     # Directory for saving the trained model
     save_dirname = "image_classification_" + net_type + ".inference.model"
 
-    train(net_type, use_cuda, save_dirname)
+    train(net_type, use_cuda, save_dirname, is_local)
     infer(use_cuda, save_dirname)
 
 

python/paddle/fluid/tests/book/test_label_semantic_roles.py

@@ -22,11 +22,12 @@ from paddle.fluid.initializer import init_on_cpu
 import contextlib
 import time
 import unittest
+import os
 
 word_dict, verb_dict, label_dict = conll05.get_dict()
 word_dict_len = len(word_dict)
 label_dict_len = len(label_dict)
-pred_len = len(verb_dict)
+pred_dict_len = len(verb_dict)
 
 mark_dict_len = 2
 word_dim = 32
@@ -53,7 +54,7 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark,
     # 8 features
     predicate_embedding = fluid.layers.embedding(
         input=predicate,
-        size=[pred_len, word_dim],
+        size=[pred_dict_len, word_dim],
         dtype='float32',
         is_sparse=IS_SPARSE,
         param_attr='vemb')
@@ -138,7 +139,7 @@ def create_random_lodtensor(lod, place, low, high):
     return res
 
 
-def train(use_cuda, save_dirname=None):
+def train(use_cuda, save_dirname=None, is_local=True):
     # define network topology
     word = fluid.layers.data(
         name='word_data', shape=[1], dtype='int64', lod_level=1)
@@ -174,7 +175,7 @@ def train(use_cuda, save_dirname=None):
             decay_steps=100000,
             decay_rate=0.5,
             staircase=True))
-    sgd_optimizer.minimize(avg_cost)
+    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
 
     # TODO(qiao)
     # add dependency track and move this config before optimizer
@@ -200,11 +201,14 @@ def train(use_cuda, save_dirname=None):
         place=place)
     exe = fluid.Executor(place)
 
+    def train_loop(main_program):
         exe.run(fluid.default_startup_program())
 
-    embedding_param = fluid.global_scope().find_var(embedding_name).get_tensor()
+        embedding_param = fluid.global_scope().find_var(
+            embedding_name).get_tensor()
         embedding_param.set(
-        load_parameter(conll05.get_embedding(), word_dict_len, word_dim), place)
+            load_parameter(conll05.get_embedding(), word_dict_len, word_dim),
+            place)
 
         start_time = time.time()
         batch_id = 0
@@ -212,7 +216,7 @@ def train(use_cuda, save_dirname=None):
             chunk_evaluator.reset(exe)
             for data in train_data():
                 cost, precision, recall, f1_score = exe.run(
-                fluid.default_main_program(),
+                    main_program,
                     feed=feeder.feed(data),
                     fetch_list=[avg_cost] + chunk_evaluator.metrics)
                 pass_precision, pass_recall, pass_f1_score = chunk_evaluator.eval(
@@ -220,16 +224,18 @@ def train(use_cuda, save_dirname=None):
 
                 if batch_id % 10 == 0:
                     print("avg_cost:" + str(cost) + " precision:" + str(
-                    precision) + " recall:" + str(recall) + " f1_score:" + str(
-                        f1_score) + " pass_precision:" + str(
-                            pass_precision) + " pass_recall:" + str(pass_recall)
-                      + " pass_f1_score:" + str(pass_f1_score))
+                        precision) + " recall:" + str(recall) + " f1_score:" +
+                          str(f1_score) + " pass_precision:" + str(
+                              pass_precision) + " pass_recall:" + str(
+                                  pass_recall) + " pass_f1_score:" + str(
+                                      pass_f1_score))
                     if batch_id != 0:
-                    print("second per batch: " + str((time.time() - start_time)
-                                                     / batch_id))
+                        print("second per batch: " + str((time.time(
+                        ) - start_time) / batch_id))
                     # Set the threshold low to speed up the CI test
                     if float(pass_precision) > 0.05:
                         if save_dirname is not None:
+                            # TODO(liuyiqun): Change the target to crf_decode
                             fluid.io.save_inference_model(save_dirname, [
                                 'word_data', 'verb_data', 'ctx_n2_data',
                                 'ctx_n1_data', 'ctx_0_data', 'ctx_p1_data',
@@ -239,6 +245,35 @@ def train(use_cuda, save_dirname=None):
 
                 batch_id = batch_id + 1
 
+    if is_local:
+        train_loop(fluid.default_main_program())
+    else:
+        port = os.getenv("PADDLE_INIT_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        eplist = []
+        for ip in pserver_ips.split(","):
+            eplist.append(':'.join([ip, port]))
+        pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
+        trainers = int(os.getenv("TRAINERS"))
+        current_endpoint = os.getenv("POD_IP") + ":" + port
+        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
+        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        t = fluid.DistributeTranspiler()
+        t.transpile(
+            optimize_ops,
+            params_grads,
+            trainer_id,
+            pservers=pserver_endpoints,
+            trainers=trainers)
+        if training_role == "PSERVER":
+            pserver_prog = t.get_pserver_program(current_endpoint)
+            pserver_startup = t.get_startup_program(current_endpoint,
+                                                    pserver_prog)
+            exe.run(pserver_startup)
+            exe.run(pserver_prog)
+        elif training_role == "TRAINER":
+            train_loop(t.get_trainer_program())
+
 
 def infer(use_cuda, save_dirname=None):
     if save_dirname is None:
@@ -247,6 +282,8 @@ def infer(use_cuda, save_dirname=None):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     exe = fluid.Executor(place)
 
+    inference_scope = fluid.core.Scope()
+    with fluid.scope_guard(inference_scope):
         # Use fluid.io.load_inference_model to obtain the inference program desc,
         # the feed_target_names (the names of variables that will be feeded
         # data using feed operators), and the fetch_targets (variables that
@@ -255,14 +292,22 @@ def infer(use_cuda, save_dirname=None):
          fetch_targets] = fluid.io.load_inference_model(save_dirname, exe)
 
         lod = [0, 4, 10]
-    ts_word = create_random_lodtensor(lod, place, low=0, high=1)
-    ts_pred = create_random_lodtensor(lod, place, low=0, high=1)
-    ts_ctx_n2 = create_random_lodtensor(lod, place, low=0, high=1)
-    ts_ctx_n1 = create_random_lodtensor(lod, place, low=0, high=1)
-    ts_ctx_0 = create_random_lodtensor(lod, place, low=0, high=1)
-    ts_ctx_p1 = create_random_lodtensor(lod, place, low=0, high=1)
-    ts_ctx_p2 = create_random_lodtensor(lod, place, low=0, high=1)
-    ts_mark = create_random_lodtensor(lod, place, low=0, high=1)
+        word = create_random_lodtensor(
+            lod, place, low=0, high=word_dict_len - 1)
+        pred = create_random_lodtensor(
+            lod, place, low=0, high=pred_dict_len - 1)
+        ctx_n2 = create_random_lodtensor(
+            lod, place, low=0, high=word_dict_len - 1)
+        ctx_n1 = create_random_lodtensor(
+            lod, place, low=0, high=word_dict_len - 1)
+        ctx_0 = create_random_lodtensor(
+            lod, place, low=0, high=word_dict_len - 1)
+        ctx_p1 = create_random_lodtensor(
+            lod, place, low=0, high=word_dict_len - 1)
+        ctx_p2 = create_random_lodtensor(
+            lod, place, low=0, high=word_dict_len - 1)
+        mark = create_random_lodtensor(
+            lod, place, low=0, high=mark_dict_len - 1)
 
