Address comments, add c++ inference example and Merge remote branch

CMakeLists.txt

@@ -137,7 +137,7 @@ include(external/openblas)  # download, build, install openblas
 include(external/mkldnn)    # download, build, install mkldnn
 include(external/swig)      # download, build, install swig
 include(external/warpctc)   # download, build, install warpctc
-include(external/boost)     # download, build, install boost
+include(external/boost)     # download boost
 include(external/any)       # download libn::any
 include(external/eigen)     # download eigen3
 include(external/pybind11)  # download pybind11
@@ -156,6 +156,7 @@ include(rdma)               # set rdma libraries
 include(flags)              # set paddle compile flags
 include(version)            # set PADDLE_VERSION
 include(coveralls)          # set code coverage
+include(inference_lib)      # add paddle fluid inference libraries
 
 
 include_directories("${PADDLE_SOURCE_DIR}")

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.
+RUN pip install paddlepaddle
+RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.cifar.train10()" | python'
+RUN pip uninstall -y paddlepaddle
+
+# 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
+ADD vgg16_fluid.py vgg16_v2.py /workspace/

benchmark/cluster/vgg16/README.md

+# Performance for Distributed vgg16
+
+## Test Result
+
+### Hardware Infomation
+
+- CPU: Intel(R) Xeon(R) CPU E5-2620 v4 @ 2.10GHz
+- cpu MHz		: 2101.000
+- cache size	: 20480 KB
+
+### Single Node Single Thread
+
+- PServer Count: 10
+- Trainer Count: 20
+- Metrics: samples / sec
+
+| Batch Size | 32 | 64 | 128 | 256 |
+| -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid | 15.44 | 16.32 | 16.74 | 16.79 |
+| PaddlePaddle v2 | 15.97 | 17.04 | 17.60 | 17.83 |
+| TensorFlow | - | - | - | - |
+
+### Different Batch Size
+
+- PServer Count: 10
+- Trainer Count: 20
+- Per trainer CPU Core: 1
+- Metrics: samples / sec
+
+| Batch Size | 32 | 64 | 128 | 256 |
+| -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid | 190.20 | 222.15 | 247.40 | 258.18 |
+| PaddlePaddle v2 | 170.96 | 233.71 | 256.14 | 329.23 |
+| TensorFlow | - | - | - | - |
+
+
+### Accelerate Rate
+
+- Pserver Count: 20
+- Batch Size: 128
+- Metrics: samples / sec
+
+| Trainer Count | 20 | 40 | 80 | 100 |
+| -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid | 263.29 (78.64%) | 518.80 (77.47%) | 836.26 (62.44%) | 1019.29 (60.89%) |
+| PaddlePaddle v2 (need more tests) | 326.85 (92.85%) | 534.58 (75.93%) | 853.30 (60.60%) | 1041.99 (59.20%) |
+| TensorFlow | - | - | - | - |
+
+### Different Pserver Count
+
+- Trainer Count: 60
+- Batch Size: 128
+- Metrics: samples/ sec
+
+| PServer Count | 3 | 6 |10 | 20 |
+| -- | -- | -- | -- | -- |
+| PaddlePaddle Fluid(should fix in next PR) | 589.1 | 592.6 | 656.4 | 655.8 |
+| PaddlePaddle v2 | 593.4 | 791.3 | 729.7 | 821.7 |
+| TensorFlow | - | - | - | - |
+
+*The performance gap between Fuild and v2 comes from the network interference.*
+
+
+## Steps to Run the Performance Test
+
+1. You must re-compile PaddlePaddle and enable `-DWITH_DISTRIBUTE` to build PaddlePaddle with distributed support.
+1. When the build finishes, copy the output `whl` package located under `build/python/dist` to current directory.
+1. Run `docker build -t [image:tag] .` to build the docker image and run `docker push [image:tag]` to push the image to reponsitory so kubernetes can find it.
+1. Run `kubectl create -f pserver.yaml && kubectl create -f trainer.yaml` to start the job on your kubernetes cluster (you must configure the `kubectl` client before this step).
+1. Run `kubectl get po` to get running pods, and run `kubectl logs [podID]` to fetch the pod log of pservers and trainers.
+
+Check the logs for the distributed training progress and analyze the performance.
+
+## Enable Verbos Logs
+
+Edit `pserver.yaml` and `trainer.yaml` and add an environment variable `GLOG_v=3` and `GLOG_logtostderr=1` to see what happend in detail.

benchmark/cluster/vgg16/fluid_pserver.yaml

+apiVersion: extensions/v1beta1
+kind: ReplicaSet
+metadata:
+  name: vgg16job-pserver
+spec:
+  replicas: 10
+  template:
+    metadata:
+      labels:
+        paddle-job-pserver: vgg16job
+    spec:
+      hostNetwork: true
+      imagePullSecrets:
+      - name: job-registry-secret
+      containers:
+      - name: pserver
+        image: "registry.baidu.com/paddlepaddle/fluid_benchmark:vgg16"
+        imagePullPolicy: Always
+        ports:
+        - name: jobport-30236
+          containerPort: 30236
+        env:
+        - name: PADDLE_JOB_NAME
+          value: vgg16job
+        - name: MKL_NUM_THREADS
+          value: "1"
+        - name: TRAINING_ROLE
+          value: "PSERVER"
+        - name: TRAINERS
+          value: "20"
+        - name: PSERVERS
+          value: "10"
+        - name: TOPOLOGY
+          value: ""
+        - name: ENTRY
+          value: "MKL_NUM_THREADS=1 python /workspace/vgg16_fluid.py --local 0"
+        - name: TRAINER_PACKAGE
+          value: "/workspace"
+        - name: PADDLE_INIT_PORT
+          value: "30236"
+        - name: PADDLE_INIT_NICS
+          value: "xgbe0"
+        - name: PADDLE_INIT_TRAINER_COUNT
+          value: "1"
+        - name: PADDLE_INIT_PORTS_NUM
+          value: "1"
+        - name: PADDLE_INIT_PORTS_NUM_FOR_SPARSE
+          value: "1"
+        - name: PADDLE_INIT_NUM_GRADIENT_SERVERS
+          value: "20"
+        - name: PADDLE_INIT_NUM_PASSES
+          value: "1"
+        - name: PADDLE_INIT_USE_GPU
+          value: "0"
+        - name: LD_LIBRARY_PATH
+          value: "/usr/local/lib:/usr/local/nvidia/lib64"
+        - name: NAMESPACE
+          valueFrom:
+            fieldRef:
+              fieldPath: "metadata.namespace"
+        - name: POD_IP
+          valueFrom:
+            fieldRef:
+              fieldPath: "status.podIP"
+        command: ["paddle_k8s", "start_fluid"]
+        resources:
+          requests:
+            memory: 10Gi
+            cpu: 4
+          limits:
+            memory: 10Gi
+            cpu: 4

benchmark/cluster/vgg16/fluid_trainer.yaml

+apiVersion: batch/v1
+kind: Job
+metadata:
+  name: vgg16job-trainer
+spec:
+  parallelism: 20
+  completions: 20
+  template:
+    metadata:
+      labels:
+        paddle-job: vgg16job
+    spec:
+      imagePullSecrets:
+        - name: job-registry-secret
+      hostNetwork: true
+      containers:
+      - name: trainer
+        image: "registry.baidu.com/paddlepaddle/fluid_benchmark:vgg16"
+        imagePullPolicy: Always
+        command: ["paddle_k8s", "start_fluid"]
+        env:
+        - name: PADDLE_JOB_NAME
+          value: vgg16job
+        - name: TRAINING_ROLE
+          value: "TRAINER"
+        - name: TRAINERS
+          value: "20"
+        - name: PSERVERS
+          value: "10"
+        - name: TOPOLOGY
+          value: ""
+        - name: ENTRY
+          value: "MKL_NUM_THREADS=1 python /workspace/vgg16_fluid.py --local 0 --batch_size 128"
+        - name: TRAINER_PACKAGE
+          value: "/workspace"
+        - name: PADDLE_INIT_PORT
+          value: "30236"
+        - name: PADDLE_INIT_NICS
+          value: "xgbe0"
+        - name: PADDLE_INIT_TRAINER_COUNT
+          value: "1"
+        - name: PADDLE_INIT_PORTS_NUM
+          value: "1"
+        - name: PADDLE_INIT_PORTS_NUM_FOR_SPARSE
+          value: "1"
+        - name: PADDLE_INIT_NUM_GRADIENT_SERVERS
+          value: "20"
+        - name: PADDLE_INIT_NUM_PASSES
+          value: "1"
+        - name: PADDLE_INIT_USE_GPU
+          value: "0"
+        - name: LD_LIBRARY_PATH
+          value: "/usr/local/lib:/usr/local/nvidia/lib64"
+        - 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/v2_pserver.yaml

+apiVersion: extensions/v1beta1
+kind: ReplicaSet
+metadata:
+  name: vgg16v2job-pserver
+spec:
+  replicas: 10
+  template:
+    metadata:
+      labels:
+        paddle-job-pserver: vgg16v2job
+    spec:
+      hostNetwork: true
+      imagePullSecrets:
+      - name: job-registry-secret
+      containers:
+      - name: pserver
+        image: "registry.baidu.com/paddlepaddle/fluid_benchmark:vgg16"
+        imagePullPolicy: Always
+        ports:
+        - name: jobport-30236
+          containerPort: 30236
+        env:
+        - name: PADDLE_JOB_NAME
+          value: vgg16v2job
+        - name: TRAINERS
+          value: "20"
+        - name: PSERVERS
+          value: "10"
+        - name: TOPOLOGY
+          value: ""
+        - name: ENTRY
+          value: "python train.py"
+        - name: TRAINER_PACKAGE
+          value: "/workspace"
+        - name: PADDLE_INIT_PORT
+          value: "30236"
+        - name: PADDLE_INIT_NICS
+          value: "xgbe0"
+        - name: PADDLE_INIT_TRAINER_COUNT
+          value: "1"
+        - name: PADDLE_INIT_PORTS_NUM
+          value: "1"
+        - name: PADDLE_INIT_PORTS_NUM_FOR_SPARSE
+          value: "1"
+        - name: PADDLE_INIT_NUM_GRADIENT_SERVERS
+          value: "20"
+        - name: PADDLE_INIT_NUM_PASSES
+          value: "1"
+        - name: PADDLE_INIT_USE_GPU
+          value: "0"
+        - name: LD_LIBRARY_PATH
+          value: "/usr/local/lib:/usr/local/nvidia/lib64"
+        - name: NAMESPACE
+          valueFrom:
+            fieldRef:
+              fieldPath: "metadata.namespace"
+        command: ["paddle_k8s", "start_pserver"]
+        resources:
+          requests:
+            memory: 10Gi
+            cpu: 4
+          limits:
+            memory: 10Gi
+            cpu: 4

benchmark/cluster/vgg16/v2_trainer.yaml

+apiVersion: batch/v1
+kind: Job
+metadata:
+  name: vgg16v2job-trainer
+spec:
+  parallelism: 20
+  completions: 20
+  template:
+    metadata:
+      labels:
+        paddle-job: vgg16v2job
+    spec:
+      imagePullSecrets:
+        - name: job-registry-secret
+      hostNetwork: true
+      containers:
+      - name: trainer
+        image: "registry.baidu.com/paddlepaddle/fluid_benchmark:vgg16"
+        imagePullPolicy: Always
+        command: ["paddle_k8s", "start_trainer", "v2"]
+        env:
+        - name: PADDLE_JOB_NAME
+          value: vgg16v2job
+        - name: BATCH_SIZE
+          value: "256"
+        - name: TRAINERS
+          value: "20"
+        - name: PSERVERS
+          value: "10"
+        - name: TOPOLOGY
+          value: ""
+        - name: ENTRY
+          value: "cd /workspace && MKL_NUM_THREADS=1 python /workspace/vgg16_v2.py"
+        - name: TRAINER_PACKAGE
+          value: "/workspace"
+        - name: PADDLE_INIT_PORT
+          value: "30236"
+        - name: PADDLE_INIT_NICS
+          value: "xgbe0"
+        - name: PADDLE_INIT_TRAINER_COUNT
+          value: "1"
+        - name: PADDLE_INIT_PORTS_NUM
+          value: "1"
+        - name: PADDLE_INIT_PORTS_NUM_FOR_SPARSE
+          value: "1"
+        - name: PADDLE_INIT_NUM_GRADIENT_SERVERS
+          value: "20"
+        - name: PADDLE_INIT_NUM_PASSES
+          value: "2"
+        - name: PADDLE_INIT_USE_GPU
+          value: "0"
+        - name: LD_LIBRARY_PATH
+          value: "/usr/local/lib:/usr/local/nvidia/lib64"
+        - name: NAMESPACE
+          valueFrom:
+            fieldRef:
+              fieldPath: "metadata.namespace"
+        resources:
+          requests:
+            memory: 40Gi
+            cpu: 2
+          limits:
+            memory: 40Gi
+            cpu: 2
+      restartPolicy: Never

benchmark/cluster/vgg16/vgg16_fluid.py

+#   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.
+"""VGG16 benchmark in Fluid"""
+from __future__ import print_function
+
+import sys
+import time
+import numpy as np
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+import paddle.v2.fluid.core as core
+import paddle.v2.fluid.profiler as profiler
+import argparse
+import functools
+import os
+
+
+def str2bool(v):
+    if v.lower() in ('yes', 'true', 't', 'y', '1'):
+        return True
+    elif v.lower() in ('no', 'false', 'f', 'n', '0'):
+        return False
+    else:
+        raise argparse.ArgumentTypeError('Boolean value expected.')
+
+
+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('--device_id', type=int, default=0, help="The device id.")
+parser.add_argument(
+    '--data_format',
+    type=str,
+    default='NCHW',
+    choices=['NCHW', 'NHWC'],
+    help='The data order, now only support NCHW.')
+parser.add_argument(
+    '--data_set',
+    type=str,
+    default='cifar10',
+    choices=['cifar10', 'flowers'],
+    help='Optional dataset for benchmark.')
+parser.add_argument(
+    '--local',
+    type=str2bool,
+    default=True,
+    help='Whether to run as local mode.')
+args = parser.parse_args()
+
+
+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
+
+
+def main():
+    if args.data_set == "cifar10":
+        classdim = 10
+        if args.data_format == 'NCHW':
+            data_shape = [3, 32, 32]
+        else:
+            data_shape = [32, 32, 3]
+    else:
+        classdim = 102
+        if args.data_format == 'NCHW':
+            data_shape = [3, 224, 224]
+        else:
+            data_shape = [224, 224, 3]
+
+    # Input data
+    images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
+    label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+
+    # Train program
+    net = vgg16_bn_drop(images)
+    predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
+    cost = fluid.layers.cross_entropy(input=predict, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+
+    # Evaluator
+    accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
+
+    # inference program
+    inference_program = fluid.default_main_program().clone()
+    with fluid.program_guard(inference_program):
+        test_target = accuracy.metrics + accuracy.states
+        inference_program = fluid.io.get_inference_program(test_target)
+
+    # Optimization
+    optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate)
+    optimize_ops, params_grads = optimizer.minimize(avg_cost)
+
+    # Initialize executor
+    place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(
+        args.device_id)
+    exe = fluid.Executor(place)
+
+    # test
+    def test(exe):
+        accuracy.reset(exe)
+        for batch_id, data in enumerate(test_reader()):
+            img_data = np.array(map(lambda x: x[0].reshape(data_shape),
+                                    data)).astype("float32")
+            y_data = np.array(map(lambda x: x[1], data)).astype("int64")
+            y_data = y_data.reshape([-1, 1])
+
+            exe.run(inference_program,
+                    feed={"pixel": img_data,
+                          "label": y_data})
+
+        return accuracy.eval(exe)
+
+    def train_loop(exe, trainer_prog):
+        iters = 0
+        ts = time.time()
+        for pass_id in range(args.num_passes):
+            # train
+            start_time = time.time()
+            num_samples = 0
+            accuracy.reset(exe)
+            with profiler.profiler("CPU", 'total') as prof:
+                for batch_id, data in enumerate(train_reader()):
+                    ts = time.time()
+                    img_data = np.array(
+                        map(lambda x: x[0].reshape(data_shape), data)).astype(
+                            "float32")
+                    y_data = np.array(map(lambda x: x[1], data)).astype("int64")
+                    y_data = y_data.reshape([-1, 1])
+
+                    loss, acc = exe.run(
+                        trainer_prog,
+                        feed={"pixel": img_data,
+                              "label": y_data},
+                        fetch_list=[avg_cost] + accuracy.metrics)
+                    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)
+                    )  # The accuracy is the accumulation of batches, but not the current batch.
+
+            pass_elapsed = time.time() - start_time
+            pass_train_acc = accuracy.eval(exe)
+            pass_test_acc = test(exe)
+            print(
+                "Pass = %d, Training performance = %f imgs/s, Train accuracy = %f, Test accuracy = %f\n"
+                % (pass_id, num_samples / pass_elapsed, pass_train_acc,
+                   pass_test_acc))
+
+    if args.local:
+        # Parameter initialization
+        exe.run(fluid.default_startup_program())
+
+        # 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.dataset.cifar.test10()
+            if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
+            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,
+            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)
+            print("starting server side startup")
+            exe.run(pserver_startup)
+            print("starting parameter server...")
+            exe.run(pserver_prog)
+        elif training_role == "TRAINER":
+            # Parameter initialization
+            exe.run(fluid.default_startup_program())
+
+            # 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.dataset.cifar.test10() if args.data_set == 'cifar10' else
+                paddle.dataset.flowers.test(),
+                batch_size=args.batch_size)
+
+            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())
+            train_loop(exe, trainer_prog)
+        else:
+            print("environment var TRAINER_ROLE should be TRAINER os PSERVER")
+
+
+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()
+    main()

benchmark/cluster/vgg16/vgg16_v2.py

+#  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.
+
+import gzip
+
+import paddle.v2.dataset.cifar as cifar
+import paddle.v2 as paddle
+import time
+import os
+
+DATA_DIM = 3 * 32 * 32
+CLASS_DIM = 10
+BATCH_SIZE = os.getenv("BATCH_SIZE")
+if BATCH_SIZE:
+    BATCH_SIZE = int(BATCH_SIZE)
+else:
+    BATCH_SIZE = 128
+print "batch_size", BATCH_SIZE
+NODE_COUNT = int(os.getenv("TRAINERS"))
+ts = 0
+
+
+def vgg(input, nums, class_dim):
+    def conv_block(input, num_filter, groups, num_channels=None):
+        return paddle.networks.img_conv_group(
+            input=input,
+            num_channels=num_channels,
+            pool_size=2,
+            pool_stride=2,
+            conv_num_filter=[num_filter] * groups,
+            conv_filter_size=3,
+            conv_act=paddle.activation.Relu(),
+            pool_type=paddle.pooling.Max())
+
+    assert len(nums) == 5
+    # the channel of input feature is 3
+    conv1 = conv_block(input, 64, nums[0], 3)
+    conv2 = conv_block(conv1, 128, nums[1])
+    conv3 = conv_block(conv2, 256, nums[2])
+    conv4 = conv_block(conv3, 512, nums[3])
+    conv5 = conv_block(conv4, 512, nums[4])
+
+    fc_dim = 512
+    fc1 = paddle.layer.fc(input=conv5,
+                          size=fc_dim,
+                          act=paddle.activation.Relu(),
+                          layer_attr=paddle.attr.Extra(drop_rate=0.5))
+    fc2 = paddle.layer.fc(input=fc1,
+                          size=fc_dim,
+                          act=paddle.activation.Relu(),
+                          layer_attr=paddle.attr.Extra(drop_rate=0.5))
+    out = paddle.layer.fc(input=fc2,
+                          size=class_dim,
+                          act=paddle.activation.Softmax())
+    return out
+
+
+def vgg13(input, class_dim):
+    nums = [2, 2, 2, 2, 2]
+    return vgg(input, nums, class_dim)
+
+
+def vgg16(input, class_dim):
+    nums = [2, 2, 3, 3, 3]
+    return vgg(input, nums, class_dim)
+
+
+def vgg19(input, class_dim):
+    nums = [2, 2, 4, 4, 4]
+    return vgg(input, nums, class_dim)
+
+
+def main():
+    global ts
+    paddle.init(use_gpu=False)
+    image = paddle.layer.data(
+        name="image", type=paddle.data_type.dense_vector(DATA_DIM))
+    lbl = paddle.layer.data(
+        name="label", type=paddle.data_type.integer_value(CLASS_DIM))
+
+    extra_layers = None
+    # NOTE: for v2 distributed training need averaging updates.
+    learning_rate = 1e-3 / NODE_COUNT
+    out = vgg16(image, class_dim=CLASS_DIM)
+    cost = paddle.layer.classification_cost(input=out, label=lbl)
+
+    # Create parameters
+    parameters = paddle.parameters.create(cost)
+
+    # Create optimizer
+    optimizer = paddle.optimizer.Momentum(
+        momentum=0.9,
+        regularization=paddle.optimizer.L2Regularization(rate=0.0005 *
+                                                         BATCH_SIZE),
+        learning_rate=learning_rate / BATCH_SIZE,
+        learning_rate_decay_a=0.1,
+        learning_rate_decay_b=128000 * 35,
+        learning_rate_schedule="discexp", )
+
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            cifar.train10(),
+            # To use other data, replace the above line with:
+            # reader.train_reader('train.list'),
+            buf_size=1000),
+        batch_size=BATCH_SIZE)
+    test_reader = paddle.batch(
+        cifar.test10(),
+        # To use other data, replace the above line with:
+        # reader.test_reader('val.list'),
+        batch_size=BATCH_SIZE)
+
+    # Create trainer
+    trainer = paddle.trainer.SGD(cost=cost,
+                                 parameters=parameters,
+                                 update_equation=optimizer,
+                                 extra_layers=extra_layers,
+                                 is_local=False)
+
+    # End batch and end pass event handler
+    def event_handler(event):
+        global ts, ts_pass
+        if isinstance(event, paddle.event.BeginPass):
+            ts_pass = time.time()
+        if isinstance(event, paddle.event.BeginIteration):
+            ts = time.time()
+        if isinstance(event, paddle.event.EndIteration):
+            if event.batch_id % 1 == 0:
+                print "\nPass %d, Batch %d, Cost %f, %s, spent: %f" % (
+                    event.pass_id, event.batch_id, event.cost, event.metrics,
+                    time.time() - ts)
+        if isinstance(event, paddle.event.EndPass):
+            print "Pass %d end, spent: %f" % (event.pass_id,
+                                              time.time() - ts_pass)
+            result = trainer.test(reader=test_reader)
+            print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics)
+
+    trainer.train(
+        reader=train_reader, num_passes=200, event_handler=event_handler)
+
+
+if __name__ == '__main__':
+    main()

cmake/external/boost.cmake

@@ -21,6 +21,7 @@ set(BOOST_URL           "http://sourceforge.net/projects/boost/files/boost/${BOO
 set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost)
 set(BOOST_DOWNLOAD_DIR  "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}")
 set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)
+set_directory_properties(PROPERTIES CLEAN_NO_CUSTOM 1)
 
 include_directories(${BOOST_INCLUDE_DIR})
 

cmake/external/eigen.cmake

@@ -28,9 +28,3 @@ endif()
 add_dependencies(eigen3 extern_eigen3)
 
 LIST(APPEND external_project_dependencies eigen3)
-
-IF(NOT WITH_C_API AND WITH_FLUID)
-    INSTALL(FILES ${EIGEN_INCLUDE_DIR}/Eigen/Core DESTINATION third_party/eigen3/Eigen)
-    INSTALL(DIRECTORY ${EIGEN_INCLUDE_DIR}/Eigen/src DESTINATION third_party/eigen3/Eigen)
-    INSTALL(DIRECTORY ${EIGEN_INCLUDE_DIR}/unsupported/Eigen DESTINATION third_party/eigen3/unsupported)
-ENDIF()

cmake/external/gflags.cmake

@@ -52,7 +52,7 @@ ADD_DEPENDENCIES(gflags extern_gflags)
 
 LIST(APPEND external_project_dependencies gflags)
 
-IF(WITH_C_API OR WITH_FLUID)
+IF(WITH_C_API)
   INSTALL(DIRECTORY ${GFLAGS_INCLUDE_DIR} DESTINATION third_party/gflags)
   IF(ANDROID)
     INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib/${ANDROID_ABI})

cmake/external/glog.cmake

@@ -68,7 +68,7 @@ LINK_LIBRARIES(glog gflags)
 
 LIST(APPEND external_project_dependencies glog)
 
-IF(WITH_C_API OR WITH_FLUID)
+IF(WITH_C_API)
   INSTALL(DIRECTORY ${GLOG_INCLUDE_DIR} DESTINATION third_party/glog)
   IF(ANDROID)
     INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib/${ANDROID_ABI})

cmake/external/protobuf.cmake

@@ -250,7 +250,7 @@ IF(NOT PROTOBUF_FOUND)
     SET(PROTOBUF_PROTOC_LIBRARY ${extern_protobuf_PROTOC_LIBRARY}
         CACHE FILEPATH "protoc library." FORCE)
 
-    IF(WITH_C_API OR WITH_FLUID)
+    IF(WITH_C_API)
         INSTALL(DIRECTORY ${PROTOBUF_INCLUDE_DIR} DESTINATION third_party/protobuf)
         IF(ANDROID)
             INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})

cmake/external/warpctc.cmake

@@ -52,6 +52,7 @@ ExternalProject_Add(
                     -DWITH_TORCH=OFF
                     -DCMAKE_DISABLE_FIND_PACKAGE_Torch=ON
                     -DBUILD_SHARED=ON
+                    -DBUILD_TESTS=OFF
                     -DCMAKE_POSITION_INDEPENDENT_CODE=ON
                     -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
                     ${EXTERNAL_OPTIONAL_ARGS}

cmake/inference_lib.cmake

+# make package for paddle fluid shared and static library
+function(copy TARGET)
+    set(options "")
+    set(oneValueArgs "")
+    set(multiValueArgs SRCS DSTS DEPS)
+    cmake_parse_arguments(copy_lib "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+
+    list(LENGTH copy_lib_SRCS copy_lib_SRCS_len)
+    list(LENGTH copy_lib_DSTS copy_lib_DSTS_len)
+    if(NOT ${copy_lib_SRCS_len} EQUAL ${copy_lib_DSTS_len})
+        message(FATAL_ERROR "${TARGET} source numbers are not equal to destination numbers")
+    endif()
+    math(EXPR len "${copy_lib_SRCS_len} - 1")
+    
+    add_custom_target(${TARGET} DEPENDS ${copy_lib_DEPS})
+    foreach(index RANGE ${len})
+        list(GET copy_lib_SRCS ${index} src)
+        list(GET copy_lib_DSTS ${index} dst)
+        add_custom_command(TARGET ${TARGET} PRE_BUILD COMMAND mkdir -p "${dst}")
+        if(IS_DIRECTORY ${src})
+            add_custom_command(TARGET ${TARGET} PRE_BUILD COMMAND cp -r "${src}" "${dst}")
+        else()
+            add_custom_command(TARGET ${TARGET} PRE_BUILD COMMAND cp "${src}" "${dst}")
+        endif()
+    endforeach()
+endfunction()
+
+# third party
+set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/eigen3")
+copy(eigen3_lib
+  SRCS ${EIGEN_INCLUDE_DIR}/Eigen/Core ${EIGEN_INCLUDE_DIR}/Eigen/src ${EIGEN_INCLUDE_DIR}/unsupported/Eigen
+  DSTS ${dst_dir}/Eigen ${dst_dir}/Eigen ${dst_dir}/unsupported
+)
+
+set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/gflags")
+copy(gflags_lib
+  SRCS ${GFLAGS_INCLUDE_DIR} ${GFLAGS_LIBRARIES}
+  DSTS ${dst_dir} ${dst_dir}/lib
+)
+
+set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/glog")
+copy(glog_lib
+  SRCS ${GLOG_INCLUDE_DIR} ${GLOG_LIBRARIES}
+  DSTS ${dst_dir} ${dst_dir}/lib
+)
+
+IF(NOT PROTOBUF_FOUND)
+    set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/protobuf")
+    copy(protobuf_lib
+      SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LITE_LIBRARY}
+      DSTS ${dst_dir} ${dst_dir}/lib
+    )
+ENDIF(NOT PROTOBUF_FOUND)
+
+# paddle fluid module
+set(src_dir "${PADDLE_SOURCE_DIR}/paddle")
+set(dst_dir "${CMAKE_INSTALL_PREFIX}/paddle")
+set(module "framework")
+copy(framework_lib DEPS framework_py_proto 
+  SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/framework/framework.pb.h
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module}
+)
+
+set(module "memory")
+copy(memory_lib
+  SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/detail/*.h
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module}/detail
+)
+
+set(module "inference")
+copy(inference_lib DEPENDS paddle_fluid_shared
+  SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/inference/libpaddle_fluid.so
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module}
+)
+
+set(module "platform")
+copy(platform_lib
+  SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/dynload/*.h ${src_dir}/${module}/details/*.h
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module}/dynload ${dst_dir}/${module}/details
+)
+
+set(module "string")
+copy(string_lib
+  SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/tinyformat/*.h
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat
+)
+
+add_custom_target(inference_lib_dist DEPENDS 
+  inference_lib framework_lib memory_lib platform_lib string_lib
+  gflags_lib glog_lib protobuf_lib eigen3_lib)

doc/CMakeLists.txt

@@ -47,3 +47,5 @@ sphinx_add_target(paddle_docs_cn
                   ${SPHINX_CACHE_DIR_CN}
                   ${CMAKE_CURRENT_SOURCE_DIR}
                   ${SPHINX_HTML_DIR_CN})
+
+add_subdirectory(api)

doc/api/CMakeLists.txt

+# configured documentation tools and intermediate build results
+set(BINARY_BUILD_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_build")
+
+# Sphinx cache with pickled ReST documents
+set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees")
+
+# HTML output director
+set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html")
+
+configure_file(
+    "${CMAKE_CURRENT_SOURCE_DIR}/../templates/conf.py.en.in"
+    "${BINARY_BUILD_DIR_EN}/conf.py"
+    @ONLY)
+
+sphinx_add_target(paddle_api_docs
+                  html
+                  ${BINARY_BUILD_DIR_EN}
+                  ${SPHINX_CACHE_DIR_EN}
+                  ${CMAKE_CURRENT_SOURCE_DIR}
+                  ${SPHINX_HTML_DIR_EN})

doc/api/v2/config/layer.rst

@@ -87,6 +87,11 @@ roi_pool
 ..  autoclass:: paddle.v2.layer.roi_pool
     :noindex:
 
+pad
+----
+..  autoclass:: paddle.v2.layer.pad
+    :noindex:
+
 Norm Layer
 ==========
 
@@ -133,6 +138,11 @@ grumemory
 ..  autoclass:: paddle.v2.layer.grumemory
     :noindex:
 
+gated_unit
+-----------
+..  autoclass:: paddle.v2.layer.gated_unit
+    :noindex:
+    
 Recurrent Layer Group
 =====================
 
@@ -340,6 +350,11 @@ bilinear_interp
 ..  autoclass:: paddle.v2.layer.bilinear_interp
     :noindex:
 
+dropout
+--------
+..  autoclass:: paddle.v2.layer.dropout
+    :noindex:
+    
 dot_prod
 ---------
 .. autoclass:: paddle.v2.layer.dot_prod
@@ -402,6 +417,11 @@ scale_shift
 ..  autoclass:: paddle.v2.layer.scale_shift
     :noindex:
 
+factorization_machine
+---------------------
+..  autoclass:: paddle.v2.layer.factorization_machine
+    :noindex:
+
 Sampling Layers
 ===============
 
@@ -420,22 +440,6 @@ multiplex
 ..  autoclass:: paddle.v2.layer.multiplex
     :noindex:
 
-Factorization Machine Layer
-============================
-
-factorization_machine
----------------------
-..  autoclass:: paddle.v2.layer.factorization_machine
-    :noindex:
-
-Slicing and Joining Layers
-==========================
-
-pad
-----
-..  autoclass:: paddle.v2.layer.pad
-    :noindex:
-
 ..  _api_v2.layer_costs:
 
 Cost Layers
@@ -526,6 +530,11 @@ multibox_loss
 ..  autoclass:: paddle.v2.layer.multibox_loss
     :noindex:
 
+detection_output
+----------------
+..  autoclass:: paddle.v2.layer.detection_output
+    :noindex:
+    
 Check Layer
 ============
 
@@ -534,31 +543,10 @@ eos
 ..  autoclass:: paddle.v2.layer.eos
     :noindex:
 
-Miscs
-=====
-
-dropout
---------
-..  autoclass:: paddle.v2.layer.dropout
-    :noindex:
-
-Activation with learnable parameter
-===================================
+Activation
+==========
 
 prelu
 --------
 ..  autoclass:: paddle.v2.layer.prelu
     :noindex:
-
-gated_unit
------------
-..  autoclass:: paddle.v2.layer.gated_unit
-    :noindex:
-
-Detection output Layer
-======================
-
-detection_output
-----------------
-..  autoclass:: paddle.v2.layer.detection_output
-    :noindex:

doc/api/v2/data/dataset.rst

@@ -73,3 +73,10 @@ wmt14
 ..  automodule:: paddle.v2.dataset.wmt14
     :members:
     :noindex:
+
+wmt16
++++++
+
+..  automodule:: paddle.v2.dataset.wmt16
+    :members:
+    :noindex:

doc/getstarted/build_and_install/index_cn.rst → doc/build_and_install/index_cn.rst

@@ -13,7 +13,7 @@ PaddlePaddle提供pip和Docker的安装方式：
 
    pip_install_cn.rst
    docker_install_cn.rst
-   ../../howto/dev/build_cn.md
+   build_cn.md
 
 编译流程
 ++++++++

doc/getstarted/build_and_install/index_en.rst → doc/build_and_install/index_en.rst

@@ -13,7 +13,7 @@ You can choose either pip or Docker to complete your install:
 
    pip_install_en.rst
    docker_install_en.rst
-   ../../howto/dev/build_en.md
+   build_en.md
 
 
 Build from Source

doc/design/auto_gradient_check.md

-## Auto Gradient Checker Design
+## Auto Gradient Check Design
 
-## Backgraound：
-- Generally, it is easy to check whether the forward computation of an Operator is correct or not. However, backpropagation is a notoriously difficult algorithm to debug and get right:
-  1. you should get the right backpropagation formula according to the forward computation.
-  2. you should implement it right in CPP.
-  3. it's difficult to prepare test data.
+## Background：
+- Generally, it is easy to check whether the forward computation of an Operator is correct or not. However, backpropagation is a notoriously difficult algorithm to debug and get right because of the following challenges:
+  1. The formula for backpropagation formula should be correct according to the forward computation.
+  2. The Implementation of the above shoule be correct in CPP.
+  3. It is difficult to prepare an unbiased test data.
 
-- Auto gradient checking gets a numerical gradient by forward Operator and use it as a reference of the backward Operator's result. It has several advantages:
-  1. numerical gradient checker only need forward operator.
-  2. user only need to prepare the input data for forward Operator.
+- Auto gradient checking gets a numerical gradient using forward Operator and uses it as a reference for the backward Operator's result. It has several advantages:
+  1. Numerical gradient checker only needs the forward operator.
+  2. The user only needs to prepare the input data for forward Operator and not worry about the backward Operator.
 
 ## Mathematical Theory
-The following two document from Stanford has a detailed explanation of how to get numerical gradient and why it's useful.
+The following documents from Stanford have a detailed explanation of how to compute the numerical gradient and why it is useful.
 
 - [Gradient checking and advanced optimization(en)](http://deeplearning.stanford.edu/wiki/index.php/Gradient_checking_and_advanced_optimization)
 - [Gradient checking and advanced optimization(cn)](http://ufldl.stanford.edu/wiki/index.php/%E6%A2%AF%E5%BA%A6%E6%A3%80%E9%AA%8C%E4%B8%8E%E9%AB%98%E7%BA%A7%E4%BC%98%E5%8C%96)
 
 
-## Numeric Gradient Implementation
+## Numerical Gradient Implementation
 ### Python Interface
 ```python
 def get_numerical_gradient(op,
@@ -27,73 +27,76 @@ def get_numerical_gradient(op,
                          delta=0.005,
                          local_scope=None):
     """
-    Get Numeric Gradient for an operator's input.
+    Get Numerical Gradient for the input of an operator.
 
-    :param op: C++ operator instance, could be an network
+    :param op: C++ operator instance, could be an network.
     :param input_values: The input variables. Should be an dictionary, whose key is
-    variable name, and value is numpy array.
+    variable name, and value is a numpy array.
     :param output_name: The final output variable name.
-    :param input_to_check: The input variable with respect to which to compute the gradient.
-    :param delta: The perturbation value for numeric gradient method. The
-    smaller delta is, the more accurate result will get. But if that delta is
-     too small, it will suffer from numerical stability problem.
+    :param input_to_check: The input variable with respect to which the gradient has to be computed.
+    :param delta: The perturbation value for numerical gradient method. The
+    smaller the delta, the more accurate the result. But if the delta is too
+    small, it will suffer from the numerical stability problem.
     :param local_scope: The local scope used for get_numeric_gradient.
     :return: The gradient array in numpy format.
     """
 ```
 
-### Explaination:
+### Explanation:
 
-- Why need `output_name`
-  - An Operator may have multiple Output, one can get independent gradient from each Output. So caller should specify the name of the output variable.
+- Why do we need an `output_name`
+  - An Operator may have multiple Outputs, one can compute an independent gradient from each Output. So the caller should specify the name of the output variable.
 
-- Why need `input_to_check`
-  - One operator may have multiple inputs. Gradient Op can calculate the gradient of these inputs at the same time. But Numeric Gradient needs to calculate them one by one. So `get_numeric_gradient` is designed to calculate the gradient for one input. If you need to compute multiple inputs, you can call `get_numeric_gradient` multiple times.
+- Why do we need `input_to_check`
+  - One operator can have multiple inputs. Gradient Op can calculate the gradient of these inputs at the same time. But Numerical Gradient needs to calculate them one by one. So `get_numeric_gradient` is designed to calculate the gradient for one input. If you need to compute multiple inputs, you can call `get_numeric_gradient` multiple times each with a different input.
 
 
 ### Core Algorithm Implementation
 
 
 ```python
-    # we only compute gradient of one element a time.
+    # we only compute the gradient of one element a time.
     # we use a for loop to compute the gradient of each element.
     for i in xrange(tensor_size):
-        # get one input element by its index i.
-        origin = tensor_to_check.get_float_element(i)
+        # get one input element using the index i.
+        original = tensor_to_check.get_float_element(i)
 
-        # add delta to it, run op and then get the new value of the result tensor.
-        x_pos = origin + delta
+        # add delta to it, run the forward op and then
+        # get the new value of the result tensor.
+        x_pos = original + delta
         tensor_to_check.set_float_element(i, x_pos)
         y_pos = get_output()
 
-        # plus delta to this element, run op and get the new value of the result tensor.
-        x_neg = origin - delta
+        # Subtract delta from this element, run the op again
+        # and get the new value of the result tensor.
+        x_neg = original - delta
         tensor_to_check.set_float_element(i, x_neg)
         y_neg = get_output()
 
         # restore old value
-        tensor_to_check.set_float_element(i, origin)
+        tensor_to_check.set_float_element(i, original)
 
-        # compute the gradient of this element and store it into a numpy array.
+        # compute the gradient of this element and store
+        # it into a numpy array.
         gradient_flat[i] = (y_pos - y_neg) / delta / 2
 
     # reshape the gradient result to the shape of the source tensor.
     return gradient_flat.reshape(tensor_to_check.get_dims())
 ```
 
-## Auto Graident Checker Framework
+## Auto Gradient Check Framework
 
 Each Operator Kernel has three kinds of Gradient:
 
 1. Numerical gradient
 2. CPU kernel gradient
-3. GPU kernel gradient (if supported)
+3. GPU kernel gradient (if supported by the device)
 
-The numerical gradient only relies on forward Operator. So we use the numerical gradient as the reference value. And the gradient checking is performed in the following three steps:
+The numerical gradient only relies on the forward Operator, so we use the numerical gradient as the reference value. The gradient checking is performed in the following three steps:
 
-1. calculate the numerical gradient
-2. calculate CPU kernel gradient with the backward Operator and compare it with the numerical gradient
-3. calculate GPU kernel gradient with the backward Operator and compare it with the numeric gradient (if supported)
+1. Calculate the numerical gradient
+2. Calculate CPU kernel gradient with the backward Operator and compare it with the numerical gradient.
+3. Calculate GPU kernel gradient with the backward Operator and compare it with the numeric gradient. (if supported)
 
 #### Python Interface
 
@@ -109,26 +112,27 @@ The numerical gradient only relies on forward Operator. So we use the numerical
         """
         :param forward_op: used to create backward_op
         :param input_vars: numpy value of input variable. The following
-            computation will use these variables.
-        :param inputs_to_check: the input variable with respect to which to compute the gradient.
+          computation will use these variables.
+        :param inputs_to_check: the input variable with respect to which the
+          gradient will be computed.
         :param output_name: The final output variable name.
         :param max_relative_error: The relative tolerance parameter.
-        :param no_grad_set: used when create backward ops
+        :param no_grad_set: used to create backward ops
         :param only_cpu: only compute and check gradient on cpu kernel.
         :return:
         """
 ```
 
-### How to check if two numpy array is close enough?
-if `abs_numerical_grad` is nearly zero, then use abs error for numerical_grad
+### How to check if two numpy arrays are close enough?
+if `abs_numerical_grad` is nearly zero, then use absolute error for numerical_grad.
 
 ```python
 numerical_grad = ...
 operator_grad = numpy.array(scope.find_var(grad_var_name(name)).get_tensor())
 
 abs_numerical_grad = numpy.abs(numerical_grad)
-# if abs_numerical_grad is nearly zero, then use abs error for numeric_grad, not relative
-# error.
+# if abs_numerical_grad is nearly zero, then use abs error for
+# numeric_grad, instead of relative error.
 abs_numerical_grad[abs_numerical_grad < 1e-3] = 1
 
 diff_mat = numpy.abs(abs_numerical_grad - operator_grad) / abs_numerical_grad
@@ -137,10 +141,10 @@ max_diff = numpy.max(diff_mat)
 
 
 #### Notes：
-The Input data for auto gradient checker should be reasonable to avoid numerical  stability problem.
+The Input data for auto gradient checker should be reasonable to avoid numerical stability problem.
 
 
-#### Refs:
+#### References:
 
 - [Gradient checking and advanced optimization(en)](http://deeplearning.stanford.edu/wiki/index.php/Gradient_checking_and_advanced_optimization)
 - [Gradient checking and advanced optimization(cn)](http://ufldl.stanford.edu/wiki/index.php/%E6%A2%AF%E5%BA%A6%E6%A3%80%E9%AA%8C%E4%B8%8E%E9%AB%98%E7%BA%A7%E4%BC%98%E5%8C%96)

doc/design/csp.md

@@ -42,7 +42,7 @@ The type *channel* is conceptually the blocking queue.  In Go, its implemented i
 
 The `select` operation has been in OS kernels long before Go language.  All Unix kernels implement system calls *poll* and *select*.  They monitor multiple file descriptors to see if I/O is possible on any of them.  This takes O(N) time.  Since Linux 2.6, a new system call, *epoll*, can do the same in O(1) time.  In BSD systems, there is a similar system call *kqueue*.  Go's Linux implementation uses epoll.
 
-It might be a good idea to implement Fluid's select using epoll too.  In this design doc, we start from the O(N) way, so we could focus on Python binding and the syntax.
+It might be a good idea to implement Fluid's select using epoll too.  In this design doc, we start from the O(N) way so that we could focus on Python binding and the syntax.
 