         # Construct feed as a dictionary of {feed_target_name: feed_target_data}
         # and results will contain a list of data corresponding to fetch_targets.
@@ -277,31 +322,30 @@ def infer(use_cuda, save_dirname=None):
 
         results = exe.run(inference_program,
                           feed={
-                          feed_target_names[0]: ts_word,
-                          feed_target_names[1]: ts_pred,
-                          feed_target_names[2]: ts_ctx_n2,
-                          feed_target_names[3]: ts_ctx_n1,
-                          feed_target_names[4]: ts_ctx_0,
-                          feed_target_names[5]: ts_ctx_p1,
-                          feed_target_names[6]: ts_ctx_p2,
-                          feed_target_names[7]: ts_mark
+                              feed_target_names[0]: word,
+                              feed_target_names[1]: pred,
+                              feed_target_names[2]: ctx_n2,
+                              feed_target_names[3]: ctx_n1,
+                              feed_target_names[4]: ctx_0,
+                              feed_target_names[5]: ctx_p1,
+                              feed_target_names[6]: ctx_p2,
+                              feed_target_names[7]: mark
                           },
                           fetch_list=fetch_targets,
                           return_numpy=False)
         print(results[0].lod())
         np_data = np.array(results[0])
         print("Inference Shape: ", np_data.shape)
-    print("Inference results: ", np_data)
 
 
-def main(use_cuda):
+def main(use_cuda, is_local=True):
     if use_cuda and not fluid.core.is_compiled_with_cuda():
         return
 
     # Directory for saving the trained model
     save_dirname = "label_semantic_roles.inference.model"
 
-    train(use_cuda, save_dirname)
+    train(use_cuda, save_dirname, is_local)
     infer(use_cuda, save_dirname)
 
 

python/paddle/fluid/tests/book/test_machine_translation.py

@@ -20,6 +20,7 @@ import paddle.fluid.framework as framework
 import paddle.fluid.layers as pd
 from paddle.fluid.executor import Executor
 import unittest
+import os
 
 dict_size = 30000
 source_dict_dim = target_dict_dim = dict_size
@@ -168,7 +169,7 @@ def to_lodtensor(data, place):
     return res
 
 
-def train_main(use_cuda, is_sparse):
+def train_main(use_cuda, is_sparse, is_local=True):
     if use_cuda and not fluid.core.is_compiled_with_cuda():
         return
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
@@ -181,7 +182,7 @@ def train_main(use_cuda, is_sparse):
     avg_cost = pd.mean(cost)
 
     optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4)
-    optimizer.minimize(avg_cost)
+    optimize_ops, params_grads = optimizer.minimize(avg_cost)
 
     train_data = paddle.batch(
         paddle.reader.shuffle(
@@ -190,6 +191,7 @@ def train_main(use_cuda, is_sparse):
 
     exe = Executor(place)
 
+    def train_loop(main_program):
         exe.run(framework.default_startup_program())
 
         batch_id = 0
@@ -198,7 +200,7 @@ def train_main(use_cuda, is_sparse):
                 word_data = to_lodtensor(map(lambda x: x[0], data), place)
                 trg_word = to_lodtensor(map(lambda x: x[1], data), place)
                 trg_word_next = to_lodtensor(map(lambda x: x[2], data), place)
-            outs = exe.run(framework.default_main_program(),
+                outs = exe.run(main_program,
                                feed={
                                    'src_word_id': word_data,
                                    'target_language_word': trg_word,
@@ -212,6 +214,35 @@ def train_main(use_cuda, is_sparse):
                     break
                 batch_id += 1
 
+    if is_local:
+        train_loop(framework.default_main_program())
+    else:
+        port = os.getenv("PADDLE_INIT_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        eplist = []
+        for ip in pserver_ips.split(","):
+            eplist.append(':'.join([ip, port]))
+        pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
+        trainers = int(os.getenv("TRAINERS"))
+        current_endpoint = os.getenv("POD_IP") + ":" + port
+        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
+        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        t = fluid.DistributeTranspiler()
+        t.transpile(
+            optimize_ops,
+            params_grads,
+            trainer_id,
+            pservers=pserver_endpoints,
+            trainers=trainers)
+        if training_role == "PSERVER":
+            pserver_prog = t.get_pserver_program(current_endpoint)
+            pserver_startup = t.get_startup_program(current_endpoint,
+                                                    pserver_prog)
+            exe.run(pserver_startup)
+            exe.run(pserver_prog)
+        elif training_role == "TRAINER":
+            train_loop(t.get_trainer_program())
+
 
 def decode_main(use_cuda, is_sparse):
     if use_cuda and not fluid.core.is_compiled_with_cuda():

python/paddle/fluid/tests/book/test_recognize_digits.py

@@ -20,27 +20,7 @@ import numpy
 import unittest
 import math
 import sys
-
-
-def parse_arg():
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "nn_type",
-        help="The neural network type, in ['mlp', 'conv']",
-        type=str,
-        choices=['mlp', 'conv'])
-    parser.add_argument(
-        "--parallel",
-        help='Run in parallel or not',
-        default=False,
-        action="store_true")
-    parser.add_argument(
-        "--use_cuda",
-        help="Run the program by using CUDA",
-        default=False,
-        action="store_true")
-    return parser.parse_args()
-
+import os
 
 BATCH_SIZE = 64
 
@@ -78,7 +58,13 @@ def conv_net(img, label):
     return loss_net(conv_pool_2, label)
 
 
-def train(nn_type, use_cuda, parallel, save_dirname, save_param_filename):
+def train(nn_type,
+          use_cuda,
+          parallel,
+          save_dirname=None,
+          model_filename=None,
+          params_filename=None,
+          is_local=True):
     if use_cuda and not fluid.core.is_compiled_with_cuda():
         return
     img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32')
@@ -109,12 +95,11 @@ def train(nn_type, use_cuda, parallel, save_dirname, save_param_filename):
     test_program = fluid.default_main_program().clone()
 
     optimizer = fluid.optimizer.Adam(learning_rate=0.001)
-    optimizer.minimize(avg_loss)
+    optimize_ops, params_grads = optimizer.minimize(avg_loss)
 
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 
     exe = fluid.Executor(place)
-    exe.run(fluid.default_startup_program())
 
     train_reader = paddle.batch(
         paddle.reader.shuffle(
@@ -124,16 +109,20 @@ def train(nn_type, use_cuda, parallel, save_dirname, save_param_filename):
         paddle.dataset.mnist.test(), batch_size=BATCH_SIZE)
     feeder = fluid.DataFeeder(feed_list=[img, label], place=place)
 
+    def train_loop(main_program):
+        exe.run(fluid.default_startup_program())
+
         PASS_NUM = 100
         for pass_id in range(PASS_NUM):
             for batch_id, data in enumerate(train_reader()):
                 # train a mini-batch, fetch nothing
-            exe.run(feed=feeder.feed(data))
+                exe.run(main_program, feed=feeder.feed(data))
                 if (batch_id + 1) % 10 == 0:
                     acc_set = []
                     avg_loss_set = []
                     for test_data in test_reader():
-                    acc_np, avg_loss_np = exe.run(program=test_program,
+                        acc_np, avg_loss_np = exe.run(
+                            program=test_program,
                             feed=feeder.feed(test_data),
                             fetch_list=[acc, avg_loss])
                         acc_set.append(float(acc_np))
@@ -141,12 +130,14 @@ def train(nn_type, use_cuda, parallel, save_dirname, save_param_filename):
                     # get test acc and loss
                     acc_val = numpy.array(acc_set).mean()
                     avg_loss_val = numpy.array(avg_loss_set).mean()
-                if float(acc_val) > 0.2:  # Smaller value to increase CI speed
+                    if float(acc_val
+                             ) > 0.2:  # Smaller value to increase CI speed
                         if save_dirname is not None:
                             fluid.io.save_inference_model(
                                 save_dirname, ["img"], [prediction],
                                 exe,
-                            save_file_name=save_param_filename)
+                                model_filename=model_filename,
+                                params_filename=params_filename)
                         return
                     else:
                         print(
@@ -157,26 +148,62 @@ def train(nn_type, use_cuda, parallel, save_dirname, save_param_filename):
                             sys.exit("got NaN loss, training failed.")
         raise AssertionError("Loss of recognize digits is too large")
 