 ### Type Channel
 
@@ -71,14 +71,14 @@ ch1 := make(chan int, 100)  // a channel that can buffer 100 ints.
 In Fluid, we should be able to do the same:
 
 ```python
-ch  = fluid.make_chan(dtype=INT)
-ch1 = fluid.make_chan(dtype=INT, 100)
+ch  = fluid.make_channel(dtype=INT)
+ch1 = fluid.make_channel(dtype=INT, 100)
 ```
 
 In addition to that, we want channels that can hold more complex element types, e.g., Tensors of float16:
 
 ```python
-ch = fluid.make_chan(dtype=Tensor, etype=float16)
+ch = fluid.make_channel(dtype=Tensor, etype=float16)
 ```
 
 or Tensors of Tensors of float16 etc.
@@ -87,8 +87,136 @@ The point here is that we need a consistent way to compose types, like in C++ we
 
 ### Send and Recv
 
+Go's CSP implementation depends on data type *channel*. There are two types of channels:
+
+1. The unblocked channel, or buffered channel, is a blocking queue with a non-zero sized buffer. The sending to buffered channel blocks if the buffer is full, and the receive operation blocks if the buffer is empty.
+1. blocked channel, or unbuffered channel, is a blocking queue with no buffer.  Both sending and receiving block with unbuffered channels.
+
+There are four types of actions with a channel:
+
+1. Create a channel
+
+   ```go
+   ch := make(chan int) // this is an unbuffered channel
+   ch := make(chan int, 100) // this is a buffered channel of 100 ints.
+   ```
+
+1. Send
+
+   ```go
+   ch <- 111
+   ```
+
+1. Recv
+
+   ```go
+   y, ok <- ch
+   ```
+
+1. Close
+
+   ```go
+   close(ch)
+   ```
+   
+   Please be aware that a closed channel is not a nil channel, which is `var ch chan int`.
+   
+There are some [axioms with channels](https://dave.cheney.net/2014/03/19/channel-axioms):
+
+1. A send to a nil channel blocks forever
+
+1. A receive from a nil channel blocks forever
+
+1. A send to a closed channel panics
+
+1. A receive from a closed channel returns the residual values and then zeros.
+
+In Fluid, we have [buffered channels](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/details/buffered_channel.h) and [unbuffered channels](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/details/unbuffered_channel.h)
+
+The following program illustrates the Python syntax for accessing Fluid buffers.
+
+```python
+import fluid
+
+buffer_size = 10
+ch = fluid.make_channel(dtype=INT, buffer_size)
+
+# Now write three elements to the channel
+with fluid.while(steps=buffer_size):
+  fluid.send(ch, step)
+
+fluid.close_channel(ch)
+
+with fluid.while(steps=buffer_size):
+  fluid.print(fluid.recv(ch))
+```
+
+The following example shows that to avoid the always-blocking behavior of unbuffered channels, we need to use Fluid's goroutines.
+
+```python
+import fluid
+
+ch = fluid.make_channel(dtype=INT)
+
+with fluid.go():
+  fluid.send(ch)
+
+y = fluid.recv(ch)
+
+fluid.close_channel(ch)
+```
+
 ### Select
 
+In Go, the `select` statement lets a goroutine wait on multiple communication operations. A `select` blocks until one of its cases can run, then it executes that case. It chooses one at random if multiple are ready.
+
+```go
+
+ch1  := make(chan int)       
+ch2  := make(chan int, 100)
+
+x := 0
+
+for {
+    select {
+    case ch1 <- x:
+      x := x + 1
+    case y <- ch2:
+      fmt.Println("Received on channel")
+    default:
+      fmt.Println("Default")
+    }
+  }
+
+```
+
+In Fluid, we should be able to do the same:
+
+```python
+ch1  = fluid.make_chan(dtype=INT)
+ch2 = fluid.make_chan(dtype=INT, 100)
+
+sel = fluid.select()
+
+with sel.case(ch1, 'w', X):
+    fluid.layers.increment(X)
+
+with sel.case(ch2, 'r', Y):
+    fluid.print("Received on Channel")
+
+with sel.default():
+    fluid.print("Default")
+
+```
+
+In the above code snippet, `X` and `Y` are variables. Now let us look at each of these statements one by one.
+
+- `sel.case(ch1, 'w', X)` : This specifies that we are writing to `ch1` and we want to write the integer in variable `X` to the channel. The character `w` is used here to make the syntax familiar to write syntax in Python I/O.
+
+- `sel.case(ch2, 'r', Y)` : This specifies that we would like to read the result from `ch2` into variable `Y`. The character `r` is used here to make the syntax familiar to read syntax in Python I/O.
+
+- `sel.default()` : This is equivalent to the default in Go `select`. If none of the channels are ready for read or write, then the fluid code in the default block will be executed.
+
 ## Example Programs
 
 ### 1. RPC between Trainers and Parameter Servers

doc/design/switch.md

+### Design Doc: Switch
+
+### Background
+
+Many programming languages provide `switch` as a generalization of `if-elif-else`.  We want to add it to Fluid.
+
+The following example shows the usage of `fluid.switch`.
+
+```python
+a = fluid.Var(10)
+b = fluid.Var(0)
+
+with switch() as switch:
+    with switch.case(fluid.less_equal(a, 10)):
+        fluid.print("Case 1")
+    with switch.case(fluid.larger(a, 0)):
+        fluid.print("Case 2")
+    with switch.default():
+        fluid.print("Case 3")
+```
+
+### The Semantics
+
+1. A `switch` control-flow checks cases one-by-one.
+1. The condition of each case is a boolean value, which is a scalar, and differs from the `fluid.if_else` control-flow, which condition could be a vector of boolean values.
+1. It runs the first matched case, or the default case if there is one.
+1. Once it matches a case, it runs the corresponding branch and only that branch.  It's like there is a C's `break` keyword at the end of each case.
+
+The above program should print and print only "Case 1".
+
+The implementation of the backward pass of the `switch` control-flow is easier than the backward of the `if_else`, because `switch` runs at most one branch, whereas `if-else` could run more than one branches.

doc/dev/contribute_to_paddle_en.md

+../../CONTRIBUTING.md
\ No newline at end of file

doc/dev/index_cn.rst

+开发标准
+========
+
+..  toctree::
+  :maxdepth: 1
+
+  contribute_to_paddle_cn.md
+  write_docs_cn.rst

doc/dev/index_en.rst

+Development
+------------
+
+..  toctree::
+  :maxdepth: 1
+
+  new_layer_en.rst
+  contribute_to_paddle_en.md
+  write_docs_en.rst

doc/howto/dev/write_docs_cn.rst → doc/dev/write_docs_cn.rst

-##################
-如何贡献/修改文档
-##################
+#############
+如何贡献文档
+#############
 
 PaddlePaddle的文档包括英文文档 ``doc`` 和中文文档 ``doc_cn`` 两个部分。文档都是通过 `cmake`_ 驱动 `sphinx`_ 编译生成，生成后的文档分别存储在编译目录的 ``doc`` 和 ``doc_cn`` 两个子目录下。
 也可以利用PaddlePaddle 工具来编译文档，这个情况下所有的文件会存在整理过的的文件目录 .ppo_workspace/content 下

doc/howto/dev/write_docs_en.rst → doc/dev/write_docs_en.rst

-##################
+########################
 Contribute Documentation
-##################
+########################
 
 PaddlePaddle supports English documentation ``doc`` and Chinese documentation ``doc_cn``.
 Both are compiled by `cmake`_ and `sphinx`_ , the compiled documentations will be stored under ``doc`` and ``doc_cn`` directories.

doc/getstarted/concepts/use_concepts_cn.rst

@@ -4,7 +4,7 @@
 
 PaddlePaddle是源于百度的一个深度学习平台。PaddlePaddle为深度学习研究人员提供了丰富的API，可以轻松地完成神经网络配置，模型训练等任务。
 这里将介绍PaddlePaddle的基本使用概念，并且展示了如何利用PaddlePaddle来解决一个经典的线性回归问题。
-在使用该文档之前，请参考 `安装文档 <../build_and_install/index_cn.html>`_ 完成PaddlePaddle的安装。
+在使用该文档之前，请参考 `安装文档 <../../build_and_install/index_cn.html>`_ 完成PaddlePaddle的安装。
 
 
 配置网络

doc/getstarted/index_cn.rst

 新手入门
 ============
 
-.. _quick_install:
-
-快速安装
-++++++++
-
-PaddlePaddle支持使用pip快速安装，目前支持CentOS 6以上, Ubuntu 14.04以及MacOS 10.12，并安装有Python2.7。
-执行下面的命令完成快速安装，版本为cpu_avx_openblas：
-
-  .. code-block:: bash
-
-     pip install paddlepaddle
-
-如果需要安装支持GPU的版本（cuda7.5_cudnn5_avx_openblas），需要执行：
-
-  .. code-block:: bash
-
-     pip install paddlepaddle-gpu
-
-更详细的安装和编译方法参考：
-
-..  toctree::
-  :maxdepth: 1
-
-  build_and_install/index_cn.rst
-
-.. _quick_start:
-
-快速开始
-++++++++
-
-创建一个 housing.py 并粘贴此Python代码：
-
-  .. code-block:: python
-
-     import paddle.v2 as paddle
-
-     # Initialize PaddlePaddle.
-     paddle.init(use_gpu=False, trainer_count=1)
-
-     # Configure the neural network.
-     x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13))
-     y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
-
-     # Infer using provided test data.
-     probs = paddle.infer(
-         output_layer=y_predict,
-         parameters=paddle.dataset.uci_housing.model(),
-         input=[item for item in paddle.dataset.uci_housing.test()()])
-
-     for i in xrange(len(probs)):
-         print 'Predicted price: ${:,.2f}'.format(probs[i][0] * 1000)
-
-执行 :code:`python housing.py` 瞧！ 它应该打印出预测住房数据的清单。
-
 ..  toctree::
   :maxdepth: 1
 
+  quickstart_cn.rst
   concepts/use_concepts_cn.rst

doc/getstarted/index_en.rst

 GET STARTED
 ============
 
-.. _quick_install:
-
-Quick Install
-----------------------
-
-You can use pip to install PaddlePaddle with a single command, supports
-CentOS 6 above, Ubuntu 14.04 above or MacOS 10.12, with Python 2.7 installed.
-Simply run the following command to install, the version is cpu_avx_openblas:
-
-  .. code-block:: bash
-
-     pip install paddlepaddle
-
-If you need to install GPU version (cuda7.5_cudnn5_avx_openblas), run:
-
-  .. code-block:: bash
-
-     pip install paddlepaddle-gpu
-
-For more details about installation and build:
-
 ..  toctree::
   :maxdepth: 1
 
-  build_and_install/index_en.rst
-
-
-.. _quick_start:
-
-Quick Start
-++++++++
-
-Create a new file called housing.py, and paste this Python
-code:
-
-
-  .. code-block:: python
-
-     import paddle.v2 as paddle
-
-     # Initialize PaddlePaddle.
-     paddle.init(use_gpu=False, trainer_count=1)
-
-     # Configure the neural network.
-     x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13))
-     y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
-
-     # Infer using provided test data.
-     probs = paddle.infer(
-         output_layer=y_predict,
-         parameters=paddle.dataset.uci_housing.model(),
-         input=[item for item in paddle.dataset.uci_housing.test()()])
-
-     for i in xrange(len(probs)):
-         print 'Predicted price: ${:,.2f}'.format(probs[i][0] * 1000)
-
-Run :code:`python housing.py` and voila! It should print out a list of predictions
-for the test housing data.
+  quickstart_en.rst

doc/getstarted/quickstart_cn.rst

+快速开始
+========
+
+PaddlePaddle支持使用pip快速安装，目前支持CentOS 6以上, Ubuntu 14.04以及MacOS 10.12，并安装有Python2.7。
+执行下面的命令完成快速安装，版本为cpu_avx_openblas：
+
+  .. code-block:: bash
+
+     pip install paddlepaddle
+
+如果需要安装支持GPU的版本（cuda7.5_cudnn5_avx_openblas），需要执行：
+
+  .. code-block:: bash
+
+     pip install paddlepaddle-gpu
+
+更详细的安装和编译方法参考：:ref:`install_steps` 。
+
+创建一个 housing.py 并粘贴此Python代码：
+
+  .. code-block:: python
+
+     import paddle.v2 as paddle
+
+     # Initialize PaddlePaddle.
+     paddle.init(use_gpu=False, trainer_count=1)
+
+     # Configure the neural network.
+     x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13))
+     y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
+
+     # Infer using provided test data.
+     probs = paddle.infer(
+         output_layer=y_predict,
+         parameters=paddle.dataset.uci_housing.model(),
+         input=[item for item in paddle.dataset.uci_housing.test()()])
+
+     for i in xrange(len(probs)):
+         print 'Predicted price: ${:,.2f}'.format(probs[i][0] * 1000)
+
+执行 :code:`python housing.py` 瞧！ 它应该打印出预测住房数据的清单。

doc/getstarted/quickstart_en.rst

+Quick Start
+============
+
+You can use pip to install PaddlePaddle with a single command, supports
+CentOS 6 above, Ubuntu 14.04 above or MacOS 10.12, with Python 2.7 installed.
+Simply run the following command to install, the version is cpu_avx_openblas:
+
+  .. code-block:: bash
+
+     pip install paddlepaddle
+
+If you need to install GPU version (cuda7.5_cudnn5_avx_openblas), run:
+
+  .. code-block:: bash
+
+     pip install paddlepaddle-gpu
+
+For more details about installation and build: :ref:`install_steps` .
+
+Create a new file called housing.py, and paste this Python
+code:
+
+
+  .. code-block:: python
+
+     import paddle.v2 as paddle
+
+     # Initialize PaddlePaddle.
+     paddle.init(use_gpu=False, trainer_count=1)
+
+     # Configure the neural network.
+     x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13))
+     y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
+
+     # Infer using provided test data.
+     probs = paddle.infer(
+         output_layer=y_predict,
+         parameters=paddle.dataset.uci_housing.model(),
+         input=[item for item in paddle.dataset.uci_housing.test()()])
+
+     for i in xrange(len(probs)):
+         print 'Predicted price: ${:,.2f}'.format(probs[i][0] * 1000)
+
+Run :code:`python housing.py` and voila! It should print out a list of predictions
+for the test housing data.

doc/howto/usage/capi/compile_paddle_lib_cn.md → doc/howto/capi/compile_paddle_lib_cn.md

-## 编译 PaddlePaddle 预测库
+## 安装与编译C-API预测库
 
 ### 概述
 

doc/howto/usage/capi/index_cn.rst → doc/howto/capi/index_cn.rst

-PaddlePaddle C-API
+C-API预测库
 ==================
 
 ..  toctree::

doc/howto/usage/capi/workflow_of_capi_cn.md → doc/howto/capi/workflow_of_capi_cn.md

-## C-API 使用流程
+## C-API使用流程
 
 这篇文档介绍 PaddlePaddle C-API 整体使用流程。
 

doc/howto/usage/cluster/cluster_train_cn.md → doc/howto/cluster/cmd_argument_cn.md

-# 分布式训练
-
-
-## 概述
-
-本文将介绍如何使用PaddlePaddle在不同的集群框架下完成分布式训练。分布式训练架构如下图所示：
-
-<img src="https://user-images.githubusercontent.com/13348433/31772175-5f419eca-b511-11e7-9db7-5231fe3d9ccb.png" width="500">
-
-- 数据分片（Data shard): 用于训练神经网络的数据，被切分成多个部分，每个部分分别给每个trainer使用。
-- 计算节点（Trainer）: 每个trainer启动后读取切分好的一部分数据，开始神经网络的“前馈”和“后馈”计算，并和参数服务器通信。在完成一定量数据的训练后，上传计算得出的梯度（gradients），然后下载优化更新后的神经网络参数（parameters）。
-- 参数服务器（Parameter server）:每个参数服务器只保存整个神经网络所有参数的一部分。参数服务器接收从计算节点上传的梯度，并完成参数优化更新，再将更新后的参数下发到每个计算节点。
-
-这样，通过计算节点和参数服务器的分布式协作，可以完成神经网络的SGD方法的训练。PaddlePaddle可以同时支持同步随机梯度下降（SGD）和异步随机梯度下降。
-
-在使用同步SGD训练神经网络时，PaddlePaddle使用同步屏障（barrier），使梯度的提交和参数的更新按照顺序方式执行。在异步SGD中，则并不会等待所有trainer提交梯度才更新参数，这样极大地提高了计算的并行性：参数服务器之间不相互依赖，并行地接收梯度和更新参数，参数服务器也不会等待计算节点全部都提交梯度之后才开始下一步，计算节点之间也不会相互依赖，并行地执行模型的训练。可以看出，虽然异步SGD方式会提高参数更新并行度, 但是并不能保证参数同步更新，在任意时间某一台参数服务器上保存的参数可能比另一台要更新，与同步SGD相比，梯度会有噪声。
-
-
-## 环境准备
-
-1. 准备您的计算集群。计算集群通常由一组（几台到几千台规模）的Linux服务器组成。服务器之间可以通过局域网（LAN）联通，每台服务器具有集群中唯一的IP地址（或者可被DNS解析的主机名）。集群中的每台计算机通常被成为一个“节点”。
-1. 我们需要在集群的所有节点上安装 PaddlePaddle。 如果要启用GPU，还需要在节点上安装对应的GPU驱动以及CUDA。PaddlePaddle的安装可以参考[build_and_install](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/build_and_install/index_cn.html)的多种安装方式。我们推荐使用[Docker](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/build_and_install/docker_install_cn.html)安装方式来快速安装PaddlePaddle。
-
-安装完成之后，执行下面的命令可以查看已经安装的版本（docker安装方式可以进入docker容器执行：`docker run -it paddlepaddle/paddle:[tag] /bin/bash`）：
-```bash
-$ paddle version
-PaddlePaddle 0.10.0, compiled with
-    with_avx: ON
-    with_gpu: OFF
-    with_double: OFF
-    with_python: ON
-    with_rdma: OFF
-    with_timer: OFF
-```
+## 启动参数说明
 
-下面以`doc/howto/usage/cluster/src/word2vec`中的代码作为实例，介绍使用PaddlePaddle v2 API完成分布式训练。
+下面以`doc/howto/cluster/src/word2vec`中的代码作为实例，介绍使用PaddlePaddle v2 API完成分布式训练。
 