-
-def infer(use_cuda, save_dirname=None, param_filename=None):
+    if is_local:
+        train_loop(fluid.default_main_program())
+    else:
+        port = os.getenv("PADDLE_INIT_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        eplist = []
+        for ip in pserver_ips.split(","):
+            eplist.append(':'.join([ip, port]))
+        pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
+        pserver_endpoints = os.getenv("PSERVERS")
+        trainers = int(os.getenv("TRAINERS"))
+        current_endpoint = os.getenv("POD_IP") + ":" + port
+        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
+        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        t = fluid.DistributeTranspiler()
+        t.transpile(
+            optimize_ops,
+            params_grads,
+            trainer_id,
+            pservers=pserver_endpoints,
+            trainers=trainers)
+        if training_role == "PSERVER":
+            pserver_prog = t.get_pserver_program(current_endpoint)
+            pserver_startup = t.get_startup_program(current_endpoint,
+                                                    pserver_prog)
+            exe.run(pserver_startup)
+            exe.run(pserver_prog)
+        elif training_role == "TRAINER":
+            train_loop(t.get_trainer_program())
+
+
+def infer(use_cuda,
+          save_dirname=None,
+          model_filename=None,
+          params_filename=None):
     if save_dirname is None:
         return
 
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     exe = fluid.Executor(place)
 
+    inference_scope = fluid.core.Scope()
+    with fluid.scope_guard(inference_scope):
         # Use fluid.io.load_inference_model to obtain the inference program desc,
         # the feed_target_names (the names of variables that will be feeded
         # data using feed operators), and the fetch_targets (variables that
         # we want to obtain data from using fetch operators).
-    [inference_program, feed_target_names, fetch_targets
-     ] = fluid.io.load_inference_model(save_dirname, exe, param_filename)
+        [inference_program, feed_target_names,
+         fetch_targets] = fluid.io.load_inference_model(
+             save_dirname, exe, model_filename, params_filename)
 
         # The input's dimension of conv should be 4-D or 5-D.
         # Use normilized image pixels as input data, which should be in the range [-1.0, 1.0].
         batch_size = 1
-    tensor_img = numpy.random.uniform(-1.0, 1.0,
-                                      [batch_size, 1, 28, 28]).astype("float32")
+        tensor_img = numpy.random.uniform(
+            -1.0, 1.0, [batch_size, 1, 28, 28]).astype("float32")
 
         # Construct feed as a dictionary of {feed_target_name: feed_target_data}
         # and results will contain a list of data corresponding to fetch_targets.
@@ -187,25 +214,28 @@ def infer(use_cuda, save_dirname=None, param_filename=None):
 
 
 def main(use_cuda, parallel, nn_type, combine):
+    save_dirname = None
+    model_filename = None
+    params_filename = None
     if not use_cuda and not parallel:
         save_dirname = "recognize_digits_" + nn_type + ".inference.model"
-        save_filename = None
         if combine == True:
-            save_filename = "__params_combined__"
-    else:
-        save_dirname = None
-        save_filename = None
+            model_filename = "__model_combined__"
+            params_filename = "__params_combined__"
 
+    # call train() with is_local argument to run distributed train
     train(
         nn_type=nn_type,
         use_cuda=use_cuda,
         parallel=parallel,
         save_dirname=save_dirname,
-        save_param_filename=save_filename)
+        model_filename=model_filename,
+        params_filename=params_filename)
     infer(
         use_cuda=use_cuda,
         save_dirname=save_dirname,
-        param_filename=save_filename)
+        model_filename=model_filename,
+        params_filename=params_filename)
 
 
 class TestRecognizeDigits(unittest.TestCase):

python/paddle/fluid/tests/book/test_recommender_system.py

@@ -14,6 +14,7 @@
 
 import math
 import sys
+import os
 import numpy as np
 import paddle.v2 as paddle
 import paddle.fluid as fluid
@@ -152,19 +153,18 @@ def model():
     return scale_infer, avg_cost
 
 
-def train(use_cuda, save_dirname):
+def train(use_cuda, save_dirname, is_local=True):
     scale_infer, avg_cost = model()
 
     # test program
     test_program = fluid.default_main_program().clone()
 
     sgd_optimizer = SGDOptimizer(learning_rate=0.2)
-    opts = sgd_optimizer.minimize(avg_cost)
+    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
 
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 
     exe = Executor(place)
-    exe.run(framework.default_startup_program())
 
     train_reader = paddle.batch(
         paddle.reader.shuffle(
@@ -212,18 +212,22 @@ def train(use_cuda, save_dirname):
             feed_tensors[key] = tensor
         return feed_tensors
 
+    def train_loop(main_program):
+        exe.run(framework.default_startup_program())
+
         PASS_NUM = 100
         for pass_id in range(PASS_NUM):
             for batch_id, data in enumerate(train_reader()):
                 # train a mini-batch
-            outs = exe.run(program=fluid.default_main_program(),
+                outs = exe.run(program=main_program,
                                feed=func_feed(feeding, data),
                                fetch_list=[avg_cost])
                 out = np.array(outs[0])
                 if (batch_id + 1) % 10 == 0:
                     avg_cost_set = []
                     for test_data in test_reader():
-                    avg_cost_np = exe.run(program=test_program,
+                        avg_cost_np = exe.run(
+                            program=test_program,
                             feed=func_feed(feeding, test_data),
                             fetch_list=[avg_cost])
                         avg_cost_set.append(avg_cost_np[0])
@@ -243,6 +247,35 @@ def train(use_cuda, save_dirname):
                 if math.isnan(float(out[0])):
                     sys.exit("got NaN loss, training failed.")
 
+    if is_local:
+        train_loop(fluid.default_main_program())
+    else:
+        port = os.getenv("PADDLE_INIT_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        eplist = []
+        for ip in pserver_ips.split(","):
+            eplist.append(':'.join([ip, port]))
+        pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
+        trainers = int(os.getenv("TRAINERS"))
+        current_endpoint = os.getenv("POD_IP") + ":" + port
+        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
+        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        t = fluid.DistributeTranspiler()
+        t.transpile(
+            optimize_ops,
+            params_grads,
+            trainer_id,
+            pservers=pserver_endpoints,
+            trainers=trainers)
+        if training_role == "PSERVER":
+            pserver_prog = t.get_pserver_program(current_endpoint)
+            pserver_startup = t.get_startup_program(current_endpoint,
+                                                    pserver_prog)
+            exe.run(pserver_startup)
+            exe.run(pserver_prog)
+        elif training_role == "TRAINER":
+            train_loop(t.get_trainer_program())
+
 
 def infer(use_cuda, save_dirname=None):
     if save_dirname is None:
@@ -251,13 +284,6 @@ def infer(use_cuda, save_dirname=None):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     exe = fluid.Executor(place)
 
-    # Use fluid.io.load_inference_model to obtain the inference program desc,
-    # the feed_target_names (the names of variables that will be feeded
-    # data using feed operators), and the fetch_targets (variables that
-    # we want to obtain data from using fetch operators).
-    [inference_program, feed_target_names,
-     fetch_targets] = fluid.io.load_inference_model(save_dirname, exe)
-
     def create_lod_tensor(data, lod=None):
         tensor = fluid.LoDTensor()
         if lod is None:
@@ -275,6 +301,15 @@ def infer(use_cuda, save_dirname=None):
         tensor.set(flattened_data, place)
         return tensor
 
+    inference_scope = fluid.core.Scope()
+    with fluid.scope_guard(inference_scope):
+        # Use fluid.io.load_inference_model to obtain the inference program desc,
+        # the feed_target_names (the names of variables that will be feeded
+        # data using feed operators), and the fetch_targets (variables that
+        # we want to obtain data from using fetch operators).
+        [inference_program, feed_target_names,
+         fetch_targets] = fluid.io.load_inference_model(save_dirname, exe)
+
         # Use the first data from paddle.dataset.movielens.test() as input
         assert feed_target_names[0] == "user_id"
         user_id = create_lod_tensor([[1]])

python/paddle/fluid/tests/book/test_understand_sentiment.py

@@ -20,6 +20,7 @@ import contextlib
 import math
 import numpy as np
 import sys
+import os
 
 
 def convolution_net(data, label, input_dim, class_dim=2, emb_dim=32,
@@ -132,7 +133,12 @@ def create_random_lodtensor(lod, place, low, high):
     return res
 