-## 启动参数说明
 ### 启动参数服务器
 执行以下的命令启动一个参数服务器并等待和计算节点的数据交互
 ```bash
@@ -167,22 +133,3 @@ test.txt-00002
 
 - `train_data_dir`：包含训练数据的目录，可以是从分布式存储挂载过来的，也可以是在任务启动前下载到本地的。
 - `test_data_dir`：包含测试数据集的目录。
-
-## 使用分布式计算平台或工具
-
-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。
-
-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。
-
-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。
-
-## 在不同集群中运行
-
-  - [fabric集群](fabric_cn.md)
-  - [openmpi集群](openmpi_cn.md)
-  - [kubernetes单机](k8s_cn.md)
-  - [kubernetes distributed分布式](k8s_distributed_cn.md)
-  - [AWS上运行kubernetes集群训练](k8s_aws_cn.md)

doc/howto/usage/cluster/cluster_train_en.md → doc/howto/cluster/cmd_argument_en.md

-# Distributed Training
-
-## Introduction
-
-In this article, we'll explain how to run distributed training jobs with PaddlePaddle on different types of clusters. The diagram below shows the main architecture of a distributed trainning job:
-
-<img src="https://user-images.githubusercontent.com/13348433/31772146-41523d84-b511-11e7-8a12-a69fd136c283.png" width="500">
-
-- Data shard: training data will be split into multiple partitions, trainers use the partitions of the whole dataset to do the training job.
-- Trainer: each trainer reads the data shard, and train the neural network. Then the trainer will upload calculated "gradients" to parameter servers, and wait for parameters to be optimized on the parameter server side. When that finishes, the trainer download optimized parameters and continues its training.
-- Parameter server: every parameter server stores part of the whole neural network model data. They will do optimization calculations when gradients are uploaded from trainers, and then send updated parameters to trainers.
-
-PaddlePaddle can support both synchronize stochastic gradient descent (SGD) and asynchronous SGD.
-
-When training with synchronize SGD, PaddlePaddle uses an internal "synchronize barrier" which makes gradients update and parameter download in strict order. On the other hand, asynchronous SGD won't wait for all trainers to finish upload at a single step, this will increase the parallelism of distributed training: parameter servers do not depend on each other, they'll do parameter optimization concurrently. Parameter servers will not wait for trainers, so trainers will also do their work concurrently. But asynchronous SGD will introduce more randomness and noises in the gradient.
-
-## Preparations
-1. Prepare your computer cluster. It's normally a bunch of Linux servers connected by LAN. Each server will be assigned a unique IP address. The computers in the cluster can be called "nodes".
-2. Install PaddlePaddle on every node. If you are going to take advantage of GPU cards, you'll also need to install proper driver and CUDA libraries. To install PaddlePaddle please read [this build and install](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html) document. We strongly recommend using [Docker installation](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/docker_install_en.html).
-
-After installation, you can check the version by typing the below command (run a docker container  if using docker: `docker run -it paddlepaddle/paddle:[tag] /bin/bash`):
-
-```bash
-$ paddle version
-PaddlePaddle 0.10.0rc, compiled with
-    with_avx: ON
-    with_gpu: OFF
-    with_double: OFF
-    with_python: ON
-    with_rdma: OFF
-    with_timer: OFF
-```
-
-We'll take `doc/howto/usage/cluster/src/word2vec` as an example to introduce distributed training using PaddlePaddle v2 API.
-
 ## Command-line arguments
 
+We'll take `doc/howto/cluster/src/word2vec` as an example to introduce distributed training using PaddlePaddle v2 API.
+
 ### Starting parameter server
 
 Type the below command to start a parameter server which will wait for trainers to connect:
@@ -171,21 +138,3 @@ Your workspace may looks like:
 
 - `train_data_dir`: containing training data. Mount from storage service or copy trainning data to here.
 - `test_data_dir`: containing testing data.
-
-## Use cluster platforms or cluster management tools
-
-PaddlePaddle supports running jobs on several platforms including:
-- [Kubernetes](http://kubernetes.io) open-source system for automating deployment, scaling, and management of containerized applications from Google.
-- [OpenMPI](https://www.open-mpi.org) Mature high performance parallel computing framework.
-- [Fabric](http://www.fabfile.org) A cluster management tool. Write scripts to submit jobs or manage the cluster.
-
-We'll introduce cluster job management on these platforms. The examples can be found under [cluster_train_v2](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/scripts/cluster_train_v2).
-
-These cluster platforms provide API or environment variables for training processes, when the job is dispatched to different nodes. Like node ID, IP or total number of nodes etc.
-
-## Use different clusters
-
-  - [fabric](fabric_en.md)
-  - [openmpi](openmpi_en.md)
-  - [kubernetes](k8s_en.md)
-  - [kubernetes on AWS](k8s_aws_en.md)

doc/howto/cluster/index_cn.rst

+分布式训练
+==========
+
+..  toctree::
+  :maxdepth: 1
+
+  introduction_cn.md
+  preparations_cn.md
+  cmd_argument_cn.md
+  multi_cluster/index_cn.rst

doc/howto/cluster/index_en.rst

+Distributed Training
+====================
+
+..  toctree::
+  :maxdepth: 1
+
+  introduction_en.md
+  preparations_en.md
+  cmd_argument_en.md
+  multi_cluster/index_en.rst

doc/howto/cluster/introduction_cn.md

+## 概述
+
+本节将介绍如何使用PaddlePaddle在不同的集群框架下完成分布式训练。分布式训练架构如下图所示：
+
+<img src="https://user-images.githubusercontent.com/13348433/31772175-5f419eca-b511-11e7-9db7-5231fe3d9ccb.png" width="500">
+
+- 数据分片（Data shard): 用于训练神经网络的数据，被切分成多个部分，每个部分分别给每个trainer使用。
+- 计算节点（Trainer）: 每个trainer启动后读取切分好的一部分数据，开始神经网络的“前馈”和“后馈”计算，并和参数服务器通信。在完成一定量数据的训练后，上传计算得出的梯度（gradients），然后下载优化更新后的神经网络参数（parameters）。
+- 参数服务器（Parameter server）:每个参数服务器只保存整个神经网络所有参数的一部分。参数服务器接收从计算节点上传的梯度，并完成参数优化更新，再将更新后的参数下发到每个计算节点。
+
+这样，通过计算节点和参数服务器的分布式协作，可以完成神经网络的SGD方法的训练。PaddlePaddle可以同时支持同步随机梯度下降（SGD）和异步随机梯度下降。
+
+在使用同步SGD训练神经网络时，PaddlePaddle使用同步屏障（barrier），使梯度的提交和参数的更新按照顺序方式执行。在异步SGD中，则并不会等待所有trainer提交梯度才更新参数，这样极大地提高了计算的并行性：参数服务器之间不相互依赖，并行地接收梯度和更新参数，参数服务器也不会等待计算节点全部都提交梯度之后才开始下一步，计算节点之间也不会相互依赖，并行地执行模型的训练。可以看出，虽然异步SGD方式会提高参数更新并行度, 但是并不能保证参数同步更新，在任意时间某一台参数服务器上保存的参数可能比另一台要更新，与同步SGD相比，梯度会有噪声。

doc/howto/cluster/introduction_en.md

+## Introduction
+
+In this section, we'll explain how to run distributed training jobs with PaddlePaddle on different types of clusters. The diagram below shows the main architecture of a distributed trainning job:
+
+<img src="https://user-images.githubusercontent.com/13348433/31772146-41523d84-b511-11e7-8a12-a69fd136c283.png" width="500">
+
+- Data shard: training data will be split into multiple partitions, trainers use the partitions of the whole dataset to do the training job.
+- Trainer: each trainer reads the data shard, and train the neural network. Then the trainer will upload calculated "gradients" to parameter servers, and wait for parameters to be optimized on the parameter server side. When that finishes, the trainer download optimized parameters and continues its training.
+- Parameter server: every parameter server stores part of the whole neural network model data. They will do optimization calculations when gradients are uploaded from trainers, and then send updated parameters to trainers.
+
+PaddlePaddle can support both synchronize stochastic gradient descent (SGD) and asynchronous SGD.
+
+When training with synchronize SGD, PaddlePaddle uses an internal "synchronize barrier" which makes gradients update and parameter download in strict order. On the other hand, asynchronous SGD won't wait for all trainers to finish upload at a single step, this will increase the parallelism of distributed training: parameter servers do not depend on each other, they'll do parameter optimization concurrently. Parameter servers will not wait for trainers, so trainers will also do their work concurrently. But asynchronous SGD will introduce more randomness and noises in the gradient.

doc/howto/cluster/multi_cluster/index_cn.rst

+在不同集群中运行
+================
+
+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。
+
+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。
+
+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。
+
+..  toctree::
+  :maxdepth: 1
+
+  fabric_cn.md
+  openmpi_cn.md
+  k8s_cn.md
+  k8s_distributed_cn.md
+  k8s_aws_cn.md

doc/howto/cluster/multi_cluster/index_en.rst

+Use different clusters
+======================
+
+PaddlePaddle supports running jobs on several platforms including:
+- `Kubernetes <http://kubernetes.io>`_ open-source system for automating deployment, scaling, and management of containerized applications from Google.
+- `OpenMPI <https://www.open-mpi.org>`_ Mature high performance parallel computing framework.
+- `Fabric <http://www.fabfile.org>`_ A cluster management tool. Write scripts to submit jobs or manage the cluster.
+
+We'll introduce cluster job management on these platforms. The examples can be found under `cluster_train_v2 <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/scripts/cluster_train_v2>`_ .
+
+These cluster platforms provide API or environment variables for training processes, when the job is dispatched to different nodes. Like node ID, IP or total number of nodes etc.
+
+..  toctree::
+  :maxdepth: 1
+
+  fabric_en.md
+  openmpi_en.md
+  k8s_en.md
+  k8s_aws_en.md

doc/howto/cluster/preparations_cn.md

+## 环境准备
+
+1. 准备您的计算集群。计算集群通常由一组（几台到几千台规模）的Linux服务器组成。服务器之间可以通过局域网（LAN）联通，每台服务器具有集群中唯一的IP地址（或者可被DNS解析的主机名）。集群中的每台计算机通常被成为一个“节点”。
+1. 我们需要在集群的所有节点上安装 PaddlePaddle。 如果要启用GPU，还需要在节点上安装对应的GPU驱动以及CUDA。PaddlePaddle的安装可以参考[build_and_install](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/build_and_install/index_cn.html)的多种安装方式。我们推荐使用[Docker](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/build_and_install/docker_install_cn.html)安装方式来快速安装PaddlePaddle。
+
+安装完成之后，执行下面的命令可以查看已经安装的版本（docker安装方式可以进入docker容器执行：`docker run -it paddlepaddle/paddle:[tag] /bin/bash`）：
+```bash
+$ paddle version
+PaddlePaddle 0.10.0, compiled with
+    with_avx: ON
+    with_gpu: OFF
+    with_double: OFF
+    with_python: ON
+    with_rdma: OFF
+    with_timer: OFF
+```

doc/howto/cluster/preparations_en.md

+## Preparations
+
+1. Prepare your computer cluster. It's normally a bunch of Linux servers connected by LAN. Each server will be assigned a unique IP address. The computers in the cluster can be called "nodes".
+2. Install PaddlePaddle on every node. If you are going to take advantage of GPU cards, you'll also need to install proper driver and CUDA libraries. To install PaddlePaddle please read [this build and install](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html) document. We strongly recommend using [Docker installation](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/docker_install_en.html).
+
+After installation, you can check the version by typing the below command (run a docker container  if using docker: `docker run -it paddlepaddle/paddle:[tag] /bin/bash`):
+
+```bash
+$ paddle version
+PaddlePaddle 0.10.0rc, compiled with
+    with_avx: ON
+    with_gpu: OFF
+    with_double: OFF
+    with_python: ON
+    with_rdma: OFF
+    with_timer: OFF
+```

doc/howto/usage/cmd_parameter/index_cn.rst → doc/howto/cmd_parameter/index_cn.rst

 ..  _cmd_line_index:
 
-设置命令行参数
+命令行参数设置
 ===============
 
 ..  toctree::

doc/howto/dev/contribute_to_paddle_en.md

-../../../CONTRIBUTING.md
\ No newline at end of file

doc/howto/index_cn.rst

-进阶指南
+进阶使用
 ========
 
-使用说明
---------
-
-..  toctree::
-  :maxdepth: 1
-
-  usage/cmd_parameter/index_cn.rst
-  usage/cluster/cluster_train_cn.md
-  usage/capi/index_cn.rst
-
-开发标准
---------
-
-..  toctree::
-  :maxdepth: 1
-
-  dev/contribute_to_paddle_cn.md
-  dev/write_docs_cn.rst
-
-模型配置
---------
-
-..  toctree::
-  :maxdepth: 1
-
-  deep_model/rnn/index_cn.rst
-
-性能优化
---------
-
 ..  toctree::
   :maxdepth: 1
 
+  cmd_parameter/index_cn.rst
+  cluster/index_cn.rst
+  capi/index_cn.rst
+  rnn/index_cn.rst
   optimization/gpu_profiling_cn.rst

doc/howto/index_en.rst

 HOW TO
 =======
 
-Usage
--------
-
-..  toctree::
-  :maxdepth: 1
-
-  usage/cmd_parameter/index_en.rst
-  usage/cluster/cluster_train_en.md
-
-Development
-------------
-
-..  toctree::
-  :maxdepth: 1
-
-  dev/new_layer_en.rst
-  dev/contribute_to_paddle_en.md
-  dev/write_docs_en.rst
-
-Configuration
--------------
-
-..  toctree::
-  :maxdepth: 1
-
-  deep_model/rnn/index_en.rst
-
-Optimization
--------------
-
 ..  toctree::
   :maxdepth: 1
 
+  cmd_parameter/index_en.rst
+  cluster/index_en.rst
+  rnn/index_en.rst
   optimization/gpu_profiling_en.rst

doc/howto/optimization/gpu_profiling_cn.rst

-==================
-GPU性能分析与调优
-==================
+============
+GPU性能调优
+============
 
 ..  contents::
 

doc/index_cn.rst

@@ -5,7 +5,8 @@ PaddlePaddle 文档
   :maxdepth: 1
 
   getstarted/index_cn.rst
+  build_and_install/index_cn.rst
   howto/index_cn.rst
+  dev/index_cn.rst
   api/index_cn.rst
   faq/index_cn.rst
-  mobile/index_cn.rst

doc/index_en.rst

@@ -5,6 +5,7 @@ PaddlePaddle Documentation
   :maxdepth: 1
 
   getstarted/index_en.rst
+  build_and_install/index_en.rst
   howto/index_en.rst
+  dev/index_en.rst
   api/index_en.rst
-  mobile/index_en.rst

doc/mobile/index_cn.rst

-MOBILE
-======
-
-..  toctree::
-  :maxdepth: 1
-
-  cross_compiling_for_android_cn.md
-  cross_compiling_for_ios_cn.md
-  cross_compiling_for_raspberry_cn.md

doc/mobile/index_en.rst

-MOBILE
-======
-
-..  toctree::
-  :maxdepth: 1
-
-  cross_compiling_for_android_en.md
-  cross_compiling_for_ios_en.md
-  cross_compiling_for_raspberry_en.md

paddle/framework/CMakeLists.txt

@@ -24,6 +24,8 @@ cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto)
 cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor paddle_memory)
 nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor init)
 
+cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
+
 cc_test(variable_test SRCS variable_test.cc)
 
 cc_library(threadpool SRCS threadpool.cc DEPS enforce)
@@ -92,11 +94,4 @@ cc_test(init_test SRCS init_test.cc DEPS init)
 cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
 cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
       
-if(NOT WITH_C_API AND WITH_FLUID)
-  file(GLOB FRAMEWORK_HEADERS *.h)
-  install(FILES ${FRAMEWORK_HEADERS} DESTINATION include/paddle/framework)
-  install(FILES ${CMAKE_CURRENT_BINARY_DIR}/framework.pb.h DESTINATION include/paddle/framework)
-  install(FILES details/cow_ptr.h details/op_registry.h DESTINATION include/paddle/framework/details)
-endif()
-
 cc_test(channel_test SRCS channel_test.cc)

paddle/framework/backward.cc

@@ -534,7 +534,7 @@ ParamGradInfoMap AppendBackward(
   auto root_block = program_desc.MutableBlock(root_block_idx);
 
   std::string fill_one_op_out = GradVarName(target.Name());
-  bool is_scalar = target.Shape() == std::vector<int64_t>{1};
+  bool is_scalar = target.GetShape() == std::vector<int64_t>{1};
   PADDLE_ENFORCE(is_scalar, "target should be scalar");
   VLOG(3) << "backward from loss=" << target.Name()
           << " data_type=" << target.GetDataType();
@@ -565,7 +565,7 @@ ParamGradInfoMap AppendBackward(
 
   auto var = root_block->Var(fill_one_op_out);
   var->SetDataType(target.GetDataType());
-  var->SetShape(target.Shape());
+  var->SetShape(target.GetShape());
   auto& target_grad = retv[target.Name()];
   target_grad.name_ = fill_one_op_out;
   target_grad.block_idx_ = root_block_idx;

paddle/framework/channel.h

@@ -23,8 +23,8 @@ namespace framework {
 template <typename T>
 class Channel {
  public:
-  virtual void Send(T*) = 0;
-  virtual void Receive(T*) = 0;
+  virtual bool Send(T*) = 0;
+  virtual bool Receive(T*) = 0;
   virtual size_t Cap() = 0;
   virtual void Close() = 0;
   virtual ~Channel() {}

paddle/framework/channel_test.cc

@@ -29,16 +29,16 @@ TEST(Channel, MakeAndClose) {
   {
     // MakeChannel should return a buffered channel is buffer_size > 0.
     auto ch = MakeChannel<int>(10);
-    EXPECT_NE(dynamic_cast<Buffered<int>*>(ch), nullptr);
-    EXPECT_EQ(dynamic_cast<UnBuffered<int>*>(ch), nullptr);
+    EXPECT_NE(dynamic_cast<Buffered<int> *>(ch), nullptr);
+    EXPECT_EQ(dynamic_cast<UnBuffered<int> *>(ch), nullptr);
     CloseChannel(ch);
     delete ch;
   }
   {
     // MakeChannel should return an un-buffered channel is buffer_size = 0.
     auto ch = MakeChannel<int>(0);
-    EXPECT_EQ(dynamic_cast<Buffered<int>*>(ch), nullptr);
-    EXPECT_NE(dynamic_cast<UnBuffered<int>*>(ch), nullptr);
+    EXPECT_EQ(dynamic_cast<Buffered<int> *>(ch), nullptr);
+    EXPECT_NE(dynamic_cast<UnBuffered<int> *>(ch), nullptr);
     CloseChannel(ch);
     delete ch;
   }
@@ -48,18 +48,59 @@ TEST(Channel, SufficientBufferSizeDoesntBlock) {
   const size_t buffer_size = 10;
   auto ch = MakeChannel<size_t>(buffer_size);
   for (size_t i = 0; i < buffer_size; ++i) {
-    ch->Send(&i);  // should not block
+    EXPECT_EQ(ch->Send(&i), true);  // should not block
   }
 
   size_t out;
   for (size_t i = 0; i < buffer_size; ++i) {
-    ch->Receive(&out);  // should not block
+    EXPECT_EQ(ch->Receive(&out), true);  // should not block
     EXPECT_EQ(out, i);
   }
   CloseChannel(ch);
   delete ch;
 }
 
+TEST(Channel, SendOnClosedChannelPanics) {
+  const size_t buffer_size = 10;
+  auto ch = MakeChannel<size_t>(buffer_size);
+  size_t i = 5;
+  EXPECT_EQ(ch->Send(&i), true);  // should not block or panic
+  CloseChannel(ch);
+  EXPECT_EQ(ch->Send(&i), false);  // should panic
+  delete ch;
+}
+
+TEST(Channel, ReceiveFromBufferedChannelReturnResidualValuesTest) {
+  const size_t buffer_size = 10;
+  auto ch = MakeChannel<size_t>(buffer_size);
+
+  for (size_t i = 0; i < buffer_size; ++i) {
+    EXPECT_EQ(ch->Send(&i), true);  // sending should not block
+  }
+
+  size_t out;
+  for (size_t i = 0; i < buffer_size / 2; ++i) {
+    EXPECT_EQ(ch->Receive(&out), true);  // receiving should not block
+    EXPECT_EQ(out, i);
+  }
+
+  CloseChannel(ch);
+
+  for (size_t i = buffer_size / 2; i < buffer_size; ++i) {
+    EXPECT_EQ(ch->Receive(&out),
+              true);  // receving should return residual values.
+    EXPECT_EQ(out, i);
+  }
+
+  for (size_t i = 0; i < buffer_size; ++i) {
+    EXPECT_EQ(ch->Receive(&out),
+              false);  // after receiving residual values, return zeros.
+    // Note: we cannot check EXPECT_EQ(out, 0), because C++ doesn't
+    // define zero values like Go does.
+  }
+  delete ch;
+}
+
 TEST(Channel, ConcurrentSendNonConcurrentReceiveWithSufficientBufferSize) {
   const size_t buffer_size = 10;
   auto ch = MakeChannel<size_t>(buffer_size);
@@ -67,7 +108,10 @@ TEST(Channel, ConcurrentSendNonConcurrentReceiveWithSufficientBufferSize) {
   std::thread t([&]() {
     // Try to write more than buffer size.
     for (size_t i = 0; i < 2 * buffer_size; ++i) {
-      ch->Send(&i);  // should not block
+      if (i < buffer_size)
+        EXPECT_EQ(ch->Send(&i), true);  // should block after 10 iterations
+      else
+        EXPECT_EQ(ch->Send(&i), false);
       sum += i;
     }
   });
@@ -78,3 +122,262 @@ TEST(Channel, ConcurrentSendNonConcurrentReceiveWithSufficientBufferSize) {
   t.join();
   delete ch;
 }
+
+TEST(Channel, SimpleUnbufferedChannelTest) {
+  auto ch = MakeChannel<int>(0);
+  unsigned sum_send = 0;
+  std::thread t([&]() {
+    for (int i = 0; i < 5; i++) {
+      EXPECT_EQ(ch->Send(&i), true);
+      sum_send += i;
+    }
+  });
+  for (int i = 0; i < 5; i++) {
+    int recv;
+    EXPECT_EQ(ch->Receive(&recv), true);
+    EXPECT_EQ(recv, i);
+  }
+
+  CloseChannel(ch);
+  t.join();
+  EXPECT_EQ(sum_send, 10U);
+  delete ch;
+}
+
+// This tests that closing a buffered channel also unblocks
+//  any receivers waiting on the channel
+TEST(Channel, BufferedChannelCloseUnblocksReceiversTest) {
+  auto ch = MakeChannel<int>(1);
+  size_t num_threads = 5;
+  std::thread t[num_threads];
+  bool thread_ended[num_threads];
+
+  // Launches threads that try to read and are blocked because of no writers
+  for (size_t i = 0; i < num_threads; i++) {
+    thread_ended[i] = false;
+    t[i] = std::thread(
+        [&](bool *p) {
+          int data;
+          // All reads should return false
+          EXPECT_EQ(ch->Receive(&data), false);
+          *p = true;
+        },
+        &thread_ended[i]);
+  }
+  std::this_thread::sleep_for(std::chrono::milliseconds(100));  // wait
+
+  // Verify that all threads are blocked
+  for (size_t i = 0; i < num_threads; i++) {
+    EXPECT_EQ(thread_ended[i], false);
+  }
+
+  // Explicitly close the channel
+  // This should unblock all receivers
+  CloseChannel(ch);
+
+  std::this_thread::sleep_for(std::chrono::milliseconds(200));  // wait
+
+  // Verify that all threads got unblocked
+  for (size_t i = 0; i < num_threads; i++) {
+    EXPECT_EQ(thread_ended[i], true);
+  }
+
+  for (size_t i = 0; i < num_threads; i++) t[i].join();
+  delete ch;
+}
+
+// This tests that closing a buffered channel also unblocks
+//  any senders waiting for channel to have write space
+TEST(Channel, BufferedChannelCloseUnblocksSendersTest) {
+  auto ch = MakeChannel<int>(1);
+  size_t num_threads = 5;
+  std::thread t[num_threads];
+  bool thread_ended[num_threads];
+  bool send_success[num_threads];
+
+  // Launches threads that try to write and are blocked because of no readers
+  for (size_t i = 0; i < num_threads; i++) {
+    thread_ended[i] = false;
+    send_success[i] = false;
+    t[i] = std::thread(
+        [&](bool *ended, bool *success) {
+          int data = 10;
+          *success = ch->Send(&data);
+          *ended = true;
+        },
+        &thread_ended[i], &send_success[i]);
+  }
+  std::this_thread::sleep_for(std::chrono::milliseconds(100));  // wait
+
+  // Verify that atleast 4 threads are blocked
+  int ct = 0;
+  for (size_t i = 0; i < num_threads; i++) {
+    if (thread_ended[i] == false) ct++;
+  }
+  // Atleast 4 threads must be blocked
+  EXPECT_GE(ct, 4);
+
+  // Explicitly close the thread
+  // This should unblock all senders
+  CloseChannel(ch);
+
+  std::this_thread::sleep_for(std::chrono::milliseconds(200));  // wait
+
+  // Verify that all threads got unblocked
+  for (size_t i = 0; i < num_threads; i++) {
+    EXPECT_EQ(thread_ended[i], true);
+  }
+
+  // Verify that only 1 send was successful
+  ct = 0;
+  for (size_t i = 0; i < num_threads; i++) {
+    if (send_success[i]) ct++;
+  }
+  // Only 1 send must be successful
+  EXPECT_EQ(ct, 1);
+
+  for (size_t i = 0; i < num_threads; i++) t[i].join();
+  delete ch;
+}
+
+// This tests that closing an unbuffered channel also unblocks
+//  unblocks any receivers waiting for senders
+TEST(Channel, UnbufferedChannelCloseUnblocksReceiversTest) {
+  auto ch = MakeChannel<int>(0);
+  size_t num_threads = 5;
+  std::thread t[num_threads];
+  bool thread_ended[num_threads];
+
+  // Launches threads that try to read and are blocked becausew of no writers
+  for (size_t i = 0; i < num_threads; i++) {
+    thread_ended[i] = false;
+    t[i] = std::thread(
+        [&](bool *p) {
+          int data;
+          EXPECT_EQ(ch->Receive(&data), false);
+          *p = true;
+        },
+        &thread_ended[i]);
+  }
+  std::this_thread::sleep_for(std::chrono::milliseconds(500));  // wait 0.5 sec
+
+  // Verify that all the threads are blocked
+  for (size_t i = 0; i < num_threads; i++) {
+    EXPECT_EQ(thread_ended[i], false);
+  }
+
+  // Explicitly close the thread
+  // This should unblock all receivers
+  CloseChannel(ch);
+
+  std::this_thread::sleep_for(std::chrono::milliseconds(500));  // wait 0.5 sec
+
+  // Verify that all threads got unblocked
+  for (size_t i = 0; i < num_threads; i++) {
+    EXPECT_EQ(thread_ended[i], true);
+  }
+
+  for (size_t i = 0; i < num_threads; i++) t[i].join();
+  delete ch;
+}
+
+// This tests that closing an unbuffered channel also unblocks
+//  unblocks any senders waiting for senders
+TEST(Channel, UnbufferedChannelCloseUnblocksSendersTest) {
+  auto ch = MakeChannel<int>(0);
+  size_t num_threads = 5;
+  std::thread t[num_threads];
+  bool thread_ended[num_threads];
+
+  // Launches threads that try to read and are blocked becausew of no writers
+  for (size_t i = 0; i < num_threads; i++) {
+    thread_ended[i] = false;
+    t[i] = std::thread(
+        [&](bool *p) {
+          int data = 10;
+          EXPECT_EQ(ch->Send(&data), false);
+          *p = true;
+        },
+        &thread_ended[i]);
+  }
+  std::this_thread::sleep_for(std::chrono::milliseconds(500));  // wait 0.5 sec
+
+  // Verify that all the threads are blocked
+  for (size_t i = 0; i < num_threads; i++) {
+    EXPECT_EQ(thread_ended[i], false);
+  }
+
+  // Explicitly close the thread
+  // This should unblock all receivers
+  CloseChannel(ch);
+
+  std::this_thread::sleep_for(std::chrono::milliseconds(500));  // wait 0.5 sec
+
+  // Verify that all threads got unblocked
+  for (size_t i = 0; i < num_threads; i++) {
+    EXPECT_EQ(thread_ended[i], true);
+  }
+
+  for (size_t i = 0; i < num_threads; i++) t[i].join();
+  delete ch;
+}
+
+TEST(Channel, UnbufferedLessReceiveMoreSendTest) {
+  auto ch = MakeChannel<int>(0);
+  unsigned sum_send = 0;
+  // Send should block after three iterations
+  // since we only have three receivers.
+  std::thread t([&]() {
+    // Try to send more number of times
+    // than receivers
+    for (int i = 0; i < 4; i++) {
+      ch->Send(&i);
+      sum_send += i;
+    }
+  });
+  for (int i = 0; i < 3; i++) {
+    int recv;
+    ch->Receive(&recv);
+    EXPECT_EQ(recv, i);
+  }
+  std::this_thread::sleep_for(std::chrono::milliseconds(100));  // wait 0.5 sec
+  EXPECT_EQ(sum_send, 3U);
+
+  CloseChannel(ch);
+  t.join();
+  delete ch;
+}
+
+TEST(Channel, UnbufferedMoreReceiveLessSendTest) {
+  auto ch = MakeChannel<int>(0);
+  unsigned sum_send = 0;
+  unsigned sum_receive = 0;
+  // The receiver should block after 5
+  // iterations, since there are only 5 senders.
+  std::thread t([&]() {
+    for (int i = 0; i < 8; i++) {
+      int recv;
+      ch->Receive(&recv);  // should block after the fifth iteration.
+      EXPECT_EQ(recv, i);
+      sum_receive += i;
+    }
+  });
+  for (int i = 0; i < 5; i++) {
+    ch->Send(&i);
+    sum_send += i;
+  }
+  std::this_thread::sleep_for(std::chrono::milliseconds(500));  // wait 0.5 sec
+  EXPECT_EQ(sum_send, 10U);
+  EXPECT_EQ(sum_receive, 10U);
+  // send three more elements
+  for (int i = 5; i < 8; i++) {
+    ch->Send(&i);
+    sum_send += i;
+  }
+
+  CloseChannel(ch);
+  t.join();
+  EXPECT_EQ(sum_send, 28U);
+  EXPECT_EQ(sum_receive, 28U);
+  delete ch;
+}

paddle/framework/details/buffered_channel.h

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <atomic>
 #include <condition_variable>
 #include <deque>
 #include <mutex>
@@ -30,8 +31,8 @@ class Buffered : public paddle::framework::Channel<T> {
   friend void paddle::framework::CloseChannel<T>(Channel<T>*);
 
  public:
-  virtual void Send(T*);
-  virtual void Receive(T*);
+  virtual bool Send(T*);
+  virtual bool Receive(T*);
   virtual size_t Cap() { return cap_; }
   virtual void Close();
   virtual ~Buffered();
@@ -42,17 +43,21 @@ class Buffered : public paddle::framework::Channel<T> {
   std::condition_variable empty_cond_var_;
   std::condition_variable full_cond_var_;
   std::deque<T> channel_;
-  bool closed_;
+  std::atomic<bool> closed_{false};
 
   Buffered(size_t cap) : cap_(cap), closed_(false) {
     PADDLE_ENFORCE_GT(cap, 0);
   }
 
-  void NotifyAllSenders(std::unique_lock<std::mutex>*);
+  void NotifyAllParticipants(std::unique_lock<std::mutex>*);
 };
 
 template <typename T>
-void Buffered<T>::Send(T* item) {
+bool Buffered<T>::Send(T* item) {
+  bool ret = false;
+  if (closed_) {
+    return ret;
+  }
   std::unique_lock<std::mutex> lock(mu_);
   full_cond_var_.wait(lock,
                       [this]() { return channel_.size() < cap_ || closed_; });
@@ -60,27 +65,33 @@ void Buffered<T>::Send(T* item) {
     channel_.push_back(std::move(*item));
     lock.unlock();
     empty_cond_var_.notify_one();
+    ret = true;
   }
+  return ret;
 }
 
 template <typename T>
-void Buffered<T>::Receive(T* item) {
+bool Buffered<T>::Receive(T* item) {
   std::unique_lock<std::mutex> lock(mu_);
   empty_cond_var_.wait(lock, [this]() { return !channel_.empty() || closed_; });
-  if (!closed_) {
+  bool ret = false;
+  if (!channel_.empty()) {
     *item = std::move(channel_.front());
     channel_.pop_front();
-    NotifyAllSenders(&lock);
-  } else {
-    item = nullptr;
+    full_cond_var_.notify_one();
+    ret = true;
   }
+  return ret;
 }
 
 template <typename T>
 void Buffered<T>::Close() {
+  if (closed_) {
+    return;
+  }
   std::unique_lock<std::mutex> lock(mu_);
   closed_ = true;
-  NotifyAllSenders(&lock);
+  NotifyAllParticipants(&lock);
 }
 
 template <typename T>
@@ -88,13 +99,14 @@ Buffered<T>::~Buffered() {
   std::unique_lock<std::mutex> lock(mu_);
   closed_ = true;
   channel_.clear();
-  NotifyAllSenders(&lock);
+  NotifyAllParticipants(&lock);
 }
 
 template <typename T>
-void Buffered<T>::NotifyAllSenders(std::unique_lock<std::mutex>* lock) {
+void Buffered<T>::NotifyAllParticipants(std::unique_lock<std::mutex>* lock) {
   lock->unlock();
   full_cond_var_.notify_all();
+  empty_cond_var_.notify_all();
 }
 
 }  // namespace details

paddle/framework/details/unbuffered_channel.h

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+/* 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.
@@ -13,8 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <atomic>
 #include <condition_variable>
-#include <deque>
 #include <mutex>
 
 #include "paddle/framework/channel.h"
@@ -29,27 +29,123 @@ class UnBuffered : public paddle::framework::Channel<T> {
   friend void paddle::framework::CloseChannel<T>(Channel<T>*);
 
  public:
-  virtual void Send(T*);
-  virtual void Receive(T*);
+  virtual bool Send(T*);
+  virtual bool Receive(T*);
   virtual size_t Cap() { return 0; }
   virtual void Close();
   virtual ~UnBuffered();
 
  private:
-  UnBuffered() {}
+  std::mutex mu_ch_;
+  // Mutex for readers and writers who are waiting for other reader
+  // and writer to complete execution
+  std::recursive_mutex mu_read_, mu_write_;
+  // reader_found_ is set true when a reader is ready to accept data
+  // writer_found_ is set true when a writer is ready to send data
+  // A transaction occurs only when both are true
+  std::atomic<bool> reader_found_{false}, writer_found_{false};
+  std::condition_variable cv_channel_;
+  std::condition_variable_any cv_reader_, cv_writer_;
+  T* item{nullptr};
+  std::atomic<bool> closed_{false};
+
+  UnBuffered() : closed_(false) {}
+
+  void NotifyAllParticipants(std::unique_lock<std::mutex>*);
 };
 
+// This function implements the concept of how data should
+// be sent from a writer to a reader.
+template <typename T>
+bool UnBuffered<T>::Send(T* data) {
+  bool ret = false;
+  if (closed_) {
+    return ret;
+  }
+  // Prevent other writers from entering
+  std::unique_lock<std::recursive_mutex> writer_lock(mu_write_);
+  writer_found_ = true;
+  std::unique_lock<std::recursive_mutex> cv_lock(mu_write_);
+  // If writer comes first, it should wait till a reader arrives
+  cv_writer_.wait(cv_lock,
+                  [this]() { return reader_found_ == true || closed_; });
+  cv_reader_.notify_one();
+  if (!closed_) {
+    std::unique_lock<std::mutex> channel_lock(mu_ch_);
+    item = data;
+    channel_lock.unlock();
+    cv_channel_.notify_one();
+    channel_lock.lock();
+    cv_channel_.wait(channel_lock,
+                     [this]() { return item == nullptr || closed_; });
+    ret = true;
+  }
+  writer_found_ = false;
+  return ret;
+}
+
+// This function implements the concept of how
+// data that was sent by a writer is read from a reader.
 template <typename T>
-void UnBuffered<T>::Send(T* channel_element) {}
+bool UnBuffered<T>::Receive(T* data) {
+  // Prevent other readers from entering
+  std::unique_lock<std::recursive_mutex> read_lock{mu_read_};
+  reader_found_ = true;
+  std::unique_lock<std::recursive_mutex> cv_lock{mu_read_};
+  // If reader comes first, it should wait till a writer arrives
+  cv_reader_.wait(cv_lock,
+                  [this]() { return writer_found_ == true || closed_; });
+  cv_writer_.notify_one();
+  bool ret = false;
+  if (!closed_) {
+    std::unique_lock<std::mutex> lock_ch{mu_ch_};
+    // Reader should wait for the writer to first write its data
+    cv_channel_.wait(lock_ch, [this]() { return item != nullptr || closed_; });
+    if (!closed_) {
+      *data = std::move(*item);
+      item = nullptr;
+      lock_ch.unlock();
+      ret = true;
+    }
+    cv_channel_.notify_one();
+  }
+  reader_found_ = false;
+  return ret;
+}
 
+// This function implements the sequence of events
+// that take place once the channel is closed.
 template <typename T>
-void UnBuffered<T>::Receive(T*) {}
+void UnBuffered<T>::Close() {
+  if (closed_) {
+    return;
+  }
+  std::unique_lock<std::mutex> lock(mu_ch_);
+  item = nullptr;
+  closed_ = true;
+  NotifyAllParticipants(&lock);
+}
 
+// This function implements the sequence of events
+// that are executed once the object of an UnBuffered
+// channel is destroyed.
 template <typename T>
-void UnBuffered<T>::Close() {}
+UnBuffered<T>::~UnBuffered() {
+  std::unique_lock<std::mutex> lock(mu_ch_);
+  item = nullptr;
+  closed_ = true;
+  NotifyAllParticipants(&lock);
+}
 
+// This function notifies all the readers, writers and
+// the channel condition variables.
 template <typename T>
-UnBuffered<T>::~UnBuffered() {}
+void UnBuffered<T>::NotifyAllParticipants(std::unique_lock<std::mutex>* lock) {
+  lock->unlock();
+  cv_writer_.notify_all();
+  cv_channel_.notify_all();
+  cv_reader_.notify_all();
+}
 
 }  // namespace details
 }  // namespace framework

paddle/framework/executor.cc

@@ -22,6 +22,7 @@ limitations under the License. */
 #include "paddle/framework/lod_rank_table.h"
 #include "paddle/framework/lod_tensor_array.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/framework/reader.h"
 #include "paddle/platform/place.h"
 #include "paddle/platform/profiler.h"
 
@@ -52,11 +53,13 @@ static void CreateTensor(Variable* var, proto::VarDesc::VarType var_type) {
     var->GetMutable<LoDTensorArray>();
   } else if (var_type == proto::VarDesc::PLACE_LIST) {
     var->GetMutable<platform::PlaceList>();
+  } else if (var_type == proto::VarDesc::READER) {
+    var->GetMutable<ReaderHolder>();
   } else {
     PADDLE_THROW(
         "Variable type %d is not in "
-        "[LoDTensor, SelectedRows, FEED_MINIBATCH, FETCH_LIST, LOD_RANK_TABLE,"
-        " PLACE_LIST]",
+        "[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, "
+        "LOD_RANK_TABLE, PLACE_LIST, READER]",
         var_type);
   }
 }

paddle/framework/framework.proto

@@ -116,6 +116,8 @@ message LoDTensorArrayDesc {
   optional int32 lod_level = 2 [ default = 0 ];
 }
 
+message ReaderDesc { repeated LoDTensorDesc lod_tensor = 1; }
+
 message VarDesc {
   enum VarType {
     LOD_TENSOR = 1;
@@ -126,13 +128,15 @@ message VarDesc {
     LOD_RANK_TABLE = 6;
     LOD_TENSOR_ARRAY = 7;
     PLACE_LIST = 8;
+    READER = 9;
   }
   required string name = 1;
   required VarType type = 2;
-  optional LoDTensorDesc lod_tensor = 3;
-  optional TensorDesc selected_rows = 4;
+  optional bool persistable = 3 [ default = false ];
+  optional LoDTensorDesc lod_tensor = 4;
+  optional TensorDesc selected_rows = 5;
   optional LoDTensorArrayDesc tensor_array = 6;
-  optional bool persistable = 5 [ default = false ];
+  optional ReaderDesc reader = 7;
 }
 
 message BlockDesc {

paddle/framework/mixed_vector.h

@@ -60,6 +60,14 @@ class Vector : public std::vector<T> {
   T *data() { return std::vector<T>::data(); }
   const T *data() const { return std::vector<T>::data(); }
 
+  T *data(const platform::Place &place) {
+    if (platform::is_cpu_place(place)) {
+      return data();
+    } else {
+      return cuda_data();
+    }
+  }
+
   /* Synchronize host vector to device vector */
   void CopyToCUDA();
   /* Synchronize device vector to host vector */

paddle/framework/op_desc.cc

@@ -39,10 +39,6 @@ class CompileTimeInferShapeContext : public InferShapeContext {
 
   bool HasOutputs(const std::string &name) const override;
 
-  DDim GetInputDim(const std::string &name) const override;
-
-  void SetOutputDim(const std::string &name, const DDim &dim) override;
-
   AttrReader Attrs() const override;
 
   const std::vector<std::string> &Inputs(
@@ -76,6 +72,11 @@ class CompileTimeInferShapeContext : public InferShapeContext {
 
   void SetDim(const std::string &name, const DDim &dim) override;
 
+  std::vector<DDim> GetRepeatedDims(const std::string &name) const override;
+
+  void SetRepeatedDims(const std::string &name,
+                       const std::vector<DDim> &dims) override;
+
   const OpDesc &op_;
   const BlockDesc &block_;
 };
@@ -124,7 +125,7 @@ OpDesc::OpDesc(const proto::OpDesc &desc, ProgramDesc *prog, BlockDesc *block)
   // restore attrs_
   for (const proto::OpDesc::Attr &attr : desc_.attrs()) {
     std::string attr_name = attr.name();
-    // The sub_block referred to by the BLOCK attr hasn't be added
+    // The sub_block referred to by the BLOCK attr hasn't been added
     // to ProgramDesc class yet, we skip setting BLOCK attr here.
     if (attr.type() != proto::AttrType::BLOCK) {
       attrs_[attr_name] = GetAttrValue(attr);
@@ -443,21 +444,6 @@ bool CompileTimeInferShapeContext::HasOutputs(const std::string &name) const {
   return true;
 }
 
-DDim CompileTimeInferShapeContext::GetInputDim(const std::string &name) const {
-  std::vector<DDim> ddims = GetInputsDim(name);
-  auto length = ddims.size();
-  PADDLE_ENFORCE_EQ(length, 1UL,
-                    "Input(%s) should have 1 value, "
-                    "but it has %d now",
-                    name, length);
-  return ddims[0];
-}
-
-void CompileTimeInferShapeContext::SetOutputDim(const std::string &name,
-                                                const DDim &dim) {
-  SetOutputsDim(name, {dim});
-}
-
 AttrReader CompileTimeInferShapeContext::Attrs() const {
   return AttrReader(op_.GetAttrMap());
 }
@@ -475,23 +461,48 @@ const std::vector<std::string> &CompileTimeInferShapeContext::Outputs(
 DDim CompileTimeInferShapeContext::GetDim(const std::string &name) const {
   auto var = block_.FindVarRecursive(name);
   PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name);
+  DDim res;
   try {
-    auto shape = var->Shape();
-    if (shape.empty()) {
-      return framework::make_ddim({0UL});
-    } else {
-      return framework::make_ddim(var->Shape());
-    }
+    auto shape = var->GetShape();
+    res = shape.empty() ? make_ddim({0UL}) : make_ddim(shape);
   } catch (...) {
     VLOG(5) << "GetDim of variable " << name << " error";
     std::rethrow_exception(std::current_exception());
   }
+  return res;
+}
+
+std::vector<DDim> CompileTimeInferShapeContext::GetRepeatedDims(
+    const std::string &name) const {
+  auto var = block_.FindVarRecursive(name);
+  PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name);
+  std::vector<DDim> res;
+  try {
+    auto shapes = var->GetShapes();
+    for (const auto &s : shapes) {
+      res.push_back(s.empty() ? make_ddim({0UL}) : make_ddim(s));
+    }
+  } catch (...) {
+    VLOG(5) << "GetRepeatedDim of variable " << name << " error.";
+    std::rethrow_exception(std::current_exception());
+  }
+  return res;
 }
 
 void CompileTimeInferShapeContext::SetDim(const std::string &name,
                                           const DDim &dim) {
-  block_.FindVarRecursive(name)->SetShape(framework::vectorize(dim));
+  block_.FindVarRecursive(name)->SetShape(vectorize(dim));
+}
+
+void CompileTimeInferShapeContext::SetRepeatedDims(
+    const std::string &name, const std::vector<DDim> &dims) {
+  auto var = block_.FindVarRecursive(name);
+  PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name);
+  std::vector<std::vector<int64_t>> dim_vec(dims.size());
+  std::transform(dims.begin(), dims.end(), dim_vec.begin(), vectorize);
+  var->SetShapes(dim_vec);
 }
+
 bool CompileTimeInferShapeContext::IsRuntime() const { return false; }
 
 proto::VarDesc::VarType CompileTimeInferShapeContext::GetVarType(

paddle/framework/operator.cc

@@ -320,8 +320,8 @@ class RuntimeInferShapeContext : public InferShapeContext {
     if (length == 0) {
       return false;
     }
-    PADDLE_ENFORCE_EQ(length, 1UL, "Input %s should have more than one inputs",
-                      name);
+    PADDLE_ENFORCE_EQ(length, 1UL,
+                      "Input %s should not have more than one inputs", name);
     auto ipt = ins[0];
     auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
     return var != nullptr;
@@ -333,8 +333,8 @@ class RuntimeInferShapeContext : public InferShapeContext {
     if (length == 0) {
       return false;
     }
-    PADDLE_ENFORCE_EQ(length, 1UL, "Output %s should have more than one inputs",
-                      name);
+    PADDLE_ENFORCE_EQ(length, 1UL,
+                      "Output %s should not have more than one inputs", name);
     auto ipt = outs[0];
     auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
     return var != nullptr;
@@ -366,14 +366,6 @@ class RuntimeInferShapeContext : public InferShapeContext {
     return true;
   }
 
-  DDim GetInputDim(const std::string& name) const override {
-    return GetDim(op_.Input(name));
-  }
-
-  void SetOutputDim(const std::string& name, const DDim& dim) override {
-    SetDim(op_.Output(name), dim);
-  }
-
   AttrReader Attrs() const override { return AttrReader(op_.Attrs()); }
 
   const std::vector<std::string>& Inputs(
@@ -429,8 +421,22 @@ class RuntimeInferShapeContext : public InferShapeContext {
     } else if (var->IsType<SelectedRows>()) {
       return var->Get<SelectedRows>().GetCompleteDims();
     } else {
-      PADDLE_THROW("Variable %s type_id %s, expect LoDTensor/SelectedRows.",
-                   name, var->Type().name());
+      PADDLE_THROW(
+          "Only LoDTensor/SelectedRows support 'GetDim', but Variable %s's "
+          "type_id is %s.",
+          name, var->Type().name());
+    }
+  }
+
+  std::vector<DDim> GetRepeatedDims(const std::string& name) const override {
+    Variable* var = scope_.FindVar(name);
+    if (var->IsType<ReaderHolder>()) {
+      return var->Get<ReaderHolder>().shapes();
+    } else {
+      PADDLE_THROW(
+          "Only ReaderHolder support 'GetRepeatedDims', but Variable %s's "
+          "type_id is %s.",
+          name, var->Type().name());
     }
   }
 
@@ -446,6 +452,19 @@ class RuntimeInferShapeContext : public InferShapeContext {
     }
   }
 
+  void SetRepeatedDims(const std::string& name,
+                       const std::vector<DDim>& dims) override {
+    Variable* var = scope_.FindVar(name);
+    if (var->IsType<ReaderHolder>()) {
+      var->GetMutable<ReaderHolder>()->set_shapes(dims);
+    } else {
+      PADDLE_THROW(
+          "Only ReaderHolder support 'SetRepeatedDims', but Variable %s's "
+          "type_id is %s.",
+          name, var->Type().name());
+    }
+  }
+
   proto::VarDesc::VarType GetVarType(const std::string& name) const override {
     auto* var = scope_.FindVar(name);
     return ToVarType(var->Type());

paddle/framework/program_desc_test.cc

@@ -53,7 +53,7 @@ TEST(ProgramDesc, copy_ctor) {
     ASSERT_NE(copy, var_before);
     ASSERT_EQ(copy->Name(), var_before->Name());
     ASSERT_EQ(copy->GetType(), var_before->GetType());
-    ASSERT_EQ(copy->Shape(), var_before->Shape());
+    ASSERT_EQ(copy->GetShape(), var_before->GetShape());
     ASSERT_EQ(copy->Proto()->SerializeAsString(),
               var_before->Proto()->SerializeAsString());
   };
@@ -117,7 +117,7 @@ TEST(ProgramDescBind, serialize_and_deserialize) {
     ASSERT_NE(restored, var_before);
     ASSERT_EQ(restored->Name(), var_before->Name());
     ASSERT_EQ(restored->GetType(), var_before->GetType());
-    ASSERT_EQ(restored->Shape(), var_before->Shape());
+    ASSERT_EQ(restored->GetShape(), var_before->GetShape());
     ASSERT_EQ(restored->Proto()->SerializeAsString(),
               var_before->Proto()->SerializeAsString());
   };

paddle/framework/reader.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/framework/reader.h"
+
+namespace paddle {
+namespace framework {
+
+DDim ReaderBase::shape(size_t idx) const {
+  PADDLE_ENFORCE_LT(
+      idx, shapes_.size(),
+      "Cannot get the %d'th shape, 'shapes_' only has %d elements.", idx,
+      shapes_.size());
+  return shapes_[idx];
+}
+
+void ShuffleReader::ReadNext(std::vector<LoDTensor>* out) {
+  if (iteration_pos_ >= buffer_.size()) {
+    // Reload buffer with new data
+    buffer_.clear();
+    buffer_.reserve(buffer_size_);
+    for (int i = 0; i < buffer_size_; ++i) {
+      if (reader_->HasNext()) {
+        buffer_.push_back(std::vector<LoDTensor>());
+        reader_->ReadNext(&buffer_.back());
+      } else {
+        break;
+      }
+    }
+    // TODO(fengjiayi): 'std::random_shuffle' can be very slow. It needs to be
+    // optimize.
+    std::random_shuffle(buffer_.begin(), buffer_.end());
+    iteration_pos_ = 0;
+  }
+  out->clear();
+  if (!buffer_.empty()) {
+    std::swap(*out, buffer_[iteration_pos_++]);
+  }
+  // if buffer_ is empty, the 'out' will return as an empty vector.
+}
+
+void BatchReader::ReadNext(std::vector<LoDTensor>* out) {
+  buffer_.clear();
+  buffer_.reserve(batch_size_);
+  for (int i = 0; i < batch_size_; ++i) {
+    if (reader_->HasNext()) {
+      buffer_.push_back(std::vector<LoDTensor>());
+      reader_->ReadNext(&buffer_.back());
+    } else {
+      break;
+    }
+  }
+  // Concat instances
+  out->clear();
+  if (buffer_.empty()) {
+    // if buffer_ is empty, the 'out' will return as an empty vector.
+    return;
+  }
+  int out_num = buffer_[0].size();
+  out->reserve(out_num);
+  for (int j = 0; j < out_num; ++j) {
+    // Merge shape and check date type
+    std::type_index batch_type = buffer_[0][j].type();
+    DDim batch_shape = buffer_[0][j].dims();
+    for (size_t i = 1; i < buffer_.size(); ++i) {
+      std::type_index ins_type = buffer_[i][j].type();
+      DDim ins_shape = buffer_[i][j].dims();
+      PADDLE_ENFORCE_EQ(batch_type, ins_type);
+      PADDLE_ENFORCE_EQ(slice_ddim(batch_shape, 1, batch_shape.size()),
+                        slice_ddim(ins_shape, 1, ins_shape.size()));
+      PADDLE_ENFORCE_GT(ins_shape[0], 0);
+      batch_shape[0] += ins_shape[0];
+    }
+
+    LoDTensor out_tensor;
+    out_tensor.Resize(batch_shape);
+    out_tensor.mutable_data(platform::CPUPlace(), batch_type);
+    int64_t dst_offset = 0;
+
+    // Merge lod and data
+    LoD batch_lod;
+    std::vector<size_t> top_level_lod({0});
+    for (size_t i = 0; i < buffer_.size(); ++i) {
+      DDim ins_shape = buffer_[i][j].dims();
+      LoD ins_lod = buffer_[i][j].lod();
+      if (i == 0) {
+        batch_lod = ins_lod;
+      } else {
+        PADDLE_ENFORCE_EQ(batch_lod.size(), ins_lod.size());
+        for (size_t level_idx = 0; level_idx < batch_lod.size(); ++level_idx) {
+          auto& lod_level = batch_lod[level_idx];
+          for (size_t k = 1; k < ins_lod[level_idx].size(); ++k) {
+            lod_level.push_back(ins_lod[level_idx][k] + lod_level.back());
+          }
+        }
+      }
+      top_level_lod.push_back(
+          top_level_lod.back() +
+          (ins_lod.empty() ? ins_shape[0] : (ins_lod[0].size() - 1)));
+
+      Tensor dst = out_tensor.Slice(dst_offset, dst_offset + ins_shape[0]);
+      Copy(buffer_[i][j], platform::CPUPlace(), &dst);
+      dst_offset += ins_shape[0];
+    }
+    batch_lod.insert(batch_lod.begin(), top_level_lod);
+    out_tensor.set_lod(batch_lod);
+    out->push_back(out_tensor);
+  }
+}
+}  // namespace framework
+}  // namespace paddle

paddle/framework/reader.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/framework/ddim.h"
+#include "paddle/framework/lod_tensor_array.h"
+
+namespace paddle {
+namespace framework {
+
+class ReaderBase {
+ public:
+  explicit ReaderBase(const std::vector<DDim>& shapes) : shapes_(shapes) {
+    PADDLE_ENFORCE(!shapes_.empty());
+  }
+  virtual void ReadNext(std::vector<LoDTensor>* out) = 0;
+  virtual bool HasNext() const = 0;
+
+  virtual void ReInit() = 0;
+
+  DDim shape(size_t idx) const;
+  std::vector<DDim> shapes() const { return shapes_; }
+  void set_shapes(const std::vector<DDim>& shapes) { shapes_ = shapes; }
+
+  virtual ~ReaderBase() {}
+
+ protected:
+  std::vector<DDim> shapes_;
+};
+
+class FileReader : public ReaderBase {
+ public:
+  explicit FileReader(const std::vector<DDim>& shapes) : ReaderBase(shapes) {}
+};
+
+class DecoratedReader : public ReaderBase {
+ public:
+  explicit DecoratedReader(ReaderBase* reader)
+      : ReaderBase(reader->shapes()), reader_(reader) {
+    PADDLE_ENFORCE_NOT_NULL(reader_);
+  }
+
+  bool HasNext() const override { return reader_->HasNext(); }
+
+  void ReInit() override { reader_->ReInit(); }
+
+ protected:
+  ReaderBase* reader_;
+};
+
+// file readers
+
+template <typename T>
+class RandomDataGenerator : public FileReader {
+ public:
+  RandomDataGenerator(const std::vector<DDim>& shapes, float min, float max)
+      : FileReader(shapes), min_(min), max_(max) {
+    PADDLE_ENFORCE_LE(
+        min, max, "'min' shouldn't be greater than 'max'.(%f vs %f)", min, max);
+    unsigned int seed = std::random_device()();
+    engine_.seed(seed);
+    dist_ = std::uniform_real_distribution<float>(min_, max_);
+  }
+
+  void ReadNext(std::vector<LoDTensor>* out) override {
+    out->clear();
+    out->reserve(shapes_.size());
+    for (const DDim& shape : shapes_) {
+      PADDLE_ENFORCE_GE(
+          shape.size(), 2,
+          "The rank of reader's output data should be 2 at least.(Now it's %d)",
+          shape.size());
+      LoDTensor out_tensor;
+      out_tensor.Resize(shape);
+      T* data = out_tensor.mutable_data<T>(platform::CPUPlace());
+      int64_t numel = product(shape);
+      for (int64_t i = 0; i < numel; ++i) {
+        data[i] = dist_(engine_);
+      }
+      out->push_back(out_tensor);
+    }
+  }
+
+  bool HasNext() const override { return true; }
+
+  void ReInit() override { return; }
+
+ private:
+  float min_;
+  float max_;
+  std::minstd_rand engine_;
+  std::uniform_real_distribution<float> dist_;
+};
+
+// decorated readers
+
+class ShuffleReader : public DecoratedReader {
+ public:
+  ShuffleReader(ReaderBase* reader, int buffer_size)
+      : DecoratedReader(reader), buffer_size_(buffer_size), iteration_pos_(0) {
+    buffer_.reserve(buffer_size);
+  }
+
+  void ReadNext(std::vector<LoDTensor>* out) override;
+
+ private:
+  int buffer_size_;
+  std::vector<std::vector<LoDTensor>> buffer_;
+  size_t iteration_pos_;
+};
+
+class BatchReader : public DecoratedReader {
+ public:
+  BatchReader(ReaderBase* reader, int batch_size)
+      : DecoratedReader(reader), batch_size_(batch_size) {
+    buffer_.reserve(batch_size_);
+  }
+
+  void ReadNext(std::vector<LoDTensor>* out) override;
+
+ private:
+  int batch_size_;
+  std::vector<std::vector<LoDTensor>> buffer_;
+};
+
+// The ReaderHolder is used as readers' unified wrapper,
+// making it easier to access different type readers in Variables.
+class ReaderHolder {
+ public:
+  void Reset(ReaderBase* reader) { reader_.reset(reader); }
+
+  ReaderBase* Get() const { return reader_.get(); }
+
+  void ReadNext(std::vector<LoDTensor>* out) { reader_->ReadNext(out); }
+  bool HasNext() const { return reader_->HasNext(); }
+  void ReInit() { reader_->ReInit(); }
+
+  DDim shape(size_t idx) const { return reader_->shape(idx); }
+  std::vector<DDim> shapes() const { return reader_->shapes(); }
+  void set_shapes(const std::vector<DDim>& shapes) {
+    reader_->set_shapes(shapes);
+  }
+
+ private:
+  std::unique_ptr<ReaderBase> reader_;
+};
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/shape_inference.cc

@@ -18,10 +18,28 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-std::vector<framework::DDim> InferShapeContext::GetInputsDim(
+DDim InferShapeContext::GetInputDim(const std::string &name) const {
+  const std::vector<std::string> &arg_names = Inputs(name);
+  PADDLE_ENFORCE_EQ(arg_names.size(), 1UL,
+                    "Input(%s) should hold one element, but now it holds %d",
+                    name, arg_names.size());
+  return this->GetDim(arg_names[0]);
+}
+
+std::vector<DDim> InferShapeContext::GetInputsDim(
     const std::string &name) const {
-  const std::vector<std::string> &names = Inputs(name);
-  return GetDims(names);
+  const std::vector<std::string> &arg_names = Inputs(name);
+  return GetDims(arg_names);
+}
+
+std::vector<DDim> InferShapeContext::GetReaderDims(
+    const std::string &name) const {
+  const std::vector<std::string> &arg_names = Inputs(name);
+  PADDLE_ENFORCE_EQ(
+      arg_names.size(), 1UL,
+      "Reader input '%s' should hold one element, but now it holds %d", name,
+      arg_names.size());
+  return this->GetRepeatedDims(arg_names[0]);
 }
 
 DDim InferShapeContext::GetInputsElementDim(const std::string &name,
@@ -30,15 +48,33 @@ DDim InferShapeContext::GetInputsElementDim(const std::string &name,
   return this->GetDim(names[idx]);
 }
 
-void InferShapeContext::SetOutputsDim(
-    const std::string &name, const std::vector<framework::DDim> &dims) {
+void InferShapeContext::SetOutputDim(const std::string &name, const DDim &dim) {
+  auto &arg_names = Outputs(name);
+  PADDLE_ENFORCE_EQ(arg_names.size(), 1UL,
+                    "Output(%s) should hold one element, but now it holds %d",
+                    name, arg_names.size());
+  SetDim(arg_names[0], dim);
+}
+
+void InferShapeContext::SetOutputsDim(const std::string &name,
+                                      const std::vector<DDim> &dims) {
   auto &names = Outputs(name);
   SetDims(names, dims);
 }
 
-std::vector<framework::DDim> InferShapeContext::GetDims(
+void InferShapeContext::SetReaderDims(const std::string &name,
+                                      const std::vector<DDim> &dims) {
+  const std::vector<std::string> &arg_names = Outputs(name);
+  PADDLE_ENFORCE_EQ(
+      arg_names.size(), 1UL,
+      "Reader output '%s' should hold one element, but now it holds %d", name,
+      arg_names.size());
+  return this->SetRepeatedDims(arg_names[0], dims);
+}
+
+std::vector<DDim> InferShapeContext::GetDims(
     const std::vector<std::string> &names) const {
-  std::vector<framework::DDim> ret;
+  std::vector<DDim> ret;
   ret.reserve(names.size());
   std::transform(
       names.begin(), names.end(), std::back_inserter(ret),
@@ -47,7 +83,7 @@ std::vector<framework::DDim> InferShapeContext::GetDims(
 }
 
 void InferShapeContext::SetDims(const std::vector<std::string> &names,
-                                const std::vector<framework::DDim> &dims) {
+                                const std::vector<DDim> &dims) {
   size_t length = names.size();
   PADDLE_ENFORCE_EQ(length, dims.size());
   for (size_t i = 0; i < length; ++i) {
@@ -57,14 +93,17 @@ void InferShapeContext::SetDims(const std::vector<std::string> &names,
     SetDim(names[i], dims[i]);
   }
 }
+
 std::vector<proto::VarDesc::VarType> InferShapeContext::GetInputsVarType(
     const std::string &name) const {
   return GetVarTypes(Inputs(name));
 }
+
 std::vector<proto::VarDesc::VarType> InferShapeContext::GetOutputsVarType(
     const std::string &name) const {
   return GetVarTypes(Outputs(name));
 }
+
 std::vector<proto::VarDesc::VarType> InferShapeContext::GetVarTypes(
     const std::vector<std::string> &names) const {
   std::vector<proto::VarDesc::VarType> retv;

paddle/framework/shape_inference.h

@@ -35,14 +35,14 @@ class InferShapeContext {
   virtual bool HasInputs(const std::string &name) const = 0;
   virtual bool HasOutputs(const std::string &name) const = 0;
 
-  virtual framework::DDim GetInputDim(const std::string &name) const = 0;
-
-  std::vector<framework::DDim> GetInputsDim(const std::string &name) const;
+  DDim GetInputDim(const std::string &name) const;
+  std::vector<DDim> GetInputsDim(const std::string &name) const;
+  std::vector<DDim> GetReaderDims(const std::string &name) const;
   DDim GetInputsElementDim(const std::string &name, int idx) const;
 
-  virtual void SetOutputDim(const std::string &name, const DDim &dim) = 0;
-  void SetOutputsDim(const std::string &name,
-                     const std::vector<framework::DDim> &dims);
+  void SetOutputDim(const std::string &name, const DDim &dim);
+  void SetOutputsDim(const std::string &name, const std::vector<DDim> &dims);
+  void SetReaderDims(const std::string &name, const std::vector<DDim> &dims);
 
   virtual AttrReader Attrs() const = 0;
   virtual const std::vector<std::string> &Inputs(
@@ -57,15 +57,16 @@ class InferShapeContext {
 
   // Note: In while op, we need this to be public
   void SetDims(const std::vector<std::string> &names,
-               const std::vector<framework::DDim> &dims);
+               const std::vector<DDim> &dims);
 
  protected:
-  virtual framework::DDim GetDim(const std::string &name) const = 0;
-  virtual void SetDim(const std::string &name, const framework::DDim &dim) = 0;
-
-  std::vector<framework::DDim> GetDims(
-      const std::vector<std::string> &names) const;
+  virtual DDim GetDim(const std::string &name) const = 0;
+  virtual void SetDim(const std::string &name, const DDim &dim) = 0;
+  virtual std::vector<DDim> GetRepeatedDims(const std::string &name) const = 0;
+  virtual void SetRepeatedDims(const std::string &name,
+                               const std::vector<DDim> &dims) = 0;
 
+  std::vector<DDim> GetDims(const std::vector<std::string> &names) const;
   std::vector<proto::VarDesc::VarType> GetVarTypes(
       const std::vector<std::string> &names) const;
 

paddle/framework/threadpool.h

@@ -21,7 +21,8 @@ limitations under the License. */
 #include <queue>
 #include <thread>
 #include <vector>
-
+#include "glog/logging.h"
+#include "paddle/platform/enforce.h"
 #include "paddle/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
 
 namespace paddle {
@@ -31,7 +32,7 @@ namespace framework {
 // number of threads.
 class ThreadPool {
  public:
-  typedef std::packaged_task<void()> Task;
+  using Task = std::packaged_task<std::unique_ptr<platform::EnforceNotMet>()>;
 
   // Returns the singleton of ThreadPool.
   static ThreadPool* GetInstance();
@@ -52,9 +53,28 @@ class ThreadPool {
   // std::future::wait().
   template <typename Callback>
   std::future<void> Run(Callback fn) {
+    auto f = this->RunAndGetException(fn);
+    return std::async(std::launch::deferred, ExceptionHandler(std::move(f)));
+  }
+
+  template <typename Callback>
+  std::future<std::unique_ptr<platform::EnforceNotMet>> RunAndGetException(
+      Callback fn) {
     std::unique_lock<std::mutex> lock(mutex_);
-    Task task(std::bind(fn));
-    std::future<void> f = task.get_future();
+    Task task([fn]() -> std::unique_ptr<platform::EnforceNotMet> {
+      try {
+        fn();
+        return nullptr;
+      } catch (platform::EnforceNotMet ex) {
+        return std::unique_ptr<platform::EnforceNotMet>(
+            new platform::EnforceNotMet(ex));
+      } catch (...) {
+        LOG(FATAL)
+            << "Unexpected exception is catched in thread pool. All "
+               "throwable exception in Fluid should be an EnforceNotMet.";
+      }
+    });
+    std::future<std::unique_ptr<platform::EnforceNotMet>> f = task.get_future();
     tasks_.push(std::move(task));
     lock.unlock();
     scheduled_.notify_one();
@@ -65,6 +85,22 @@ class ThreadPool {
   void Wait();
 
  private:
+  struct ExceptionHandler {
+    mutable std::future<std::unique_ptr<platform::EnforceNotMet>> future_;
+    explicit ExceptionHandler(
+        std::future<std::unique_ptr<platform::EnforceNotMet>>&& f)
+        : future_(std::move(f)) {}
+    void operator()() const {
+      auto ex = this->future_.get();
+      if (ex != nullptr) {
+        LOG(FATAL) << "The exception is thrown inside the thread pool. You "
+                      "should use RunAndGetException to handle the exception.\n"
+                      "The default exception handler is LOG(FATAL)."
+                   << ex->what();
+      }
+    }
+  };
+
   DISABLE_COPY_AND_ASSIGN(ThreadPool);
 
   explicit ThreadPool(int num_threads);

paddle/framework/var_desc.cc

@@ -26,18 +26,98 @@ void VarDesc::SetShape(const std::vector<int64_t> &dims) {
   VectorToRepeated(dims, mutable_tensor_desc()->mutable_dims());
 }
 
+void VarDesc::SetTensorDescNum(size_t num) {
+  switch (desc_.type()) {
+    case proto::VarDesc::READER: {
+      auto *lod_tensors_ptr = desc_.mutable_reader()->mutable_lod_tensor();
+      lod_tensors_ptr->Clear();
+      for (size_t i = 0; i < num; ++i) {
+        lod_tensors_ptr->Add();
+      }
+      return;
+    } break;
+    default:
+      PADDLE_THROW(
+          "Setting 'sub_tensor_number' is not supported by the type of var %s.",
+          this->Name());
+  }
+}
+
+size_t VarDesc::GetTensorDescNum() const {
+  switch (desc_.type()) {
+    case proto::VarDesc::READER:
+      return desc_.reader().lod_tensor_size();
+      break;
+    default:
+      PADDLE_THROW(
+          "Getting 'sub_tensor_number' is not supported by the type of var %s.",
+          this->Name());
+  }
+}
+
+void VarDesc::SetShapes(
+    const std::vector<std::vector<int64_t>> &multiple_dims) {
+  if (multiple_dims.size() != GetTensorDescNum()) {
+    VLOG(3) << "WARNING: The number of given shapes(" << multiple_dims.size()
+            << ") doesn't match the existing tensor number("
+            << GetTensorDescNum()
+            << "). The Reader is going to be reinitialized.";
+    SetTensorDescNum(multiple_dims.size());
+  }
+  std::vector<proto::TensorDesc *> tensors = mutable_tensor_descs();
+  for (size_t i = 0; i < multiple_dims.size(); ++i) {
+    VectorToRepeated(multiple_dims[i], tensors[i]->mutable_dims());
+  }
+}
+
+std::vector<int64_t> VarDesc::GetShape() const {
+  return RepeatedToVector(tensor_desc().dims());
+}
+
+std::vector<std::vector<int64_t>> VarDesc::GetShapes() const {
+  std::vector<proto::TensorDesc> descs = tensor_descs();
+  std::vector<std::vector<int64_t>> res;
+  res.reserve(descs.size());
+  for (const auto &tensor_desc : descs) {
+    res.push_back(RepeatedToVector(tensor_desc.dims()));
+  }
+  return res;
+}
+
 void VarDesc::SetDataType(proto::DataType data_type) {
   mutable_tensor_desc()->set_data_type(data_type);
 }
 
-std::vector<int64_t> VarDesc::Shape() const {
-  return RepeatedToVector(tensor_desc().dims());
+void VarDesc::SetDataTypes(
+    const std::vector<proto::DataType> &multiple_data_type) {
+  if (multiple_data_type.size() != GetTensorDescNum()) {
+    VLOG(3) << "WARNING: The number of given data types("
+            << multiple_data_type.size()
+            << ") doesn't match the existing tensor number("
+            << GetTensorDescNum()
+            << "). The Reader is going to be reinitialized.";
+    SetTensorDescNum(multiple_data_type.size());
+  }
+  std::vector<proto::TensorDesc *> tensor_descs = mutable_tensor_descs();
+  for (size_t i = 0; i < multiple_data_type.size(); ++i) {
+    tensor_descs[i]->set_data_type(multiple_data_type[i]);
+  }
 }
 
 proto::DataType VarDesc::GetDataType() const {
   return tensor_desc().data_type();
 }
 
+std::vector<proto::DataType> VarDesc::GetDataTypes() const {
+  std::vector<proto::TensorDesc> descs = tensor_descs();
+  std::vector<proto::DataType> res;
+  res.reserve(descs.size());
+  for (const auto &tensor_desc : descs) {
+    res.push_back(tensor_desc.data_type());
+  }
+  return res;
+}
+
 void VarDesc::SetLoDLevel(int32_t lod_level) {
   switch (desc_.type()) {
     case proto::VarDesc::LOD_TENSOR:
@@ -47,8 +127,32 @@ void VarDesc::SetLoDLevel(int32_t lod_level) {
       desc_.mutable_tensor_array()->set_lod_level(lod_level);
       break;
     default:
-      PADDLE_THROW("Tensor type=%d does not support LoDLevel",
-                   desc_.tensor_array().lod_level());
+      PADDLE_THROW(
+          "Setting 'lod_level' is not supported by the type of var %s.",
+          this->Name());
+  }
+}
+
+void VarDesc::SetLoDLevels(const std::vector<int32_t> &multiple_lod_level) {
+  if (multiple_lod_level.size() != GetTensorDescNum()) {
+    VLOG(3) << "WARNING: The number of given lod_levels("
+            << multiple_lod_level.size()
+            << ") doesn't match the existing tensor number("
+            << GetTensorDescNum()
+            << "). The Reader is going to be reinitialized.";
+    SetTensorDescNum(multiple_lod_level.size());
+  }
+  switch (desc_.type()) {
+    case proto::VarDesc::READER: {
+      size_t i = 0;
+      for (auto &lod_tensor : *desc_.mutable_reader()->mutable_lod_tensor()) {
+        lod_tensor.set_lod_level(multiple_lod_level[i++]);
+      }
+    } break;
+    default:
+      PADDLE_THROW(
+          "Setting 'lod_levels' is not supported by the type of var %s.",
+          this->Name());
   }
 }
 
@@ -59,13 +163,31 @@ int32_t VarDesc::GetLoDLevel() const {
     case proto::VarDesc::LOD_TENSOR_ARRAY:
       return desc_.tensor_array().lod_level();
     default:
-      PADDLE_THROW("Tensor type=%d does not support LoDLevel",
-                   desc_.tensor_array().lod_level());
+      PADDLE_THROW(
+          "Getting 'lod_level' is not supported by the type of var %s.",
+          this->Name());
+  }
+}
+
+std::vector<int32_t> VarDesc::GetLoDLevels() const {
+  std::vector<int32_t> res;
+  switch (desc_.type()) {
+    case proto::VarDesc::READER:
+      res.reserve(desc_.reader().lod_tensor_size());
+      for (auto &lod_tensor : desc_.reader().lod_tensor()) {
+        res.push_back(lod_tensor.lod_level());
+      }
+      return res;
+      break;
+    default:
+      PADDLE_THROW(
+          "Getting 'lod_levels' is not supported by the type of var %s.",
+          this->Name());
   }
 }
 
 const proto::TensorDesc &VarDesc::tensor_desc() const {
-  PADDLE_ENFORCE(desc_.has_type(), "invoke TensorDesc must after set type");
+  PADDLE_ENFORCE(desc_.has_type(), "The var's type hasn't been set.");
   switch (desc_.type()) {
     case proto::VarDesc::SELECTED_ROWS:
       return desc_.selected_rows();
@@ -74,13 +196,32 @@ const proto::TensorDesc &VarDesc::tensor_desc() const {
     case proto::VarDesc::LOD_TENSOR_ARRAY:
       return desc_.tensor_array().tensor();
     default:
-      PADDLE_THROW("The type of var %s is unsupported.", this->Name());
+      PADDLE_THROW(
+          "Getting 'tensor_desc' is not supported by the type of var %s.",
+          this->Name());
+  }
+}
+
+std::vector<proto::TensorDesc> VarDesc::tensor_descs() const {
+  PADDLE_ENFORCE(desc_.has_type(), "The var type hasn't been set.");
+  std::vector<proto::TensorDesc> res;
+  res.reserve(GetTensorDescNum());
+  switch (desc_.type()) {
+    case proto::VarDesc::READER:
+      for (const auto &lod_tensor : desc_.reader().lod_tensor()) {
+        res.push_back(lod_tensor.tensor());
+      }
+      return res;
+    default:
+      PADDLE_THROW(
+          "Getting 'tensor_descs' is not supported by the type of var "
+          "%s.",
+          this->Name());
   }
 }
 
 proto::TensorDesc *VarDesc::mutable_tensor_desc() {
-  PADDLE_ENFORCE(desc_.has_type(),
-                 "invoke MutableTensorDesc must after set type");
+  PADDLE_ENFORCE(desc_.has_type(), "The var type hasn't been set.");
   switch (desc_.type()) {
     case proto::VarDesc::SELECTED_ROWS:
       return desc_.mutable_selected_rows();
@@ -89,8 +230,30 @@ proto::TensorDesc *VarDesc::mutable_tensor_desc() {
     case proto::VarDesc::LOD_TENSOR_ARRAY:
       return desc_.mutable_tensor_array()->mutable_tensor();
     default:
-      PADDLE_THROW("Unexpected branch.");
+      PADDLE_THROW(
+          "Getting 'mutable_tensor_desc' is not supported by the type of var "
+          "%s.",
+          this->Name());
   }
 }
+
+std::vector<proto::TensorDesc *> VarDesc::mutable_tensor_descs() {
+  PADDLE_ENFORCE(desc_.has_type(), "The var type hasn't been set.");
+  std::vector<proto::TensorDesc *> res;
+  res.reserve(GetTensorDescNum());
+  switch (desc_.type()) {
+    case proto::VarDesc::READER:
+      for (auto &lod_tensor : *desc_.mutable_reader()->mutable_lod_tensor()) {
+        res.push_back(lod_tensor.mutable_tensor());
+      }
+      return res;
+    default:
+      PADDLE_THROW(
+          "Getting 'tensor_descs' is not supported by the type of var "
+          "%s.",
+          this->Name());
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/var_desc.h

@@ -68,18 +68,34 @@ class VarDesc {
 
   void SetName(std::string name) { desc_.set_name(name); }
 
+  void SetTensorDescNum(size_t num);
+
+  size_t GetTensorDescNum() const;
+
   void SetShape(const std::vector<int64_t> &dims);
 
+  void SetShapes(const std::vector<std::vector<int64_t>> &multiple_dims);
+
+  std::vector<int64_t> GetShape() const;
+
+  std::vector<std::vector<int64_t>> GetShapes() const;
+
   void SetDataType(proto::DataType data_type);
 
-  std::vector<int64_t> Shape() const;
+  void SetDataTypes(const std::vector<proto::DataType> &multiple_data_type);
 
   proto::DataType GetDataType() const;
 
+  std::vector<proto::DataType> GetDataTypes() const;
+
   void SetLoDLevel(int32_t lod_level);
 
+  void SetLoDLevels(const std::vector<int32_t> &multiple_lod_level);
+
   int32_t GetLoDLevel() const;
 
+  std::vector<int32_t> GetLoDLevels() const;
+
   proto::VarDesc::VarType GetType() const;
 
   void SetType(proto::VarDesc::VarType type);
@@ -90,7 +106,9 @@ class VarDesc {
 
  private:
   const proto::TensorDesc &tensor_desc() const;
+  std::vector<proto::TensorDesc> tensor_descs() const;
   proto::TensorDesc *mutable_tensor_desc();
+  std::vector<proto::TensorDesc *> mutable_tensor_descs();
 
   proto::VarDesc desc_;
 };

paddle/framework/var_type.h

@@ -17,6 +17,7 @@ limitations under the License. */
 #include "paddle/framework/lod_rank_table.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/lod_tensor_array.h"
+#include "paddle/framework/reader.h"
 #include "paddle/framework/selected_rows.h"
 #include "paddle/framework/variable.h"
 
@@ -31,6 +32,8 @@ inline proto::VarDesc::VarType ToVarType(std::type_index type) {
     return proto::VarDesc_VarType_LOD_TENSOR_ARRAY;
   } else if (type.hash_code() == typeid(SelectedRows).hash_code()) {
     return proto::VarDesc_VarType_SELECTED_ROWS;
+  } else if (type.hash_code() == typeid(ReaderHolder).hash_code()) {
+    return proto::VarDesc_VarType_READER;
   } else {
     PADDLE_THROW("ToVarType:Unsupported type %s", type.name());
   }
@@ -40,7 +43,7 @@ template <typename Visitor>
 inline void VisitVarType(const framework::Variable& var, Visitor visitor) {
   switch (ToVarType(var.Type())) {
     case proto::VarDesc_VarType_LOD_TENSOR:
-      visitor(var.Get<framework::LoDTensor>());
+      visitor(var.Get<LoDTensor>());
       return;
     case proto::VarDesc_VarType_LOD_RANK_TABLE:
       visitor(var.Get<LoDRankTable>());
@@ -51,6 +54,9 @@ inline void VisitVarType(const framework::Variable& var, Visitor visitor) {
     case proto::VarDesc_VarType_SELECTED_ROWS:
       visitor(var.Get<SelectedRows>());
       return;
+    case proto::VarDesc_VarType_READER:
+      visitor(var.Get<ReaderHolder>());
+      return;
     default:
       PADDLE_THROW("Not supported visit type, %d", ToVarType(var.Type()));
   }

paddle/function/GemmConvOp.cpp

@@ -178,19 +178,22 @@ public:
     real* inputData = inputs[0].data<real>();
     real* filterData = inputs[1].data<real>();
     real* outputData = outputs[0].data<real>();
+    real* colData = NULL;
     bool needIm2col = isNeedIm2col(filter);
 
     TensorShape imShape =
         TensorShape({inputChannels / groups_, inputHeight, inputWidth});
-
     TensorShape colShape;
-    real* colData = NULL;
 
-    size_t colHeight = inputChannels / groups_ * filterHeight * filterWidth;
-    size_t colWidth = outputHeight * outputWidth;
-    // Max col matrix height 256, Max col matrix width 1024
-    size_t stepColHeight = std::min(colHeight, static_cast<size_t>(256));
-    size_t stepColWidth = std::min(colWidth, static_cast<size_t>(2048));
+    // Max col matrix width 4096, Max col matrix size 4M.
+    size_t outputHeightSteps =
+        std::min(std::max(4096 / outputWidth, (size_t)1), outputHeight);
+    size_t maxColWidth = outputHeightSteps * outputWidth;
+    size_t channelSteps =
+        std::min(std::max((1048576 / maxColWidth) / filterHeight * filterWidth,
+                          (size_t)1),
+                 inputChannels / groups_);
+    size_t maxColHeight = channelSteps * filterHeight * filterWidth;
 
     if (needIm2col) {
       colShape = TensorShape({inputChannels / groups_,
@@ -199,7 +202,7 @@ public:
                               outputHeight,
                               outputWidth});
 
-      resizeBuffer<Device>(stepColHeight * stepColWidth * sizeof(real));
+      resizeBuffer<Device>(maxColHeight * maxColWidth * sizeof(real));
       colData = reinterpret_cast<real*>(memory_->getBuf());
     }
 
@@ -209,20 +212,24 @@ public:
         (outputChannels / groups_) * outputHeight * outputWidth;
     size_t filterOffset = filter.getElements() / groups_;
 
-    int nStride = colWidth;
-    int kStride = colHeight;
+    int nStride = outputHeight * outputWidth;
+    int kStride = inputChannels / groups_ * filterHeight * filterWidth;
     for (size_t i = 0; i < batchSize; i++) {
+      filterData = inputs[1].data<real>();
       for (size_t g = 0; g < groups_; g++) {
         if (needIm2col) {
           real beta_ = beta;
-          for (size_t colHeightStart = 0; colHeightStart < colHeight;
-               colHeightStart += stepColHeight) {
-            for (size_t colWidthStart = 0; colWidthStart < colWidth;
-                 colWidthStart += stepColWidth) {
-              int N = std::min(colWidth - colWidthStart, stepColWidth);
-              int K = std::min(colHeight - colHeightStart, stepColHeight);
+          for (size_t ic = 0; ic < inputChannels / groups_;
+               ic += channelSteps) {
+            int channels = std::min(inputChannels / groups_ - ic, channelSteps);
+            for (size_t oh = 0; oh < outputHeight; oh += outputHeightSteps) {
+              int height = std::min(outputHeight - oh, outputHeightSteps);
+
+              int M = outputChannels / groups_;
+              int N = height * outputWidth;
+              int K = channels * filterHeight * filterWidth;
               // im2col
-              im2col(inputData + g * inputOffset,
+              im2col(inputData,
                      imShape,
                      colData,
                      colShape,
@@ -232,13 +239,12 @@ public:
                      paddingW(),
                      dilationH(),
                      dilationW(),
-                     colHeightStart,
-                     K,
-                     colWidthStart,
+                     channels,
+                     oh,
+                     height,
                      N);
 
               // gemm
-              int M = outputChannels / groups_;
               BlasGemm<Device, real>::compute(
                   false,
                   false,
@@ -246,12 +252,12 @@ public:
                   N,
                   K,
                   1.0f,
-                  filterData + g * filterOffset + colHeightStart,
+                  filterData + ic * filterHeight * filterWidth,
                   kStride,
                   colData,
                   N,
                   beta_,
-                  outputData + g * outputOffset + colWidthStart,
+                  outputData + oh * outputWidth,
                   nStride);
             }
             beta_ = 1.0;
@@ -266,17 +272,18 @@ public:
                                           N,
                                           K,
                                           1.0f,
-                                          filterData + g * filterOffset,
+                                          filterData,
                                           K,
-                                          inputData + g * inputOffset,
+                                          inputData,
                                           N,
                                           beta,
-                                          outputData + g * outputOffset,
+                                          outputData,
                                           N);
         }
+        inputData += inputOffset;
+        outputData += outputOffset;
+        filterData += filterOffset;
       }
-      inputData += inputChannels * inputHeight * inputWidth;
-      outputData += outputChannels * outputHeight * outputWidth;
     }
 
     memory_.reset();

paddle/function/Im2Col.h

@@ -111,39 +111,42 @@ public:
                   int paddingWidth,
                   int dilationHeight,
                   int dilationWidth,
-                  int colHeightStart,
-                  int colHeightSize,
-                  int colWidthStart,
-                  int colWidthSize) {
+                  int inputChannels,
+                  int colOffset,
+                  int colOutputHeight,
+                  int colWidth) {
     int inputHeight = imShape[1];
     int inputWidth = imShape[2];
     int filterHeight = colShape[1];
     int filterWidth = colShape[2];
     int outputWidth = colShape[4];
 
-    for (int colh = 0; colh < colHeightSize; colh++) {
-      int wOffset = (colHeightStart + colh) % filterWidth;
-      int hOffset = ((colHeightStart + colh) / filterWidth) % filterHeight;
-      int c_im = (colHeightStart + colh) / filterWidth / filterHeight;
-
-      for (int colw = 0; colw < colWidthSize; colw++) {
-        int h = (colWidthStart + colw) / outputWidth;
-        int w = (colWidthStart + colw) % outputWidth;
-
-        int imRowIdx = h * strideHeight + hOffset * dilationHeight;
-        int imColIdx = w * strideWidth + wOffset * dilationWidth;
-        if ((imRowIdx - paddingHeight) < 0 ||
-            (imRowIdx - paddingHeight) >= inputHeight ||
-            (imColIdx - paddingWidth) < 0 ||
-            (imColIdx - paddingWidth) >= inputWidth) {
-          colData[colh * colWidthSize + colw] = static_cast<T>(0);
-        } else {
-          imRowIdx += c_im * inputHeight - paddingHeight;
-          imColIdx -= paddingWidth;
-          colData[colh * colWidthSize + colw] =
-              imData[imRowIdx * inputWidth + imColIdx];
+    for (int ic = 0; ic < inputChannels; ic++) {
+      for (int oh = 0; oh < colOutputHeight; oh++) {
+        T* dstData = colData + oh * outputWidth;
+        for (int fh = 0; fh < filterHeight; fh++) {
+          for (int fw = 0; fw < filterWidth; fw++) {
+            int imRowIdx = (oh + colOffset) * strideHeight +
+                           fh * dilationHeight - paddingHeight;
+            if (imRowIdx < 0 || imRowIdx >= inputHeight) {
+              memset(dstData, 0, outputWidth * sizeof(T));
+            } else {
+              for (int ow = 0; ow < outputWidth; ow++) {
+                int imColIdx =
+                    ow * strideWidth + fw * dilationWidth - paddingWidth;
+                if (imColIdx < 0 || imColIdx >= inputWidth) {
+                  dstData[ow] = T(0);
+                } else {
+                  dstData[ow] = imData[imRowIdx * inputWidth + imColIdx];
+                }
+              }
+            }
+            dstData += colWidth;
+          }
         }
       }
+      colData += filterHeight * filterWidth * colWidth;
+      imData += inputHeight * inputWidth;
     }
   }
 };

paddle/function/Im2ColTest.cpp

@@ -202,10 +202,10 @@ void TestIm2ColMobileFunctor() {
                           padding,
                           dilation,
                           dilation,
+                          channels,
                           0,
-                          height,
-                          0,
-                          width);
+                          outputHeight,
+                          outputHeight * outputWidth);
 
                   autotest::TensorCheckEqual(*output1, *output2);
                 }

paddle/inference/CMakeLists.txt

@@ -13,17 +13,11 @@ add_library(paddle_fluid_shared SHARED io.cc)
 target_circle_link_libraries(paddle_fluid_shared
   ARCHIVE_START
   ${GLOB_OP_LIB}
-  ARCHIVE_END
-  ${FLUID_CORE_MODULES})
+  ${FLUID_CORE_MODULES}
+  ARCHIVE_END)
 
 SET_TARGET_PROPERTIES(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
 
-# install library & headers
-if(NOT WITH_C_API AND WITH_FLUID)
-  install(FILES io.h DESTINATION include/paddle/inference)
-  install(TARGETS paddle_fluid_shared DESTINATION lib)
-endif()
-
 if(WITH_TESTING)
   add_subdirectory(tests/book)
 endif()

paddle/inference/io.cc

@@ -55,7 +55,7 @@ void LoadPersistables(framework::Executor& executor,
       VLOG(3) << "parameter's name: " << var->Name();
 
       framework::VarDesc* new_var = load_block->Var(var->Name());
-      new_var->SetShape(var->Shape());
+      new_var->SetShape(var->GetShape());
       new_var->SetDataType(var->GetDataType());
       new_var->SetType(var->GetType());
       new_var->SetLoDLevel(var->GetLoDLevel());

paddle/inference/tests/book/CMakeLists.txt

@@ -3,5 +3,29 @@ cc_test(test_inference_recognize_digits_mlp
     SRCS test_inference_recognize_digits.cc
     DEPS ARCHIVE_START paddle_fluid ARCHIVE_END
     ARGS --dirname=${PYTHON_TESTS_DIR}/book/recognize_digits_mlp.inference.model)
+cc_test(test_inference_image_classification_vgg
+    SRCS test_inference_image_classification.cc
+    DEPS ARCHIVE_START paddle_fluid ARCHIVE_END
+    ARGS --dirname=${PYTHON_TESTS_DIR}/book/image_classification_vgg.inference.model)
+cc_test(test_inference_image_classification_resnet
+    SRCS test_inference_image_classification.cc
+    DEPS ARCHIVE_START paddle_fluid ARCHIVE_END
+    ARGS --dirname=${PYTHON_TESTS_DIR}/book/image_classification_resnet.inference.model)
+cc_test(test_inference_label_semantic_roles
+    SRCS test_inference_label_semantic_roles.cc
+    DEPS ARCHIVE_START paddle_fluid ARCHIVE_END
+    ARGS --dirname=${PYTHON_TESTS_DIR}/book/label_semantic_roles.inference.model)
+cc_test(test_inference_rnn_encoder_decoder
+    SRCS test_inference_rnn_encoder_decoder.cc
+    DEPS ARCHIVE_START paddle_fluid ARCHIVE_END
+    ARGS --dirname=${PYTHON_TESTS_DIR}/book/rnn_encoder_decoder.inference.model)
 set_tests_properties(test_inference_recognize_digits_mlp
-    PROPERTIES DEPENDS test_recognize_digits_mlp_cpu)
+    PROPERTIES DEPENDS test_recognize_digits)
+set_tests_properties(test_inference_image_classification_vgg
+    PROPERTIES DEPENDS test_image_classification_train)
+set_tests_properties(test_inference_image_classification_resnet
+    PROPERTIES DEPENDS test_image_classification_train)
+set_tests_properties(test_inference_label_semantic_roles
+    PROPERTIES DEPENDS test_label_semantic_roles) 
+set_tests_properties(test_inference_rnn_encoder_decoder
+    PROPERTIES DEPENDS test_rnn_encoder_decoder) 

paddle/inference/tests/book/test_helper.h

+/* 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/framework/lod_tensor.h"
+#include "paddle/inference/io.h"
+
+template <typename T>
+void SetupTensor(paddle::framework::LoDTensor& input,
+                 paddle::framework::DDim dims,
+                 T lower,
+                 T upper) {
+  srand(time(0));
+  T* input_ptr = input.mutable_data<T>(dims, paddle::platform::CPUPlace());
+  for (int i = 0; i < input.numel(); ++i) {
+    input_ptr[i] =
+        (static_cast<T>(rand()) / static_cast<T>(RAND_MAX)) * (upper - lower) +
+        lower;
+  }
+}
+
+template <typename T>
+void SetupLoDTensor(paddle::framework::LoDTensor& input,
+                    paddle::framework::LoD& lod,
+                    T lower,
+                    T upper) {
+  input.set_lod(lod);
+  int dim = lod[0][lod[0].size() - 1];
+  SetupTensor(input, {dim, 1}, lower, upper);
+}
+
+template <typename T>
+void CheckError(paddle::framework::LoDTensor& output1,
+                paddle::framework::LoDTensor& output2) {
+  // Check lod information
+  EXPECT_EQ(output1.lod(), output2.lod());
+
+  EXPECT_EQ(output1.dims(), output2.dims());
+  EXPECT_EQ(output1.numel(), output2.numel());
+
+  T err = static_cast<T>(0);
+  if (typeid(T) == typeid(float)) {
+    err = 1E-3;
+  } else if (typeid(T) == typeid(double)) {
+    err = 1E-6;
+  } else {
+    err = 0;
+  }
+
+  size_t count = 0;
+  for (int64_t i = 0; i < output1.numel(); ++i) {
+    if (fabs(output1.data<T>()[i] - output2.data<T>()[i]) > err) {
+      count++;
+    }
+  }
+  EXPECT_EQ(count, 0) << "There are " << count << " different elements.";
+}
+
+template <typename Place, typename T>
+void TestInference(const std::string& dirname,
+                   const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
+                   std::vector<paddle::framework::LoDTensor*>& cpu_fetchs) {
+  // 1. Define place, executor and scope
+  auto place = Place();
+  auto executor = paddle::framework::Executor(place);
+  auto* scope = new paddle::framework::Scope();
+
+  // 2. Initialize the inference_program and load all parameters from file
+  auto inference_program = paddle::inference::Load(executor, *scope, dirname);
+
+  // 3. Get the feed_target_names and fetch_target_names
+  const std::vector<std::string>& feed_target_names =
+      inference_program->GetFeedTargetNames();
+  const std::vector<std::string>& fetch_target_names =
+      inference_program->GetFetchTargetNames();
+
+  // 4. Prepare inputs: set up maps for feed targets
+  std::map<std::string, const paddle::framework::LoDTensor*> feed_targets;
+  for (size_t i = 0; i < feed_target_names.size(); ++i) {
+    // Please make sure that cpu_feeds[i] is right for feed_target_names[i]
+    feed_targets[feed_target_names[i]] = cpu_feeds[i];
+  }
+
+  // 5. Define Tensor to get the outputs: set up maps for fetch targets
+  std::map<std::string, paddle::framework::LoDTensor*> fetch_targets;
+  for (size_t i = 0; i < fetch_target_names.size(); ++i) {
+    fetch_targets[fetch_target_names[i]] = cpu_fetchs[i];
+  }
+
+  // 6. Run the inference program
+  executor.Run(*inference_program, scope, feed_targets, fetch_targets);
+
+  delete scope;
+}

paddle/inference/tests/book/test_inference_image_classification.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 <gtest/gtest.h>
+#include <time.h>
+#include <sstream>
+#include "gflags/gflags.h"
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/inference/io.h"
+
+DEFINE_string(dirname, "", "Directory of the inference model.");
+
+template <typename Place, typename T>
+void TestInference(const std::string& dirname,
+                   const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
+                   std::vector<paddle::framework::LoDTensor*>& cpu_fetchs) {
+  // 1. Define place, executor and scope
+  auto place = Place();
+  auto executor = paddle::framework::Executor(place);
+  auto* scope = new paddle::framework::Scope();
+
+  // 2. Initialize the inference_program and load all parameters from file
+  auto inference_program = paddle::inference::Load(executor, *scope, dirname);
+
+  // 3. Get the feed_target_names and fetch_target_names
+  const std::vector<std::string>& feed_target_names =
+      inference_program->GetFeedTargetNames();
+  const std::vector<std::string>& fetch_target_names =
+      inference_program->GetFetchTargetNames();
+
+  // 4. Prepare inputs: set up maps for feed targets
+  std::map<std::string, const paddle::framework::LoDTensor*> feed_targets;
+  for (size_t i = 0; i < feed_target_names.size(); ++i) {
+    // Please make sure that cpu_feeds[i] is right for feed_target_names[i]
+    feed_targets[feed_target_names[i]] = cpu_feeds[i];
+  }
+
+  // 5. Define Tensor to get the outputs: set up maps for fetch targets
+  std::map<std::string, paddle::framework::LoDTensor*> fetch_targets;
+  for (size_t i = 0; i < fetch_target_names.size(); ++i) {
+    fetch_targets[fetch_target_names[i]] = cpu_fetchs[i];
+  }
+
+  // 6. Run the inference program
+  executor.Run(*inference_program, scope, feed_targets, fetch_targets);
+
+  delete scope;
+}
+
+TEST(inference, image_classification) {
+  if (FLAGS_dirname.empty()) {
+    LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model";
+  }
+
+  LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
+  std::string dirname = FLAGS_dirname;
+
+  // 0. Call `paddle::framework::InitDevices()` initialize all the devices
+  // In unittests, this is done in paddle/testing/paddle_gtest_main.cc
+
+  paddle::framework::LoDTensor input;
+  srand(time(0));
+  float* input_ptr =
+      input.mutable_data<float>({1, 3, 32, 32}, paddle::platform::CPUPlace());
+  for (int i = 0; i < 3072; ++i) {
+    input_ptr[i] = rand() / (static_cast<float>(RAND_MAX));
+  }
+  std::vector<paddle::framework::LoDTensor*> cpu_feeds;
+  cpu_feeds.push_back(&input);
+
+  paddle::framework::LoDTensor output1;
+  std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
+  cpu_fetchs1.push_back(&output1);
+
+  // Run inference on CPU
+  TestInference<paddle::platform::CPUPlace, float>(
+      dirname, cpu_feeds, cpu_fetchs1);
+  LOG(INFO) << output1.dims();
+
+#ifdef PADDLE_WITH_CUDA
+  paddle::framework::LoDTensor output2;
+  std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
+  cpu_fetchs2.push_back(&output2);
+
+  // Run inference on CUDA GPU
+  TestInference<paddle::platform::CUDAPlace, float>(
+      dirname, cpu_feeds, cpu_fetchs2);
+  LOG(INFO) << output2.dims();
+
+  EXPECT_EQ(output1.dims(), output2.dims());
+  EXPECT_EQ(output1.numel(), output2.numel());
+
+  float err = 1E-3;
+  int count = 0;
+  for (int64_t i = 0; i < output1.numel(); ++i) {
+    if (fabs(output1.data<float>()[i] - output2.data<float>()[i]) > err) {
+      count++;
+    }
+  }
+  EXPECT_EQ(count, 0) << "There are " << count << " different elements.";
+#endif
+}

paddle/inference/tests/book/test_inference_label_semantic_roles.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 <gtest/gtest.h>
+#include <time.h>
+#include <sstream>
+#include "gflags/gflags.h"
+#include "test_helper.h"
+
+DEFINE_string(dirname, "", "Directory of the inference model.");
+
+TEST(inference, label_semantic_roles) {
+  if (FLAGS_dirname.empty()) {
+    LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model";
+  }
+
+  LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
+  std::string dirname = FLAGS_dirname;
+
+  // 0. Call `paddle::framework::InitDevices()` initialize all the devices
+  // In unittests, this is done in paddle/testing/paddle_gtest_main.cc
+
+  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));
+
+  std::vector<paddle::framework::LoDTensor*> cpu_feeds;
+  cpu_feeds.push_back(&word);
+  cpu_feeds.push_back(&predicate);
+  cpu_feeds.push_back(&ctx_n2);
+  cpu_feeds.push_back(&ctx_n1);
+  cpu_feeds.push_back(&ctx_0);
+  cpu_feeds.push_back(&ctx_p1);
+  cpu_feeds.push_back(&ctx_p2);
+  cpu_feeds.push_back(&mark);
+
+  paddle::framework::LoDTensor output1;
+  std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
+  cpu_fetchs1.push_back(&output1);
+
+  // Run inference on CPU
+  TestInference<paddle::platform::CPUPlace, float>(
+      dirname, cpu_feeds, cpu_fetchs1);
+  LOG(INFO) << output1.lod();
+  LOG(INFO) << output1.dims();
+
+#ifdef PADDLE_WITH_CUDA
+  paddle::framework::LoDTensor output2;
+  std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
+  cpu_fetchs2.push_back(&output2);
+
+  // Run inference on CUDA GPU
+  TestInference<paddle::platform::CUDAPlace, float>(
+      dirname, cpu_feeds, cpu_fetchs2);
+  LOG(INFO) << output2.lod();
+  LOG(INFO) << output2.dims();
+
+  CheckError<float>(output1, output2);
+#endif
+}

paddle/inference/tests/book/test_inference_recognize_digits.cc

@@ -16,48 +16,10 @@ limitations under the License. */
 #include <time.h>
 #include <sstream>
 #include "gflags/gflags.h"
-#include "paddle/framework/lod_tensor.h"
-#include "paddle/inference/io.h"
+#include "test_helper.h"
 
 DEFINE_string(dirname, "", "Directory of the inference model.");
 
-template <typename Place, typename T>
-void TestInference(const std::string& dirname,
-                   const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
-                   std::vector<paddle::framework::LoDTensor*>& cpu_fetchs) {
-  // 1. Define place, executor and scope
-  auto place = Place();
-  auto executor = paddle::framework::Executor(place);
-  auto* scope = new paddle::framework::Scope();
-
-  // 2. Initialize the inference_program and load all parameters from file
-  auto inference_program = paddle::inference::Load(executor, *scope, dirname);
-
-  // 3. Get the feed_target_names and fetch_target_names
-  const std::vector<std::string>& feed_target_names =
-      inference_program->GetFeedTargetNames();
-  const std::vector<std::string>& fetch_target_names =
-      inference_program->GetFetchTargetNames();
-
-  // 4. Prepare inputs: set up maps for feed targets
-  std::map<std::string, const paddle::framework::LoDTensor*> feed_targets;
-  for (size_t i = 0; i < feed_target_names.size(); ++i) {
-    // Please make sure that cpu_feeds[i] is right for feed_target_names[i]
-    feed_targets[feed_target_names[i]] = cpu_feeds[i];
-  }
-
-  // 5. Define Tensor to get the outputs: set up maps for fetch targets
-  std::map<std::string, paddle::framework::LoDTensor*> fetch_targets;
-  for (size_t i = 0; i < fetch_target_names.size(); ++i) {
-    fetch_targets[fetch_target_names[i]] = cpu_fetchs[i];
-  }
-
-  // 6. Run the inference program
-  executor.Run(*inference_program, scope, feed_targets, fetch_targets);
-
-  delete scope;
-}
-
 TEST(inference, recognize_digits) {
   if (FLAGS_dirname.empty()) {
     LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model";
@@ -70,12 +32,10 @@ TEST(inference, recognize_digits) {
   // In unittests, this is done in paddle/testing/paddle_gtest_main.cc
 
   paddle::framework::LoDTensor input;
-  srand(time(0));
-  float* input_ptr =
-      input.mutable_data<float>({1, 28, 28}, paddle::platform::CPUPlace());
-  for (int i = 0; i < 784; ++i) {
-    input_ptr[i] = rand() / (static_cast<float>(RAND_MAX));
-  }
+  // Use normilized image pixels as input data,
+  // which should be in the range [-1.0, 1.0].
+  SetupTensor<float>(
+      input, {1, 28, 28}, static_cast<float>(-1), static_cast<float>(1));
   std::vector<paddle::framework::LoDTensor*> cpu_feeds;
   cpu_feeds.push_back(&input);
 
@@ -98,16 +58,6 @@ TEST(inference, recognize_digits) {
       dirname, cpu_feeds, cpu_fetchs2);
   LOG(INFO) << output2.dims();
 
-  EXPECT_EQ(output1.dims(), output2.dims());
-  EXPECT_EQ(output1.numel(), output2.numel());
-
-  float err = 1E-3;
-  int count = 0;
-  for (int64_t i = 0; i < output1.numel(); ++i) {
-    if (fabs(output1.data<float>()[i] - output2.data<float>()[i]) > err) {
-      count++;
-    }
-  }
-  EXPECT_EQ(count, 0) << "There are " << count << " different elements.";
+  CheckError<float>(output1, output2);
 #endif
 }

paddle/inference/tests/book/test_inference_rnn_encoder_decoder.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 <gtest/gtest.h>
+#include "gflags/gflags.h"
+#include "test_helper.h"
+
+DEFINE_string(dirname, "", "Directory of the inference model.");
+
+TEST(inference, rnn_encoder_decoder) {
+  if (FLAGS_dirname.empty()) {
+    LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model";
+  }
+
+  LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
+  std::string dirname = FLAGS_dirname;
+
+  // 0. Call `paddle::framework::InitDevices()` initialize all the devices
+  // In unittests, this is done in paddle/testing/paddle_gtest_main.cc
+
+  paddle::framework::LoDTensor word_data, trg_word;
+  paddle::framework::LoD lod{{0, 4, 10}};
+
+  SetupLoDTensor(
+      word_data, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
+  SetupLoDTensor(
+      trg_word, lod, static_cast<int64_t>(0), static_cast<int64_t>(1));
+
+  std::vector<paddle::framework::LoDTensor*> cpu_feeds;
+  cpu_feeds.push_back(&word_data);
+  cpu_feeds.push_back(&trg_word);
+
+  paddle::framework::LoDTensor output1;
+  std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
+  cpu_fetchs1.push_back(&output1);
+
+  // Run inference on CPU
+  TestInference<paddle::platform::CPUPlace, float>(
+      dirname, cpu_feeds, cpu_fetchs1);
+  LOG(INFO) << output1.lod();
+  LOG(INFO) << output1.dims();
+
+#ifdef PADDLE_WITH_CUDA
+  paddle::framework::LoDTensor output2;
+  std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
+  cpu_fetchs2.push_back(&output2);
+
+  // Run inference on CUDA GPU
+  TestInference<paddle::platform::CUDAPlace, float>(
+      dirname, cpu_feeds, cpu_fetchs2);
+  LOG(INFO) << output2.lod();
+  LOG(INFO) << output2.dims();
+
+  CheckError<float>(output1, output2);
+#endif
+}

paddle/math/Matrix.cpp

@@ -2015,13 +2015,6 @@ void CpuMatrix::maxPoolForward(Matrix& inputMat,
     CHECK_EQ(channels * outLength, maskMatP->getWidth());
   }
 
-  /* initialize the data_ */
-  for (size_t i = 0; i < height_; i++) {
-    for (size_t j = 0; j < width_; j++) {
-      outData[i * outStride + j] = -(real)FLT_MAX;
-    }
-  }
-
   /* pool max one by one */
   for (size_t n = 0; n < num; ++n) {  // frame by frame
     if (!isContiguous()) {
@@ -2030,19 +2023,24 @@ void CpuMatrix::maxPoolForward(Matrix& inputMat,
     for (size_t c = 0; c < channels; ++c) {  // channel by channel
       for (size_t ph = 0; ph < outputH; ++ph) {
         int hstart = ph * strideH - paddingH;
-        int hend = std::min(hstart + sizeY, imgSizeH);
-        hstart = std::max(hstart, 0);
+        int hend = hstart + sizeY;
+        hstart = hstart < 0 ? 0 : hstart;
+        hend = hend < (int)imgSizeH ? hend : (int)imgSizeH;
         for (size_t pw = 0; pw < outputW; ++pw) {
           int wstart = pw * strideW - paddingW;
-          int wend = std::min(wstart + sizeX, imgSizeW);
-          wstart = std::max(wstart, 0);
+          int wend = wstart + sizeX;
+          wstart = wstart < 0 ? 0 : wstart;
+          wend = wend < (int)imgSizeW ? wend : (int)imgSizeW;
           if (maskData == NULL) {
+            real tmp = -(real)FLT_MAX;
             for (int h = hstart; h < hend; ++h) {
               for (int w = wstart; w < wend; ++w) {
-                outData[ph * outputW + pw] = std::max(
-                    outData[ph * outputW + pw], inputData[h * imgSizeW + w]);
+                tmp = tmp < inputData[h * imgSizeW + w]
+                          ? inputData[h * imgSizeW + w]
+                          : tmp;
               }
             }
+            outData[ph * outputW + pw] = tmp;
           } else {
             for (int h = hstart; h < hend; ++h) {