 
-def train(word_dict, net_method, use_cuda, parallel=False, save_dirname=None):
+def train(word_dict,
+          net_method,
+          use_cuda,
+          parallel=False,
+          save_dirname=None,
+          is_local=True):
     BATCH_SIZE = 128
     PASS_NUM = 5
     dict_dim = len(word_dict)
@@ -164,7 +170,7 @@ def train(word_dict, net_method, use_cuda, parallel=False, save_dirname=None):
         assert save_dirname is None
 
     adagrad = fluid.optimizer.Adagrad(learning_rate=0.002)
-    adagrad.minimize(cost)
+    optimize_ops, params_grads = adagrad.minimize(cost)
 
     train_data = paddle.batch(
         paddle.reader.shuffle(
@@ -174,11 +180,12 @@ def train(word_dict, net_method, use_cuda, parallel=False, save_dirname=None):
     exe = fluid.Executor(place)
     feeder = fluid.DataFeeder(feed_list=[data, label], place=place)
 
+    def train_loop(main_program):
         exe.run(fluid.default_startup_program())
 
         for pass_id in xrange(PASS_NUM):
             for data in train_data():
-            cost_val, acc_val = exe.run(fluid.default_main_program(),
+                cost_val, acc_val = exe.run(main_program,
                                             feed=feeder.feed(data),
                                             fetch_list=[cost, acc_out])
                 print("cost=" + str(cost_val) + " acc=" + str(acc_val))
@@ -192,14 +199,45 @@ def train(word_dict, net_method, use_cuda, parallel=False, save_dirname=None):
         raise AssertionError("Cost is too large for {0}".format(
             net_method.__name__))
 
-
-def infer(use_cuda, save_dirname=None):
+    if is_local:
+        train_loop(fluid.default_main_program())
+    else:
+        port = os.getenv("PADDLE_INIT_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        eplist = []
+        for ip in pserver_ips.split(","):
+            eplist.append(':'.join([ip, port]))
+        pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
+        trainers = int(os.getenv("TRAINERS"))
+        current_endpoint = os.getenv("POD_IP") + ":" + port
+        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
+        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        t = fluid.DistributeTranspiler()
+        t.transpile(
+            optimize_ops,
+            params_grads,
+            trainer_id,
+            pservers=pserver_endpoints,
+            trainers=trainers)
+        if training_role == "PSERVER":
+            pserver_prog = t.get_pserver_program(current_endpoint)
+            pserver_startup = t.get_startup_program(current_endpoint,
+                                                    pserver_prog)
+            exe.run(pserver_startup)
+            exe.run(pserver_prog)
+        elif training_role == "TRAINER":
+            train_loop(t.get_trainer_program())
+
+
+def infer(word_dict, use_cuda, save_dirname=None):
     if save_dirname is None:
         return
 
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     exe = fluid.Executor(place)
 
+    inference_scope = fluid.core.Scope()
+    with fluid.scope_guard(inference_scope):
         # Use fluid.io.load_inference_model to obtain the inference program desc,
         # the feed_target_names (the names of variables that will be feeded
         # data using feed operators), and the fetch_targets (variables that
@@ -207,10 +245,11 @@ def infer(use_cuda, save_dirname=None):
         [inference_program, feed_target_names,
          fetch_targets] = fluid.io.load_inference_model(save_dirname, exe)
 
+        word_dict_len = len(word_dict)
+
         lod = [0, 4, 10]
-    word_dict = paddle.dataset.imdb.word_dict()
         tensor_words = create_random_lodtensor(
-        lod, place, low=0, high=len(word_dict) - 1)
+            lod, place, low=0, high=word_dict_len - 1)
 
         # Construct feed as a dictionary of {feed_target_name: feed_target_data}
         # and results will contain a list of data corresponding to fetch_targets.
@@ -258,7 +297,7 @@ class TestUnderstandSentiment(unittest.TestCase):
                 self.word_dict,
                 net_method=convolution_net,
                 use_cuda=False,
-                save_dirname="understand_sentiment.inference.model")
+                save_dirname="understand_sentiment_conv.inference.model")
 
     def test_conv_cpu_parallel(self):
         with self.new_program_scope():
@@ -271,7 +310,11 @@ class TestUnderstandSentiment(unittest.TestCase):
     @unittest.skip(reason="make CI faster")
     def test_stacked_lstm_cpu(self):
         with self.new_program_scope():
-            main(self.word_dict, net_method=stacked_lstm_net, use_cuda=False)
+            main(
+                self.word_dict,
+                net_method=stacked_lstm_net,
+                use_cuda=False,
+                save_dirname="understand_sentiment_stacked_lstm.inference.model")
 
     def test_stacked_lstm_cpu_parallel(self):
         with self.new_program_scope():
@@ -287,7 +330,7 @@ class TestUnderstandSentiment(unittest.TestCase):
                 self.word_dict,
                 net_method=convolution_net,
                 use_cuda=True,
-                save_dirname="understand_sentiment.inference.model")
+                save_dirname="understand_sentiment_conv.inference.model")
 
     def test_conv_gpu_parallel(self):
         with self.new_program_scope():
@@ -300,7 +343,11 @@ class TestUnderstandSentiment(unittest.TestCase):
     @unittest.skip(reason="make CI faster")
     def test_stacked_lstm_gpu(self):
         with self.new_program_scope():
-            main(self.word_dict, net_method=stacked_lstm_net, use_cuda=True)
+            main(
+                self.word_dict,
+                net_method=stacked_lstm_net,
+                use_cuda=True,
+                save_dirname="understand_sentiment_stacked_lstm.inference.model")
 
     def test_stacked_lstm_gpu_parallel(self):
         with self.new_program_scope():

python/paddle/fluid/tests/book/test_word2vec.py

-#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-# # Licensed under the Apache License, Version 2.0 (the "License");
+#   Copyright (c) 2018 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
 #
@@ -21,6 +22,7 @@ import sys
 
 
 def create_random_lodtensor(lod, place, low, high):
+    # The range of data elements is [low, high]
     data = np.random.random_integers(low, high, [lod[-1], 1]).astype("int64")
     res = fluid.LoDTensor()
     res.set(data, place)
@@ -28,54 +30,7 @@ def create_random_lodtensor(lod, place, low, high):
     return res
 
 
-def infer(use_cuda, save_dirname=None):
-    if save_dirname is None:
-        return
-
-    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-    exe = fluid.Executor(place)
-
-    # Use fluid.io.load_inference_model to obtain the inference program desc,
-    # the feed_target_names (the names of variables that will be feeded 
-    # data using feed operators), and the fetch_targets (variables that 
-    # we want to obtain data from using fetch operators).
-    [inference_program, feed_target_names,
-     fetch_targets] = fluid.io.load_inference_model(save_dirname, exe)
-
-    word_dict = paddle.dataset.imikolov.build_dict()
-    dict_size = len(word_dict) - 1
-
-    # Setup input, by creating 4 words, and setting up lod required for 
-    # lookup_table_op
-    lod = [0, 1]
-    first_word = create_random_lodtensor(lod, place, low=0, high=dict_size)
-    second_word = create_random_lodtensor(lod, place, low=0, high=dict_size)
-    third_word = create_random_lodtensor(lod, place, low=0, high=dict_size)
-    fourth_word = create_random_lodtensor(lod, place, low=0, high=dict_size)
-
-    assert feed_target_names[0] == 'firstw'
-    assert feed_target_names[1] == 'secondw'
-    assert feed_target_names[2] == 'thirdw'
-    assert feed_target_names[3] == 'forthw'
-
-    # Construct feed as a dictionary of {feed_target_name: feed_target_data}
-    # and results will contain a list of data corresponding to fetch_targets.
-    results = exe.run(inference_program,
-                      feed={
-                          feed_target_names[0]: first_word,
-                          feed_target_names[1]: second_word,
-                          feed_target_names[2]: third_word,
-                          feed_target_names[3]: fourth_word
-                      },
-                      fetch_list=fetch_targets,
-                      return_numpy=False)
-    print(results[0].lod())
-    np_data = np.array(results[0])
-    print("Inference Shape: ", np_data.shape)
-    print("Inference results: ", np_data)
-
-
-def train(use_cuda, is_sparse, parallel, save_dirname):
+def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True):
     PASS_NUM = 100
     EMBED_SIZE = 32
     HIDDEN_SIZE = 256
@@ -130,7 +85,7 @@ def train(use_cuda, is_sparse, parallel, save_dirname):
     forth_word = fluid.layers.data(name='forthw', shape=[1], dtype='int64')
     next_word = fluid.layers.data(name='nextw', shape=[1], dtype='int64')
 