               for (int w = wstart; w < wend; ++w) {

paddle/memory/CMakeLists.txt

@@ -14,10 +14,3 @@ cc_library(paddle_memory
     system_allocator)
 
 cc_test(memory_test SRCS memory_test.cc DEPS place paddle_memory)
-
-if(NOT WITH_C_API AND WITH_FLUID)
-  file(GLOB MEMORY_HEADERS *.h)
-  file(GLOB MEMORY_DETAIL_HEADERS detail/*.h)
-  install(FILES ${MEMORY_HEADERS} DESTINATION include/paddle/memory)
-  install(FILES ${MEMORY_DETAIL_HEADERS} DESTINATION include/paddle/memory/detail)
-endif()

paddle/operators/CMakeLists.txt

@@ -62,7 +62,7 @@ function(op_library TARGET)
     endif()
 
     # Define operators that don't need pybind here.
-    foreach(manual_pybind_op "net_op" "compare_op" "logical_op" "nccl_op" "tensor_array_read_write_op")
+    foreach(manual_pybind_op "net_op" "compare_op" "logical_op" "nccl_op" "tensor_array_read_write_op" "create_reader_op")
         if ("${TARGET}" STREQUAL "${manual_pybind_op}")
             set(pybind_flag 1)
         endif()
@@ -155,10 +155,14 @@ op_library(recurrent_op DEPS executor)
 op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale math_function)
 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)
+
+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
@@ -182,7 +186,7 @@ list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 foreach(src ${GENERAL_OPS})
     op_library(${src})
 endforeach()
-file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")
+file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\nUSE_NO_KERNEL_OP(create_random_data_generator);\n")
 
 set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
 

paddle/operators/bipartite_match_op.cc

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+/* 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.
@@ -28,12 +28,18 @@ class BipartiteMatchOp : public framework::OperatorWithKernel {
   void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("DistMat"),
                    "Input(DistMat) of BipartiteMatch should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("ColToRowMatchIndices"),
+        "Output(ColToRowMatchIndices) of BipartiteMatch should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("ColToRowMatchDist"),
+        "Output(ColToRowMatchDist) of BipartiteMatch should not be null.");
 
     auto dims = ctx->GetInputDim("DistMat");
     PADDLE_ENFORCE_EQ(dims.size(), 2, "The rank of Input(DistMat) must be 2.");
 
     ctx->SetOutputDim("ColToRowMatchIndices", dims);
-    ctx->SetOutputDim("ColToRowMatchDis", dims);
+    ctx->SetOutputDim("ColToRowMatchDist", dims);
   }
 };
 
@@ -91,7 +97,7 @@ class BipartiteMatchKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& context) const override {
     auto* dist_mat = context.Input<LoDTensor>("DistMat");
     auto* match_indices = context.Output<Tensor>("ColToRowMatchIndices");
-    auto* match_dist = context.Output<Tensor>("ColToRowMatchDis");
+    auto* match_dist = context.Output<Tensor>("ColToRowMatchDist");
 
     auto& dev_ctx = context.device_context<platform::CPUDeviceContext>();
 
@@ -148,13 +154,13 @@ class BipartiteMatchOpMaker : public framework::OpProtoAndCheckerMaker {
               "Otherwise, it means B[j] is matched to row "
               "ColToRowMatchIndices[i][j] in i-th instance. The row number of "
               "i-th instance is saved in ColToRowMatchIndices[i][j].");
-    AddOutput("ColToRowMatchDis",
+    AddOutput("ColToRowMatchDist",
               "(Tensor) A 2-D Tensor with shape [N, M] in float type. "
               "N is batch size. If ColToRowMatchIndices[i][j] is -1, "
-              "ColToRowMatchDis[i][j] is also -1.0. Otherwise, assumed "
+              "ColToRowMatchDist[i][j] is also -1.0. Otherwise, assumed "
               "ColToRowMatchIndices[i][j] = d, and the row offsets of each "
               "instance are called LoD. Then "
-              "ColToRowMatchDis[i][j] = DistMat[d+LoD[i]][j]");
+              "ColToRowMatchDist[i][j] = DistMat[d+LoD[i]][j]");
     AddComment(R"DOC(
 This operator is a greedy bipartite matching algorithm, which is used to
 obtain the matching with the maximum distance based on the input

paddle/operators/box_coder_op.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/operators/box_coder_op.h"
+
+namespace paddle {
+namespace operators {
+
+class BoxCoderOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("PriorBox"),
+                   "Input(PriorBox) of BoxCoderOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("PriorBoxVar"),
+                   "Input(PriorBoxVar) of BoxCoderOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("TargetBox"),
+                   "Input(TargetBox) of BoxCoderOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("OutputBox"),
+                   "Output(OutputBox) of BoxCoderOp should not be null.");
+
+    auto prior_box_dims = ctx->GetInputDim("PriorBox");
+    auto prior_box_var_dims = ctx->GetInputDim("PriorBoxVar");
+    auto target_box_dims = ctx->GetInputDim("TargetBox");
+
+    PADDLE_ENFORCE_EQ(prior_box_dims.size(), 2,
+                      "The rank of Input of PriorBoxVar must be 2");
+    PADDLE_ENFORCE_EQ(prior_box_dims[1], 4, "The shape of PriorBox is [N, 4]");
+    PADDLE_ENFORCE_EQ(prior_box_dims, prior_box_var_dims);
+    PADDLE_ENFORCE_EQ(target_box_dims.size(), 2,
+                      "The rank of Input of TargetBox must be 2");
+    PADDLE_ENFORCE_EQ(target_box_dims[1], 4,
+                      "The shape of TargetBox is [M, 4]");
+
+    GetBoxCodeType(ctx->Attrs().Get<std::string>("code_type"));
+
+    ctx->SetOutputDim(
+        "OutputBox",
+        framework::make_ddim({target_box_dims[0], prior_box_dims[0], 4}));
+    ctx->ShareLoD("TargetBox", /*->*/ "OutputBox");
+  }
+};
+
+class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  BoxCoderOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput(
+        "PriorBox",
+        "(Tensor, default Tensor<float>) "
+        "Box list PriorBox is a 2-D Tensor with shape [M, 4] holds M boxes, "
+        "each box is represented as [xmin, ymin, xmax, ymax], "
+        "[xmin, ymin] is the left top coordinate of the anchor box, "
+        "if the input is image feature map, they are close to the origin "
+        "of the coordinate system. [xmax, ymax] is the right bottom "
+        "coordinate of the anchor box.");
+    AddInput("PriorBoxVar",
+             "(Tensor, default Tensor<float>) "
+             "PriorBoxVar is a 2-D Tensor with shape [M, 4] holds M group "
+             "of variance.");
+    AddInput(
+        "TargetBox",
+        "(LoDTensor or Tensor) this input is a 2-D LoDTensor with shape "
+        "[N, 4], each box is represented as [xmin, ymin, xmax, ymax], "
+        "[xmin, ymin] is the left top coordinate of the box if the input "
+        "is image feature map, they are close to the origin of the coordinate "
+        "system. [xmax, ymax] is the right bottom coordinate of the box. "
+        "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>("code_type",
+                         "(string, default encode_center_size) "
+                         "the code type used with the target box")
+        .SetDefault("encode_center_size")
+        .InEnum({"encode_center_size", "decode_center_size"});
+    AddOutput(
+        "OutputBox",
+        "(LoDTensor or Tensor) "
+        "(Tensor) The output of box_coder_op, a tensor with shape [N, M, 4] "
+        "representing the result of N target boxes encoded/decoded with "
+        "M Prior boxes and variances.");
+
+    AddComment(R"DOC(
+Bounding Box Coder Operator.
+Encode/Decode the target bounding box with the priorbox information.
+The Encoding schema described below:
+ox = (tx - px) / pw / pxv
+oy = (ty - py) / ph / pyv
+ow = log(abs(tw / pw)) / pwv 
+oh = log(abs(th / ph)) / phv 
+The Decoding schema described below:
+ox = (pw * pxv * tx * + px) - tw / 2
+oy = (ph * pyv * ty * + py) - th / 2
+ow = exp(pwv * tw) * pw + tw / 2
+oh = exp(phv * th) * ph + th / 2
+where tx, ty, tw, th denote the target box's center coordinates, width and
+height respectively. Similarly, px, py, pw, ph denote the priorbox's(anchor)
+center coordinates, width and height. pxv, pyv, pwv, phv denote the variance
+of the priorbox and ox, oy, ow, oh denote the encoded/decoded coordinates,
+width and height.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(box_coder, ops::BoxCoderOp, ops::BoxCoderOpMaker);
+REGISTER_OP_CPU_KERNEL(box_coder, ops::BoxCoderKernel<float>,
+                       ops::BoxCoderKernel<double>);

paddle/operators/box_coder_op.cu

+/* 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/operators/box_coder_op.h"
+#include "paddle/platform/cuda_helper.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+__global__ void EncodeCenterSizeKernel(const T* prior_box_data,
+                                       const T* prior_box_var_data,
+                                       const T* target_box_data, const int row,
+                                       const int col, const int len,
+                                       T* output) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < row * col) {
+    const int row_idx = idx / col;
+    const int col_idx = idx % col;
+    T prior_box_width =
+        prior_box_data[col_idx * len + 2] - prior_box_data[col_idx * len];
+    T prior_box_height =
+        prior_box_data[col_idx * len + 3] - prior_box_data[col_idx * len + 1];
+    T prior_box_center_x =
+        (prior_box_data[col_idx * len + 2] + prior_box_data[col_idx * len]) / 2;
+    T prior_box_center_y = (prior_box_data[col_idx * len + 3] +
+                            prior_box_data[col_idx * len + 1]) /
+                           2;
+
+    T target_box_center_x =
+        (target_box_data[row_idx * len + 2] + target_box_data[row_idx * len]) /
+        2;
+    T target_box_center_y = (target_box_data[row_idx * len + 3] +
+                             target_box_data[row_idx * len + 1]) /
+                            2;
+    T target_box_width =
+        target_box_data[row_idx * len + 2] - target_box_data[row_idx * len];
+    T target_box_height =
+        target_box_data[row_idx * len + 3] - target_box_data[row_idx * len + 1];
+
+    output[idx * len] = (target_box_center_x - prior_box_center_x) /
+                        prior_box_width / prior_box_var_data[col_idx * len];
+    output[idx * len + 1] = (target_box_center_y - prior_box_center_y) /
+                            prior_box_height /
+                            prior_box_var_data[col_idx * len + 1];
+    output[idx * len + 2] = log(fabs(target_box_width / prior_box_width)) /
+                            prior_box_var_data[col_idx * len + 2];
+    output[idx * len + 3] = log(fabs(target_box_height / prior_box_height)) /
+                            prior_box_var_data[col_idx * len + 3];
+  }
+}
+
+template <typename T>
+__global__ void DecodeCenterSizeKernel(const T* prior_box_data,
+                                       const T* prior_box_var_data,
+                                       const T* target_box_data, const int row,
+                                       const int col, const int len,
+                                       T* output) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < row * col) {
+    const int row_idx = idx / col;
+    const int col_idx = idx % col;
+    T prior_box_width =
+        prior_box_data[col_idx * len + 2] - prior_box_data[col_idx * len];
+    T prior_box_height =
+        prior_box_data[col_idx * len + 3] - prior_box_data[col_idx * len + 1];
+    T prior_box_center_x =
+        (prior_box_data[col_idx * len + 2] + prior_box_data[col_idx * len]) / 2;
+    T prior_box_center_y = (prior_box_data[col_idx * len + 3] +
+                            prior_box_data[col_idx * len + 1]) /
+                           2;
+
+    T target_box_width = exp(prior_box_var_data[col_idx * len + 2] *
+                             target_box_data[row_idx * len + 2]) *
+                         prior_box_width;
+    T target_box_height = exp(prior_box_var_data[col_idx * len + 3] *
+                              target_box_data[row_idx * len + 3]) *
+                          prior_box_height;
+    T target_box_center_x = prior_box_var_data[col_idx * len] *
+                                target_box_data[row_idx * len] *
+                                prior_box_width +
+                            prior_box_center_x;
+    T target_box_center_y = prior_box_var_data[col_idx * len + 1] *
+                                target_box_data[row_idx * len + 1] *
+                                prior_box_height +
+                            prior_box_center_y;
+
+    output[idx * len] = target_box_center_x - target_box_width / 2;
+    output[idx * len + 1] = target_box_center_y - target_box_height / 2;
+    output[idx * len + 2] = target_box_center_x + target_box_width / 2;
+    output[idx * len + 3] = target_box_center_y + target_box_height / 2;
+  }
+}
+
+template <typename T>
+class BoxCoderCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(context.GetPlace()),
+                   "This kernel only runs on GPU device.");
+    auto* prior_box = context.Input<framework::Tensor>("PriorBox");
+    auto* prior_box_var = context.Input<framework::Tensor>("PriorBoxVar");
+    auto* target_box = context.Input<framework::LoDTensor>("TargetBox");
+    auto* output_box = context.Output<framework::Tensor>("OutputBox");
+
+    if (target_box->lod().size()) {
+      PADDLE_ENFORCE_EQ(target_box->lod().size(), 1,
+                        "Only support 1 level of LoD.");
+    }
+    auto row = target_box->dims()[0];
+    auto col = prior_box->dims()[0];
+    auto len = prior_box->dims()[1];
+    int block = 512;
+    int grid = (row * col + block - 1) / block;
+    auto& device_ctx = context.cuda_device_context();
+
+    const T* prior_box_data = prior_box->data<T>();
+    const T* prior_box_var_data = prior_box_var->data<T>();
+    const T* target_box_data = target_box->data<T>();
+
+    output_box->mutable_data<T>({row, col, len}, context.GetPlace());
+    T* output = output_box->data<T>();
+
+    auto code_type = GetBoxCodeType(context.Attr<std::string>("code_type"));
+    if (code_type == BoxCodeType::kEncodeCenterSize) {
+      EncodeCenterSizeKernel<T><<<grid, block, 0, device_ctx.stream()>>>(
+          prior_box_data, prior_box_var_data, target_box_data, row, col, len,
+          output);
+    } else if (code_type == BoxCodeType::kDecodeCenterSize) {
+      DecodeCenterSizeKernel<T><<<grid, block, 0, device_ctx.stream()>>>(
+          prior_box_data, prior_box_var_data, target_box_data, row, col, len,
+          output);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(box_coder, ops::BoxCoderCUDAKernel<float>,
+                        ops::BoxCoderCUDAKernel<double>);

paddle/operators/box_coder_op.h

+/* 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. */
+
+#pragma once
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+enum class BoxCodeType { kEncodeCenterSize = 0, kDecodeCenterSize = 1 };
+
+inline BoxCodeType GetBoxCodeType(const std::string& type) {
+  if (type == "encode_center_size") {
+    return BoxCodeType::kEncodeCenterSize;
+  } else if (type == "decode_center_size") {
+    return BoxCodeType::kDecodeCenterSize;
+  }
+  PADDLE_THROW("Not support type %s.", type);
+}
+
+template <typename T>
+class BoxCoderKernel : public framework::OpKernel<T> {
+ public:
+  void EncodeCenterSize(const framework::Tensor& target_box,
+                        const framework::Tensor& prior_box,
+                        const framework::Tensor& prior_box_var,
+                        T* output) const {
+    int64_t row = target_box.dims()[0];
+    int64_t col = prior_box.dims()[0];
+    int64_t len = prior_box.dims()[1];
+    auto* target_box_data = target_box.data<T>();
+    auto* prior_box_data = prior_box.data<T>();
+    auto* prior_box_var_data = prior_box_var.data<T>();
+
+    for (int64_t i = 0; i < row; ++i) {
+      for (int64_t j = 0; j < col; ++j) {
+        T prior_box_width =
+            prior_box_data[j * len + 2] - prior_box_data[j * len];
+        T prior_box_height =
+            prior_box_data[j * len + 3] - prior_box_data[j * len + 1];
+        T prior_box_center_x =
+            (prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2;
+        T prior_box_center_y =
+            (prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2;
+
+        T target_box_center_x =
+            (target_box_data[i * len + 2] + target_box_data[i * len]) / 2;
+        T target_box_center_y =
+            (target_box_data[i * len + 3] + target_box_data[i * len + 1]) / 2;
+        T target_box_width =
+            target_box_data[i * len + 2] - target_box_data[i * len];
+        T target_box_height =
+            target_box_data[i * len + 3] - target_box_data[i * len + 1];
+
+        size_t offset = i * col * len + j * len;
+        output[offset] = (target_box_center_x - prior_box_center_x) /
+                         prior_box_width / prior_box_var_data[j * len];
+        output[offset + 1] = (target_box_center_y - prior_box_center_y) /
+                             prior_box_height / prior_box_var_data[j * len + 1];
+        output[offset + 2] =
+            std::log(std::fabs(target_box_width / prior_box_width)) /
+            prior_box_var_data[j * len + 2];
+        output[offset + 3] =
+            std::log(std::fabs(target_box_height / prior_box_height)) /
+            prior_box_var_data[j * len + 3];
+      }
+    }
+  }
+  void DecodeCenterSize(const framework::Tensor& target_box,
+                        const framework::Tensor& prior_box,
+                        const framework::Tensor& prior_box_var,
+                        T* output) const {
+    int64_t row = target_box.dims()[0];
+    int64_t col = prior_box.dims()[0];
+    int64_t len = prior_box.dims()[1];
+
+    auto* target_box_data = target_box.data<T>();
+    auto* prior_box_data = prior_box.data<T>();
+    auto* prior_box_var_data = prior_box_var.data<T>();
+
+    for (int64_t i = 0; i < row; ++i) {
+      for (int64_t j = 0; j < col; ++j) {
+        T prior_box_width =
+            prior_box_data[j * len + 2] - prior_box_data[j * len];
+        T prior_box_height =
+            prior_box_data[j * len + 3] - prior_box_data[j * len + 1];
+        T prior_box_center_x =
+            (prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2;
+        T prior_box_center_y =
+            (prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2;
+
+        T target_box_center_x = prior_box_var_data[j * len] *
+                                    target_box_data[i * len] * prior_box_width +
+                                prior_box_center_x;
+        T target_box_center_y = prior_box_var_data[j * len + 1] *
+                                    target_box_data[i * len + 1] *
+                                    prior_box_height +
+                                prior_box_center_y;
+        T target_box_width = std::exp(prior_box_var_data[j * len + 2] *
+                                      target_box_data[i * len + 2]) *
+                             prior_box_width;
+        T target_box_height = std::exp(prior_box_var_data[j * len + 3] *
+                                       target_box_data[i * len + 3]) *
+                              prior_box_height;
+
+        size_t offset = i * col * len + j * len;
+        output[offset] = target_box_center_x - target_box_width / 2;
+        output[offset + 1] = target_box_center_y - target_box_height / 2;
+        output[offset + 2] = target_box_center_x + target_box_width / 2;
+        output[offset + 3] = target_box_center_y + target_box_height / 2;
+      }
+    }
+  }
+
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* prior_box = context.Input<framework::Tensor>("PriorBox");
+    auto* prior_box_var = context.Input<framework::Tensor>("PriorBoxVar");
+    auto* target_box = context.Input<framework::LoDTensor>("TargetBox");
+    auto* output_box = context.Output<framework::Tensor>("OutputBox");
+
+    if (target_box->lod().size()) {
+      PADDLE_ENFORCE_EQ(target_box->lod().size(), 1UL,
+                        "Only support 1 level of LoD.");
+    }
+    auto row = target_box->dims()[0];
+    auto col = prior_box->dims()[0];
+    auto len = prior_box->dims()[1];
+
+    output_box->mutable_data<T>({row, col, len}, context.GetPlace());
+
+    auto code_type = GetBoxCodeType(context.Attr<std::string>("code_type"));
+    T* output = output_box->data<T>();
+    if (code_type == BoxCodeType::kEncodeCenterSize) {
+      EncodeCenterSize(*target_box, *prior_box, *prior_box_var, output);
+    } else if (code_type == BoxCodeType::kDecodeCenterSize) {
+      DecodeCenterSize(*target_box, *prior_box, *prior_box_var, output);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/compare_op.h

@@ -54,7 +54,15 @@ class CompareOpKernel
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     using T = typename Functor::ELEM_TYPE;
-    ElementwiseComputeEx<Functor, DeviceContext, T, bool>(context);
+    using Tensor = framework::Tensor;
+
+    auto* x = context.Input<Tensor>("X");
+    auto* y = context.Input<Tensor>("Y");
+    auto* z = context.Output<Tensor>("Out");
+    z->mutable_data<T>(context.GetPlace());
+    int axis = context.Attr<int>("axis");
+    ElementwiseComputeEx<Functor, DeviceContext, T, bool>(context, x, y, axis,
+                                                          z);
   }
 };
 

paddle/operators/conditional_block_op.cc

@@ -41,6 +41,21 @@ class ConditionalOp : public framework::OperatorBase {
         });
     return retv;
   }
+
+  bool ScalarCondition(
+      const std::vector<const framework::LoDTensor *> &ips) const {
+    if (!(ips.size() == 1UL && ips[0]->IsInitialized())) {
+      PADDLE_THROW("should have one initialized input as condition");
+    }
+    if (!(ips[0]->type().hash_code() == typeid(bool).hash_code() &&
+          ips[0]->numel() == 1)) {
+      PADDLE_THROW(
+          "condition input's data type should be bool, "
+          "numel should be 1, actual numel is %d",
+          ips[0]->numel());
+    }
+    return ips[0]->data<bool>()[0];
+  }
 };
 
 class ConditionalBlockOp : public ConditionalOp {
@@ -53,9 +68,15 @@ class ConditionalBlockOp : public ConditionalOp {
   void Run(const framework::Scope &scope,
            const platform::Place &dev_place) const override {
     auto xs = InputTensors(scope);
-    bool need_run = std::all_of(
-        xs.begin(), xs.end(),
-        [](const framework::LoDTensor *t) { return t->numel() != 0; });
+
+    bool need_run;
+    if (Attr<bool>("is_scalar_condition")) {
+      need_run = ScalarCondition(xs);
+    } else {
+      need_run = std::all_of(
+          xs.begin(), xs.end(),
+          [](const framework::LoDTensor *t) { return t->numel() != 0; });
+    }
 
     if (need_run) {
       auto *scope_var = scope.FindVar(Output("Scope"));
@@ -88,6 +109,10 @@ class ConditionalBlockOpProtoMaker : public framework::OpProtoAndCheckerMaker {
               "scope is std::vector<Scope*>");
     AddAttr<framework::BlockDesc *>(
         "sub_block", "The step block of conditional block operator");
+    AddAttr<bool>("is_scalar_condition",
+                  "the input X is used as scalar "
+                  "condition")
+        .SetDefault(false);
     AddComment(R"DOC(Conditional block operator
 
 Run the sub-block if X is not empty. Params is the other inputs and Out is the
@@ -106,9 +131,15 @@ class ConditionalBlockGradOp : public ConditionalOp {
   void Run(const framework::Scope &scope,
            const platform::Place &dev_place) const override {
     auto xs = this->InputTensors(scope);
-    bool need_run = std::all_of(
-        xs.begin(), xs.end(),
-        [](const framework::LoDTensor *t) { return t->numel() != 0; });
+
+    bool need_run;
+    if (Attr<bool>("is_scalar_condition")) {
+      need_run = ScalarCondition(xs);
+    } else {
+      need_run = std::all_of(
+          xs.begin(), xs.end(),
+          [](const framework::LoDTensor *t) { return t->numel() != 0; });
+    }
 
     if (need_run) {
       auto *scope_var = scope.FindVar(Input("Scope"));
@@ -182,6 +213,7 @@ class ConditionalBlockGradMaker : public framework::SingleGradOpDescMaker {
     grad_op->SetOutput(framework::GradVarName("Params"),
                        InputGrad("Params", false));
     grad_op->SetBlockAttr("sub_block", *this->grad_block_[0]);
+    grad_op->SetAttr("is_scalar_condition", GetAttr("is_scalar_condition"));
     return std::unique_ptr<framework::OpDesc>(grad_op);
   }
 };

paddle/operators/conv_op.cc

@@ -318,9 +318,25 @@ framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
 namespace ops = paddle::operators;
 REGISTER_OP(conv2d, ops::ConvOp, ops::Conv2DOpMaker, conv2d_grad,
             ops::ConvOpGrad);
+
+// depthwise convolution op
+REGISTER_OP(depthwise_conv2d, ops::ConvOp, ops::Conv2DOpMaker,
+            depthwise_conv2d_grad, ops::ConvOpGrad);
 REGISTER_OP(conv3d, ops::ConvOp, ops::Conv3DOpMaker, conv3d_grad,
             ops::ConvOpGrad);
 
+// depthwise conv kernel
+// TODO(xingzhaolong): neon kernel for mobile
+REGISTER_OP_CPU_KERNEL(
+    depthwise_conv2d,
+    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);
+
+REGISTER_OP_CPU_KERNEL(
+    depthwise_conv2d_grad,
+    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);
+
 REGISTER_OP_CPU_KERNEL(
     conv2d, ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
     ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);

paddle/operators/conv_op.cu.cc

@@ -16,6 +16,16 @@ limitations under the License. */
 
 namespace ops = paddle::operators;
 
+REGISTER_OP_CUDA_KERNEL(
+    depthwise_conv2d,
+    ops::DepthwiseConvKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::DepthwiseConvKernel<paddle::platform::CUDADeviceContext, double>);
+
+REGISTER_OP_CUDA_KERNEL(
+    depthwise_conv2d_grad,
+    ops::DepthwiseConvGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::DepthwiseConvGradKernel<paddle::platform::CUDADeviceContext, double>);
+
 REGISTER_OP_CUDA_KERNEL(
     conv2d, ops::GemmConvKernel<paddle::platform::CUDADeviceContext, float>,
     ops::GemmConvKernel<paddle::platform::CUDADeviceContext, double>);

paddle/operators/conv_op.h

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include "paddle/framework/eigen.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/depthwise_conv.h"
 #include "paddle/operators/math/im2col.h"
 #include "paddle/operators/math/math_function.h"
 #include "paddle/operators/math/vol2col.h"
@@ -350,5 +351,72 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
     }
   }
 };
+
+template <typename DeviceContext, typename T>
+class DepthwiseConvKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const Tensor* input = context.Input<Tensor>("Input");
+    Tensor filter = *context.Input<Tensor>("Filter");
+    Tensor* output = context.Output<Tensor>("Output");
+    output->mutable_data<T>(context.GetPlace());
+
+    PADDLE_ENFORCE_EQ(
+        output->dims()[1] % input->dims()[1], 0,
+        "The output channels must be a multiple of the input channels");
+    std::vector<int> strides = context.Attr<std::vector<int>>("strides");
+    std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
+    std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
+
+    math::DepthwiseConvFunctor<DeviceContext, T> depthwiseConv;
+
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    depthwiseConv(dev_ctx, *input, filter, strides, paddings, output);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class DepthwiseConvGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const Tensor* input = context.Input<Tensor>("Input");
+    const Tensor* output_grad =
+        context.Input<Tensor>(framework::GradVarName("Output"));
+    Tensor* input_grad =
+        context.Output<Tensor>(framework::GradVarName("Input"));
+    Tensor* filter_grad =
+        context.Output<Tensor>(framework::GradVarName("Filter"));
+    Tensor filter = *context.Input<Tensor>("Filter");
+
+    if (!input_grad && !filter_grad) return;
+
+    std::vector<int> strides = context.Attr<std::vector<int>>("strides");
+    std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
+    std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
+
+    math::SetConstant<DeviceContext, T> set_zero;
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+
+    math::DepthwiseConvInputGradFunctor<DeviceContext, T>
+        depthwiseConvInputGrad;
+    math::DepthwiseConvFilterGradFunctor<DeviceContext, T>
+        depthwiseConvFilterGrad;
+
+    if (input_grad) {
+      input_grad->mutable_data<T>(context.GetPlace());
+      set_zero(dev_ctx, input_grad, static_cast<T>(0));
+      depthwiseConvInputGrad(dev_ctx, *input, filter, *output_grad, strides,
+                             paddings, input_grad);
+    }
+
+    if (filter_grad) {
+      filter_grad->mutable_data<T>(context.GetPlace());
+      set_zero(dev_ctx, filter_grad, static_cast<T>(0));
+      depthwiseConvFilterGrad(dev_ctx, *input, *output_grad, strides, paddings,
+                              filter_grad);
+    }
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle

paddle/operators/create_reader_op.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/framework/op_registry.h"
+#include "paddle/framework/reader.h"
+
+namespace paddle {
+namespace operators {
+
+static std::vector<framework::DDim> RestoreShapes(
+    const std::vector<int>& shape_concat, const std::vector<int>& ranks) {
+  std::vector<framework::DDim> res;
+  int offset = 0;
+  for (int len : ranks) {
+    auto start_it = shape_concat.begin() + offset;
+    auto end_it = start_it + len;
+    res.push_back(framework::make_ddim(std::vector<int>(start_it, end_it)));
+    offset += len;
+  }
+  return res;
+}
+
+// general infershape for file readers
+class CreateFileReaderInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "The output file reader should not be null.");
+    const auto shape_concat =
+        ctx->Attrs().Get<std::vector<int>>("shape_concat");
+    const auto ranks = ctx->Attrs().Get<std::vector<int>>("ranks");
+    std::vector<framework::DDim> shapes = RestoreShapes(shape_concat, ranks);
+    ctx->SetReaderDims("Out", shapes);
+  }
+};
+
+// general infershape for decorated readers
+class CreateDecoratedReaderInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("UnderlyingReader"),
+                   "Input(UnderlyingReader) should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "The output decorated reader should not be null.");
+    ctx->SetReaderDims("Out", ctx->GetReaderDims("UnderlyingReader"));
+  }
+};
+
+// general var type inference for all readers
+class CreateReaderInferVarType : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc& op_desc,
+                  framework::BlockDesc* block) const override {
+    std::string reader_name = op_desc.Output("Out")[0];
+    framework::VarDesc* reader = block->FindVarRecursive(reader_name);
+    reader->SetType(framework::proto::VarDesc::READER);
+  }
+};
+
+template <typename T>
+class CreateRandomDataGeneratorOp : public framework::OperatorBase {
+ public:
+  using framework::OperatorBase::OperatorBase;
+  void Run(const framework::Scope& scope,
+           const platform::Place& dev_place) const override {
+    const auto& shape_concat = Attr<std::vector<int>>("shape_concat");
+    const auto& ranks = Attr<std::vector<int>>("ranks");
+    PADDLE_ENFORCE(!shape_concat.empty() && !ranks.empty());
+    PADDLE_ENFORCE_EQ(std::accumulate(ranks.begin(), ranks.end(), 0),
+                      int(shape_concat.size()),
+                      "The accumulate of all ranks should be equal to the "
+                      "shape concat's length.");
+    std::vector<framework::DDim> shapes = RestoreShapes(shape_concat, ranks);
+    auto* out = scope.FindVar(Output("Out"))
+                    ->template GetMutable<framework::ReaderHolder>();
+    out->Reset(new framework::RandomDataGenerator<T>(shapes, Attr<float>("min"),
+                                                     Attr<float>("max")));
+  }
+};
+
+class CreateRandomDataGeneratorOpMaker
+    : public framework::OpProtoAndCheckerMaker {
+ public:
+  CreateRandomDataGeneratorOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(op_proto, op_checker) {
+    AddOutput("Out", "(ReaderHolder) The created random reader.");
+    AddAttr<std::vector<int>>("shape_concat",
+                              "The concat of all data's shapes.");
+    AddAttr<std::vector<int>>(
+        "ranks",
+        "The ranks of each data."
+        "e.g."
+        "shape_concat = [2,3,4,5,6]"
+        "ranks = [3,2]"
+        "It means the reader will generate two data each time,"
+        "whose shapes are [2,3,4] and [5,6] respectively.");
+    AddAttr<float>("min", "The lower bound of reader's uniform distribution.");
+    AddAttr<float>("max", "The upper bound of reader's uniform distribution.");
+    AddComment(R"DOC(
+      CreateRandomDataGenerator Operator
+
+      This Op creates a random reader. 
+      The reader generates random data instead of really reading from files.
+      Generated data follow an uniform distribution between 'min' and 'max'.
+    )DOC");
+  }
+};
+
+class CreateShuffleReaderOp : public framework::OperatorBase {
+ public:
+  using framework::OperatorBase::OperatorBase;
+  void Run(const framework::Scope& scope,
+           const platform::Place& dev_place) const override {
+    const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader"))
+                                        ->Get<framework::ReaderHolder>();
+    auto* out = scope.FindVar(Output("Out"))
+                    ->template GetMutable<framework::ReaderHolder>();
+    out->Reset(new framework::ShuffleReader(underlying_reader.Get(),
+                                            Attr<int>("buffer_size")));
+  }
+};
+
+class CreateShuffleReaderOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  CreateShuffleReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(op_proto, op_checker) {
+    AddInput(
+        "UnderlyingReader",
+        "(ReaderHolder) The underlying reader for creating a shuffle reader.");
+    AddOutput("Out", "(ReaderHolder) The created shuffle reader.");
+    AddAttr<int>("buffer_size", "The shuffle buffer size.").GreaterThan(0);
+    AddComment(R"DOC(
+      CreateShuffleReader Operator
+
+      A shuffle reader takes another reader as its 'underlying reader'
+      and yields the underlying reader's outputs in a shuffled order. 
+    )DOC");
+  }
+};
+
+class CreateBatchReaderOp : public framework::OperatorBase {
+ public:
+  using framework::OperatorBase::OperatorBase;
+  void Run(const framework::Scope& scope,
+           const platform::Place& dev_place) const override {
+    const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader"))
+                                        ->Get<framework::ReaderHolder>();
+    auto* out = scope.FindVar(Output("Out"))
+                    ->template GetMutable<framework::ReaderHolder>();
+    out->Reset(new framework::BatchReader(underlying_reader.Get(),
+                                          Attr<int>("batch_size")));
+  }
+};
+
+class CreateBatchReaderOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  CreateBatchReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(op_proto, op_checker) {
+    AddInput(
+        "UnderlyingReader",
+        "(ReaderHolder) The underlying reader for creating a batch reader.");
+    AddOutput("Out", "(ReaderHolder) The created batch reader.");
+    AddAttr<int>("batch_size",
+                 "How many instances the batch reader yields each time.")
+        .GreaterThan(0);
+    AddComment(R"DOC(
+      CreateBatchReader Operator
+
+      A batch reader takes another reader as its 'underlying reader', 
+      gathers the underlying reader's outputs and then yields them in batches. 
+    )DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(create_random_data_generator,
+                  ops::CreateRandomDataGeneratorOp<float>,
+                  ops::CreateFileReaderInferShape,
+                  ops::CreateRandomDataGeneratorOpMaker,
+                  paddle::framework::EmptyGradOpMaker,
+                  ops::CreateReaderInferVarType);
+REGISTER_OPERATOR(create_shuffle_reader, ops::CreateShuffleReaderOp,
+                  ops::CreateDecoratedReaderInferShape,
+                  ops::CreateShuffleReaderOpMaker,
+                  paddle::framework::EmptyGradOpMaker,
+                  ops::CreateReaderInferVarType);
+REGISTER_OPERATOR(create_batch_reader, ops::CreateBatchReaderOp,
+                  ops::CreateDecoratedReaderInferShape,
+                  ops::CreateBatchReaderOpMaker,
+                  paddle::framework::EmptyGradOpMaker,
+                  ops::CreateReaderInferVarType);

paddle/operators/elementwise_add_op.h

@@ -28,7 +28,14 @@ template <typename DeviceContext, typename T>
 class ElementwiseAddKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseComputeEx<AddFunctor<T>, DeviceContext, T>(ctx);
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* z = ctx.Output<Tensor>("Out");
+    z->mutable_data<T>(ctx.GetPlace());
+    int axis = ctx.Attr<int>("axis");
+    ElementwiseComputeEx<AddFunctor<T>, DeviceContext, T>(ctx, x, y, axis, z);
   }
 };
 
@@ -92,9 +99,19 @@ template <typename DeviceContext, typename T>
 class ElementwiseAddGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* out = ctx.Input<Tensor>("Out");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    int axis = ctx.Attr<int>("axis");
     ElementwiseGradCompute<DeviceContext, T, ElementwiseAddGradFunctor<T>,
                            ElementwiseAddBroadCastGradFunctor<T>,
-                           ElementwiseAddBroadCast2GradFunctor<T>>(ctx);
+                           ElementwiseAddBroadCast2GradFunctor<T>>(
+        ctx, x, y, out, dout, axis, dx, dy);
   }
 };
 

paddle/operators/elementwise_div_op.h

@@ -28,7 +28,14 @@ template <typename DeviceContext, typename T>
 class ElementwiseDivKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseComputeEx<DivFunctor<T>, DeviceContext, T>(ctx);
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* z = ctx.Output<Tensor>("Out");
+    z->mutable_data<T>(ctx.GetPlace());
+    int axis = ctx.Attr<int>("axis");
+    ElementwiseComputeEx<DivFunctor<T>, DeviceContext, T>(ctx, x, y, axis, z);
   }
 };
 
@@ -111,9 +118,19 @@ template <typename DeviceContext, typename T>
 class ElementwiseDivGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* out = ctx.Input<Tensor>("Out");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    int axis = ctx.Attr<int>("axis");
     ElementwiseGradCompute<DeviceContext, T, ElementwiseDivGradFunctor<T>,
                            ElementwiseDivBroadCastGradFunctor<T>,
-                           ElementwiseDivBroadCast2GradFunctor<T>>(ctx);
+                           ElementwiseDivBroadCast2GradFunctor<T>>(
+        ctx, x, y, out, dout, axis, dx, dy);
   }
 };
 

paddle/operators/elementwise_max_op.h

@@ -28,7 +28,14 @@ template <typename DeviceContext, typename T>
 class ElementwiseMaxKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseComputeEx<MaxFunctor<T>, DeviceContext, T>(ctx);
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* z = ctx.Output<Tensor>("Out");
+    z->mutable_data<T>(ctx.GetPlace());
+    int axis = ctx.Attr<int>("axis");
+    ElementwiseComputeEx<MaxFunctor<T>, DeviceContext, T>(ctx, x, y, axis, z);
   }
 };
 
@@ -110,9 +117,19 @@ template <typename DeviceContext, typename T>
 class ElementwiseMaxGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* out = ctx.Input<Tensor>("Out");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    int axis = ctx.Attr<int>("axis");
     ElementwiseGradCompute<DeviceContext, T, ElementwiseMaxGradFunctor<T>,
                            ElementwiseMaxBroadCastGradFunctor<T>,
-                           ElementwiseMaxBroadCast2GradFunctor<T>>(ctx);
+                           ElementwiseMaxBroadCast2GradFunctor<T>>(
+        ctx, x, y, out, dout, axis, dx, dy);
   }
 };
 

paddle/operators/elementwise_min_op.h

@@ -28,7 +28,14 @@ template <typename DeviceContext, typename T>
 class ElementwiseMinKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseComputeEx<MinFunctor<T>, DeviceContext, T>(ctx);
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* z = ctx.Output<Tensor>("Out");
+    z->mutable_data<T>(ctx.GetPlace());
+    int axis = ctx.Attr<int>("axis");
+    ElementwiseComputeEx<MinFunctor<T>, DeviceContext, T>(ctx, x, y, axis, z);
   }
 };
 
@@ -110,9 +117,19 @@ template <typename DeviceContext, typename T>
 class ElementwiseMinGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* out = ctx.Input<Tensor>("Out");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    int axis = ctx.Attr<int>("axis");
     ElementwiseGradCompute<DeviceContext, T, ElementwiseMinGradFunctor<T>,
                            ElementwiseMinBroadCastGradFunctor<T>,
-                           ElementwiseMinBroadCast2GradFunctor<T>>(ctx);
+                           ElementwiseMinBroadCast2GradFunctor<T>>(
+        ctx, x, y, out, dout, axis, dx, dy);
   }
 };
 

paddle/operators/elementwise_mul_op.h

@@ -27,7 +27,14 @@ template <typename DeviceContext, typename T>
 class ElementwiseMulKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseComputeEx<MulFunctor<T>, DeviceContext, T>(ctx);
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* z = ctx.Output<Tensor>("Out");
+    z->mutable_data<T>(ctx.GetPlace());
+    int axis = ctx.Attr<int>("axis");
+    ElementwiseComputeEx<MulFunctor<T>, DeviceContext, T>(ctx, x, y, axis, z);
   }
 };
 
@@ -110,9 +117,19 @@ template <typename DeviceContext, typename T>
 class ElementwiseMulGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* out = ctx.Input<Tensor>("Out");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    int axis = ctx.Attr<int>("axis");
     ElementwiseGradCompute<DeviceContext, T, ElementwiseMulGradFunctor<T>,
                            ElementwiseMulBroadCastGradFunctor<T>,
-                           ElementwiseMulBroadCast2GradFunctor<T>>(ctx);
+                           ElementwiseMulBroadCast2GradFunctor<T>>(
+        ctx, x, y, out, dout, axis, dx, dy);
   }
 };
 

paddle/operators/elementwise_op_function.h

@@ -313,21 +313,18 @@ EIGEN_FUNCTOR(Div, EIGEN_DIV);
 
 template <typename DeviceContext, typename T, typename functor,
           typename broadcastfunctor, typename broadcast2functor>
-void ElementwiseGradCompute(const framework::ExecutionContext& ctx) {
-  using Tensor = framework::Tensor;
-
-  auto* x = ctx.Input<Tensor>("X");
-  auto* y = ctx.Input<Tensor>("Y");
-  auto* out = ctx.Input<Tensor>("Out");
-  auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+void ElementwiseGradCompute(const framework::ExecutionContext& ctx,
 
+                            const framework::Tensor* x,
+                            const framework::Tensor* y,
+                            const framework::Tensor* out,
+                            const framework::Tensor* dout, int axis,
+                            framework::Tensor* dx, framework::Tensor* dy) {
   auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
 
   auto x_dims = x->dims();
   auto y_dims = y->dims();
 
-  auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
-  auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
   if (dx) {
     dx->mutable_data<T>(ctx.GetPlace());
   }
@@ -348,7 +345,6 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) {
     x_dims = framework::make_ddim(extended_dims);
   }
 
-  int axis = ctx.Attr<int>("axis");
   axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
 
   int pre, n, post;
@@ -367,13 +363,10 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) {
 
 template <typename Functor, typename DeviceContext, typename T,
           typename OutType = T>
-void ElementwiseComputeEx(const framework::ExecutionContext& ctx) {
-  using Tensor = framework::Tensor;
-
-  auto* x = ctx.Input<Tensor>("X");
-  auto* y = ctx.Input<Tensor>("Y");
-  auto* z = ctx.Output<Tensor>("Out");
-  z->mutable_data<OutType>(ctx.GetPlace());
+void ElementwiseComputeEx(const framework::ExecutionContext& ctx,
+                          const framework::Tensor* x,
+                          const framework::Tensor* y, int axis,
+                          framework::Tensor* z) {
   TransformFunctor<Functor, T, DeviceContext, OutType> functor(
       x, y, z, ctx.template device_context<DeviceContext>(), Functor());
 
@@ -394,7 +387,6 @@ void ElementwiseComputeEx(const framework::ExecutionContext& ctx) {
     x_dims = framework::make_ddim(extended_dims);
   }
 
-  int axis = ctx.Attr<int>("axis");
   axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
   PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
                  "Axis should be in range [0, x_dims)");

paddle/operators/elementwise_pow_op.h

@@ -29,7 +29,14 @@ template <typename DeviceContext, typename T>
 class ElementwisePowKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseComputeEx<PowFunctor<T>, DeviceContext, T>(ctx);
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* z = ctx.Output<Tensor>("Out");
+    z->mutable_data<T>(ctx.GetPlace());
+    int axis = ctx.Attr<int>("axis");
+    ElementwiseComputeEx<PowFunctor<T>, DeviceContext, T>(ctx, x, y, axis, z);
   }
 };
 

paddle/operators/elementwise_sub_op.h

@@ -27,7 +27,14 @@ template <typename DeviceContext, typename T>
 class ElementwiseSubKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    ElementwiseComputeEx<SubFunctor<T>, DeviceContext, T>(ctx);
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* z = ctx.Output<Tensor>("Out");
+    z->mutable_data<T>(ctx.GetPlace());
+    int axis = ctx.Attr<int>("axis");
+    ElementwiseComputeEx<SubFunctor<T>, DeviceContext, T>(ctx, x, y, axis, z);
   }
 };
 
@@ -93,9 +100,19 @@ template <typename DeviceContext, typename T>
 class ElementwiseSubGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* out = ctx.Input<Tensor>("Out");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    int axis = ctx.Attr<int>("axis");
     ElementwiseGradCompute<DeviceContext, T, ElementwiseSubGradFunctor<T>,
                            ElementwiseSubBroadCastGradFunctor<T>,
-                           ElementwiseSubBroadCast2GradFunctor<T>>(ctx);
+                           ElementwiseSubBroadCast2GradFunctor<T>>(
+        ctx, x, y, out, dout, axis, dx, dy);
   }
 };
 

paddle/operators/math/CMakeLists.txt

@@ -8,6 +8,7 @@ if(WITH_GPU)
     nv_library(softmax SRCS softmax.cc softmax.cu DEPS device_context)
     nv_library(cross_entropy SRCS cross_entropy.cc cross_entropy.cu DEPS device_context)
     nv_library(pooling SRCS pooling.cc pooling.cu DEPS device_context)
+    nv_library(depthwise_conv SRCS depthwise_conv.cu DEPS device_context)
     nv_library(sequence_pooling SRCS sequence_pooling.cc sequence_pooling.cu DEPS device_context math_function)
     nv_library(vol2col SRCS vol2col.cc vol2col.cu DEPS device_context tensor)
     nv_library(context_project SRCS context_project.cc context_project.cu DEPS device_context math_function)

paddle/operators/math/depthwise_conv.cu

+/* Copyright (c) 2016 paddlepaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/math/depthwise_conv.h"
+#include "paddle/platform/cuda_helper.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+// A Cuda kernel to compute the depthwise convolution forward pass
+// in NCHW format.
+template <typename T>
+__global__ void KernelDepthwiseConv(
+    const int nthreads, const T* const input_data, const T* const filter_data,
+    const int batch_size, const int output_channels, const int output_height,
+    const int output_width, const int input_channels, const int input_height,
+    const int input_width, const int filter_multiplier, const int filter_height,
+    const int filter_width, const int stride_height, const int stride_width,
+    const int padding_height, const int padding_width, T* const output_data) {
+  int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
+
+  if (index < nthreads) {
+    const int batch = index / output_channels / output_height / output_width;
+    const int c_out = (index / output_height / output_width) % output_channels;
+    const int h_out = (index / output_width) % output_height;
+    const int w_out = index % output_width;
+
+    const int c_in = c_out / filter_multiplier;
+    const T* weight = filter_data + c_out * filter_height * filter_width;
+    T value = 0;
+    const int h_in_start = -padding_height + h_out * stride_height;
+    const int w_in_start = -padding_width + w_out * stride_width;
+    const int h_in_end = h_in_start + filter_height;
+    const int w_in_end = w_in_start + filter_width;
+
+    const int in_offset =
+        ((batch * input_channels + c_in) * input_height) * input_width;
+
+    const int h_end = h_in_end < input_height ? h_in_end : input_height;
+    const int w_end = w_in_end < input_width ? w_in_end : input_width;
+    const int h_start = h_in_start > 0 ? h_in_start : 0;
+    const int w_start = w_in_start > 0 ? w_in_start : 0;
+
+    for (int h_in = h_start; h_in < h_end; h_in++) {
+      for (int w_in = w_start; w_in < w_end; w_in++) {
+        const int offset = in_offset + h_in * input_width + w_in;
+        value +=
+            weight[(h_in - h_in_start) * filter_width + (w_in - w_in_start)] *
+            input_data[offset];
+      }
+    }
+    output_data[index] = value;
+  }
+}
+
+// CUDA kernel to compute the depthwise convolution backprop w.r.t input.
+template <typename T>
+__global__ void KernelDepthwiseConvInputGrad(
+    const int nthreads, const T* const output_grad_data,
+    const T* const filter_data, const int batch_size, const int output_channels,
+    const int output_height, const int output_width, const int input_channels,
+    const int input_height, const int input_width, const int filter_multiplier,
+    const int filter_height, const int filter_width, const int stride_height,
+    const int stride_width, const int padding_height, const int padding_width,
+    T* const input_grad_data) {
+  int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
+  if (index < nthreads) {
+    const int batch = index / input_channels / input_height / input_width;
+    const int c_in = (index / input_height / input_width) % input_channels;
+    const int h_in = (index / input_width) % input_height;
+    const int w_in = index % input_width;
+
+    const int c_out_start = c_in * filter_multiplier;
+
+    int h_out_start =
+        (h_in - filter_height + padding_height + stride_height) / stride_height;
+    h_out_start = 0 > h_out_start ? 0 : h_out_start;
+
+    int h_out_end = (h_in + padding_height) / stride_height;
+    h_out_end = output_height - 1 < h_out_end ? output_height - 1 : h_out_end;
+
+    int w_out_start =
+        (w_in - filter_width + padding_width + stride_width) / stride_width;
+    w_out_start = 0 > w_out_start ? 0 : w_out_start;
+
+    int w_out_end = (w_in + padding_width) / stride_width;
+    w_out_end = output_width - 1 < w_out_end ? output_width - 1 : w_out_end;
+
+    T value = 0;
+
+    for (int c_out = c_out_start; c_out < c_out_start + filter_multiplier;
+         c_out++) {
+      for (int h_out = h_out_start; h_out <= h_out_end; ++h_out) {
+        const int filter_h = h_in + padding_height - h_out * stride_height;
+        for (int w_out = w_out_start; w_out <= w_out_end; ++w_out) {
+          const int filter_w = w_in + padding_width - w_out * stride_width;
+          const int filter_offset = c_out * filter_height * filter_width +
+                                    filter_h * filter_width + filter_w;
+          const int output_grad_offset =
+              ((batch * output_channels + c_out) * output_height + h_out) *
+                  output_width +
+              w_out;
+          value +=
+              output_grad_data[output_grad_offset] * filter_data[filter_offset];
+        }
+      }
+    }
+    input_grad_data[index] += value;
+  }
+}
+
+// Cuda kernel to compute the depthwise convolution backprop w.r.t. filter.
+template <typename T>
+__global__ void KernelDepthwiseConvFilterGrad(
+    const int nthreads, const T* const output_grad_data,
+    const T* const input_data, const int num, const int output_channels,
+    const int output_height, const int output_width, const int input_channels,
+    const int input_height, const int input_width, const int filter_multiplier,
+    const int filter_height, const int filter_width, const int stride_height,
+    const int stride_width, const int padding_height, const int padding_width,
+    T* const filter_grad_data) {
+  int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
+  if (index < nthreads) {
+    const int w_out = index % output_width;
+    const int h_out = (index / output_width) % output_height;
+    const int c_out = (index / output_width / output_height) % output_channels;
+    const int batch = (index / output_width / output_height / output_channels);
+    const int c_in = c_out / filter_multiplier;
+    const int h_in_start = -padding_height + h_out * stride_height;
+    const int w_in_start = -padding_width + w_out * stride_width;
+    const int h_in_end =
+        -padding_height + h_out * stride_height + filter_height;
+    const int w_in_end = -padding_width + w_out * stride_width + filter_width;
+    const int in_offset =
+        (batch * input_channels + c_in) * input_height * input_width;
+
+    T* addr_offset = filter_grad_data + c_out * filter_height * filter_width;
+    const int h_end = h_in_end < input_height ? h_in_end : input_height;
+    const int w_end = w_in_end < input_width ? w_in_end : input_width;
+    const int h_start = h_in_start > 0 ? h_in_start : 0;
+    const int w_start = w_in_start > 0 ? w_in_start : 0;
+
+    for (int h_in = h_start; h_in < h_end; h_in++) {
+      for (int w_in = w_start; w_in < w_end; w_in++) {
+        const int offset = in_offset + h_in * input_width + w_in;
+        const T diff_temp = output_grad_data[index] * input_data[offset];
+        T* addr = addr_offset + (h_in - h_in_start) * filter_width +
+                  (w_in - w_in_start);
+        paddle::platform::CudaAtomicAdd(addr, diff_temp);
+      }
+    }
+  }
+}
+
+/*
+ * All tensors are in NCHW format.
+ * Ksize, strides, paddings are two elements. These two elements represent
+ * height and width, respectively.
+ */
+template <class T>
+class DepthwiseConvFunctor<platform::CUDADeviceContext, T> {
+ public:
+  void operator()(const platform::CUDADeviceContext& context,
+                  const framework::Tensor& input,
+                  const framework::Tensor& filter,
+                  const std::vector<int>& strides,
+                  const std::vector<int>& paddings, framework::Tensor* output) {
+    const int batch_size = input.dims()[0];
+    const int input_channels = input.dims()[1];
+    const int input_height = input.dims()[2];
+    const int input_width = input.dims()[3];
+    const int output_channels = output->dims()[1];
+    const int output_height = output->dims()[2];
+    const int output_width = output->dims()[3];
+    const int ksize_height = filter.dims()[2];
+    const int ksize_width = filter.dims()[3];
+    const int stride_height = strides[0];
+    const int stride_width = strides[1];
+    const int padding_height = paddings[0];
+    const int padding_width = paddings[1];
+
+    const T* input_data = input.data<T>();
+    const T* filter_data = filter.data<T>();
+    T* output_data = output->mutable_data<T>(context.GetPlace());
+
+    int nthreads = batch_size * output_channels * output_height * output_width;
+    int blocks = (nthreads + 1024 - 1) / 1024;
+    dim3 threads(1024, 1);
+    dim3 grid(blocks, 1);
+
+    KernelDepthwiseConv<T><<<grid, threads, 0, context.stream()>>>(
+        nthreads, input_data, filter_data, batch_size, output_channels,
+        output_height, output_width, input_channels, input_height, input_width,
+        output_channels / input_channels, ksize_height, ksize_width,
+        stride_height, stride_width, padding_height, padding_width,
+        output_data);
+  }
+};
+
+template <typename T>
+class DepthwiseConvInputGradFunctor<platform::CUDADeviceContext, T> {
+ public:
+  void operator()(const platform::CUDADeviceContext& context,
+                  const framework::Tensor& input,
+                  const framework::Tensor& filter,
+                  const framework::Tensor& output_grad,
+                  const std::vector<int>& strides,
+                  const std::vector<int>& paddings,
+                  framework::Tensor* input_grad) {
+    const int batch_size = input.dims()[0];
+    const int input_channels = input.dims()[1];
+    const int input_height = input.dims()[2];
+    const int input_width = input.dims()[3];
+    const int output_channels = output_grad.dims()[1];
+    const int output_height = output_grad.dims()[2];
+    const int output_width = output_grad.dims()[3];
+    const int ksize_height = filter.dims()[2];
+    const int ksize_width = filter.dims()[3];
+    const int stride_height = strides[0];
+    const int stride_width = strides[1];
+    const int padding_height = paddings[0];
+    const int padding_width = paddings[1];
+
+    const T* filter_data = filter.data<T>();
+    const T* output_grad_data = output_grad.data<T>();
+    T* input_grad_data = input_grad->mutable_data<T>(context.GetPlace());
+
+    int nthreads = batch_size * input_channels * input_height * input_width;
+    int blocks = (nthreads + 1024 - 1) / 1024;
+    dim3 threads(1024, 1);
+    dim3 grid(blocks, 1);
+
+    KernelDepthwiseConvInputGrad<T><<<grid, threads, 0, context.stream()>>>(
+        nthreads, output_grad_data, filter_data, batch_size, output_channels,
+        output_height, output_width, input_channels, input_height, input_width,
+        output_channels / input_channels, ksize_height, ksize_width,
+        stride_height, stride_width, padding_height, padding_width,
+        input_grad_data);
+  }
+};
+
+template <typename T>
+class DepthwiseConvFilterGradFunctor<platform::CUDADeviceContext, T> {
+ public:
+  void operator()(const platform::CUDADeviceContext& context,
+                  const framework::Tensor& input,
+                  const framework::Tensor& output_grad,
+                  const std::vector<int>& strides,
+                  const std::vector<int>& paddings,
+                  framework::Tensor* filter_grad) {
+    const int batch_size = input.dims()[0];
+    const int input_channels = input.dims()[1];
+    const int input_height = input.dims()[2];
+    const int input_width = input.dims()[3];
+    const int output_channels = output_grad.dims()[1];
+    const int output_height = output_grad.dims()[2];
+    const int output_width = output_grad.dims()[3];
+    const int ksize_height = filter_grad->dims()[2];
+    const int ksize_width = filter_grad->dims()[3];
+    const int stride_height = strides[0];
+    const int stride_width = strides[1];
+    const int padding_height = paddings[0];
+    const int padding_width = paddings[1];
+
+    const T* input_data = input.data<T>();
+    const T* output_grad_data = output_grad.data<T>();
+    T* filter_grad_data = filter_grad->mutable_data<T>(context.GetPlace());
+
+    int nthreads = batch_size * output_channels * output_height * output_width;
+
+    int blocks = (nthreads + 1024 - 1) / 1024;
+    dim3 threads(1024, 1);
+    dim3 grid(blocks, 1);
+
+    KernelDepthwiseConvFilterGrad<T><<<grid, threads, 0, context.stream()>>>(
+        nthreads, output_grad_data, input_data, batch_size, output_channels,
+        output_height, output_width, input_channels, input_height, input_width,
+        output_channels / input_channels, ksize_height, ksize_width,
+        stride_height, stride_width, padding_height, padding_width,
+        filter_grad_data);
+  }
+};
+
+template class DepthwiseConvFunctor<platform::CUDADeviceContext, float>;
+template class DepthwiseConvFunctor<platform::CUDADeviceContext, double>;
+
+template class DepthwiseConvInputGradFunctor<platform::CUDADeviceContext,
+                                             float>;
+template class DepthwiseConvInputGradFunctor<platform::CUDADeviceContext,
+                                             double>;
+
+template class DepthwiseConvFilterGradFunctor<platform::CUDADeviceContext,
+                                              float>;
+template class DepthwiseConvFilterGradFunctor<platform::CUDADeviceContext,
+                                              double>;
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/depthwise_conv.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/device_context.h"
+#include "paddle/platform/hostdevice.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+/*
+ * \brief Compute the depthwise convolution which include
+ * forward process and backpropagation process
+ */
+template <typename DeviceContext, typename T>
+class DepthwiseConvFunctor {
+ public:
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  const framework::Tensor& filter,
+                  const std::vector<int>& strides,
+                  const std::vector<int>& paddings, framework::Tensor* output);
+};
+
+template <typename DeviceContext, typename T>
+class DepthwiseConvInputGradFunctor {
+ public:
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  const framework::Tensor& filter,
+                  const framework::Tensor& output_grad,
+                  const std::vector<int>& strides,
+                  const std::vector<int>& paddings,
+                  framework::Tensor* input_grad);
+};
+
+template <typename DeviceContext, typename T>
+class DepthwiseConvFilterGradFunctor {
+ public:
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  const framework::Tensor& output_grad,
+                  const std::vector<int>& strides,
+                  const std::vector<int>& paddings,
+                  framework::Tensor* filter_grad);
+};
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/mine_hard_examples_op.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/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+enum MiningType { kNone = 0, kMaxNegative, kHardExample };
+
+template <typename T>
+bool SortScoreDescend(const std::pair<float, T>& pair1,
+                      const std::pair<float, T>& pair2) {
+  return pair1.first > pair2.first;
+}
+
+inline bool IsEligibleMining(const MiningType mining_type, const int match_idx,
+                             const float match_dist,
+                             const float neg_dist_threshold) {
+  if (mining_type == MiningType::kMaxNegative) {
+    return match_idx == -1 && match_dist < neg_dist_threshold;
+  } else if (mining_type == MiningType::kHardExample) {
+    return true;
+  } else {
+    return false;
+  }
+}
+
+inline MiningType GetMiningType(std::string str) {
+  if (str == "max_negative") {
+    return MiningType::kMaxNegative;
+  } else if (str == "hard_example") {
+    return MiningType::kHardExample;
+  } else {
+    return MiningType::kNone;
+  }
+}
+
+template <typename DeviceContext, typename T>
+class MineHardExamplesKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in_cls_loss = ctx.Input<framework::Tensor>("ClsLoss");
+    auto* in_loc_loss = ctx.Input<framework::Tensor>("LocLoss");
+    auto* in_matched_indices = ctx.Input<framework::Tensor>("MatchIndices");
+    auto* in_match_dist = ctx.Input<framework::Tensor>("MatchDist");
+    float neg_pos_ratio = ctx.Attr<float>("neg_pos_ratio");
+    T neg_dist_threshold =
+        static_cast<T>(ctx.Attr<float>("neg_dist_threshold"));
+    int sample_size = ctx.Attr<int>("sample_size");
+    MiningType mining_type =
+        GetMiningType(ctx.Attr<std::string>("mining_type"));
+
+    auto out_neg_indices = ctx.Output<framework::LoDTensor>("NegIndices");
+    auto out_match_indices =
+        ctx.Output<framework::Tensor>("UpdatedMatchIndices");
+
+    framework::Copy(*in_matched_indices, ctx.GetPlace(), out_match_indices);
+
+    int batch_size = in_matched_indices->dims()[0];
+    int prior_num = in_matched_indices->dims()[1];
+
+    auto match_indices = framework::EigenMatrix<int>::From(*in_matched_indices);
+
+    auto match_indices_et =
+        framework::EigenMatrix<int>::From(*out_match_indices);
+
+    auto match_dist = framework::EigenMatrix<T>::From(*in_match_dist);
+
+    const T* cls_loss = in_cls_loss->data<T>();
+    const T* loc_loss = nullptr;
+    if (in_loc_loss) {
+      loc_loss = in_loc_loss->data<T>();
+    }
+
+    std::vector<std::vector<int>> all_neg_indices;
+    std::vector<size_t> batch_starts = {0};
+    for (int n = 0; n < batch_size; ++n) {
+      std::vector<std::pair<T, size_t>> loss_idx;
+      int neg_sel = 0;
+      for (int m = 0; m < prior_num; ++m) {
+        if (IsEligibleMining(mining_type, match_indices(n, m), match_dist(n, m),
+                             neg_dist_threshold)) {
+          T loss = cls_loss[n * prior_num + m];
+          if (mining_type == MiningType::kHardExample && loc_loss != nullptr) {
+            loss = cls_loss[n * prior_num + m] + loc_loss[n * prior_num + m];
+          }
+          loss_idx.push_back(std::make_pair(loss, m));
+          ++neg_sel;
+        }
+      }
+
+      if (mining_type == MiningType::kMaxNegative) {
+        int num_pos = 0;
+        for (int m = 0; m < prior_num; ++m) {
+          if (match_indices(n, m) != -1) ++num_pos;
+        }
+        neg_sel = std::min(static_cast<int>(num_pos * neg_pos_ratio), neg_sel);
+      } else if (mining_type == MiningType::kHardExample) {
+        neg_sel = std::min(sample_size, neg_sel);
+      }
+
+      std::sort(loss_idx.begin(), loss_idx.end(), SortScoreDescend<size_t>);
+      std::set<int> sel_indices;
+      std::vector<int> neg_indices;
+      std::transform(loss_idx.begin(), loss_idx.begin() + neg_sel,
+                     std::inserter(sel_indices, sel_indices.begin()),
+                     [](std::pair<T, size_t>& l) -> int {
+                       return static_cast<int>(l.second);
+                     });
+
+      if (mining_type == MiningType::kHardExample) {
+        for (int m = 0; m < prior_num; ++m) {
+          if (match_indices(n, m) > -1) {
+            if (sel_indices.find(m) == sel_indices.end()) {
+              match_indices_et(n, m) = -1;
+            }
+          } else {
+            if (sel_indices.find(m) != sel_indices.end()) {
+              neg_indices.push_back(m);
+            }
+          }
+        }
+      } else {
+        neg_indices.resize(sel_indices.size());
+        std::copy(sel_indices.begin(), sel_indices.end(), neg_indices.begin());
+      }
+
+      all_neg_indices.push_back(neg_indices);
+      batch_starts.push_back(batch_starts.back() + neg_indices.size());
+    }
+
+    framework::LoD out_neg_indices_lod;
+    out_neg_indices_lod.emplace_back(batch_starts);
+    int neg_offset = 0;
+    auto neg_data = out_neg_indices->mutable_data<int>(
+        framework::make_ddim({static_cast<int>(batch_starts.back()), 1}),
+        ctx.GetPlace());
+
+    for (auto neg_indices : all_neg_indices) {
+      std::copy(neg_indices.begin(), neg_indices.end(), neg_data + neg_offset);
+      neg_offset += neg_indices.size();
+    }
+    out_neg_indices->set_lod(out_neg_indices_lod);
+    return;
+  }
+};
+
+class MineHardExamplesOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("ClsLoss"),
+                   "Input(ClsLoss) of MineHardExamplesOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasInput("MatchIndices"),
+        "Input(MatchIndices) of MineHardExamplesOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasInput("MatchDist"),
+        "Input(MatchDist) of MineHardExamplesOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("NegIndices"),
+        "Output(NegIndices) of MineHardExamplesOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("UpdatedMatchIndices"),
+                   "Output(UpdatedMatchIndices) of MineHardExamplesOp should "
+                   "not be null.");
+
+    auto cls_loss_dims = ctx->GetInputDim("ClsLoss");
+    auto idx_dims = ctx->GetInputDim("MatchIndices");
+    auto dis_dims = ctx->GetInputDim("MatchDist");
+
+    PADDLE_ENFORCE_EQ(cls_loss_dims.size(), 2UL,
+                      "The shape of ClsLoss is [N, Np].");
+    PADDLE_ENFORCE_EQ(idx_dims.size(), 2UL,
+                      "The shape of MatchIndices is [N, Np].");
+    PADDLE_ENFORCE_EQ(dis_dims.size(), 2UL,
+                      "The shape of MatchDist is [N, Np].");
+
+    if (ctx->HasInput("LocLoss")) {
+      auto loc_loss_dims = ctx->GetInputDim("LocLoss");
+      PADDLE_ENFORCE_EQ(loc_loss_dims.size(), 2UL,
+                        "The shape of LocLoss is [N, Np].");
+      PADDLE_ENFORCE_EQ(cls_loss_dims[0], loc_loss_dims[0],
+                        "Batch size of ClsLoss and LocLoss must be the same.");
+      PADDLE_ENFORCE_EQ(
+          cls_loss_dims[1], loc_loss_dims[1],
+          "Prior box number of ClsLoss and LocLoss must be the same.");
+    }
+
+    PADDLE_ENFORCE_EQ(
+        cls_loss_dims[0], idx_dims[0],
+        "Batch size of ClsLoss and MatchIndices must be the same.");
+    PADDLE_ENFORCE_EQ(
+        cls_loss_dims[1], idx_dims[1],
+        "Prior box number of ClsLoss and MatchIndices must be the same.");
+
+    PADDLE_ENFORCE_EQ(cls_loss_dims[0], dis_dims[0],
+                      "Batch size of ClsLoss and MatchDist must be the same.");
+    PADDLE_ENFORCE_EQ(
+        cls_loss_dims[1], idx_dims[1],
+        "Prior box number of ClsLoss and MatchDist must be the same.");
+
+    auto mining_type =
+        GetMiningType(ctx->Attrs().Get<std::string>("mining_type"));
+
+    PADDLE_ENFORCE_NE(mining_type, MiningType::kNone,
+                      "mining_type must be hard_example or max_negative");
+
+    if (mining_type == MiningType::kMaxNegative) {
+      auto neg_pos_ratio = ctx->Attrs().Get<float>("neg_pos_ratio");
+      auto neg_dist_threshold = ctx->Attrs().Get<float>("neg_dist_threshold");
+      PADDLE_ENFORCE_GT(
+          neg_pos_ratio, 0.0f,
+          "neg_pos_ratio must greater than zero in max_negative mode");
+      PADDLE_ENFORCE_GT(
+          neg_dist_threshold, 0.0f,
+          "neg_dist_threshold must greater than zero in max_negative mode");
+    } else if (mining_type == MiningType::kHardExample) {
+      auto sample_size = ctx->Attrs().Get<int>("sample_size");
+      PADDLE_ENFORCE_GT(
+          sample_size, 0,
+          "sample_size must greater than zero in hard_example mode");
+    }
+
+    ctx->SetOutputDim("UpdatedMatchIndices", idx_dims);
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<framework::Tensor>("ClsLoss")->type()),
+        ctx.device_context());
+  }
+};
+
+class MineHardExamplesOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  MineHardExamplesOpMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput(
+        "ClsLoss",
+        "(Tensor, default Tensor<float>), The classification loss with shape "
+        "[N, Np], N is the batch size and Np is the number of prior box.");
+    AddInput("LocLoss",
+             "(Tensor, optional, default Tensor<float>), The localization loss "
+             "with shape [N, Np], N is the batch size and Np is the number of "
+             "prior box.")
+        .AsDispensable();
+    AddInput("MatchIndices",
+             "(Tensor, Tensor<int>), Matched indices with shape [N, Np], N is "
+             "the batch size and Np is the number of prior box. "
+             "MatchIndices[i][j] equal -1 means the j-th prior box in i-th "
+             "instance does not match any entity, otherwise means it is "
+             "matched to row.");
+    AddInput("MatchDist",
+             "(Tensor, default Tensor<float>) Matched indices with shape [N, "
+             "Np], N is the batch size and Np is the number of prior box.");
+    AddAttr<float>("neg_pos_ratio",
+                   "(float) The ratio of the negative box to the positive "
+                   "box. Use only when mining_type is max_negative.")
+        .SetDefault(1.0);
+    AddAttr<float>("neg_dist_threshold",
+                   "(float) The negative overlap upper bound for the unmatched "
+                   "predictions. Use only when mining_type is max_negative.")
+        .SetDefault(0.5);
+    AddAttr<int>("sample_size",
+                 "(float) The max sample size of negative box. Use only when "
+                 "mining_type is hard_example.")
+        .SetDefault(0);
+    AddAttr<std::string>("mining_type",
+                         "(float) The mining algorithm name, the value is "
+                         "hard_example or max_negative.")
+        .SetDefault("max_negative")
+        .InEnum({"hard_example", "max_negative"});
+
+    AddOutput(
+        "NegIndices",
+        "(LoDTensor<int>) The output of negative example indices. a LoDTensor "
+        "with shape [Neg, 1]. The size of lod[0] minus 1 is batch size, "
+        "and each element is the prior box index. "
+        "For example, the batch size is 2, the lod is [[0, 1, 2]], "
+        "the sample 0's box 1(MatchIndices[0][1]) is selected, "
+        "and sample 1's box 0 is selected. The output NegIndices is "
+        "[[1], [0]].");
+
+    AddOutput("UpdatedMatchIndices",
+              "(Tensor<int>) The output of updated MatchIndices, a tensor with "
+              "shape [N, Np]. Only update when mining_type is "
+              "hard_example. The input MatchIndices elements will be update to "
+              "-1 when it is not in the candidate high loss list of negative "
+              "examples.");
+
+    AddComment(R"DOC(
+Mine hard examples Operator.
+This operator implements hard example mining to select a subset of negative box indices.
+For each image, selects the box with highest losses. subject to the condition that the 
+box cannot have an Matcht > neg_dist_threshold when mining_type is max_negative. 
+The selected number is min(sample_size, max_negative_box_number) when mining_type is 
+hard_example, or min(neg_pos_ratio * positive_box_number, max_negative_box_number) 
+when mining_type is max_negative, where the max_negative_box_number is the count of 
+MatchIndices elements with value -1.
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(mine_hard_examples, ops::MineHardExamplesOp,
+                             ops::MineHardExamplesOpMaker);
+
+REGISTER_OP_CPU_KERNEL(
+    mine_hard_examples,
+    ops::MineHardExamplesKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::MineHardExamplesKernel<paddle::platform::CPUDeviceContext, double>);

paddle/operators/multiclass_nms_op.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/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+
+constexpr int64_t kOutputDim = 6;
+constexpr int64_t kBBoxSize = 4;
+
+class MultiClassNMSOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("BBoxes"),
+                   "Input(BBoxes) of MultiClassNMS should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Scores"),
+                   "Input(Scores) of MultiClassNMS should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of MultiClassNMS should not be null.");
+
+    auto box_dims = ctx->GetInputDim("BBoxes");
+    auto score_dims = ctx->GetInputDim("Scores");
+
+    PADDLE_ENFORCE_EQ(box_dims.size(), 2,
+                      "The rank of Input(BBoxes) must be 2.");
+    PADDLE_ENFORCE_EQ(score_dims.size(), 3,
+                      "The rank of Input(Scores) must be 3.");
+    PADDLE_ENFORCE_EQ(box_dims[1], 4,
+                      "The 2nd dimension of Input(BBoxes) must be 4, "
+                      "represents the layout of coordinate "
+                      "[xmin, ymin, xmax, ymax]");
+    PADDLE_ENFORCE_EQ(box_dims[0], score_dims[2],
+                      "The 1st dimensiong of Input(BBoxes) must be equal to "
+                      "3rd dimension of Input(Scores), which represents the "
+                      "predicted bboxes.");
+
+    // Here the box_dims[0] is not the real dimension of output.
+    // It will be rewritten in the computing kernel.
+    ctx->SetOutputDim("Out", {box_dims[0], 6});
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(
+            ctx.Input<framework::LoDTensor>("Scores")->type()),
+        ctx.device_context());
+  }
+};
+
+template <class T>
+bool SortScorePairDescend(const std::pair<float, T>& pair1,
+                          const std::pair<float, T>& pair2) {
+  return pair1.first > pair2.first;
+}
+
+template <class T>
+static inline void GetMaxScoreIndex(
+    const std::vector<T>& scores, const T threshold, int top_k,
+    std::vector<std::pair<T, int>>* sorted_indices) {
+  for (size_t i = 0; i < scores.size(); ++i) {
+    if (scores[i] > threshold) {
+      sorted_indices->push_back(std::make_pair(scores[i], i));
+    }
+  }
+  // Sort the score pair according to the scores in descending order
+  std::stable_sort(sorted_indices->begin(), sorted_indices->end(),
+                   SortScorePairDescend<int>);
+  // Keep top_k scores if needed.
+  if (top_k > -1 && top_k < static_cast<int>(sorted_indices->size())) {
+    sorted_indices->resize(top_k);
+  }
+}
+
+template <class T>
+static inline T BBoxArea(const T* box, const bool normalized) {
+  if (box[2] < box[0] || box[3] < box[1]) {
+    // If coordinate values are is invalid
+    // (e.g. xmax < xmin or ymax < ymin), return 0.
+    return static_cast<T>(0.);
+  } else {
+    const T w = box[2] - box[0];
+    const T h = box[3] - box[1];
+    if (normalized) {
+      return w * h;
+    } else {
+      // If coordinate values are not within range [0, 1].
+      return (w + 1) * (h + 1);
+    }
+  }
+}
+
+template <class T>
+static inline T JaccardOverlap(const T* box1, const T* box2,
+                               const bool normalized) {
+  if (box2[0] > box1[2] || box2[2] < box1[0] || box2[1] > box1[3] ||
+      box2[3] < box1[1]) {
+    return static_cast<T>(0.);
+  } else {
+    const T inter_xmin = std::max(box1[0], box2[0]);
+    const T inter_ymin = std::max(box1[1], box2[1]);
+    const T inter_xmax = std::min(box1[2], box2[2]);
+    const T inter_ymax = std::min(box1[3], box2[3]);
+    const T inter_w = inter_xmax - inter_xmin;
+    const T inter_h = inter_ymax - inter_ymin;
+    const T inter_area = inter_w * inter_h;
+    const T bbox1_area = BBoxArea<T>(box1, normalized);
+    const T bbox2_area = BBoxArea<T>(box2, normalized);
+    return inter_area / (bbox1_area + bbox2_area - inter_area);
+  }
+}
+
+template <typename T>
+class MultiClassNMSKernel : public framework::OpKernel<T> {
+ public:
+  void NMSFast(const Tensor& bbox, const Tensor& scores,
+               const T score_threshold, const T nms_threshold, const T eta,
+               const int64_t top_k, std::vector<int>* selected_indices) const {
+    // The total boxes for each instance.
+    int64_t num_boxes = bbox.dims()[0];
+    // 4: [xmin ymin xmax ymax]
+    int64_t box_size = bbox.dims()[1];
+
+    std::vector<T> scores_data(num_boxes);
+    std::copy_n(scores.data<T>(), num_boxes, scores_data.begin());
+    std::vector<std::pair<T, int>> sorted_indices;
+    GetMaxScoreIndex(scores_data, score_threshold, top_k, &sorted_indices);
+
+    selected_indices->clear();
+    T adaptive_threshold = nms_threshold;
+    const T* bbox_data = bbox.data<T>();
+
+    while (sorted_indices.size() != 0) {
+      const int idx = sorted_indices.front().second;
+      bool keep = true;
+      for (size_t k = 0; k < selected_indices->size(); ++k) {
+        if (keep) {
+          const int kept_idx = (*selected_indices)[k];
+          T overlap = JaccardOverlap<T>(bbox_data + idx * box_size,
+                                        bbox_data + kept_idx * box_size, true);
+          keep = overlap <= adaptive_threshold;
+        } else {
+          break;
+        }
+      }
+      if (keep) {
+        selected_indices->push_back(idx);
+      }
+      sorted_indices.erase(sorted_indices.begin());
+      if (keep && eta < 1 && adaptive_threshold > 0.5) {
+        adaptive_threshold *= eta;
+      }
+    }
+  }
+
+  void MultiClassNMS(const framework::ExecutionContext& ctx,
+                     const Tensor& scores, const Tensor& bboxes,
+                     std::map<int, std::vector<int>>& indices,
+                     int& num_nmsed_out) const {
+    int64_t background_label = ctx.Attr<int>("background_label");
+    int64_t nms_top_k = ctx.Attr<int>("nms_top_k");
+    int64_t keep_top_k = ctx.Attr<int>("keep_top_k");
+    T nms_threshold = static_cast<T>(ctx.Attr<float>("nms_threshold"));
+    T nms_eta = static_cast<T>(ctx.Attr<float>("nms_eta"));
+    T score_threshold = static_cast<T>(ctx.Attr<float>("score_threshold"));
+
+    int64_t class_num = scores.dims()[0];
+    int64_t predict_dim = scores.dims()[1];
+    int num_det = 0;
+    for (int64_t c = 0; c < class_num; ++c) {
+      if (c == background_label) continue;
+      Tensor score = scores.Slice(c, c + 1);
+      NMSFast(bboxes, score, score_threshold, nms_threshold, nms_eta, nms_top_k,
+              &(indices[c]));
+      num_det += indices[c].size();
+    }
+
+    num_nmsed_out = num_det;
+    const T* scores_data = scores.data<T>();
+    if (keep_top_k > -1 && num_det > keep_top_k) {
+      std::vector<std::pair<float, std::pair<int, int>>> score_index_pairs;
+      for (const auto& it : indices) {
+        int label = it.first;
+        const T* sdata = scores_data + label * predict_dim;
+        const std::vector<int>& label_indices = it.second;
+        for (size_t j = 0; j < label_indices.size(); ++j) {
+          int idx = label_indices[j];
+          PADDLE_ENFORCE_LT(idx, predict_dim);
+          score_index_pairs.push_back(
+              std::make_pair(sdata[idx], std::make_pair(label, idx)));
+        }
+      }
+      // Keep top k results per image.
+      std::stable_sort(score_index_pairs.begin(), score_index_pairs.end(),
+                       SortScorePairDescend<std::pair<int, int>>);
+      score_index_pairs.resize(keep_top_k);
+
+      // Store the new indices.
+      std::map<int, std::vector<int>> new_indices;
+      for (size_t j = 0; j < score_index_pairs.size(); ++j) {
+        int label = score_index_pairs[j].second.first;
+        int idx = score_index_pairs[j].second.second;
+        new_indices[label].push_back(idx);
+      }
+      new_indices.swap(indices);
+      num_nmsed_out = keep_top_k;
+    }
+  }
+
+  void MultiClassOutput(const Tensor& scores, const Tensor& bboxes,
+                        std::map<int, std::vector<int>>& selected_indices,
+                        Tensor* outs) const {
+    int predict_dim = scores.dims()[1];
+    auto* scores_data = scores.data<T>();
+    auto* bboxes_data = bboxes.data<T>();
+    auto* odata = outs->data<T>();
+
+    int count = 0;
+    for (const auto& it : selected_indices) {
+      int label = it.first;
+      const T* sdata = scores_data + label * predict_dim;
+      const std::vector<int>& indices = it.second;
+      for (size_t j = 0; j < indices.size(); ++j) {
+        int idx = indices[j];
+        const T* bdata = bboxes_data + idx * kBBoxSize;
+        odata[count * kOutputDim] = label;           // label
+        odata[count * kOutputDim + 1] = sdata[idx];  // score
+        // xmin, ymin, xmax, ymax
+        std::memcpy(odata + count * kOutputDim + 2, bdata, 4 * sizeof(T));
+        count++;
+      }
+    }
+  }
+
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* boxes = ctx.Input<Tensor>("BBoxes");
+    auto* scores = ctx.Input<Tensor>("Scores");
+    auto* outs = ctx.Output<LoDTensor>("Out");
+
+    auto score_dims = scores->dims();
+
+    int64_t batch_size = score_dims[0];
+    int64_t class_num = score_dims[1];
+    int64_t predict_dim = score_dims[2];
+
+    std::vector<std::map<int, std::vector<int>>> all_indices;
+    std::vector<size_t> batch_starts = {0};
+    for (int64_t i = 0; i < batch_size; ++i) {
+      Tensor ins_score = scores->Slice(i, i + 1);
+      ins_score.Resize({class_num, predict_dim});
+      std::map<int, std::vector<int>> indices;
+      int num_nmsed_out = 0;
+      MultiClassNMS(ctx, ins_score, *boxes, indices, num_nmsed_out);
+      all_indices.push_back(indices);
+      batch_starts.push_back(batch_starts.back() + num_nmsed_out);
+    }
+
+    int num_kept = batch_starts.back();
+    if (num_kept == 0) {
+      T* od = outs->mutable_data<T>({1}, ctx.GetPlace());
+      od[0] = -1;
+    } else {
+      outs->mutable_data<T>({num_kept, kOutputDim}, ctx.GetPlace());
+      for (int64_t i = 0; i < batch_size; ++i) {
+        Tensor ins_score = scores->Slice(i, i + 1);
+        ins_score.Resize({class_num, predict_dim});
+        int64_t s = batch_starts[i];
+        int64_t e = batch_starts[i + 1];
+        if (e > s) {
+          Tensor out = outs->Slice(s, e);
+          MultiClassOutput(ins_score, *boxes, all_indices[i], &out);
+        }
+      }
+    }
+
+    framework::LoD lod;
+    lod.emplace_back(batch_starts);
+
+    outs->set_lod(lod);
+  }
+};
+
+class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  MultiClassNMSOpMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("BBoxes",
+             "(Tensor) A 2-D Tensor with shape [M, 4] represents the "
+             "predicted locations of M bounding bboxes. Each bounding box "
+             "has four coordinate values and the layout is "
+             "[xmin, ymin, xmax, ymax].");
+    AddInput("Scores",
+             "(Tensor) A 3-D Tensor with shape [N, C, M] represents the "
+             "predicted confidence predictions. N is the batch size, C is the "
+             "class number, M is number of bounding boxes. For each category "
+             "there are total M scores which corresponding M bounding boxes. "
+             " Please note, M is equal to the 1st dimension of BBoxes. ");
+    AddAttr<int>(
+        "background_label",
+        "(int64_t, defalut: 0) "
+        "The index of background label, the background label will be ignored. "
+        "If set to -1, then all categories will be considered.")
+        .SetDefault(0);
+    AddAttr<float>("score_threshold",
+                   "(float) "
+                   "Threshold to filter out bounding boxes with low "
+                   "confidence score. If not provided, consider all boxes.");
+    AddAttr<int>("nms_top_k",
+                 "(int64_t) "
+                 "Maximum number of detections to be kept according to the "
+                 "confidences aftern the filtering detections based on "
+                 "score_threshold");
+    AddAttr<float>("nms_threshold",
+                   "(float, defalut: 0.3) "
+                   "The threshold to be used in NMS.")
+        .SetDefault(0.3);
+    AddAttr<float>("nms_eta",
+                   "(float) "
+                   "The parameter for adaptive NMS.")
+        .SetDefault(1.0);
+    AddAttr<int>("keep_top_k",
+                 "(int64_t) "
+                 "Number of total bboxes to be kept per image after NMS "
+                 "step. -1 means keeping all bboxes after NMS step.");
+    AddOutput("Out",
+              "(LoDTensor) A 2-D LoDTensor with shape [No, 6] represents the "
+              "detections. Each row has 6 values: "
+              "[label, confidence, xmin, ymin, xmax, ymax], No is the total "
+              "number of detections in this mini-batch. For each instance, "
+              "the offsets in first dimension are called LoD, the number of "
+              "offset is N + 1, if LoD[i + 1] - LoD[i] == 0, means there is "
+              "no detected bbox.");
+    AddComment(R"DOC(
+This operator is to do multi-class non maximum suppression (NMS) on a batched
+of boxes and scores.
+
+In the NMS step, this operator greedily selects a subset of detection bounding
+boxes that have high scores larger than score_threshold, if providing this
+threshold, then selects the largest nms_top_k confidences scores if nms_top_k
+is larger than -1. Then this operator pruns away boxes that have high IOU
+(intersection over union) overlap with already selected boxes by adaptive
+threshold NMS based on parameters of nms_threshold and nms_eta.
+
+Aftern NMS step, at most keep_top_k number of total bboxes are to be kept
+per image if keep_top_k is larger than -1.
+
+This operator support multi-class and batched inputs. It applying NMS
+independently for each class. The outputs is a 2-D LoDTenosr, for each
+image, the offsets in first dimension of LoDTensor are called LoD, the number
+of offset is N + 1, where N is the batch size. If LoD[i + 1] - LoD[i] == 0,
+means there is no detected bbox for this image. If there is no detected boxes
+for all images, all the elements in LoD are 0, and the Out only contains one
+value which is -1.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(multiclass_nms, ops::MultiClassNMSOp,
+                  ops::MultiClassNMSOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+REGISTER_OP_CPU_KERNEL(multiclass_nms, ops::MultiClassNMSKernel<float>,
+                       ops::MultiClassNMSKernel<double>);

paddle/operators/prior_box_op.cc

@@ -44,12 +44,6 @@ class PriorBoxOp : public framework::OperatorWithKernel {
     auto aspect_ratios = ctx->Attrs().Get<std::vector<float>>("aspect_ratios");
     bool flip = ctx->Attrs().Get<bool>("flip");
 
-    PADDLE_ENFORCE_GT(min_sizes.size(), 0,
-                      "Size of min_sizes must be at least 1.");
-    for (size_t i = 0; i < min_sizes.size(); ++i) {
-      PADDLE_ENFORCE_GT(min_sizes[i], 0, "min_sizes[%d] must be positive.", i);
-    }
-
     std::vector<float> aspect_ratios_vec;
     ExpandAspectRatios(aspect_ratios, flip, aspect_ratios_vec);
 
@@ -65,17 +59,6 @@ class PriorBoxOp : public framework::OperatorWithKernel {
       }
     }
 
-    PADDLE_ENFORCE_EQ(variances.size(), 4, "Must and only provide 4 variance.");
-    for (size_t i = 0; i < variances.size(); ++i) {
-      PADDLE_ENFORCE_GT(variances[i], 0.0,
-                        "variance[%d] must be greater than 0.", i);
-    }
-
-    const float step_h = ctx->Attrs().Get<float>("step_h");
-    PADDLE_ENFORCE_GT(step_h, 0.0, "step_h should be larger than 0.");
-    const float step_w = ctx->Attrs().Get<float>("step_w");
-    PADDLE_ENFORCE_GT(step_w, 0.0, "step_w should be larger than 0.");
-
     std::vector<int64_t> dim_vec(4);
     dim_vec[0] = input_dims[2];
     dim_vec[1] = input_dims[3];
@@ -106,26 +89,54 @@ class PriorBoxOpMaker : public framework::OpProtoAndCheckerMaker {
               "PriorBoxOp. The layout is [H, W, num_priors, 4]. "
               "H is the height of input, W is the width of input, num_priors "
               "is the box count of each position.");
-    AddAttr<std::vector<int>>("min_sizes", "(vector<int>) ",
-                              "List of min sizes of generated prior boxes.");
-    AddAttr<std::vector<int>>("max_sizes", "(vector<int>) ",
-                              "List of max sizes of generated prior boxes.");
+
+    AddAttr<std::vector<int>>("min_sizes",
+                              "(vector<int>) List of min sizes "
+                              "of generated prior boxes.")
+        .AddCustomChecker([](const std::vector<int>& min_sizes) {
+          PADDLE_ENFORCE_GT(min_sizes.size(), 0,
+                            "Size of min_sizes must be at least 1.");
+          for (size_t i = 0; i < min_sizes.size(); ++i) {
+            PADDLE_ENFORCE_GT(min_sizes[i], 0,
+                              "min_sizes[%d] must be positive.", i);
+          }
+        });
+    AddAttr<std::vector<int>>(
+        "max_sizes",
+        "(vector<int>) List of max sizes of generated prior boxes.");
     AddAttr<std::vector<float>>(
-        "aspect_ratios", "(vector<float>) ",
-        "List of aspect ratios of generated prior boxes.");
+        "aspect_ratios",
+        "(vector<float>) List of aspect ratios of generated prior boxes.");
+
     AddAttr<std::vector<float>>(
-        "variances", "(vector<float>) ",
-        "List of variances to be encoded in prior boxes.");
-    AddAttr<bool>("flip", "(bool) ", "Whether to flip aspect ratios.")
+        "variances",
+        "(vector<float>) List of variances to be encoded in prior boxes.")
+        .AddCustomChecker([](const std::vector<float>& variances) {
+          PADDLE_ENFORCE_EQ(variances.size(), 4,
+                            "Must and only provide 4 variance.");
+          for (size_t i = 0; i < variances.size(); ++i) {
+            PADDLE_ENFORCE_GT(variances[i], 0.0,
+                              "variance[%d] must be greater than 0.", i);
+          }
+        });
+    AddAttr<bool>("flip", "(bool) Whether to flip aspect ratios.")
         .SetDefault(true);
-    AddAttr<bool>("clip", "(bool) ", "Whether to clip out-of-boundary boxes.")
+    AddAttr<bool>("clip", "(bool) Whether to clip out-of-boundary boxes.")
         .SetDefault(true);
+
     AddAttr<float>("step_w",
                    "Prior boxes step across width, 0 for auto calculation.")
-        .SetDefault(0.0);
+        .SetDefault(0.0)
+        .AddCustomChecker([](const float& step_w) {
+          PADDLE_ENFORCE_GT(step_w, 0.0, "step_w should be larger than 0.");
+        });
     AddAttr<float>("step_h",
                    "Prior boxes step across height, 0 for auto calculation.")
-        .SetDefault(0.0);
+        .SetDefault(0.0)
+        .AddCustomChecker([](const float& step_h) {
+          PADDLE_ENFORCE_GT(step_h, 0.0, "step_h should be larger than 0.");
+        });
+
     AddAttr<float>("offset",
                    "(float) "
                    "Prior boxes center offset.")

paddle/operators/prior_box_op.h

@@ -25,7 +25,7 @@ inline void ExpandAspectRatios(const std::vector<float>& input_aspect_ratior,
                                std::vector<float>& output_aspect_ratior) {
   constexpr float epsilon = 1e-6;
   output_aspect_ratior.clear();
-  output_aspect_ratior.push_back(1.);
+  output_aspect_ratior.push_back(1.0f);
   for (size_t i = 0; i < input_aspect_ratior.size(); ++i) {
     float ar = input_aspect_ratior[i];
     bool already_exist = false;
@@ -38,7 +38,7 @@ inline void ExpandAspectRatios(const std::vector<float>& input_aspect_ratior,
     if (!already_exist) {
       output_aspect_ratior.push_back(ar);
       if (flip) {
-        output_aspect_ratior.push_back(1. / ar);
+        output_aspect_ratior.push_back(1.0f / ar);
       }
     }
   }
@@ -46,7 +46,7 @@ inline void ExpandAspectRatios(const std::vector<float>& input_aspect_ratior,
 
 template <typename T>
 struct ClipFunctor {
-  HOSTDEVICE T operator()(T in) const {
+  HOSTDEVICE inline T operator()(T in) const {
     return std::min<T>(std::max<T>(in, 0.), 1.);
   }
 };
@@ -97,6 +97,9 @@ class PriorBoxOpKernel : public framework::OpKernel<T> {
     boxes->mutable_data<T>(ctx.GetPlace());
     vars->mutable_data<T>(ctx.GetPlace());
 
+    T inv_img_width = 1.0 / img_width;
+    T inv_img_height = 1.0 / img_height;
+
     auto e_boxes = framework::EigenTensor<T, 4>::From(*boxes);
     for (int h = 0; h < feature_height; ++h) {
       for (int w = 0; w < feature_width; ++w) {
@@ -109,13 +112,15 @@ class PriorBoxOpKernel : public framework::OpKernel<T> {
           // first prior: aspect_ratio = 1, size = min_size
           box_width = box_height = min_size;
           // xmin
-          e_boxes(h, w, idx, 0) = (center_x - box_width / 2.) / img_width;
+          e_boxes(h, w, idx, 0) = (center_x - box_width * 0.5) * inv_img_width;
           // ymin
-          e_boxes(h, w, idx, 1) = (center_y - box_height / 2.) / img_height;
+          e_boxes(h, w, idx, 1) =
+              (center_y - box_height * 0.5) * inv_img_height;
           // xmax
-          e_boxes(h, w, idx, 2) = (center_x + box_width / 2.) / img_width;
+          e_boxes(h, w, idx, 2) = (center_x + box_width * 0.5) * inv_img_width;
           // ymax
-          e_boxes(h, w, idx, 3) = (center_y + box_height / 2.) / img_height;
+          e_boxes(h, w, idx, 3) =
+              (center_y + box_height * 0.5) * inv_img_height;
 
           idx++;
           if (max_sizes.size() > 0) {
@@ -124,13 +129,17 @@ class PriorBoxOpKernel : public framework::OpKernel<T> {
             // size = sqrt(min_size * max_size)
             box_width = box_height = sqrt(min_size * max_size);
             // xmin
-            e_boxes(h, w, idx, 0) = (center_x - box_width / 2.) / img_width;
+            e_boxes(h, w, idx, 0) =
+                (center_x - box_width * 0.5) * inv_img_width;
             // ymin
-            e_boxes(h, w, idx, 1) = (center_y - box_height / 2.) / img_height;
+            e_boxes(h, w, idx, 1) =
+                (center_y - box_height * 0.5) * inv_img_height;
             // xmax
-            e_boxes(h, w, idx, 2) = (center_x + box_width / 2.) / img_width;
+            e_boxes(h, w, idx, 2) =
+                (center_x + box_width * 0.5) * inv_img_width;
             // ymax
-            e_boxes(h, w, idx, 3) = (center_y + box_height / 2.) / img_height;
+            e_boxes(h, w, idx, 3) =
+                (center_y + box_height * 0.5) * inv_img_height;
             idx++;
           }
 
@@ -143,13 +152,17 @@ class PriorBoxOpKernel : public framework::OpKernel<T> {
             box_width = min_size * sqrt(ar);
             box_height = min_size / sqrt(ar);
             // xmin
-            e_boxes(h, w, idx, 0) = (center_x - box_width / 2.) / img_width;
+            e_boxes(h, w, idx, 0) =
+                (center_x - box_width * 0.5) * inv_img_width;
             // ymin
-            e_boxes(h, w, idx, 1) = (center_y - box_height / 2.) / img_height;
+            e_boxes(h, w, idx, 1) =
+                (center_y - box_height * 0.5) * inv_img_height;
             // xmax
-            e_boxes(h, w, idx, 2) = (center_x + box_width / 2.) / img_width;
+            e_boxes(h, w, idx, 2) =
+                (center_x + box_width * 0.5) * inv_img_width;
             // ymax
-            e_boxes(h, w, idx, 3) = (center_y + box_height / 2.) / img_height;
+            e_boxes(h, w, idx, 3) =
+                (center_y + box_height * 0.5) * inv_img_height;
             idx++;
           }
         }

paddle/operators/read_op.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/framework/op_registry.h"
+#include "paddle/framework/reader.h"
+
+namespace paddle {
+namespace operators {
+
+class ReadInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Reader"),
+                   "The ReadOp must take a reader as input.");
+    PADDLE_ENFORCE(ctx->HasOutputs("Out"),
+                   "The ReadOp should be assigned with output.");
+    std::vector<framework::DDim> reader_dims = ctx->GetReaderDims("Reader");