-    if not parallel:
+    if not is_parallel:
         avg_cost, predict_word = __network__(
             [first_word, second_word, third_word, forth_word, next_word])
     else:
@@ -146,7 +101,7 @@ def train(use_cuda, is_sparse, parallel, save_dirname):
         avg_cost = fluid.layers.mean(pd())
 
     sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
-    sgd_optimizer.minimize(avg_cost)
+    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
 
     train_reader = paddle.batch(
         paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE)
@@ -157,11 +112,12 @@ def train(use_cuda, is_sparse, parallel, save_dirname):
         feed_list=[first_word, second_word, third_word, forth_word, next_word],
         place=place)
 
+    def train_loop(main_program):
         exe.run(fluid.default_startup_program())
 
         for pass_id in range(PASS_NUM):
             for data in train_reader():
-            avg_cost_np = exe.run(fluid.default_main_program(),
+                avg_cost_np = exe.run(main_program,
                                       feed=feeder.feed(data),
                                       fetch_list=[avg_cost])
                 if avg_cost_np[0] < 5.0:
@@ -175,12 +131,97 @@ def train(use_cuda, is_sparse, parallel, save_dirname):
 
         raise AssertionError("Cost is too large {0:2.2}".format(avg_cost_np[0]))
 
+    if is_local:
+        train_loop(fluid.default_main_program())
+    else:
+        port = os.getenv("PADDLE_INIT_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        eplist = []
+        for ip in pserver_ips.split(","):
+            eplist.append(':'.join([ip, port]))
+        pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
+        trainers = int(os.getenv("TRAINERS"))
+        current_endpoint = os.getenv("POD_IP") + ":" + port
+        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
+        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        t = fluid.DistributeTranspiler()
+        t.transpile(
+            optimize_ops,
+            params_grads,
+            trainer_id,
+            pservers=pserver_endpoints,
+            trainers=trainers)
+        if training_role == "PSERVER":
+            pserver_prog = t.get_pserver_program(current_endpoint)
+            pserver_startup = t.get_startup_program(current_endpoint,
+                                                    pserver_prog)
+            exe.run(pserver_startup)
+            exe.run(pserver_prog)
+        elif training_role == "TRAINER":
+            train_loop(t.get_trainer_program())
 
-def main(use_cuda, is_sparse, parallel):
+
+def infer(use_cuda, save_dirname=None):
+    if save_dirname is None:
+        return
+
+    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
+    exe = fluid.Executor(place)
+
+    inference_scope = fluid.core.Scope()
+    with fluid.scope_guard(inference_scope):
+        # Use fluid.io.load_inference_model to obtain the inference program desc,
+        # the feed_target_names (the names of variables that will be feeded
+        # data using feed operators), and the fetch_targets (variables that
+        # we want to obtain data from using fetch operators).
+        [inference_program, feed_target_names,
+         fetch_targets] = fluid.io.load_inference_model(save_dirname, exe)
+
+        word_dict = paddle.dataset.imikolov.build_dict()
+        dict_size = len(word_dict)
+
+        # Setup inputs, by creating 4 words, the lod of which should be [0, 1]
+        lod = [0, 1]
+        first_word = create_random_lodtensor(
+            lod, place, low=0, high=dict_size - 1)
+        second_word = create_random_lodtensor(
+            lod, place, low=0, high=dict_size - 1)
+        third_word = create_random_lodtensor(
+            lod, place, low=0, high=dict_size - 1)
+        fourth_word = create_random_lodtensor(
+            lod, place, low=0, high=dict_size - 1)
+
+        assert feed_target_names[0] == 'firstw'
+        assert feed_target_names[1] == 'secondw'
+        assert feed_target_names[2] == 'thirdw'
+        assert feed_target_names[3] == 'forthw'
+
+        # Construct feed as a dictionary of {feed_target_name: feed_target_data}
+        # and results will contain a list of data corresponding to fetch_targets.
+        results = exe.run(inference_program,
+                          feed={
+                              feed_target_names[0]: first_word,
+                              feed_target_names[1]: second_word,
+                              feed_target_names[2]: third_word,
+                              feed_target_names[3]: fourth_word
+                          },
+                          fetch_list=fetch_targets,
+                          return_numpy=False)
+        print(results[0].lod())
+        np_data = np.array(results[0])
+        print("Inference Shape: ", np_data.shape)
+
+
+def main(use_cuda, is_sparse, is_parallel):
     if use_cuda and not fluid.core.is_compiled_with_cuda():
         return
+
+    if not is_parallel:
         save_dirname = "word2vec.inference.model"
-    train(use_cuda, is_sparse, parallel, save_dirname)
+    else:
+        save_dirname = None
+
+    train(use_cuda, is_sparse, is_parallel, save_dirname)
     infer(use_cuda, save_dirname)
 
 
@@ -193,10 +234,10 @@ class W2VTest(unittest.TestCase):
     pass
 
 
-def inject_test_method(use_cuda, is_sparse, parallel):
+def inject_test_method(use_cuda, is_sparse, is_parallel):
     fn_name = "test_{0}_{1}_{2}".format("cuda" if use_cuda else "cpu", "sparse"
                                         if is_sparse else "dense", "parallel"
-                                        if parallel else "normal")
+                                        if is_parallel else "normal")
 
     def __impl__(*args, **kwargs):
         prog = fluid.Program()
@@ -204,10 +245,12 @@ def inject_test_method(use_cuda, is_sparse, parallel):
         scope = fluid.core.Scope()
         with fluid.scope_guard(scope):
             with fluid.program_guard(prog, startup_prog):
-                main(use_cuda=use_cuda, is_sparse=is_sparse, parallel=parallel)
+                main(
+                    use_cuda=use_cuda,
+                    is_sparse=is_sparse,
+                    is_parallel=is_parallel)
 
-    # run only 2 cases: use_cuda is either True or False
-    if is_sparse == False and parallel == False:
+    if use_cuda and is_sparse:
         fn = __impl__
     else:
         # skip the other test when on CI server
@@ -219,8 +262,8 @@ def inject_test_method(use_cuda, is_sparse, parallel):
 
 for use_cuda in (False, True):
     for is_sparse in (False, True):
-        for parallel in (False, True):
-            inject_test_method(use_cuda, is_sparse, parallel)
+        for is_parallel in (False, True):
+            inject_test_method(use_cuda, is_sparse, is_parallel)
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/book_distribute/CMakeLists.txt

-file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
-string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
-foreach(src ${TEST_OPS})
-    py_test(${src} SRCS ${src}.py)
-endforeach()

python/paddle/fluid/tests/book_distribute/notest_dist_fit_a_line.py

-#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import numpy as np
-import paddle.v2 as paddle
-import paddle.fluid as fluid
-import os
-
-x = fluid.layers.data(name='x', shape=[13], dtype='float32')
-
-y_predict = fluid.layers.fc(input=x, size=1, act=None)
-
-y = fluid.layers.data(name='y', shape=[1], dtype='float32')
-
-cost = fluid.layers.square_error_cost(input=y_predict, label=y)
-avg_cost = fluid.layers.mean(cost)
-
-sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
-optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
-
-BATCH_SIZE = 20
-
-train_reader = paddle.batch(
-    paddle.reader.shuffle(
-        paddle.dataset.uci_housing.train(), buf_size=500),
-    batch_size=BATCH_SIZE)
-
-place = fluid.CPUPlace()
-feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
-exe = fluid.Executor(place)
-
-t = fluid.DistributeTranspiler()
-# all parameter server endpoints list for spliting parameters
-pserver_endpoints = os.getenv("PSERVERS")
-# server endpoint for current node
-current_endpoint = os.getenv("SERVER_ENDPOINT")
-# run as trainer or parameter server
-training_role = os.getenv("TRAINING_ROLE",
-                          "TRAINER")  # get the training role: trainer/pserver
-t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
-
-if training_role == "PSERVER":
-    if not current_endpoint:
-        print("need env SERVER_ENDPOINT")
-        exit(1)
-    pserver_prog = t.get_pserver_program(current_endpoint)
-    pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
-    exe.run(pserver_startup)
-    exe.run(pserver_prog)
-else:
-    trainer_prog = t.get_trainer_program()
-
-    exe.run(fluid.default_startup_program())
-
-    PASS_NUM = 100
-    for pass_id in range(PASS_NUM):
-        fluid.io.save_persistables(exe, "./fit_a_line.model/")
-        fluid.io.load_persistables(exe, "./fit_a_line.model/")
-        for data in train_reader():
-            avg_loss_value = exe.run(trainer_prog,
-                                     feed=feeder.feed(data),
-                                     fetch_list=[avg_cost])
-            print("loss:" + str(avg_loss_value))
-            if avg_loss_value[0] < 10.0:
-                exit(0)
-exit(1)

python/paddle/fluid/tests/book_distribute/notest_dist_image_classification.py

-#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import print_function
-
-import paddle.v2 as paddle
-import paddle.fluid as fluid
-import os
-import sys
-
-TRAINERS = 5
-BATCH_SIZE = 128
-PASS_NUM = 100
-
-
-def resnet_cifar10(input, depth=32):
-    def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'):
-        tmp = fluid.layers.conv2d(
-            input=input,
-            filter_size=filter_size,
-            num_filters=ch_out,
-            stride=stride,
-            padding=padding,
-            act=None,
-            bias_attr=False)
-        return fluid.layers.batch_norm(input=tmp, act=act)
-
-    def shortcut(input, ch_in, ch_out, stride):
-        if ch_in != ch_out:
-            return conv_bn_layer(input, ch_out, 1, stride, 0, None)
-        else:
-            return input
-
-    def basicblock(input, ch_in, ch_out, stride):
-        tmp = conv_bn_layer(input, ch_out, 3, stride, 1)
-        tmp = conv_bn_layer(tmp, ch_out, 3, 1, 1, act=None)
-        short = shortcut(input, ch_in, ch_out, stride)
-        return fluid.layers.elementwise_add(x=tmp, y=short, act='relu')
-
-    def layer_warp(block_func, input, ch_in, ch_out, count, stride):
-        tmp = block_func(input, ch_in, ch_out, stride)
-        for i in range(1, count):
-            tmp = block_func(tmp, ch_out, ch_out, 1)
-        return tmp
-
-    assert (depth - 2) % 6 == 0
-    n = (depth - 2) / 6
-    conv1 = conv_bn_layer(
-        input=input, ch_out=16, filter_size=3, stride=1, padding=1)
-    res1 = layer_warp(basicblock, conv1, 16, 16, n, 1)
-    res2 = layer_warp(basicblock, res1, 16, 32, n, 2)
-    res3 = layer_warp(basicblock, res2, 32, 64, n, 2)
-    pool = fluid.layers.pool2d(
-        input=res3, pool_size=8, pool_type='avg', pool_stride=1)
-    return pool
-
-
-def vgg16_bn_drop(input):
-    def conv_block(input, num_filter, groups, dropouts):
-        return fluid.nets.img_conv_group(
-            input=input,
-            pool_size=2,
-            pool_stride=2,
-            conv_num_filter=[num_filter] * groups,
-            conv_filter_size=3,
-            conv_act='relu',
-            conv_with_batchnorm=True,
-            conv_batchnorm_drop_rate=dropouts,
-            pool_type='max')
-
-    conv1 = conv_block(input, 64, 2, [0.3, 0])
-    conv2 = conv_block(conv1, 128, 2, [0.4, 0])
-    conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0])
-    conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0])
-    conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0])
-
-    drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5)
-    fc1 = fluid.layers.fc(input=drop, size=512, act=None)
-    bn = fluid.layers.batch_norm(input=fc1, act='relu')
-    drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5)
-    fc2 = fluid.layers.fc(input=drop2, size=512, act=None)
-    return fc2
-
-
-classdim = 10
-data_shape = [3, 32, 32]
-
-images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
-label = fluid.layers.data(name='label', shape=[1], dtype='int64')
-
-net_type = "vgg"
-if len(sys.argv) >= 2:
-    net_type = sys.argv[1]
-
-if net_type == "vgg":
-    print("training vgg net")
-    net = vgg16_bn_drop(images)
-elif net_type == "resnet":
-    print("training resnet")
-    net = resnet_cifar10(images, 32)
-else:
-    raise ValueError("%s network is not supported" % net_type)
-
-predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
-cost = fluid.layers.cross_entropy(input=predict, label=label)
-avg_cost = fluid.layers.mean(cost)
-
-optimizer = fluid.optimizer.Adam(learning_rate=0.001)
-optimize_ops, params_grads = optimizer.minimize(avg_cost)
-
-accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
-
-train_reader = paddle.batch(
-    paddle.reader.shuffle(
-        paddle.dataset.cifar.train10(), buf_size=128 * 10),
-    batch_size=BATCH_SIZE)
-
-place = fluid.CPUPlace()
-feeder = fluid.DataFeeder(place=place, feed_list=[images, label])
-exe = fluid.Executor(place)
-
-t = fluid.DistributeTranspiler()
-# all parameter server endpoints list for spliting parameters
-pserver_endpoints = os.getenv("PSERVERS")
-# server endpoint for current node
-current_endpoint = os.getenv("SERVER_ENDPOINT")
-# run as trainer or parameter server
-training_role = os.getenv("TRAINING_ROLE",
-                          "TRAINER")  # get the training role: trainer/pserver
-t.transpile(
-    optimize_ops, params_grads, pservers=pserver_endpoints, trainers=TRAINERS)
-
-if training_role == "PSERVER":
-    if not current_endpoint:
-        print("need env SERVER_ENDPOINT")
-        exit(1)
-    pserver_prog = t.get_pserver_program(current_endpoint)
-    pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
-    exe.run(pserver_startup)
-    exe.run(pserver_prog)
-elif training_role == "TRAINER":
-    trainer_prog = t.get_trainer_program()
-    exe.run(fluid.default_startup_program())
-
-    for pass_id in range(PASS_NUM):
-        accuracy.reset(exe)
-        for data in train_reader():
-            loss, acc = exe.run(trainer_prog,
-                                feed=feeder.feed(data),
-                                fetch_list=[avg_cost] + accuracy.metrics)
-            pass_acc = accuracy.eval(exe)
-            print("pass_id:" + str(pass_id) + "loss:" + str(loss) + " pass_acc:"
-                  + str(pass_acc))
-            # this model is slow, so if we can train two mini batches,
-            # we think it works properly.
-    print("trainer run end")
-else:
-    print("environment var TRAINER_ROLE should be TRAINER os PSERVER")
-exit(1)