+    std::vector<std::string> out_names = ctx->Outputs("Out");
+    PADDLE_ENFORCE_EQ(
+        reader_dims.size(), out_names.size(),
+        "The reader's dim number doesn't match the output number.");
+    ctx->SetOutputsDim("Out", reader_dims);
+  }
+};
+
+class ReadInferVarType : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc& op_desc,
+                  framework::BlockDesc* block) const override {
+    std::string reader_name = op_desc.Input("Reader")[0];
+    std::vector<std::string> out_names = op_desc.Output("Out");
+    framework::VarDesc* reader = block->FindVarRecursive(reader_name);
+    auto dtypes = reader->GetDataTypes();
+    PADDLE_ENFORCE_EQ(dtypes.size(), out_names.size());
+    for (size_t i = 0; i < dtypes.size(); ++i) {
+      framework::VarDesc& out = block->FindRecursiveOrCreateVar(out_names[i]);
+      out.SetType(framework::proto::VarDesc::LOD_TENSOR);
+      out.SetDataType(dtypes[i]);
+    }
+  }
+};
+
+class ReadOp : public framework::OperatorBase {
+ public:
+  using framework::OperatorBase::OperatorBase;
+  void Run(const framework::Scope& scope,
+           const platform::Place& dev_place) const override {
+    framework::ReaderHolder* reader =
+        scope.FindVar(Input("Reader"))->GetMutable<framework::ReaderHolder>();
+    if (!reader->HasNext()) {
+      reader->ReInit();
+      PADDLE_ENFORCE(
+          reader->HasNext(),
+          "Reader can not read the next data even it has been re-initialized.");
+    }
+    std::vector<std::string> out_arg_names = Outputs("Out");
+    std::vector<framework::LoDTensor> ins;
+    reader->ReadNext(&ins);
+    PADDLE_ENFORCE_EQ(ins.size(), out_arg_names.size());
+    for (size_t i = 0; i < ins.size(); ++i) {
+      auto* out =
+          scope.FindVar(out_arg_names[i])->GetMutable<framework::LoDTensor>();
+      out->ShareDataWith(ins[i]);
+      out->set_lod(ins[i].lod());
+    }
+  }
+};
+
+class ReadOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ReadOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(op_proto, op_checker) {
+    AddInput("Reader", "(ReaderHolder) The executed reader.");
+    AddOutput("Out", "(LoDTensor) The output data.").AsDuplicable();
+    AddComment(R"DOC(
+      Read Operator
+
+      Execute a given reader once and output data.
+    )DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(read, ops::ReadOp, ops::ReadInferShape, ops::ReadOpMaker,
+                  paddle::framework::EmptyGradOpMaker, ops::ReadInferVarType);

paddle/operators/target_assign_op.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/operators/target_assign_op.h"
+
+namespace paddle {
+namespace operators {
+
+class TargetAssignOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    // checkout inputs
+    PADDLE_ENFORCE(ctx->HasInput("EncodedGTBBox"),
+                   "Input(EncodedGTBBox) of TargetAssignOp should not be null");
+    PADDLE_ENFORCE(ctx->HasInput("GTScoreLabel"),
+                   "Input(GTScoreLabel) of TargetAssignOp should not be null");
+    PADDLE_ENFORCE(ctx->HasInput("MatchIndices"),
+                   "Input(MatchIndices) of TargetAssignOp should not be null");
+    PADDLE_ENFORCE(ctx->HasInput("NegIndices"),
+                   "Input(NegIndices) of TargetAssignOp should not be null");
+
+    // checkout outputs
+    PADDLE_ENFORCE(
+        ctx->HasOutput("PredBBoxLabel"),
+        "Output(PredBBoxLabel) of TargetAssignOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("PredBBoxWeight"),
+        "Output(PredBBoxWeight) of TargetAssignOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("PredScoreLabel"),
+        "Output(PredScoreLabel) of TargetAssignOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("PredScoreWeight"),
+        "Output(PredScoreWeight) of TargetAssignOp should not be null.");
+
+    auto blabel_dims = ctx->GetInputDim("EncodedGTBBox");
+    auto slabel_dims = ctx->GetInputDim("GTScoreLabel");
+    auto mi_dims = ctx->GetInputDim("MatchIndices");
+    auto neg_dims = ctx->GetInputDim("NegIndices");
+
+    PADDLE_ENFORCE_EQ(blabel_dims.size(), 3UL,
+                      "The rank of Input(EncodedGTBBox) must be 3.");
+    PADDLE_ENFORCE_EQ(slabel_dims.size(), 2UL,
+                      "The rank of Input(GTScoreLabel) must be 2.");
+    PADDLE_ENFORCE_EQ(mi_dims.size(), 2UL,
+                      "The rank of Input(MatchIndices) must be 2.");
+    PADDLE_ENFORCE_EQ(neg_dims.size(), 2UL,
+                      "The rank of Input(NegIndices) must be 2.");
+
+    PADDLE_ENFORCE_EQ(blabel_dims[0], slabel_dims[0],
+                      "The 1st dimension (means the total number of "
+                      "ground-truth bounding boxes) of Input(EncodedGTBBox) "
+                      "and Input(GTScoreLabel) must be the same.");
+    PADDLE_ENFORCE_EQ(blabel_dims[1], mi_dims[1],
+                      "The 2nd dimension (means the number of priod boxes) "
+                      "of Input(EncodedGTBBox) and "
+                      "Input(MatchIndices) must be the same.");
+    PADDLE_ENFORCE_EQ(blabel_dims[2], 4,
+                      "The 3rd dimension of Input(EncodedGTBBox) must be 4.");
+
+    auto n = mi_dims[0];
+    auto np = mi_dims[1];
+    ctx->SetOutputDim("PredBBoxLabel", {n, np, 4});
+    ctx->SetOutputDim("PredBBoxWeight", {n, np, 1});
+    ctx->SetOutputDim("PredScoreLabel", {n, np, 1});
+    ctx->SetOutputDim("PredScoreWeight", {n, np, 1});
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(
+            ctx.Input<framework::LoDTensor>("EncodedGTBBox")->type()),
+        ctx.device_context());
+  }
+};
+
+class TargetAssignOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  TargetAssignOpMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("EncodedGTBBox",
+             "(LoDTensor), The encoded ground-truth bounding boxes with shape "
+             "[Ng, Np, 4], where Ng is the total number of ground-truth boxes "
+             "in this mini-batch, Np the number of predictions, 4 is the "
+             "number of coordinate in [xmin, ymin, xmax, ymax] layout.");
+    AddInput("GTScoreLabel",
+             "(LoDTensor, default LoDTensor<int>),  The input ground-truth "
+             "labels with shape [Ng, 1], where the Ng is the same as it in "
+             "the input of EncodedGTBBox.");
+    AddInput("MatchIndices",
+             "(Tensor, default Tensor<int>), The input matched indices "
+             "with shape [N, Np], where N is the batch size, Np is the same "
+             "as it in the input of EncodedGTBBox. If MatchIndices[i][j] "
+             "is -1, the j-th prior box is not matched to any ground-truh "
+             "box in i-th instance.");
+    AddInput("NegIndices",
+             "(LoDTensor, default LoDTensor<int>), The input negative example "
+             "indices with shape [Neg, 1], where is the total number of "
+             "negative example indices.");
+    AddAttr<int>("background_label",
+                 "(int, default 0), Label index of background class.")
+        .SetDefault(0);
+    AddOutput("PredBBoxLabel",
+              "(Tensor), The output encoded ground-truth labels "
+              "with shape [N, Np, 4], N is the batch size and Np, 4 is the "
+              "same as they in input of EncodedGTBBox. If MatchIndices[i][j] "
+              "is -1, the PredBBoxLabel[i][j][:] is the encoded ground-truth "
+              "box for background_label in i-th instance.");
+    AddOutput("PredBBoxWeight",
+              "(Tensor), The weight for PredBBoxLabel with the shape "
+              "of [N, Np, 1]");
+    AddOutput("PredScoreLabel",
+              "(Tensor, default Tensor<int>), The output score labels for "
+              "each predictions with shape [N, Np, 1]. If MatchIndices[i][j] "
+              "is -1, PredScoreLabel[i][j] = background_label.");
+    AddOutput("PredScoreWeight",
+              "(Tensor), The weight for PredScoreLabel with the shape "
+              "of [N, Np, 1]");
+    AddComment(R"DOC(
+This operator is, for given the encoded boxes between prior boxes and
+ground-truth boxes and ground-truth class labels, to assign classification
+and regression targets to each prior box as well as weights to each
+prior box. The weights is used to specify which prior box would not contribute
+to training loss.
+
+For each instance, the output `PredBBoxLabel`, `PredBBoxWeight`,
+`PredScoreLabel` and `PredScoreWeight` are assigned based on `MatchIndices`.
+Assumed that the row offset for each instance in `EncodedGTBBox` is called lod,
+this operato assigns classification/regression targets by performing the
+following steps:
+
+1. Assigning all outpts based on `MatchIndices`:
+
+If id = MatchIndices[i][j] > 0,
+
+    PredBBoxLabel[i][j] = EncodedGTBBox[lod[i] + id][j]
+    PredBBoxWeight[i][j] = 1.
+    PredScoreLabel[i][j] = GTScoreLabel[lod[i] + id]
+    PredScoreWeight[i][j] = 1.
+
+Otherwise, 
+
+    PredBBoxLabel[j][j] = [0., 0., 0., 0.]
+    PredBBoxWeight[i][j] = 0.
+    PredScoreLabel[i][j] = background_label
+    PredScoreWeight[i][j] = 0.
+
+2. Assigning PredScoreWeight based on `NegIndices`:
+
+Assumed that the row offset for each instance in `NegIndices` is caleed neg_lod,
+for i-th instance and all ids of NegIndices in this instance:
+
+    PredScoreLabel[i][id] = background_label
+    PredScoreWeight[i][id] = 1.0
+
+    )DOC");
+  }
+};
+
+template <typename T>
+struct NegTargetAssignFunctor<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& ctx, const int* neg_indices,
+                  const size_t* lod, const int num, const int num_prior_box,
+                  const int background_label, int* out_label, T* out_label_wt) {
+    for (int i = 0; i < num; ++i) {
+      for (size_t j = lod[i]; j < lod[i + 1]; ++j) {
+        int id = neg_indices[j];
+        out_label[i * num_prior_box + id] = background_label;
+        out_label_wt[i * num_prior_box + id] = static_cast<T>(1.0);
+      }
+    }
+  }
+};
+
+template struct NegTargetAssignFunctor<platform::CPUDeviceContext, float>;
+template struct NegTargetAssignFunctor<platform::CPUDeviceContext, double>;
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(target_assign, ops::TargetAssignOp,
+                             ops::TargetAssignOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    target_assign,
+    ops::TargetAssignKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::TargetAssignKernel<paddle::platform::CPUDeviceContext, double>);

paddle/operators/target_assign_op.cu

+/* 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/operators/target_assign_op.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+__global__ void NegTargetAssignKernel(const int* neg_indices, const size_t* lod,
+                                      const int num, const int num_prior_box,
+                                      const int background_label,
+                                      int* out_label, T* out_label_wt) {
+  int bidx = blockIdx.x;
+  int st = lod[bidx];
+  int ed = lod[bidx + 1];
+
+  int row_start = bidx * num_prior_box;
+  for (int i = st + threadIdx.x; i < ed; i += blockDim.x) {
+    int id = row_start + neg_indices[i];
+    out_label[id] = background_label;
+    out_label_wt[id] = 1.;
+  }
+}
+
+template <typename T>
+struct NegTargetAssignFunctor<platform::CUDADeviceContext, T> {
+  void operator()(const platform::CUDADeviceContext& ctx,
+                  const int* neg_indices, const size_t* lod, const int num,
+                  const int num_prior_box, const int background_label,
+                  int* out_label, T* out_label_wt) {
+    const int block_size = 256;
+    const int grid_size = num;
+    NegTargetAssignKernel<T><<<grid_size, block_size, 0, ctx.stream()>>>(
+        neg_indices, lod, num, num_prior_box, background_label, out_label,
+        out_label_wt);
+  }
+};
+
+template struct NegTargetAssignFunctor<platform::CUDADeviceContext, float>;
+template struct NegTargetAssignFunctor<platform::CUDADeviceContext, double>;
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    target_assign,
+    ops::TargetAssignKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::TargetAssignKernel<paddle::platform::CUDADeviceContext, double>);

paddle/operators/target_assign_op.h

+/* 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. */
+
+#pragma once
+#include "paddle/framework/op_registry.h"
+#include "paddle/platform/assert.h"
+#include "paddle/platform/for_range.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+struct TargetAssignFunctor {
+  const T* gt_box_;
+  const int* gt_label_;
+  const int* match_indices_;
+  const size_t* lod_;
+  const int background_label_;
+  const int64_t num_;
+  const int64_t num_prior_box_;
+
+  T* out_box_;
+  T* out_box_wt_;
+  int* out_label_;
+  T* out_label_wt_;
+
+  TargetAssignFunctor(const T* gt_box, const int* gt_label,
+                      const int* match_indices, const size_t* lod,
+                      const int background_label, const int64_t num,
+                      const int64_t np, T* out_box, T* out_box_wt,
+                      int* out_label, T* out_label_wt)
+      : gt_box_(gt_box),
+        gt_label_(gt_label),
+        match_indices_(match_indices),
+        lod_(lod),
+        background_label_(background_label),
+        num_(num),
+        num_prior_box_(np),
+        out_box_(out_box),
+        out_box_wt_(out_box_wt),
+        out_label_(out_label),
+        out_label_wt_(out_label_wt) {}
+
+  HOSTDEVICE void operator()(size_t i) const {
+    int row = i / num_prior_box_;
+    int col = i - row * num_prior_box_;
+
+    size_t row_off = lod_[row];
+    int offset = row * num_prior_box_ + col;
+
+    int id = match_indices_[offset];
+    T* obox = out_box_ + offset * 4;
+    int* olabel = out_label_ + offset;
+    T* obox_wt = out_box_wt_ + offset;
+    T* olabel_wt = out_label_wt_ + offset;
+
+    if (id > -1) {
+      const T* gtbox = gt_box_ + ((row_off + id) * num_prior_box_ + col) * 4;
+
+      obox[0] = gtbox[0];
+      obox[1] = gtbox[1];
+      obox[2] = gtbox[2];
+      obox[3] = gtbox[3];
+
+      olabel[0] = gt_label_[row_off + id];
+      obox_wt[0] = static_cast<T>(1.);
+      olabel_wt[0] = static_cast<T>(1.);
+    } else {
+      obox[0] = static_cast<T>(0.);
+      obox[1] = static_cast<T>(0.);
+      obox[2] = static_cast<T>(0.);
+      obox[3] = static_cast<T>(0.);
+
+      olabel[0] = background_label_;
+      obox_wt[0] = static_cast<T>(0.);
+      olabel_wt[0] = static_cast<T>(0.);
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+struct NegTargetAssignFunctor {
+  void operator()(const platform::DeviceContext& ctx, const int* neg_indices,
+                  const size_t* lod, const int num, const int num_prior_box,
+                  const int background_label, int* out_label,
+                  T* out_label_wt) const;
+};
+
+template <typename DeviceContext, typename T>
+class TargetAssignKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* enc_gt_box = ctx.Input<framework::LoDTensor>("EncodedGTBBox");
+    auto* gt_label = ctx.Input<framework::LoDTensor>("GTScoreLabel");
+    auto* match_indices = ctx.Input<framework::Tensor>("MatchIndices");
+    auto* neg_indices = ctx.Input<framework::LoDTensor>("NegIndices");
+
+    auto* out_box = ctx.Output<framework::Tensor>("PredBBoxLabel");
+    auto* out_box_wt = ctx.Output<framework::Tensor>("PredBBoxWeight");
+    auto* out_label = ctx.Output<framework::Tensor>("PredScoreLabel");
+    auto* out_label_wt = ctx.Output<framework::Tensor>("PredScoreWeight");
+
+    PADDLE_ENFORCE_EQ(enc_gt_box->lod().size(), 1UL);
+    PADDLE_ENFORCE_EQ(gt_label->lod().size(), 1UL);
+    PADDLE_ENFORCE_EQ(neg_indices->lod().size(), 1UL);
+
+    int background_label = ctx.Attr<int>("background_label");
+
+    const T* box_data = enc_gt_box->data<T>();
+    const int* label_data = gt_label->data<int>();
+    const int* match_idx_data = match_indices->data<int>();
+    const int* neg_idx_data = neg_indices->data<int>();
+
+    T* obox_data = out_box->mutable_data<T>(ctx.GetPlace());
+    T* obox_wt_data = out_box_wt->mutable_data<T>(ctx.GetPlace());
+    int* olabel_data = out_label->mutable_data<int>(ctx.GetPlace());
+    T* olabel_wt_data = out_label_wt->mutable_data<T>(ctx.GetPlace());
+
+    int64_t num = match_indices->dims()[0];
+    int64_t num_prior_box = match_indices->dims()[1];
+
+    auto gt_lod = enc_gt_box->lod().back();
+    auto gt_label_lod = gt_label->lod().back();
+    auto neg_lod = neg_indices->lod().back();
+    for (size_t i = 0; i < gt_lod.size(); ++i) {
+      PADDLE_ENFORCE_EQ(gt_lod.data()[i], gt_label_lod.data()[i]);
+    }
+
+    size_t* gt_lod_data = gt_lod.data(ctx.GetPlace());
+    size_t* neg_lod_data = neg_lod.data(ctx.GetPlace());
+
+    TargetAssignFunctor<T> functor(box_data, label_data, match_idx_data,
+                                   gt_lod_data, background_label, num,
+                                   num_prior_box, obox_data, obox_wt_data,
+                                   olabel_data, olabel_wt_data);
+
+    auto& device_ctx = ctx.template device_context<DeviceContext>();
+    platform::ForRange<DeviceContext> for_range(device_ctx,
+                                                num * num_prior_box);
+    for_range(functor);
+
+    NegTargetAssignFunctor<DeviceContext, T> neg_trg_functor;
+    neg_trg_functor(device_ctx, neg_idx_data, neg_lod_data, num, num_prior_box,
+                    background_label, olabel_data, olabel_wt_data);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/while_op.cc

@@ -53,6 +53,8 @@ class WhileOp : public framework::OperatorBase {
     auto step_scopes =
         scope.FindVar(Output(kStepScopes))->GetMutable<StepScopeVar>();
 
+    PADDLE_ENFORCE(platform::is_cpu_place(cond.place()),
+                   "Condition of while op must in CPU memory.");
     while (cond.data<bool>()[0]) {
       auto &current_scope = scope.NewScope();
       step_scopes->push_back(&current_scope);
@@ -99,6 +101,9 @@ class WhileGradOp : public framework::OperatorBase {
 
   void Run(const framework::Scope &scope,
            const platform::Place &dev_place) const override {
+    // get device context from pool
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+    auto &dev_ctx = *pool.Get(dev_place);
     framework::Executor executor(dev_place);
     auto *block = Attr<framework::BlockDesc *>(kStepBlock);
     auto *program = block->Program();
@@ -205,6 +210,8 @@ class WhileGradOp : public framework::OperatorBase {
         sum_op->Run(cur_scope, dev_place);
         cur_scope.Rename(new_inside_name, inside_grad_name);
       }
+      dev_ctx.Wait();
+      const_cast<framework::Scope &>(scope).DeleteScope(&cur_scope);
     }
   }
 };

paddle/platform/CMakeLists.txt

@@ -39,11 +39,3 @@ 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_test(profiler_test SRCS profiler_test.cc DEPS profiler)
-
-if(NOT WITH_C_API AND WITH_FLUID)
-  file(GLOB PLATFORM_HEADERS *.h)
-  file(GLOB PLATFORM_dynload_HEADERS dynload/*.h)
-  install(FILES ${PLATFORM_HEADERS} DESTINATION include/paddle/platform)
-  install(FILES ${PLATFORM_HEADERS} DESTINATION include/paddle/platform/dynload)
-  install(FILES details/device_ptr_cast.h DESTINATION include/paddle/platform/details)
-endif()

paddle/platform/assert.h

-//   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.
+/*   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. */
 
 #pragma once
 

paddle/platform/profiler.cc

@@ -233,7 +233,7 @@ void ParseEvents(std::vector<std::vector<Event>>& events,
       };
       break;
     default:
-      sorted_domain = "event end time";
+      sorted_domain = "event first end time";
   }
 
   std::vector<std::vector<EventItem>> events_table;

paddle/pybind/protobuf.cc

@@ -214,11 +214,18 @@ void BindVarDsec(py::module &m) {
            py::return_value_policy::reference)
       .def("set_name", &VarDesc::SetName)
       .def("set_shape", &VarDesc::SetShape)
+      .def("set_shapes", &VarDesc::SetShapes)
       .def("set_dtype", &VarDesc::SetDataType)
-      .def("shape", &VarDesc::Shape, py::return_value_policy::reference)
+      .def("set_dtypes", &VarDesc::SetDataTypes)
+      .def("shape", &VarDesc::GetShape, py::return_value_policy::reference)
+      .def("shapes", &VarDesc::GetShapes, py::return_value_policy::reference)
       .def("dtype", &VarDesc::GetDataType, py::return_value_policy::reference)
+      .def("dtypes", &VarDesc::GetDataTypes, py::return_value_policy::reference)
       .def("lod_level", &VarDesc::GetLoDLevel)
+      .def("lod_levels", &VarDesc::GetLoDLevels,
+           py::return_value_policy::reference)
       .def("set_lod_level", &VarDesc::SetLoDLevel)
+      .def("set_lod_levels", &VarDesc::SetLoDLevels)
       .def("type", &VarDesc::GetType)
       .def("set_type", &VarDesc::SetType)
       .def("serialize_to_string", SerializeMessage<VarDesc>)
@@ -233,7 +240,8 @@ void BindVarDsec(py::module &m) {
       .value("STEP_SCOPES", proto::VarDesc::STEP_SCOPES)
       .value("LOD_RANK_TABLE", proto::VarDesc::LOD_RANK_TABLE)
       .value("LOD_TENSOR_ARRAY", proto::VarDesc::LOD_TENSOR_ARRAY)
-      .value("PLACE_LIST", proto::VarDesc::PLACE_LIST);
+      .value("PLACE_LIST", proto::VarDesc::PLACE_LIST)
+      .value("READER", proto::VarDesc::READER);
 }
 
 void BindOpDesc(py::module &m) {

paddle/scripts/docker/build.sh

@@ -79,6 +79,7 @@ function run_build() {
     Building in /paddle/build ...
     ============================================
 EOF
+    make clean
     make -j `nproc`
 }
 
@@ -116,7 +117,7 @@ EOF
             -DWITH_STYLE_CHECK=OFF
         make -j `nproc` gen_proto_py
         make -j `nproc` paddle_python
-        make -j `nproc` paddle_docs paddle_docs_cn
+        make -j `nproc` paddle_docs paddle_docs_cn paddle_api_docs
         make -j `nproc` print_operators_doc
         paddle/pybind/print_operators_doc > doc/en/html/operators.json
         popd

paddle/scripts/travis/build_doc.sh

@@ -9,13 +9,14 @@ cd $TRAVIS_BUILD_DIR/build
 cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON
 make -j `nproc` gen_proto_py
 make -j `nproc` paddle_python
-make -j `nproc` paddle_docs paddle_docs_cn
+make -j `nproc` paddle_docs paddle_docs_cn paddle_api_docs
 make -j `nproc` print_operators_doc
 paddle/pybind/print_operators_doc > doc/en/html/operators.json
 
 # check websites for broken links
 linkchecker doc/en/html/index.html
 linkchecker doc/cn/html/index.html
+linkchecker doc/api/en/html/index.html
 
 # Parse Github URL
 REPO=`git config remote.origin.url`
@@ -54,10 +55,11 @@ function deploy_docs() {
   mkdir -p ${DIR}
   # remove old docs. mv new docs.
   set +e
-  rm -rf ${DIR}/doc ${DIR}/doc_cn
+  rm -rf ${DIR}/doc ${DIR}/doc_cn ${DIR}/api_doc
   set -e
   cp -r ../doc/cn/html ${DIR}/doc_cn
   cp -r ../doc/en/html ${DIR}/doc
+  cp -r ../doc/api/en/html ${DIR}/api_doc
   git add .
 }
 

paddle/string/CMakeLists.txt

@@ -2,9 +2,3 @@ cc_library(stringpiece SRCS piece.cc)
 cc_test(stringpiece_test SRCS piece_test.cc DEPS stringpiece glog gflags)
 cc_test(stringprintf_test SRCS printf_test.cc DEPS glog gflags)
 cc_test(to_string_test SRCS to_string_test.cc)
-
-if(NOT WITH_C_API AND WITH_FLUID)
-  file(GLOB STRING_HEADERS *.h)
-  install(FILES ${STRING_HEADERS} DESTINATION include/paddle/string)
-  install(FILES tinyformat/tinyformat.h DESTINATION include/paddle/string/tinyformat)
-endif()

python/paddle/v2/fluid/debuger.py

+#   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.
+
+import sys
+import re
+from graphviz import GraphPreviewGenerator
+import proto.framework_pb2 as framework_pb2
+
+_vartype2str_ = [
+    "UNK",
+    "LoDTensor",
+    "SelectedRows",
+    "FeedMinibatch",
+    "FetchList",
+    "StepScopes",
+    "LodRankTable",
+    "LoDTensorArray",
+    "PlaceList",
+]
+_dtype2str_ = [
+    "bool",
+    "int16",
+    "int32",
+    "int64",
+    "float16",
+    "float32",
+    "float64",
+]
+
+
+def repr_data_type(type):
+    return _dtype2str_[type]
+
+
+def repr_tensor(proto):
+    return "tensor(type={}, shape={})".format(_dtype2str_[int(proto.data_type)],
+                                              str(proto.dims))
+
+
+reprtpl = "{ttype} {name} ({reprs})"
+
+
+def repr_lodtensor(proto):
+    if not proto.lod_tensor: return
+    level = proto.lod_tensor.lod_level
+    reprs = repr_tensor(proto.lod_tensor.tensor)
+    return reprtpl.format(
+        ttype="LoDTensor" if level > 0 else "Tensor",
+        name=proto.name,
+        reprs="level=%d, %s" % (level, reprs) if level > 0 else reprs)
+
+
+def repr_selected_rows(proto):
+    if not proto.selected_rows: return
+    return reprtpl.format(
+        ttype="SelectedRows",
+        name=proto.name,
+        reprs=repr_tensor(proto.selected_rows))
+
+
+def repr_tensor_array(proto):
+    if not proto.tensor_array: return
+    return reprtpl.format(
+        ttype="TensorArray",
+        name=proto.name,
+        reprs="level=%d, %s" % (proto.tensor_array.lod_level,
+                                repr_tensor(proto.lod_tensor)))
+
+
+type_handlers = [
+    repr_lodtensor,
+    repr_selected_rows,
+    repr_tensor_array,
+]
+
+
+def repr_var(vardesc):
+    for handler in type_handlers:
+        res = handler(vardesc)
+        if res:
+            return res
+
+
+def pprint_program_codes(program_desc):
+    reprs = []
+    for block_idx in range(program_desc.num_blocks()):
+        block_desc = program_desc.block(block_idx)
+        block_repr = pprint_block_codes(block_desc)
+        reprs.append(block_repr)
+    return '\n'.join(reprs)
+
+
+def pprint_block_codes(block_desc, show_backward=False):
+    def is_op_backward(op_desc):
+        if op_desc.type.endswith('_grad'): return True
+
+        def is_var_backward(var):
+            if "@GRAD" in var.parameter: return True
+            for arg in var.arguments:
+                if "@GRAD" in arg: return True
+
+        for var in op_desc.inputs:
+            if is_var_backward(var): return True
+        for var in op_desc.outputs:
+            if is_var_backward(var): return True
+        return False
+
+    def is_var_backward(var_desc):
+        return "@GRAD" in var_desc.name
+
+    if type(block_desc) is not framework_pb2.BlockDesc:
+        block_desc = framework_pb2.BlockDesc.FromString(
+            block_desc.serialize_to_string())
+    var_reprs = []
+    op_reprs = []
+    for var in block_desc.vars:
+        if not show_backward and is_var_backward(var):
+            continue
+        var_reprs.append(repr_var(var))
+
+    for op in block_desc.ops:
+        if not show_backward and is_op_backward(op): continue
+        op_reprs.append(repr_op(op))
+
+    tpl = "// block-{idx}  parent-{pidx}\n// variables\n{vars}\n\n// operators\n{ops}\n"
+    return tpl.format(
+        idx=block_desc.idx,
+        pidx=block_desc.parent_idx,
+        vars='\n'.join(var_reprs),
+        ops='\n'.join(op_reprs), )
+
+
+def repr_attr(desc):
+    tpl = "{key}={value}"
+    valgetter = [
+        lambda attr: attr.i,
+        lambda attr: attr.f,
+        lambda attr: attr.s,
+        lambda attr: attr.ints,
+        lambda attr: attr.floats,
+        lambda attr: attr.strings,
+        lambda attr: attr.b,
+        lambda attr: attr.bools,
+        lambda attr: attr.block_idx,
+        lambda attr: attr.l,
+    ]
+    key = desc.name
+    value = valgetter[desc.type](desc)
+    if key == "dtype":
+        value = repr_data_type(value)
+    return tpl.format(key=key, value=str(value)), (key, value)
+
+
+def _repr_op_fill_constant(optype, inputs, outputs, attrs):
+    if optype == "fill_constant":
+        return "{output} = {data} [shape={shape}]".format(
+            output=','.join(outputs),
+            data=attrs['value'],
+            shape=str(attrs['shape']))
+
+
+op_repr_handlers = [_repr_op_fill_constant, ]
+
+
+def repr_op(opdesc):
+    optype = None
+    attrs = []
+    attr_dict = {}
+    is_target = None
+    inputs = []
+    outputs = []
+
+    tpl = "{outputs} = {optype}({inputs}{is_target}) [{attrs}]"
+    args2value = lambda args: args[0] if len(args) == 1 else str(list(args))
+    for var in opdesc.inputs:
+        key = var.parameter
+        value = args2value(var.arguments)
+        inputs.append("%s=%s" % (key, value))
+    for var in opdesc.outputs:
+        value = args2value(var.arguments)
+        outputs.append(value)
+    for attr in opdesc.attrs:
+        attr_repr, attr_pair = repr_attr(attr)
+        attrs.append(attr_repr)
+        attr_dict[attr_pair[0]] = attr_pair[1]
+
+    is_target = opdesc.is_target
+
+    for handler in op_repr_handlers:
+        res = handler(opdesc.type, inputs, outputs, attr_dict)
+        if res: return res
+
+    return tpl.format(
+        outputs=', '.join(outputs),
+        optype=opdesc.type,
+        inputs=', '.join(inputs),
+        attrs="{%s}" % ','.join(attrs),
+        is_target=", is_target" if is_target else "")
+
+
+def draw_block_graphviz(block, highlights=None, path="./temp.dot"):
+    '''
+    Generate a debug graph for block.
+    Args:
+        block(Block): a block.
+    '''
+    graph = GraphPreviewGenerator("some graph")
+    # collect parameters and args
+    protostr = block.desc.serialize_to_string()
+    desc = framework_pb2.BlockDesc.FromString(str(protostr))
+
+    def need_highlight(name):
+        if highlights is None: return False
+        for pattern in highlights:
+            assert type(pattern) is str
+            if re.match(pattern, name):
+                return True
+        return False
+
+    # draw parameters and args
+    vars = {}
+    for var in desc.vars:
+        shape = [str(i) for i in var.lod_tensor.tensor.dims]
+        if not shape:
+            shape = ['null']
+        # create var
+        if var.persistable:
+            varn = graph.add_param(
+                var.name, var.type, shape, highlight=need_highlight(var.name))
+        else:
+            varn = graph.add_arg(var.name, highlight=need_highlight(var.name))
+        vars[var.name] = varn
+
+    def add_op_link_var(op, var, op2var=False):
+        for arg in var.arguments:
+            if arg not in vars:
+                # add missing variables as argument
+                vars[arg] = graph.add_arg(arg, highlight=need_highlight(arg))
+            varn = vars[arg]
+            highlight = need_highlight(op.description) or need_highlight(
+                varn.description)
+            if op2var:
+                graph.add_edge(op, varn, highlight=highlight)
+            else:
+                graph.add_edge(varn, op, highlight=highlight)
+
+    for op in desc.ops:
+        opn = graph.add_op(op.type, highlight=need_highlight(op.type))
+        for var in op.inputs:
+            add_op_link_var(opn, var, False)
+        for var in op.outputs:
+            add_op_link_var(opn, var, True)
+
+    graph(path, show=True)

python/paddle/v2/fluid/executor.py

@@ -17,7 +17,9 @@ import contextlib
 from framework import Program, default_main_program
 from . import core
 
-__all__ = ['Executor', 'global_scope', 'scope_guard', 'switch_scope']
+__all__ = [
+    'Executor', 'global_scope', 'scope_guard', 'switch_scope', 'fetch_var'
+]
 
 g_scope = core.Scope()
 
@@ -80,12 +82,12 @@ def has_feed_operators(block, feed_targets, feed_holder_name):
     Args:
         block: a block instance (typically global block of a program)
         feed_targets: a dictionary of {feed_target_name: feed_target_data}
-        feed_holder_name: the name of the variable that holds the data of 
-            all feed targets. The type of this feed_holder variable is 
+        feed_holder_name: the name of the variable that holds the data of
+            all feed targets. The type of this feed_holder variable is
             FEED_MINIBATCH, which is essentially vector<LoDTensor>.
 
     Returns:
-        A boolean value that indicates whether a block has feed operators 
+        A boolean value that indicates whether a block has feed operators
         that match the info contained in feed_targets and feed_holder_name.
     """
 
@@ -108,7 +110,7 @@ def has_feed_operators(block, feed_targets, feed_holder_name):
 
 def has_fetch_operators(block, fetch_targets, fetch_holder_name):
     """ Check whether the block already has fetch operators.
-    
+
     Return false if the block does not have any fetch operators.
     If some fetch operators have been appended to the block, check that
     the info contained in these fetch operators matches the fetch_targets
@@ -118,13 +120,13 @@ def has_fetch_operators(block, fetch_targets, fetch_holder_name):
     Args:
         block: a block instance (typically global block of a program)
         fetch_targets: a dictionary of {fetch_target_name: fetch_target_data}
-        fetch_holder_name: the name of the variable that holds the data of 
-            all fetch targets. The type of this fetch_holder variable is 
-            FETCH_LIST, which is essentially vector<LoDTensor>.    
+        fetch_holder_name: the name of the variable that holds the data of
+            all fetch targets. The type of this fetch_holder variable is
+            FETCH_LIST, which is essentially vector<LoDTensor>.
 
-    Return:    
-        A boolean value that indicates whether a block has fetch operators 
-        that match the info contained in fetch_targets and fetch_holder_name.     
+    Return:
+        A boolean value that indicates whether a block has fetch operators
+        that match the info contained in fetch_targets and fetch_holder_name.
     """
 
     fetch_count = 0
@@ -146,6 +148,35 @@ def has_fetch_operators(block, fetch_targets, fetch_holder_name):
     return fetch_count > 0
 
 
+def fetch_var(name, scope=None, return_numpy=True):
+    """
+    Fetch the value of the variable with the given name from the given scope
+    Args:
+        name(str): name of the variable. Typically, only persistable variables
+            can be found in the scope used for running the program.
+        scope(core.Scope|None): scope object. It should be the scope where
+            you pass to Executor.run() when running your program.
+            If None, global_scope() will be used.
+        return_numpy(bool): whether convert the tensor to numpy.ndarray
+    Returns:
+       LodTensor|numpy.ndarray
+    """
+    assert isinstance(name, str)
+    if scope is None:
+        scope = global_scope()
+    assert isinstance(scope, core.Scope)
+
+    var = global_scope().find_var(name)
+    assert var is not None, (
+        "Cannot find " + name + " in scope. Perhaps you need to make the"
+        " variable persistable by using var.persistable = True in your"
+        " program.")
+    tensor = var.get_tensor()
+    if return_numpy:
+        tensor = as_numpy(tensor)
+    return tensor
+
+
 class Executor(object):
     def __init__(self, places):
         if not isinstance(places, list) and not isinstance(places, tuple):

python/paddle/v2/fluid/framework.py

@@ -31,6 +31,7 @@ __all__ = [
     'program_guard',
     'switch_startup_program',
     'switch_main_program',
+    'get_var',
 ]
 
 EMPTY_VAR_NAME = core.kEmptyVarName()
@@ -451,9 +452,8 @@ class Operator(object):
             if not given == need:
                 raise ValueError(("Incorrect setting for output(s) of "
                                   "operator \"%s\". Need: [%s] Given: [%s]") %
-                                 (type, ", ".join(str(e)
-                                                  for e in need), ", ".join(
-                                                      str(e) for e in given)))
+                                 (type, ", ".join(str(e) for e in need),
+                                  ", ".join(str(e) for e in given)))
 
             for out_proto in proto.outputs:
                 out_args = outputs[out_proto.name]
@@ -1124,3 +1124,22 @@ def program_guard(main_program, startup_program=None):
     switch_main_program(main_program)
     if startup_program is not None:
         switch_startup_program(startup_program)
+
+
+def get_var(name, program=None):
+    """
+    Get a variable by name from the global block of a program
+    Args:
+        name(str): name of the variable
+        program(Program|None): program object.
+             If None, default_global_program() will be used.
+
+    Returns:
+        Variable
+    """
+    if program is None:
+        program = default_main_program()
+    assert isinstance(name, str)
+    assert isinstance(name, Program)
+
+    return program.global_block().var(name)

python/paddle/v2/fluid/graphviz.py

+#   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.
+
+import os
+import random
+import subprocess
+import logging
+
+
+def crepr(v):
+    if type(v) is str or type(v) is unicode:
+        return '"%s"' % v
+    return str(v)
+
+
+class Rank(object):
+    def __init__(self, kind, name, priority):
+        '''
+        kind: str
+        name: str
+        priority: int
+        '''
+        self.kind = kind
+        self.name = name
+        self.priority = priority
+        self.nodes = []
+
+    def __str__(self):
+        if not self.nodes:
+            return ''
+
+        return '{' + 'rank={};'.format(self.kind) + \
+               ','.join([node.name for node in self.nodes]) + '}'
+
+
+class Graph(object):
+    rank_counter = 0
+
+    def __init__(self, title, **attrs):
+        self.title = title
+        self.attrs = attrs
+        self.nodes = []
+        self.edges = []
+        self.rank_groups = {}
+
+    def code(self):
+        return self.__str__()
+
+    def rank_group(self, kind, priority):
+        name = "rankgroup-%d" % Graph.rank_counter
+        Graph.rank_counter += 1
+        rank = Rank(kind, name, priority)
+        self.rank_groups[name] = rank
+        return name
+
+    def node(self, label, prefix, description="", **attrs):
+        node = Node(label, prefix, description, **attrs)
+
+        if 'rank' in attrs:
+            rank = self.rank_groups[attrs['rank']]
+            del attrs['rank']
+            rank.nodes.append(node)
+        self.nodes.append(node)
+        return node
+
+    def edge(self, source, target, **attrs):
+        edge = Edge(source, target, **attrs)
+        self.edges.append(edge)
+        return edge
+
+    def compile(self, dot_path):
+        file = open(dot_path, 'w')
+        file.write(self.__str__())
+        image_path = os.path.join(
+            os.path.dirname(__file__), dot_path[:-3] + "pdf")
+        cmd = ["dot", "-Tpdf", dot_path, "-o", image_path]
+        subprocess.Popen(
+            cmd,
+            stdin=subprocess.PIPE,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE)
+        logging.warning("write block debug graph to {}".format(image_path))
+        return image_path
+
+    def show(self, dot_path):
+        image = self.compile(dot_path)
+        cmd = ["open", image]
+        subprocess.Popen(
+            cmd,
+            stdin=subprocess.PIPE,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE)
+
+    def _rank_repr(self):
+        ranks = sorted(
+            self.rank_groups.items(),
+            cmp=lambda a, b: a[1].priority > b[1].priority)
+        repr = []
+        for x in ranks:
+            repr.append(str(x[1]))
+        return '\n'.join(repr) + '\n'
+
+    def __str__(self):
+        reprs = [
+            'digraph G {',
+            'title = {}'.format(crepr(self.title)),
+        ]
+
+        for attr in self.attrs:
+            reprs.append("{key}={value};".format(
+                key=attr, value=crepr(self.attrs[attr])))
+
+        reprs.append(self._rank_repr())
+
+        random.shuffle(self.nodes)
+        reprs += [str(node) for node in self.nodes]
+
+        for x in self.edges:
+            reprs.append(str(x))
+
+        reprs.append('}')
+        return '\n'.join(reprs)
+
+
+class Node(object):
+    counter = 1
+
+    def __init__(self, label, prefix, description="", **attrs):
+        self.label = label
+        self.name = "%s_%d" % (prefix, Node.counter)
+        self.description = description
+        self.attrs = attrs
+        Node.counter += 1
+
+    def __str__(self):
+        reprs = '{name} [label={label} {extra} ];'.format(
+            name=self.name,
+            label=self.label,
+            extra=',' + ','.join("%s=%s" % (key, crepr(value))
+                                 for key, value in self.attrs.items())
+            if self.attrs else "")
+        return reprs
+
+
+class Edge(object):
+    def __init__(self, source, target, **attrs):
+        '''
+        Link source to target.
+        :param source: Node
+        :param target: Node
+        :param graph: Graph
+        :param attrs: dic
+        '''
+        self.source = source
+        self.target = target
+        self.attrs = attrs
+
+    def __str__(self):
+        repr = "{source} -> {target} {extra}".format(
+            source=self.source.name,
+            target=self.target.name,
+            extra="" if not self.attrs else
+            "[" + ','.join("{}={}".format(attr[0], crepr(attr[1]))
+                           for attr in self.attrs.items()) + "]")
+        return repr
+
+
+class GraphPreviewGenerator(object):
+    '''
+    Generate a graph image for ONNX proto.
+    '''
+
+    def __init__(self, title):
+        # init graphviz graph
+        self.graph = Graph(
+            title,
+            layout="dot",
+            concentrate="true",
+            rankdir="TB", )
+
+        self.op_rank = self.graph.rank_group('same', 2)
+        self.param_rank = self.graph.rank_group('same', 1)
+        self.arg_rank = self.graph.rank_group('same', 0)
+
+    def __call__(self, path='temp.dot', show=False):
+        if not show:
+            self.graph.compile(path)
+        else:
+            self.graph.show(path)
+
+    def add_param(self, name, data_type, shape, highlight=False):
+        label = '\n'.join([
+            '<<table cellpadding="5">',
+            '  <tr>',
+            '    <td bgcolor="#2b787e">',
+            '    <b>',
+            name,
+            '    </b>',
+            '    </td>',
+            '  </tr>',
+            '  <tr>',
+            '    <td>',
+            str(data_type),
+            '    </td>'
+            '  </tr>',
+            '  <tr>',
+            '    <td>',
+            '[%s]' % 'x'.join(shape),
+            '    </td>'
+            '  </tr>',
+            '</table>>',
+        ])
+        return self.graph.node(
+            label,
+            prefix="param",
+            description=name,
+            shape="none",
+            style="rounded,filled,bold",
+            width="1.3",
+            color="#148b97" if not highlight else "orange",
+            fontcolor="#ffffff",
+            fontname="Arial")
+
+    def add_op(self, opType, **kwargs):
+        highlight = False
+        if 'highlight' in kwargs:
+            highlight = kwargs['highlight']
+            del kwargs['highlight']
+        return self.graph.node(
+            "<<B>%s</B>>" % opType,
+            prefix="op",
+            description=opType,
+            shape="box",
+            style="rounded, filled, bold",
+            color="#303A3A" if not highlight else "orange",
+            fontname="Arial",
+            fontcolor="#ffffff",
+            width="1.3",
+            height="0.84", )
+
+    def add_arg(self, name, highlight=False):
+        return self.graph.node(
+            crepr(name),
+            prefix="arg",
+            description=name,
+            shape="box",
+            style="rounded,filled,bold",
+            fontname="Arial",
+            fontcolor="#999999",
+            color="#dddddd" if not highlight else "orange")
+
+    def add_edge(self, source, target, **kwargs):
+        highlight = False
+        if 'highlight' in kwargs:
+            highlight = kwargs['highlight']
+            del kwargs['highlight']
+        return self.graph.edge(
+            source,
+            target,
+            color="#00000" if not highlight else "orange",
+            **kwargs)

python/paddle/v2/fluid/layers/control_flow.py

@@ -18,6 +18,7 @@ from tensor import assign, fill_constant
 from .. import core
 from ..framework import Program, Variable, Operator
 from ..layer_helper import LayerHelper, unique_name
+from ops import logical_and, logical_not, logical_or
 
 __all__ = [
     'split_lod_tensor',
@@ -27,6 +28,7 @@ __all__ = [
     'StaticRNNMemoryLink',
     'WhileGuard',
     'While',
+    'Switch',
     'lod_rank_table',
     'max_sequence_len',
     'topk',
@@ -1063,11 +1065,12 @@ class ConditionalBlockGuard(BlockGuard):
 
 
 class ConditionalBlock(object):
-    def __init__(self, inputs, name=None):
+    def __init__(self, inputs, is_scalar_condition=False, name=None):
         for each_input in inputs:
             if not isinstance(each_input, Variable):
                 raise TypeError("Each input should be variable")
         self.inputs = inputs
+        self.is_scalar_condition = is_scalar_condition
         self.helper = LayerHelper('conditional_block', name=name)
 
     def block(self):
@@ -1112,7 +1115,66 @@ class ConditionalBlock(object):
             },
             outputs={'Out': out_list,
                      'Scope': [step_scope]},
-            attrs={'sub_block': inside_block})
+            attrs={
+                'sub_block': inside_block,
+                'is_scalar_condition': self.is_scalar_condition
+            })
+
+
+class Switch(object):
+    def __init__(self, name=None):
+        self.helper = LayerHelper('switch', name=name)
+        self.inside_scope = False
+        self.pre_not_conditions = []
+
+    def case(self, condition):
+        """create a new block for this condition
+        """
+        if not self.inside_scope:
+            raise ValueError("case should be called inside with")
+
+        if len(self.pre_not_conditions) == 0:
+            cond_block = ConditionalBlock([condition], is_scalar_condition=True)
+            not_cond = logical_not(x=condition)
+            self.pre_not_conditions.append(not_cond)
+        else:
+            pre_cond_num = len(self.pre_not_conditions)
+            pre_not_cond = self.pre_not_conditions[pre_cond_num - 1]
+            new_not_cond = logical_and(
+                x=pre_not_cond, y=logical_not(x=condition))
+            self.pre_not_conditions.append(new_not_cond)
+            cond_block = ConditionalBlock(
+                [logical_and(
+                    x=pre_not_cond, y=condition)],
+                is_scalar_condition=True)
+
+        return ConditionalBlockGuard(cond_block)
+
+    def default(self):
+        """create a default case for this switch
+        """
+        pre_cond_num = len(self.pre_not_conditions)
+        if pre_cond_num == 0:
+            raise ValueError("there should be at least one condition")
+        cond_block = ConditionalBlock(
+            [self.pre_not_conditions[pre_cond_num - 1]],
+            is_scalar_condition=True)
+        return ConditionalBlockGuard(cond_block)
+
+    def __enter__(self):
+        """
+        set flag that now is inside switch.block {}
+        :return:
+        """
+        self.inside_scope = True
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        self.inside_scope = False
+        if exc_type is not None:
+            return False  # re-raise exception
+
+        return True
 
 
 class IfElseBlockGuard(object):

python/paddle/v2/fluid/layers/nn.py

@@ -1231,10 +1231,17 @@ def conv2d(input,
     """
     if stride is None:
         stride = [1, 1]
-    helper = LayerHelper('conv2d', **locals())
-    dtype = helper.input_dtype()
 
     num_channels = input.shape[1]
+
+    l_type = 'conv2d'
+    if (num_channels == groups and num_filters % num_channels == 0 and
+            not use_cudnn):
+        l_type = 'depthwise_conv2d'
+
+    helper = LayerHelper(l_type, **locals())
+    dtype = helper.input_dtype()
+
     if groups is None:
         num_filter_channels = num_channels
     else:
@@ -1267,7 +1274,7 @@ def conv2d(input,
     pre_bias = helper.create_tmp_variable(dtype)
 
     helper.append_op(
-        type='conv2d',
+        type=l_type,
         inputs={
             'Input': input,
             'Filter': filter_param,
@@ -1478,7 +1485,9 @@ def batch_norm(input,
                param_attr=None,
                bias_attr=None,
                data_layout='NCHW',
-               name=None):
+               name=None,
+               moving_mean_name=None,
+               moving_variance_name=None):
     """
     This function helps create an operator to implement
     the BatchNorm layer using the configurations from the input parameters.
@@ -1508,6 +1517,7 @@ def batch_norm(input,
         attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
 
     mean = helper.create_global_variable(
+        name=moving_mean_name,
         dtype=input.dtype,
         shape=param_shape,
         persistable=True,
@@ -1515,6 +1525,7 @@ def batch_norm(input,
     helper.set_variable_initializer(var=mean, initializer=Constant(0.0))
 
     variance = helper.create_global_variable(
+        name=moving_variance_name,
         dtype=input.dtype,
         shape=param_shape,
         persistable=True,

python/paddle/v2/fluid/layers/ops.py

@@ -59,7 +59,12 @@ __all__ = [
     'elementwise_pow',
     'clip',
     'clip_by_norm',
+    'softmax',
     'sequence_softmax',
+    'logical_and',
+    'logical_or',
+    'logical_xor',
+    'logical_not',
 ] + __activations__
 