python/paddle/fluid/tests/book_distribute/notest_dist_label_semantic_roles.py

-#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import math
-
-import numpy as np
-import paddle.v2 as paddle
-import paddle.v2.dataset.conll05 as conll05
-import paddle.fluid as fluid
-import time
-import os
-
-word_dict, verb_dict, label_dict = conll05.get_dict()
-word_dict_len = len(word_dict)
-label_dict_len = len(label_dict)
-pred_len = len(verb_dict)
-
-mark_dict_len = 2
-word_dim = 32
-mark_dim = 5
-hidden_dim = 512
-depth = 8
-mix_hidden_lr = 1e-3
-
-IS_SPARSE = True
-PASS_NUM = 10
-BATCH_SIZE = 20
-
-embedding_name = 'emb'
-
-
-def load_parameter(file_name, h, w):
-    with open(file_name, 'rb') as f:
-        f.read(16)  # skip header.
-        return np.fromfile(f, dtype=np.float32).reshape(h, w)
-
-
-def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark,
-            **ignored):
-    # 8 features
-    predicate_embedding = fluid.layers.embedding(
-        input=predicate,
-        size=[pred_len, word_dim],
-        dtype='float32',
-        is_sparse=IS_SPARSE,
-        param_attr='vemb')
-
-    mark_embedding = fluid.layers.embedding(
-        input=mark,
-        size=[mark_dict_len, mark_dim],
-        dtype='float32',
-        is_sparse=IS_SPARSE)
-
-    word_input = [word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2]
-    emb_layers = [
-        fluid.layers.embedding(
-            size=[word_dict_len, word_dim],
-            input=x,
-            param_attr=fluid.ParamAttr(
-                name=embedding_name, trainable=False)) for x in word_input
-    ]
-    emb_layers.append(predicate_embedding)
-    emb_layers.append(mark_embedding)
-
-    hidden_0_layers = [
-        fluid.layers.fc(input=emb, size=hidden_dim) for emb in emb_layers
-    ]
-
-    hidden_0 = fluid.layers.sums(input=hidden_0_layers)
-
-    lstm_0 = fluid.layers.dynamic_lstm(
-        input=hidden_0,
-        size=hidden_dim,
-        candidate_activation='relu',
-        gate_activation='sigmoid',
-        cell_activation='sigmoid')
-
-    # stack L-LSTM and R-LSTM with direct edges
-    input_tmp = [hidden_0, lstm_0]
-
-    for i in range(1, depth):
-        mix_hidden = fluid.layers.sums(input=[
-            fluid.layers.fc(input=input_tmp[0], size=hidden_dim),
-            fluid.layers.fc(input=input_tmp[1], size=hidden_dim)
-        ])
-
-        lstm = fluid.layers.dynamic_lstm(
-            input=mix_hidden,
-            size=hidden_dim,
-            candidate_activation='relu',
-            gate_activation='sigmoid',
-            cell_activation='sigmoid',
-            is_reverse=((i % 2) == 1))
-
-        input_tmp = [mix_hidden, lstm]
-
-    feature_out = fluid.layers.sums(input=[
-        fluid.layers.fc(input=input_tmp[0], size=label_dict_len),
-        fluid.layers.fc(input=input_tmp[1], size=label_dict_len)
-    ])
-
-    return feature_out
-
-
-def to_lodtensor(data, place):
-    seq_lens = [len(seq) for seq in data]
-    cur_len = 0
-    lod = [cur_len]
-    for l in seq_lens:
-        cur_len += l
-        lod.append(cur_len)
-    flattened_data = np.concatenate(data, axis=0).astype("int64")
-    flattened_data = flattened_data.reshape([len(flattened_data), 1])
-    res = fluid.LoDTensor()
-    res.set(flattened_data, place)
-    res.set_lod([lod])
-    return res
-
-
-def main():
-    # define network topology
-    word = fluid.layers.data(
-        name='word_data', shape=[1], dtype='int64', lod_level=1)
-    predicate = fluid.layers.data(
-        name='verb_data', shape=[1], dtype='int64', lod_level=1)
-    ctx_n2 = fluid.layers.data(
-        name='ctx_n2_data', shape=[1], dtype='int64', lod_level=1)
-    ctx_n1 = fluid.layers.data(
-        name='ctx_n1_data', shape=[1], dtype='int64', lod_level=1)
-    ctx_0 = fluid.layers.data(
-        name='ctx_0_data', shape=[1], dtype='int64', lod_level=1)
-    ctx_p1 = fluid.layers.data(
-        name='ctx_p1_data', shape=[1], dtype='int64', lod_level=1)
-    ctx_p2 = fluid.layers.data(
-        name='ctx_p2_data', shape=[1], dtype='int64', lod_level=1)
-    mark = fluid.layers.data(
-        name='mark_data', shape=[1], dtype='int64', lod_level=1)
-    feature_out = db_lstm(**locals())
-    target = fluid.layers.data(
-        name='target', shape=[1], dtype='int64', lod_level=1)
-    crf_cost = fluid.layers.linear_chain_crf(
-        input=feature_out,
-        label=target,
-        param_attr=fluid.ParamAttr(
-            name='crfw', learning_rate=mix_hidden_lr))
-    avg_cost = fluid.layers.mean(crf_cost)
-
-    # TODO(qiao)
-    # check other optimizers and check why out will be NAN
-    sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.0001)
-    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
-
-    # TODO(qiao)
-    # add dependency track and move this config before optimizer
-    crf_decode = fluid.layers.crf_decoding(
-        input=feature_out, param_attr=fluid.ParamAttr(name='crfw'))
-
-    chunk_evaluator = fluid.evaluator.ChunkEvaluator(
-        input=crf_decode,
-        label=target,
-        chunk_scheme="IOB",
-        num_chunk_types=int(math.ceil((label_dict_len - 1) / 2.0)))
-
-    train_data = paddle.batch(
-        paddle.reader.shuffle(
-            paddle.dataset.conll05.test(), buf_size=8192),
-        batch_size=BATCH_SIZE)
-    place = fluid.CPUPlace()
-    feeder = fluid.DataFeeder(
-        feed_list=[
-            word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, predicate, mark, target
-        ],
-        place=place)
-    exe = fluid.Executor(place)
-
-    t = fluid.DistributeTranspiler()
-    pserver_endpoints = os.getenv("PSERVERS")
-    # server endpoint for current node
-    current_endpoint = os.getenv("SERVER_ENDPOINT")
-    # run as trainer or parameter server
-    training_role = os.getenv(
-        "TRAINING_ROLE", "TRAINER")  # get the training role: trainer/pserver
-    t.transpile(
-        optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
-
-    if training_role == "PSERVER":
-        if not current_endpoint:
-            print("need env SERVER_ENDPOINT")
-            exit(1)
-        pserver_prog = t.get_pserver_program(current_endpoint)
-        pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
-        exe.run(pserver_startup)
-        exe.run(pserver_prog)
-    elif training_role == "TRAINER":
-        trainer_prog = t.get_trainer_program()
-        start_time = time.time()
-        batch_id = 0
-        exe.run(fluid.default_startup_program())
-        embedding_param = fluid.global_scope().find_var(
-            embedding_name).get_tensor()
-        embedding_param.set(
-            load_parameter(conll05.get_embedding(), word_dict_len, word_dim),
-            place)
-        for pass_id in xrange(PASS_NUM):
-            chunk_evaluator.reset(exe)
-            for data in train_data():
-                cost, precision, recall, f1_score = exe.run(
-                    trainer_prog,
-                    feed=feeder.feed(data),
-                    fetch_list=[avg_cost] + chunk_evaluator.metrics)
-                pass_precision, pass_recall, pass_f1_score = chunk_evaluator.eval(
-                    exe)
-
-                if batch_id % 10 == 0:
-                    print("avg_cost:" + str(cost) + " precision:" + str(
-                        precision) + " recall:" + str(recall) + " f1_score:" +
-                          str(f1_score) + " pass_precision:" + str(
-                              pass_precision) + " pass_recall:" + str(
-                                  pass_recall) + " pass_f1_score:" + str(
-                                      pass_f1_score))
-                    if batch_id != 0:
-                        print("second per batch: " + str((time.time(
-                        ) - start_time) / batch_id))
-
-                batch_id = batch_id + 1
-
-
-if __name__ == '__main__':
-    main()

python/paddle/fluid/tests/book_distribute/notest_dist_word2vec.py

-#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import print_function
-import numpy as np
-import paddle.v2 as paddle
-import paddle.fluid as fluid
-import os
-
-PASS_NUM = 100
-EMBED_SIZE = 32
-HIDDEN_SIZE = 256
-N = 5
-BATCH_SIZE = 32
-IS_SPARSE = True
-TRAINERS = 2
-
-word_dict = paddle.dataset.imikolov.build_dict()
-dict_size = len(word_dict)
-
-first_word = fluid.layers.data(name='firstw', shape=[1], dtype='int64')
-second_word = fluid.layers.data(name='secondw', shape=[1], dtype='int64')
-third_word = fluid.layers.data(name='thirdw', shape=[1], dtype='int64')
-forth_word = fluid.layers.data(name='forthw', shape=[1], dtype='int64')
-next_word = fluid.layers.data(name='nextw', shape=[1], dtype='int64')
-
-embed_first = fluid.layers.embedding(
-    input=first_word,
-    size=[dict_size, EMBED_SIZE],
-    dtype='float32',
-    is_sparse=IS_SPARSE,
-    param_attr='shared_w')
-embed_second = fluid.layers.embedding(
-    input=second_word,
-    size=[dict_size, EMBED_SIZE],
-    dtype='float32',
-    is_sparse=IS_SPARSE,
-    param_attr='shared_w')
-embed_third = fluid.layers.embedding(
-    input=third_word,
-    size=[dict_size, EMBED_SIZE],
-    dtype='float32',
-    is_sparse=IS_SPARSE,
-    param_attr='shared_w')
-embed_forth = fluid.layers.embedding(
-    input=forth_word,
-    size=[dict_size, EMBED_SIZE],
-    dtype='float32',
-    is_sparse=IS_SPARSE,
-    param_attr='shared_w')
-
-concat_embed = fluid.layers.concat(
-    input=[embed_first, embed_second, embed_third, embed_forth], axis=1)
-hidden1 = fluid.layers.fc(input=concat_embed, size=HIDDEN_SIZE, act='sigmoid')
-predict_word = fluid.layers.fc(input=hidden1, size=dict_size, act='softmax')
-cost = fluid.layers.cross_entropy(input=predict_word, label=next_word)
-avg_cost = fluid.layers.mean(cost)
-sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
-optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
-train_reader = paddle.batch(
-    paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE)
-
-place = fluid.CPUPlace()
-exe = fluid.Executor(place)
-
-t = fluid.DistributeTranspiler()
-# all parameter server endpoints list for spliting parameters
-pserver_endpoints = os.getenv("PSERVERS")
-# server endpoint for current node
-current_endpoint = os.getenv("SERVER_ENDPOINT")
-# run as trainer or parameter server
-training_role = os.getenv("TRAINING_ROLE",
-                          "TRAINER")  # get the training role: trainer/pserver
-t.transpile(
-    optimize_ops, params_grads, pservers=pserver_endpoints, trainers=TRAINERS)
-if training_role == "PSERVER":
-    if not current_endpoint:
-        print("need env SERVER_ENDPOINT")
-        exit(1)
-    pserver_prog = t.get_pserver_program(current_endpoint)
-    pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
-    exe.run(pserver_startup)
-    exe.run(pserver_prog)
-elif training_role == "TRAINER":
-    feeder = fluid.DataFeeder(
-        feed_list=[first_word, second_word, third_word, forth_word, next_word],
-        place=place)
-    exe.run(fluid.default_startup_program())
-    for pass_id in range(PASS_NUM):
-        for data in train_reader():
-            avg_cost_np = exe.run(t.get_trainer_program(),
-                                  feed=feeder.feed(data),
-                                  fetch_list=[avg_cost])
-            print("avg_cost_np", avg_cost_np)
-            if avg_cost_np[0] < 5.0:
-                exit(
-                    0)  # if avg cost less than 10.0, we think our code is good.
-else:
-    print("environment var TRAINER_ROLE should be TRAINER os PSERVER")
-exit(1)

python/paddle/fluid/tests/book_distribute/notest_recognize_digits_conv_dist.py

-#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from __future__ import print_function
-import numpy as np
-import paddle.v2 as paddle
-import paddle.fluid as fluid
-import os
-
-images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype='float32')
-label = fluid.layers.data(name='label', shape=[1], dtype='int64')
-conv_pool_1 = fluid.nets.simple_img_conv_pool(
-    input=images,
-    filter_size=5,
-    num_filters=20,
-    pool_size=2,
-    pool_stride=2,
-    act="relu")
-conv_pool_2 = fluid.nets.simple_img_conv_pool(
-    input=conv_pool_1,
-    filter_size=5,
-    num_filters=50,
-    pool_size=2,
-    pool_stride=2,
-    act="relu")
-
-predict = fluid.layers.fc(input=conv_pool_2, size=10, act="softmax")
-cost = fluid.layers.cross_entropy(input=predict, label=label)
-avg_cost = fluid.layers.mean(cost)
-optimizer = fluid.optimizer.Adam(learning_rate=0.01)
-optimize_ops, params_grads = optimizer.minimize(avg_cost)
-
-accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
-
-BATCH_SIZE = 50
-PASS_NUM = 3
-train_reader = paddle.batch(
-    paddle.reader.shuffle(
-        paddle.dataset.mnist.train(), buf_size=500),
-    batch_size=BATCH_SIZE)
-
-place = fluid.CPUPlace()
-exe = fluid.Executor(place)
-
-pserver_endpoints = os.getenv("PSERVERS")  # all pserver endpoints
-trainers = int(os.getenv("TRAINERS"))  # total trainer count
-current_endpoint = os.getenv("SERVER_ENDPOINT")  # current pserver endpoint
-training_role = os.getenv("TRAINING_ROLE",
-                          "TRAINER")  # get the training role: trainer/pserver
-if not current_endpoint:
-    print("need env SERVER_ENDPOINT")
-    exit(1)
-
-t = fluid.DistributeTranspiler()
-t.transpile(
-    optimize_ops,
-    params_grads,
-    0,
-    pservers=pserver_endpoints,
-    trainers=trainers)
-
-if training_role == "PSERVER":
-    pserver_prog = t.get_pserver_program(current_endpoint)
-    pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
-    exe.run(pserver_startup)
-    exe.run(pserver_prog)
-elif training_role == "TRAINER":
-    trainer_prog = t.get_trainer_program()
-    feeder = fluid.DataFeeder(feed_list=[images, label], place=place)
-    # TODO(typhoonzero): change trainer startup program to fetch parameters from pserver
-    exe.run(fluid.default_startup_program())
-
-    for pass_id in range(PASS_NUM):
-        accuracy.reset(exe)
-        batch_id = 0
-        for data in train_reader():
-            loss, acc = exe.run(trainer_prog,
-                                feed=feeder.feed(data),
-                                fetch_list=[avg_cost] + accuracy.metrics)
-            pass_acc = accuracy.eval(exe)
-            if batch_id % 100 == 0:
-                print("batch_id %d, loss: %f, acc: %f" %
-                      (batch_id, loss, pass_acc))
-            batch_id += 1
-
-        pass_acc = accuracy.eval(exe)
-        print("pass_id=" + str(pass_id) + " pass_acc=" + str(pass_acc))
-else:
-    print("environment var TRAINER_ROLE should be TRAINER os PSERVER")

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -41,6 +41,7 @@ list(REMOVE_ITEM TEST_OPS test_while_op)
 list(REMOVE_ITEM TEST_OPS test_lod_array_length_op)
 list(REMOVE_ITEM TEST_OPS test_reorder_lod_tensor)
 list(REMOVE_ITEM TEST_OPS test_profiler)
+list(REMOVE_ITEM TEST_OPS test_nvprof)
 list(REMOVE_ITEM TEST_OPS test_normalization_wrapper)
 list(REMOVE_ITEM TEST_OPS test_executor_and_mul)
 list(REMOVE_ITEM TEST_OPS test_assign_value_op)
@@ -75,6 +76,7 @@ py_test_modules(test_while_op MODULES test_while_op)
 py_test_modules(test_lod_array_length_op MODULES test_lod_array_length_op)
 py_test_modules(test_reorder_lod_tensor MODULES test_reorder_lod_tensor)
 py_test_modules(test_profiler MODULES test_profiler)
+py_test_modules(test_nvprof MODULES test_nvprof)
 py_test_modules(test_normalization_wrapper MODULES test_normalization_wrapper)
 py_test_modules(test_executor_and_mul MODULES test_executor_and_mul)
 py_test_modules(test_assign_value_op MODULES test_assign_value_op)

python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py

@@ -200,7 +200,8 @@ class TestCUDNNWithStride(TestWithStride):
         self.op_type = "conv2d_transpose"
 
 
-# #cudnn v5 does not support dilation conv.
+# Please Don't remove the following code.
+# Currently, CI use cudnn V5.0 which not support dilation conv.
 # class TestCUDNNWithDilation(TestWithDilation):
 #     def init_test_case(self):
 #         self.pad = [1, 1]

python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py

@@ -207,7 +207,8 @@ class TestCUDNNWithStride(TestWithStride):
         self.op_type = "conv3d_transpose"
 
 
-# #cudnn v5 does not support dilation conv.
+# Please Don't remove the following code.
+# Currently, CI use cudnn V5.0 which not support dilation conv.
 # class TestCUDNNWithDilation(TestWithDilation):
 #     def init_test_case(self):
 #         self.pad = [1, 1, 1]