 for _OP in set(__all__):

python/paddle/v2/fluid/layers/tensor.py

@@ -35,13 +35,15 @@ __all__ = [
 ]
 
 
-def create_tensor(dtype, name=None):
+def create_tensor(dtype, name=None, persistable=False):
     helper = LayerHelper("create_tensor", **locals())
-    return helper.create_variable(name=helper.name, dtype=dtype)
+    return helper.create_variable(
+        name=helper.name, dtype=dtype, persistable=persistable)
 
 
 def create_parameter(shape,
                      dtype,
+                     name=None,
                      attr=None,
                      is_bias=False,
                      default_initializer=None):
@@ -62,7 +64,7 @@ def create_parameter(shape,
     """
     helper = LayerHelper("create_parameter", **locals())
     if attr is None:
-        attr = ParamAttr()
+        attr = ParamAttr(name=name)
     return helper.create_parameter(attr, shape, dtype, is_bias,
                                    default_initializer)
 
@@ -295,7 +297,7 @@ def fill_constant_batch_size_like(input,
     return out
 
 
-def ones(shape, dtype):
+def ones(shape, dtype, force_cpu=False):
     """
     **ones**
 
@@ -319,7 +321,7 @@ def ones(shape, dtype):
     return fill_constant(value=1.0, **locals())
 
 
-def zeros(shape, dtype):
+def zeros(shape, dtype, force_cpu=False):
     """
     **zeros**
 

python/paddle/v2/fluid/memory_optimization_transpiler.py

@@ -31,7 +31,7 @@ dtype_to_size = {
 
 
 class ControlFlowGraph(object):
-    def __init__(self, Program, ops, forward_num):
+    def __init__(self, Program, ops, forward_num, skip_opt):
         self._program = Program
         self._ops = ops
         self._forward_num = forward_num
@@ -41,6 +41,7 @@ class ControlFlowGraph(object):
         self._defs = defaultdict(set)
         self._live_in = defaultdict(set)
         self._live_out = defaultdict(set)
+        self._skip_opt = skip_opt
 
     def _add_connections(self, connections):
         for node1, node2 in connections:
@@ -130,6 +131,10 @@ class ControlFlowGraph(object):
                     block_desc, x,
                     is_forward).type() != core.VarDesc.VarType.LOD_TENSOR:
                 return False
+            if x in self._skip_opt:
+                return False
+            if not self._find_var(block_desc, x, is_forward).shape():
+                return False
             return True
 
         self._build_graph()
@@ -150,6 +155,9 @@ class ControlFlowGraph(object):
                     for x in defs_can_optimize
                 ]
                 for x, x_shape in out_pair:
+                    # If x is both in uses and defs, it can not be optimized!
+                    if x in self._uses[i]:
+                        continue
                     for index, cache_pair in enumerate(self.pool):
                         cache_var = cache_pair[0]
                         cache_shape = cache_pair[1]
@@ -197,28 +205,32 @@ def get_cfgs(input_program):
     block_desc = pdesc.block(0)
     op_size = block_desc.op_size()
     # Get global block ops
-    ops_list.append(([block_desc.op(i) for i in range(op_size)], op_size))
+    ops_list.append(
+        ([block_desc.op(i) for i in range(op_size)], op_size, set()))
 
     while_sub_block_ids = []
     while_grad_sub_block_ids = []
-    while_pair = []
+    while_block_id_pair = []
+    while_op_dict = {}
 
     for i in range(op_size):
         op = block_desc.op(i)
         if op.type() == "while":
             while_sub_block_ids.append(op.attr("sub_block").id)
+            while_op_dict[op.attr("sub_block").id] = op
         elif op.type() == "while_grad":
             while_grad_sub_block_ids.append(op.attr("sub_block").id)
+            while_op_dict[op.attr("sub_block").id] = op
 
     # Find while/while_grad block pair
     for grad_id in while_grad_sub_block_ids:
         parent_id = pdesc.block(grad_id).parent
         if parent_id in while_sub_block_ids:
-            while_pair.append((parent_id, grad_id))
+            while_block_id_pair.append((parent_id, grad_id))
             while_sub_block_ids.remove(parent_id)
 
     # Get while/while_grad block ops
-    for parent_id, grad_id in while_pair:
+    for parent_id, grad_id in while_block_id_pair:
         while_block_ops = []
         while_block = pdesc.block(parent_id)
         while_block_op_size = while_block.op_size()
@@ -230,7 +242,11 @@ def get_cfgs(input_program):
         for i in range(while_grad_block_op_size):
             while_block_ops.append(while_grad_block.op(i))
 
-        ops_list.append((while_block_ops, while_block_op_size))
+        while_op_output = set()
+        while_op_output.update(while_op_dict[parent_id].output_arg_names())
+        while_op_output.update(while_op_dict[grad_id].output_arg_names())
+
+        ops_list.append((while_block_ops, while_block_op_size, while_op_output))
 
     # Process rest while block ops
     for parent_id in while_sub_block_ids:
@@ -240,9 +256,15 @@ def get_cfgs(input_program):
         for i in range(while_block_op_size):
             while_block_ops.append(while_block.op(i))
 
-        ops_list.append((while_block_ops, while_block_op_size))
+        while_op_output = set()
+        while_op_output.update(while_op_dict[parent_id].output_arg_names())
+
+        ops_list.append((while_block_ops, while_block_op_size, while_op_output))
 
-    cfgs = [ControlFlowGraph(input_program, i, j) for i, j in ops_list]
+    cfgs = [
+        ControlFlowGraph(input_program, ops, forward_num, skip_opt)
+        for ops, forward_num, skip_opt in ops_list
+    ]
     return cfgs
 
 

python/paddle/v2/fluid/profiler.py

@@ -103,10 +103,10 @@ def profiler(state, sorted_key=None):
     core.enable_profiler(prof_state)
     yield
 
-    if sorted_key not in ['calls', 'total', 'max', 'min', 'ave']:
-        raise ValueError("The state must be in 'calls', 'total', "
-                         "'max', 'min', 'ave'")
     sorted_key = 'default' if sorted_key is None else sorted_key
+    if sorted_key not in ['default', 'calls', 'total', 'max', 'min', 'ave']:
+        raise ValueError("The sorted_key must be None or in 'calls', 'total', "
+                         "'max', 'min' and 'ave'")
     key_map = {
         'default': core.EventSortingKey.kDefault,
         'calls': core.EventSortingKey.kCalls,

python/paddle/v2/fluid/tests/book/.gitignore

+recognize_digits_*.inference.model

python/paddle/v2/fluid/tests/book/CMakeLists.txt

 file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
 string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
 
-list(REMOVE_ITEM TEST_OPS test_recognize_digits)
-py_test(test_recognize_digits_mlp_cpu
-  SRCS test_recognize_digits.py
-  ARGS mlp)
-py_test(test_recognize_digits_mlp_cuda
-  SRCS test_recognize_digits.py
-  ARGS mlp --use_cuda)
-py_test(test_recognize_digits_conv_cpu
-  SRCS test_recognize_digits.py
-  ARGS conv)
-py_test(test_recognize_digits_conv_cuda
-  SRCS test_recognize_digits.py
-  ARGS conv --use_cuda)
-py_test(test_recognize_digits_mlp_cpu_parallel
-  SRCS test_recognize_digits.py
-  ARGS mlp --parallel)
-py_test(test_recognize_digits_mlp_cuda_parallel
-  SRCS test_recognize_digits.py
-  ARGS mlp --use_cuda --parallel)
-py_test(test_recognize_digits_conv_cpu_parallel
-  SRCS test_recognize_digits.py
-  ARGS conv --parallel)
-py_test(test_recognize_digits_conv_cuda_parallel
-  SRCS test_recognize_digits.py
-  ARGS conv --use_cuda --parallel)
-
 # default test
 foreach(src ${TEST_OPS})
     py_test(${src} SRCS ${src}.py)

python/paddle/v2/fluid/tests/book/test_image_classification_train.py

@@ -16,8 +16,9 @@ from __future__ import print_function
 
 import paddle.v2 as paddle
 import paddle.v2.fluid as fluid
-import unittest
 import contextlib
+import numpy
+import unittest
 
 
 def resnet_cifar10(input, depth=32):
@@ -89,10 +90,7 @@ def vgg16_bn_drop(input):
     return fc2
 
 
-def main(net_type, use_cuda):
-    if use_cuda and not fluid.core.is_compiled_with_cuda():
-        return
-
+def train(net_type, use_cuda, save_dirname):
     classdim = 10
     data_shape = [3, 32, 32]
 
@@ -111,12 +109,14 @@ def main(net_type, use_cuda):
     predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
     cost = fluid.layers.cross_entropy(input=predict, label=label)
     avg_cost = fluid.layers.mean(x=cost)
+    acc = fluid.layers.accuracy(input=predict, label=label)
+
+    # Test program 
+    test_program = fluid.default_main_program().clone()
 
     optimizer = fluid.optimizer.Adam(learning_rate=0.001)
     optimizer.minimize(avg_cost)
 
-    accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
-
     BATCH_SIZE = 128
     PASS_NUM = 1
 
@@ -125,6 +125,9 @@ def main(net_type, use_cuda):
             paddle.dataset.cifar.train10(), buf_size=128 * 10),
         batch_size=BATCH_SIZE)
 
+    test_reader = paddle.batch(
+        paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE)
+
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     exe = fluid.Executor(place)
     feeder = fluid.DataFeeder(place=place, feed_list=[images, label])
@@ -132,18 +135,68 @@ def main(net_type, use_cuda):
 
     loss = 0.0
     for pass_id in range(PASS_NUM):
-        accuracy.reset(exe)
-        for data in train_reader():
-            loss, acc = exe.run(fluid.default_main_program(),
-                                feed=feeder.feed(data),
-                                fetch_list=[avg_cost] + accuracy.metrics)
-            pass_acc = accuracy.eval(exe)
-            print("loss:" + str(loss) + " acc:" + str(acc) + " pass_acc:" + str(
-                pass_acc))
-            return
-
-    raise AssertionError(
-        "Image classification loss is too large, {0:2.2}".format(loss))
+        for batch_id, data in enumerate(train_reader()):
+            exe.run(feed=feeder.feed(data))
+
+            if (batch_id % 10) == 0:
+                acc_list = []
+                avg_loss_list = []
+                for tid, test_data in enumerate(test_reader()):
+                    loss_t, acc_t = exe.run(program=test_program,
+                                            feed=feeder.feed(test_data),
+                                            fetch_list=[avg_cost, acc])
+                    acc_list.append(float(acc_t))
+                    avg_loss_list.append(float(loss_t))
+                    break  # Use 1 segment for speeding up CI
+
+                acc_value = numpy.array(acc_list).mean()
+                avg_loss_value = numpy.array(avg_loss_list).mean()
+
+                print(
+                    'PassID {0:1}, BatchID {1:04}, Test Loss {2:2.2}, Acc {3:2.2}'.
+                    format(pass_id, batch_id + 1,
+                           float(avg_loss_value), float(acc_value)))
+
+                if acc_value > 0.01:  # Low threshold for speeding up CI
+                    fluid.io.save_inference_model(save_dirname, ["pixel"],
+                                                  [predict], exe)
+                    return
+
+
+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)
+
+    # The input's dimension of conv should be 4-D or 5-D.
+    tensor_img = numpy.random.rand(1, 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.
+    results = exe.run(inference_program,
+                      feed={feed_target_names[0]: tensor_img},
+                      fetch_list=fetch_targets)
+    print("infer results: ", results[0])
+
+
+def main(net_type, use_cuda):
+    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)
+    infer(use_cuda, save_dirname)
 
 
 class TestImageClassification(unittest.TestCase):

python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py

@@ -18,7 +18,9 @@ import numpy as np
 import paddle.v2 as paddle
 import paddle.v2.dataset.conll05 as conll05
 import paddle.v2.fluid as fluid
+import contextlib
 import time
+import unittest
 
 word_dict, verb_dict, label_dict = conll05.get_dict()
 word_dict_len = len(word_dict)
@@ -127,7 +129,15 @@ def to_lodtensor(data, place):
     return res
 
 
-def main():
+def create_random_lodtensor(lod, place, low, high):
+    data = np.random.random_integers(low, high, [lod[-1], 1]).astype("int64")
+    res = fluid.LoDTensor()
+    res.set(data, place)
+    res.set_lod([lod])
+    return res
+
+
+def train(use_cuda, save_dirname=None):
     # define network topology
     word = fluid.layers.data(
         name='word_data', shape=[1], dtype='int64', lod_level=1)
@@ -175,8 +185,8 @@ def main():
         paddle.reader.shuffle(
             paddle.dataset.conll05.test(), buf_size=8192),
         batch_size=BATCH_SIZE)
-    # place = fluid.CPUPlace()
-    place = fluid.CUDAPlace(0)
+
+    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     feeder = fluid.DataFeeder(
         feed_list=[
             word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, predicate, mark, target
@@ -211,12 +221,102 @@ def main():
                 if batch_id != 0:
                     print("second per batch: " + str((time.time() - start_time)
                                                      / batch_id))
-
-            # exit early for CI
-            exit(0)
+                # Set the threshold low to speed up the CI test
+                if float(pass_precision) > 0.05:
+                    if save_dirname is not None:
+                        fluid.io.save_inference_model(save_dirname, [
+                            'word_data', 'verb_data', 'ctx_n2_data',
+                            'ctx_n1_data', 'ctx_0_data', 'ctx_p1_data',
+                            'ctx_p2_data', 'mark_data'
+                        ], [feature_out], exe)
+                    return
 
             batch_id = batch_id + 1
 
 
+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)
+
+    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)
+
+    # Construct feed as a dictionary of {feed_target_name: feed_target_data}
+    # and results will contain a list of data corresponding to fetch_targets.
+    assert feed_target_names[0] == 'word_data'
+    assert feed_target_names[1] == 'verb_data'
+    assert feed_target_names[2] == 'ctx_n2_data'
+    assert feed_target_names[3] == 'ctx_n1_data'
+    assert feed_target_names[4] == 'ctx_0_data'
+    assert feed_target_names[5] == 'ctx_p1_data'
+    assert feed_target_names[6] == 'ctx_p2_data'
+    assert feed_target_names[7] == 'mark_data'
+
+    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
+                      },
+                      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):
+    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)
+    infer(use_cuda, save_dirname)
+
+
+class TestLabelSemanticRoles(unittest.TestCase):
+    def test_cuda(self):
+        with self.scope_prog_guard():
+            main(use_cuda=True)
+
+    def test_cpu(self):
+        with self.scope_prog_guard():
+            main(use_cuda=False)
+
+    @contextlib.contextmanager
+    def scope_prog_guard(self):
+        prog = fluid.Program()
+        startup_prog = fluid.Program()
+        scope = fluid.core.Scope()
+        with fluid.scope_guard(scope):
+            with fluid.program_guard(prog, startup_prog):
+                yield
+
+
 if __name__ == '__main__':
-    main()
+    unittest.main()

python/paddle/v2/fluid/tests/book/test_machine_translation.py

@@ -11,21 +11,20 @@
 # 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 contextlib
 
 import numpy as np
 import paddle.v2 as paddle
 import paddle.v2.fluid as fluid
-import paddle.v2.fluid.core as core
 import paddle.v2.fluid.framework as framework
 import paddle.v2.fluid.layers as pd
 from paddle.v2.fluid.executor import Executor
+import unittest
 
 dict_size = 30000
 source_dict_dim = target_dict_dim = dict_size
-src_dict, trg_dict = paddle.dataset.wmt14.get_dict(dict_size)
 hidden_dim = 32
 word_dim = 16
-IS_SPARSE = True
 batch_size = 2
 max_length = 8
 topk_size = 50
@@ -34,10 +33,8 @@ beam_size = 2
 
 decoder_size = hidden_dim
 
-place = core.CPUPlace()
 
-
-def encoder():
+def encoder(is_sparse):
     # encoder
     src_word_id = pd.data(
         name="src_word_id", shape=[1], dtype='int64', lod_level=1)
@@ -45,7 +42,7 @@ def encoder():
         input=src_word_id,
         size=[dict_size, word_dim],
         dtype='float32',
-        is_sparse=IS_SPARSE,
+        is_sparse=is_sparse,
         param_attr=fluid.ParamAttr(name='vemb'))
 
     fc1 = pd.fc(input=src_embedding, size=hidden_dim * 4, act='tanh')
@@ -54,7 +51,7 @@ def encoder():
     return encoder_out
 
 
-def decoder_train(context):
+def decoder_train(context, is_sparse):
     # decoder
     trg_language_word = pd.data(
         name="target_language_word", shape=[1], dtype='int64', lod_level=1)
@@ -62,7 +59,7 @@ def decoder_train(context):
         input=trg_language_word,
         size=[dict_size, word_dim],
         dtype='float32',
-        is_sparse=IS_SPARSE,
+        is_sparse=is_sparse,
         param_attr=fluid.ParamAttr(name='vemb'))
 
     rnn = pd.DynamicRNN()
@@ -82,10 +79,10 @@ def decoder_train(context):
     return rnn()
 
 
-def decoder_decode(context):
+def decoder_decode(context, is_sparse):
     init_state = context
     array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length)
-    counter = pd.zeros(shape=[1], dtype='int64')
+    counter = pd.zeros(shape=[1], dtype='int64', force_cpu=True)
 
     # fill the first element with init_state
     state_array = pd.create_array('float32')
@@ -117,7 +114,7 @@ def decoder_decode(context):
             input=pre_ids,
             size=[dict_size, word_dim],
             dtype='float32',
-            is_sparse=IS_SPARSE)
+            is_sparse=is_sparse)
 
         # use rnn unit to update rnn
         current_state = pd.fc(input=[pre_ids_emb, pre_state_expanded],
@@ -150,7 +147,7 @@ def decoder_decode(context):
 
 
 def set_init_lod(data, lod, place):
-    res = core.LoDTensor()
+    res = fluid.LoDTensor()
     res.set(data, place)
     res.set_lod(lod)
     return res
@@ -165,15 +162,19 @@ def to_lodtensor(data, place):
         lod.append(cur_len)
     flattened_data = np.concatenate(data, axis=0).astype("int64")
     flattened_data = flattened_data.reshape([len(flattened_data), 1])
-    res = core.LoDTensor()
+    res = fluid.LoDTensor()
     res.set(flattened_data, place)
     res.set_lod([lod])
     return res
 
 
-def train_main():
-    context = encoder()
-    rnn_out = decoder_train(context)
+def train_main(use_cuda, is_sparse):
+    if use_cuda and not fluid.core.is_compiled_with_cuda():
+        return
+    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
+
+    context = encoder(is_sparse)
+    rnn_out = decoder_train(context, is_sparse)
     label = pd.data(
         name="target_language_next_word", shape=[1], dtype='int64', lod_level=1)
     cost = pd.cross_entropy(input=rnn_out, label=label)
@@ -212,9 +213,13 @@ def train_main():
             batch_id += 1
 
 
-def decode_main():
-    context = encoder()
-    translation_ids, translation_scores = decoder_decode(context)
+def decode_main(use_cuda, is_sparse):
+    if use_cuda and not fluid.core.is_compiled_with_cuda():
+        return
+    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
+
+    context = encoder(is_sparse)
+    translation_ids, translation_scores = decoder_decode(context, is_sparse)
 
     exe = Executor(place)
     exe.run(framework.default_startup_program())
@@ -250,6 +255,60 @@ def decode_main():
         break
 
 
+class TestMachineTranslation(unittest.TestCase):
+    pass
+
+
+@contextlib.contextmanager
+def scope_prog_guard():
+    prog = fluid.Program()
+    startup_prog = fluid.Program()
+    scope = fluid.core.Scope()
+    with fluid.scope_guard(scope):
+        with fluid.program_guard(prog, startup_prog):
+            yield
+
+
+def inject_test_train(use_cuda, is_sparse):
+    f_name = 'test_{0}_{1}_train'.format('cuda' if use_cuda else 'cpu', 'sparse'
+                                         if is_sparse else 'dense')
+
+    def f(*args):
+        with scope_prog_guard():
+            train_main(use_cuda, is_sparse)
+
+    setattr(TestMachineTranslation, f_name, f)
+
+
+def inject_test_decode(use_cuda, is_sparse, decorator=None):
+    f_name = 'test_{0}_{1}_decode'.format('cuda'
+                                          if use_cuda else 'cpu', 'sparse'
+                                          if is_sparse else 'dense')
+
+    def f(*args):
+        with scope_prog_guard():
+            decode_main(use_cuda, is_sparse)
+
+    if decorator is not None:
+        f = decorator(f)
+
+    setattr(TestMachineTranslation, f_name, f)
+
+
+for _use_cuda_ in (False, True):
+    for _is_sparse_ in (False, True):
+        inject_test_train(_use_cuda_, _is_sparse_)
+
+for _use_cuda_ in (False, True):
+    for _is_sparse_ in (False, True):
+
+        _decorator_ = None
+        if _use_cuda_:
+            _decorator_ = unittest.skip(
+                reason='Beam Search does not support CUDA!')
+
+        inject_test_decode(
+            is_sparse=_is_sparse_, use_cuda=_use_cuda_, decorator=_decorator_)
+
 if __name__ == '__main__':
-    # train_main()
-    decode_main()
+    unittest.main()

python/paddle/v2/fluid/tests/book/test_recognize_digits.py

@@ -17,6 +17,7 @@ import paddle.v2.fluid as fluid
 import paddle.v2 as paddle
 import sys
 import numpy
+import unittest
 
 
 def parse_arg():
@@ -74,18 +75,18 @@ def conv_net(img, label):
     return loss_net(conv_pool_2, label)
 
 
-def train(args, save_dirname=None):
-    print("recognize digits with args: {0}".format(" ".join(sys.argv[1:])))
-
+def train(nn_type, use_cuda, parallel, save_dirname):
+    if use_cuda and not fluid.core.is_compiled_with_cuda():
+        return
     img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32')
     label = fluid.layers.data(name='label', shape=[1], dtype='int64')
 
-    if args.nn_type == 'mlp':
+    if nn_type == 'mlp':
         net_conf = mlp
     else:
         net_conf = conv_net
 
-    if args.parallel:
+    if parallel:
         places = fluid.layers.get_places()
         pd = fluid.layers.ParallelDo(places)
         with pd.do():
@@ -107,7 +108,7 @@ def train(args, save_dirname=None):
     optimizer = fluid.optimizer.Adam(learning_rate=0.001)
     optimizer.minimize(avg_loss)
 
-    place = fluid.CUDAPlace(0) if args.use_cuda else fluid.CPUPlace()
+    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 
     exe = fluid.Executor(place)
     exe.run(fluid.default_startup_program())
@@ -147,13 +148,14 @@ def train(args, save_dirname=None):
                         'PassID {0:1}, BatchID {1:04}, Test Loss {2:2.2}, Acc {3:2.2}'.
                         format(pass_id, batch_id + 1,
                                float(avg_loss_val), float(acc_val)))
+    raise AssertionError("Loss of recognize digits is too large")
 
 
-def infer(args, save_dirname=None):
+def infer(use_cuda, save_dirname=None):
     if save_dirname is None:
         return
 
-    place = fluid.CUDAPlace(0) if args.use_cuda else fluid.CPUPlace()
+    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,
@@ -164,7 +166,9 @@ def infer(args, 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, 1, 28, 28).astype("float32")
+    # Use normilized image pixels as input data, which should be in the range [-1.0, 1.0].
+    tensor_img = numpy.random.uniform(-1.0, 1.0,
+                                      [1, 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.
@@ -174,11 +178,48 @@ def infer(args, save_dirname=None):
     print("infer results: ", results[0])
 
 
-if __name__ == '__main__':
-    args = parse_arg()
-    if not args.use_cuda and not args.parallel:
-        save_dirname = "recognize_digits_" + args.nn_type + ".inference.model"
+def main(use_cuda, parallel, nn_type):
+    if not use_cuda and not parallel:
+        save_dirname = "recognize_digits_" + nn_type + ".inference.model"
     else:
         save_dirname = None
-    train(args, save_dirname)
-    infer(args, save_dirname)
+
+    train(
+        nn_type=nn_type,
+        use_cuda=use_cuda,
+        parallel=parallel,
+        save_dirname=save_dirname)
+    infer(use_cuda=use_cuda, save_dirname=save_dirname)
+
+
+class TestRecognizeDigits(unittest.TestCase):
+    pass
+
+
+def inject_test_method(use_cuda, parallel, nn_type):
+    def __impl__(self):
+        prog = fluid.Program()
+        startup_prog = fluid.Program()
+        scope = fluid.core.Scope()
+        with fluid.scope_guard(scope):
+            with fluid.program_guard(prog, startup_prog):
+                main(use_cuda, parallel, nn_type)
+
+    fn = 'test_{0}_{1}_{2}'.format(nn_type, 'cuda'
+                                   if use_cuda else 'cpu', 'parallel'
+                                   if parallel else 'normal')
+
+    setattr(TestRecognizeDigits, fn, __impl__)
+
+
+def inject_all_tests():
+    for use_cuda in (False, True):
+        for parallel in (False, True):
+            for nn_type in ('mlp', 'conv'):
+                inject_test_method(use_cuda, parallel, nn_type)
+
+
+inject_all_tests()
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/fluid/tests/book/test_rnn_encoder_decoder.py

@@ -18,6 +18,8 @@ import paddle.v2.fluid as fluid
 import paddle.v2.fluid.core as core
 import paddle.v2.fluid.framework as framework
 import paddle.v2.fluid.layers as layers
+import contextlib
+import unittest
 from paddle.v2.fluid.executor import Executor
 
 dict_size = 30000
@@ -163,7 +165,15 @@ def to_lodtensor(data, place):
     return res
 
 
-def train(save_dirname=None):
+def create_random_lodtensor(lod, place, low, high):
+    data = np.random.random_integers(low, high, [lod[-1], 1]).astype("int64")
+    res = fluid.LoDTensor()
+    res.set(data, place)
+    res.set_lod([lod])
+    return res
+
+
+def train(use_cuda, save_dirname=None):
     [avg_cost, prediction] = seq_to_seq_net()
 
     optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4)
@@ -174,7 +184,7 @@ def train(save_dirname=None):
             paddle.dataset.wmt14.train(dict_size), buf_size=1000),
         batch_size=batch_size)
 
-    place = core.CPUPlace()
+    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     exe = Executor(place)
 
     exe.run(framework.default_startup_program())
@@ -185,6 +195,7 @@ def train(save_dirname=None):
             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(),
                            feed={
                                'source_sequence': word_data,
@@ -192,24 +203,26 @@ def train(save_dirname=None):
                                'label_sequence': trg_word_next
                            },
                            fetch_list=[avg_cost])
+
             avg_cost_val = np.array(outs[0])
             print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) +
                   " avg_cost=" + str(avg_cost_val))
+
             if batch_id > 3:
                 if save_dirname is not None:
                     fluid.io.save_inference_model(
                         save_dirname, ['source_sequence',
                                        'target_sequence'], [prediction], exe)
-                    return
-                exit(0)
+                return
+
             batch_id += 1
 
 
-def infer(save_dirname=None):
+def infer(use_cuda, save_dirname=None):
     if save_dirname is None:
         return
 
-    place = fluid.CPUPlace()
+    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,
@@ -219,9 +232,9 @@ def infer(save_dirname=None):
     [inference_program, feed_target_names,
      fetch_targets] = fluid.io.load_inference_model(save_dirname, exe)
 
-    data = [[0, 1, 0, 1], [0, 1, 1, 0, 0, 1]]
-    word_data = to_lodtensor(data, place)
-    trg_word = to_lodtensor(data, place)
+    lod = [0, 4, 10]
+    word_data = create_random_lodtensor(lod, place, low=0, high=1)
+    trg_word = create_random_lodtensor(lod, place, low=0, high=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.
@@ -240,7 +253,35 @@ def infer(save_dirname=None):
     print("Inference results: ", np_data)
 
 
-if __name__ == '__main__':
+def main(use_cuda):
+    if use_cuda and not fluid.core.is_compiled_with_cuda():
+        return
+
+    # Directory for saving the trained model
     save_dirname = "rnn_encoder_decoder.inference.model"
-    train(save_dirname)
-    infer(save_dirname)
+
+    train(use_cuda, save_dirname)
+    infer(use_cuda, save_dirname)
+
+
+class TestRnnEncoderDecoder(unittest.TestCase):
+    def test_cuda(self):
+        with self.scope_prog_guard():
+            main(use_cuda=True)
+
+    def test_cpu(self):
+        with self.scope_prog_guard():
+            main(use_cuda=False)
+
+    @contextlib.contextmanager
+    def scope_prog_guard(self):
+        prog = fluid.Program()
+        startup_prog = fluid.Program()
+        scope = fluid.core.Scope()
+        with fluid.scope_guard(scope):
+            with fluid.program_guard(prog, startup_prog):
+                yield
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/fluid/tests/test_bipartite_match_op.py

@@ -62,7 +62,7 @@ def batch_bipartite_match(distance, lod):
     return match_indices, match_dist
 
 
-class TestBipartiteMatchOpForWithLoD(OpTest):
+class TestBipartiteMatchOpWithLoD(OpTest):
     def setUp(self):
         self.op_type = 'bipartite_match'
         lod = [[0, 5, 11, 23]]
@@ -72,7 +72,7 @@ class TestBipartiteMatchOpForWithLoD(OpTest):
         self.inputs = {'DistMat': (dist, lod)}
         self.outputs = {
             'ColToRowMatchIndices': (match_indices),
-            'ColToRowMatchDis': (match_dist),
+            'ColToRowMatchDist': (match_dist),
         }
 
     def test_check_output(self):
@@ -89,7 +89,7 @@ class TestBipartiteMatchOpWithoutLoD(OpTest):
         self.inputs = {'DistMat': dist}
         self.outputs = {
             'ColToRowMatchIndices': match_indices,
-            'ColToRowMatchDis': match_dist,
+            'ColToRowMatchDist': match_dist,
         }
 
     def test_check_output(self):

python/paddle/v2/fluid/tests/test_box_coder_op.py

+#   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.
+
+import unittest
+import numpy as np
+import sys
+import math
+from op_test import OpTest
+
+
+def box_coder(target_box, prior_box, prior_box_var, output_box, code_type):
+    prior_box_x = (
+        (prior_box[:, 2] + prior_box[:, 0]) / 2).reshape(1, prior_box.shape[0])
+    prior_box_y = (
+        (prior_box[:, 3] + prior_box[:, 1]) / 2).reshape(1, prior_box.shape[0])
+    prior_box_width = (
+        (prior_box[:, 2] - prior_box[:, 0])).reshape(1, prior_box.shape[0])
+    prior_box_height = (
+        (prior_box[:, 3] - prior_box[:, 1])).reshape(1, prior_box.shape[0])
+    prior_box_var = prior_box_var.reshape(1, prior_box_var.shape[0],
+                                          prior_box_var.shape[1])
+
+    if (code_type == "EncodeCenterSize"):
+        target_box_x = ((target_box[:, 2] + target_box[:, 0]) / 2).reshape(
+            target_box.shape[0], 1)
+        target_box_y = ((target_box[:, 3] + target_box[:, 1]) / 2).reshape(
+            target_box.shape[0], 1)
+        target_box_width = ((target_box[:, 2] - target_box[:, 0])).reshape(
+            target_box.shape[0], 1)
+        target_box_height = ((target_box[:, 3] - target_box[:, 1])).reshape(
+            target_box.shape[0], 1)
+
+        output_box[:,:,0] = (target_box_x - prior_box_x) / prior_box_width / \
+                prior_box_var[:,:,0]
+        output_box[:,:,1] = (target_box_y - prior_box_y) / prior_box_height / \
+                prior_box_var[:,:,1]
+        output_box[:,:,2] = np.log(np.fabs(target_box_width / prior_box_width)) / \
+                prior_box_var[:,:,2]
+        output_box[:,:,3] = np.log(np.fabs(target_box_height / prior_box_height)) / \
+                prior_box_var[:,:,3]
+
+    elif (code_type == "DecodeCenterSize"):
+        target_box = target_box.reshape(target_box.shape[0], 1,
+                                        target_box.shape[1])
+        target_box_x = prior_box_var[:,:,0] * target_box[:,:,0] * \
+                       prior_box_width + prior_box_x
+        target_box_y = prior_box_var[:,:,1] * target_box[:,:,1] * \
+                       prior_box_height + prior_box_y
+        target_box_width = np.exp(prior_box_var[:,:,2] * target_box[:,:,2]) * \
+                           prior_box_width
+        target_box_height = np.exp(prior_box_var[:,:,3] * target_box[:,:,3]) * \
+                            prior_box_height
+        output_box[:, :, 0] = target_box_x - target_box_width / 2
+        output_box[:, :, 1] = target_box_y - target_box_height / 2
+        output_box[:, :, 2] = target_box_x + target_box_width / 2
+        output_box[:, :, 3] = target_box_y + target_box_height / 2
+
+
+def batch_box_coder(prior_box, prior_box_var, target_box, lod, code_type):
+    n = target_box.shape[0]
+    m = prior_box.shape[0]
+    output_box = np.zeros((n, m, 4), dtype=np.float32)
+    for i in range(len(lod) - 1):
+        box_coder(target_box[lod[i]:lod[i + 1], :], prior_box, prior_box_var,
+                  output_box[lod[i]:lod[i + 1], :, :], code_type)
+    return output_box
+
+
+class TestBoxCoderOp(OpTest):
+    def test_check_output(self):
+        self.check_output()
+
+    def setUp(self):
+        self.op_type = "box_coder"
+        lod = [[0, 20]]
+        prior_box = np.random.random((10, 4)).astype('float32')
+        prior_box_var = np.random.random((10, 4)).astype('float32')
+        target_box = np.random.random((20, 4)).astype('float32')
+        code_type = "DecodeCenterSize"
+        output_box = batch_box_coder(prior_box, prior_box_var, target_box,
+                                     lod[0], code_type)
+
+        self.inputs = {
+            'PriorBox': prior_box,
+            'PriorBoxVar': prior_box_var,
+            'TargetBox': target_box,
+        }
+        self.attrs = {'code_type': 'decode_center_size'}
+        self.outputs = {'OutputBox': output_box}
+
+
+class TestBoxCoderOpWithLoD(OpTest):
+    def test_check_output(self):
+        self.check_output()
+
+    def setUp(self):
+        self.op_type = "box_coder"
+        lod = [[0, 4, 12, 20]]
+        prior_box = np.random.random((10, 4)).astype('float32')
+        prior_box_var = np.random.random((10, 4)).astype('float32')
+        target_box = np.random.random((20, 4)).astype('float32')
+        code_type = "EncodeCenterSize"
+        output_box = batch_box_coder(prior_box, prior_box_var, target_box,
+                                     lod[0], code_type)
+
+        self.inputs = {
+            'PriorBox': prior_box,
+            'PriorBoxVar': prior_box_var,
+            'TargetBox': (target_box, lod),
+        }
+        self.attrs = {'code_type': 'encode_center_size'}
+        self.outputs = {'OutputBox': output_box}
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/fluid/tests/test_conv2d_op.py

@@ -241,6 +241,30 @@ class TestCUDNNWith1x1(TestWith1x1):
         self.op_type = "conv2d"
 
 
+class TestDepthwiseConv(TestConv2dOp):
+    def init_test_case(self):
+        self.pad = [1, 1]
+        self.stride = [2, 2]
+        self.input_size = [2, 3, 5, 5]  # NCHW
+        self.groups = 3
+        assert np.mod(self.input_size[1], self.groups) == 0
+        f_c = self.input_size[1] / self.groups
+        self.filter_size = [6, f_c, 3, 3]
+        self.op_type = "depthwise_conv2d"
+
+
+class TestDepthwiseConv2(TestConv2dOp):
+    def init_test_case(self):
+        self.pad = [1, 1]
+        self.stride = [1, 1]
+        self.input_size = [2, 3, 5, 5]  # NCHW
+        self.groups = 3
+        assert np.mod(self.input_size[1], self.groups) == 0
+        f_c = self.input_size[1] / self.groups
+        self.filter_size = [6, f_c, 3, 3]
+        self.op_type = "depthwise_conv2d"
+
+
 #  cudnn v5 does not support dilation conv.
 # class TestCUDNNWithDilation(TestWithDilation):
 #     def init_op_type(self):

python/paddle/v2/fluid/tests/test_cpp_reader.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 paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+import numpy as np
+
+prog = fluid.framework.Program()
+block = prog.current_block()
+
+random_reader = block.create_var(
+    type=fluid.core.VarDesc.VarType.READER, name="RandomDataGenerator")
+random_reader.desc.set_lod_levels([0, 0])
+
+create_random_data_generator_op = block.append_op(
+    type="create_random_data_generator",
+    outputs={"Out": random_reader},
+    attrs={
+        "shape_concat": [1, 2, 1, 1],
+        "ranks": [2, 2],
+        "min": 0.0,
+        "max": 1.0
+    })
+
+out1 = block.create_var(
+    type=fluid.core.VarDesc.VarType.LOD_TENSOR,
+    name="Out1",
+    shape=[10, 2],
+    dtype="float32",
+    lod_level=1)
+out2 = block.create_var(
+    type=fluid.core.VarDesc.VarType.LOD_TENSOR,
+    name="Out2",
+    shape=[10, 1],
+    dtype="float32",
+    lod_level=1)
+
+read_op = block.append_op(
+    type="read",
+    inputs={"Reader": random_reader},
+    outputs={"Out": [out1, out2]})
+
+place = fluid.CPUPlace()
+exe = fluid.Executor(place)
+
+[res1, res2] = exe.run(prog, fetch_list=[out1, out2])
+
+if len(res1) == 0 or len(res2) == 0:
+    exit(1)
+
+exit(0)

python/paddle/v2/fluid/tests/test_fetch_var.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 paddle.v2.fluid as fluid
+import paddle.v2.fluid.layers as layers
+import op_test
+import numpy
+import unittest
+
+
+class TestFetchVar(op_test.OpTest):
+    def test_fetch_var(self):
+        val = numpy.array([1, 3, 5]).astype(numpy.int32)
+        x = layers.create_tensor(dtype="int32", persistable=True, name="x")
+        layers.assign(input=val, output=x)
+        exe = fluid.Executor(fluid.CPUPlace())
+        exe.run(fluid.default_main_program(), feed={}, fetch_list=[])
+        fetched_x = fluid.fetch_var("x")
+        self.assertTrue(
+            numpy.array_equal(fetched_x, val),
+            "fetch_x=%s val=%s" % (fetched_x, val))
+        self.assertEqual(fetched_x.dtype, val.dtype)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/fluid/tests/test_layers.py

@@ -223,6 +223,14 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(layers.sequence_softmax(x=seq))
         print(str(program))
 
+    def test_softmax(self):
+        program = Program()
+        with program_guard(program):
+            data = layers.data(name='data', shape=[10], dtype='float32')
+            hid = layers.fc(input=data, size=20)
+            self.assertIsNotNone(layers.softmax(x=hid))
+        print(str(program))
+
     def test_get_places(self):
         program = Program()
         with program_guard(program):

python/paddle/v2/fluid/tests/test_mine_hard_examples_op.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 unittest
+import numpy as np
+import sys
+import math
+from op_test import OpTest
+
+
+class TestMineHardExamplesOp(OpTest):
+    def set_data(self):
+        self.init_test_data()
+        self.inputs = {
+            'ClsLoss': self.cls_loss,
+            'LocLoss': self.loc_loss,
+            'MatchIndices': self.match_indices,
+            'MatchDist': self.match_dis
+        }
+
+        self.attrs = {
+            'neg_pos_ratio': self.neg_pos_ratio,
+            'neg_overlap': self.neg_overlap,
+            'sample_size': self.sample_size,
+            'mining_type': self.mining_type
+        }
+
+        self.outputs = {
+            'NegIndices': (self.neg_indices, self.neg_indices_lod),
+            'UpdatedMatchIndices': self.updated_match_indices
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        return
+
+    def setUp(self):
+        self.op_type = "mine_hard_examples"
+        self.set_data()
+
+    def init_test_data(self):
+        self.neg_pos_ratio = 1.0
+        self.neg_overlap = 0.5
+        self.sample_size = 0
+        self.mining_type = "max_negative"
+        self.cls_loss = np.array([[0.1, 0.1, 0.3],
+                                  [0.3, 0.1, 0.1]]).astype('float32')
+
+        self.loc_loss = np.array([[0.1, 0.2, 0.3],
+                                  [0.3, 0.4, 0.1]]).astype('float32')
+
+        self.match_dis = np.array([[0.2, 0.4, 0.8],
+                                   [0.1, 0.9, 0.3]]).astype('float32')
+
+        self.match_indices = np.array([[0, -1, -1],
+                                       [-1, 0, -1]]).astype('int32')
+
+        self.updated_match_indices = self.match_indices
+
+        self.neg_indices_lod = [[0, 1, 2]]
+        self.neg_indices = np.array([[1], [0]]).astype('int32')
+
+
+class TestMineHardExamplesOpHardExample(TestMineHardExamplesOp):
+    def init_test_data(self):
+        super(TestMineHardExamplesOpHardExample, self).init_test_data()
+        self.mining_type = "hard_example"
+        self.sample_size = 2
+
+        self.cls_loss = np.array([[0.5, 0.1, 0.3],
+                                  [0.3, 0.1, 0.1]]).astype('float32')
+
+        self.loc_loss = np.array([[0.2, 0.2, 0.3],
+                                  [0.3, 0.1, 0.2]]).astype('float32')
+
+        self.match_indices = np.array([[0, -1, -1],
+                                       [-1, 0, -1]]).astype('int32')
+
+        self.updated_match_indices = np.array([[0, -1, -1],
+                                               [-1, -1, -1]]).astype('int32')
+
+        self.neg_indices_lod = [[0, 1, 3]]
+        self.neg_indices = np.array([[2], [0], [2]]).astype('int32')
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/fluid/tests/test_multiclass_nms_op.py

+#  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.
+import unittest
+import numpy as np
+import copy
+from op_test import OpTest
+
+
+def iou(box_a, box_b):
+    """Apply intersection-over-union overlap between box_a and box_b
+    """
+    xmin_a = min(box_a[0], box_a[2])
+    ymin_a = min(box_a[1], box_a[3])
+    xmax_a = max(box_a[0], box_a[2])
+    ymax_a = max(box_a[1], box_a[3])
+
+    xmin_b = min(box_b[0], box_b[2])
+    ymin_b = min(box_b[1], box_b[3])
+    xmax_b = max(box_b[0], box_b[2])
+    ymax_b = max(box_b[1], box_b[3])
+
+    area_a = (ymax_a - ymin_a) * (xmax_a - xmin_a)
+    area_b = (ymax_b - ymin_b) * (xmax_b - xmin_b)
+    if area_a <= 0 and area_b <= 0:
+        return 0.0
+
+    xa = max(xmin_a, xmin_b)
+    ya = max(ymin_a, ymin_b)
+    xb = min(xmax_a, xmax_b)
+    yb = min(ymax_a, ymax_b)
+
+    inter_area = max(xb - xa, 0.0) * max(yb - ya, 0.0)
+
+    box_a_area = (box_a[2] - box_a[0]) * (box_a[3] - box_a[1])
+    box_b_area = (box_b[2] - box_b[0]) * (box_b[3] - box_b[1])
+
+    iou_ratio = inter_area / (area_a + area_b - inter_area)
+
+    return iou_ratio
+
+
+def nms(boxes, scores, score_threshold, nms_threshold, top_k=200, eta=1.0):
+    """Apply non-maximum suppression at test time to avoid detecting too many
+    overlapping bounding boxes for a given object.
+    Args:
+        boxes: (tensor) The location preds for the img, Shape: [num_priors,4].
+        scores: (tensor) The class predscores for the img, Shape:[num_priors].
+        score_threshold: (float) The confidence thresh for filtering low
+            confidence boxes.
+        nms_threshold: (float) The overlap thresh for suppressing unnecessary
+            boxes.
+        top_k: (int) The maximum number of box preds to consider.
+        eta: (float) The parameter for adaptive NMS.
+    Return:
+        The indices of the kept boxes with respect to num_priors.
+    """
+    all_scores = copy.deepcopy(scores)
+    all_scores = all_scores.flatten()
+    selected_indices = np.argwhere(all_scores > score_threshold)
+    selected_indices = selected_indices.flatten()
+    all_scores = all_scores[selected_indices]
+
+    sorted_indices = np.argsort(-all_scores, axis=0, kind='mergesort')
+    sorted_scores = all_scores[sorted_indices]
+    if top_k > -1 and top_k < sorted_indices.shape[0]:
+        sorted_indices = sorted_indices[:top_k]
+        sorted_scores = sorted_scores[:top_k]
+
+    selected_indices = []
+    adaptive_threshold = nms_threshold
+    for i in range(sorted_scores.shape[0]):
+        idx = sorted_indices[i]
+        keep = True
+        for k in range(len(selected_indices)):
+            if keep:
+                kept_idx = selected_indices[k]
+                overlap = iou(boxes[idx], boxes[kept_idx])
+                keep = True if overlap <= adaptive_threshold else False
+            else:
+                break
+        if keep:
+            selected_indices.append(idx)
+        if keep and eta < 1 and adaptive_threshold > 0.5:
+            adaptive_threshold *= eta
+    return selected_indices
+
+
+def multiclass_nms(boxes, scores, background, score_threshold, nms_threshold,
+                   nms_top_k, keep_top_k):
+    class_num = scores.shape[0]
+    priorbox_num = scores.shape[1]
+
+    selected_indices = {}
+    num_det = 0
+    for c in range(class_num):
+        if c == background: continue
+        indices = nms(boxes, scores[c], score_threshold, nms_threshold,
+                      nms_top_k)
+        selected_indices[c] = indices
+        num_det += len(indices)
+
+    if keep_top_k > -1 and num_det > keep_top_k:
+        score_index = []
+        for c, indices in selected_indices.iteritems():
+            for idx in indices:
+                score_index.append((scores[c][idx], c, idx))
+
+        sorted_score_index = sorted(
+            score_index, key=lambda tup: tup[0], reverse=True)
+        sorted_score_index = sorted_score_index[:keep_top_k]
+        selected_indices = {}
+
+        for _, c, _ in sorted_score_index:
+            selected_indices[c] = []
+        for s, c, idx in sorted_score_index:
+            selected_indices[c].append(idx)
+        num_det = keep_top_k
+
+    return selected_indices, num_det
+
+
+def batched_multiclass_nms(boxes, scores, background, score_threshold,
+                           nms_threshold, nms_top_k, keep_top_k):
+    batch_size = scores.shape[0]
+
+    det_outs = []
+    lod = [0]
+    for n in range(batch_size):
+        nmsed_outs, nmsed_num = multiclass_nms(boxes, scores[n], background,
+                                               score_threshold, nms_threshold,
+                                               nms_top_k, keep_top_k)
+        lod.append(lod[-1] + nmsed_num)
+        if nmsed_num == 0: continue
+
+        for c, indices in nmsed_outs.iteritems():
+            for idx in indices:
+                xmin, ymin, xmax, ymax = boxes[idx][:]
+                det_outs.append([c, scores[n][c][idx], xmin, ymin, xmax, ymax])
+
+    return det_outs, lod
+
+
+class TestMulticlassNMSOp(OpTest):
+    def set_argument(self):
+        self.score_threshold = 0.01
+
+    def setUp(self):
+        self.set_argument()
+        N = 7
+        M = 1200
+        C = 21
+        BOX_SIZE = 4
+
+        background = 0
+        nms_threshold = 0.3
+        nms_top_k = 400
+        keep_top_k = 200
+        score_threshold = self.score_threshold
+
+        scores = np.random.random((N * M, C)).astype('float32')
+
+        def softmax(x):
+            shiftx = x - np.max(x).clip(-64.)
+            exps = np.exp(shiftx)
+            return exps / np.sum(exps)
+
+        scores = np.apply_along_axis(softmax, 1, scores)
+        scores = np.reshape(scores, (N, M, C))
+        scores = np.transpose(scores, (0, 2, 1))
+
+        boxes = np.random.random((M, BOX_SIZE)).astype('float32')
+        boxes[:, 0:2] = boxes[:, 0:2] * 0.5
+        boxes[:, 2:4] = boxes[:, 2:4] * 0.5 + 0.5
+
+        nmsed_outs, lod = batched_multiclass_nms(boxes, scores, background,
+                                                 score_threshold, nms_threshold,
+                                                 nms_top_k, keep_top_k)
+        nmsed_outs = [-1] if not nmsed_outs else nmsed_outs
+        nmsed_outs = np.array(nmsed_outs).astype('float32')
+
+        self.op_type = 'multiclass_nms'
+        self.inputs = {'BBoxes': boxes, 'Scores': scores}
+        self.outputs = {'Out': (nmsed_outs, [lod])}
+        self.attrs = {
+            'background_label': 0,
+            'nms_threshold': nms_threshold,
+            'nms_top_k': nms_top_k,
+            'keep_top_k': keep_top_k,
+            'score_threshold': score_threshold,
+            'nms_eta': 1.0,
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestMulticlassNMSOpNoOutput(TestMulticlassNMSOp):
+    def set_argument(self):
+        # Here set 2.0 to test the case there is no outputs.
+        # In practical use, 0.0 < score_threshold < 1.0 
+        self.score_threshold = 2.0
+
+
+class TestIOU(unittest.TestCase):
+    def test_iou(self):
+        box1 = np.array([4.0, 3.0, 7.0, 5.0]).astype('float32')
+        box2 = np.array([3.0, 4.0, 6.0, 8.0]).astype('float32')
+
+        expt_output = np.array([2.0 / 16.0]).astype('float32')
+        calc_output = np.array([iou(box1, box2)]).astype('float32')
+        self.assertTrue(np.allclose(calc_output, expt_output))
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/fluid/tests/test_protobuf_descs.py

@@ -115,6 +115,17 @@ class TestVarDesc(unittest.TestCase):
         self.assertEqual(src_shape, res_shape)
         self.assertEqual(core.VarDesc.VarType.SELECTED_ROWS, var.type())
 
+    def test_multiple_shape(self):
+        program_desc = core.ProgramDesc()
+        block = program_desc.block(0)
+        var = block.var('my_reader')
+        var.set_type(core.VarDesc.VarType.READER)
+        src_shapes = [[2, 3, 3], [4, 5], [6, 7, 8, 9]]
+        var.set_shapes(src_shapes)
+        res_shapes = var.shapes()
+        self.assertEqual(src_shapes, res_shapes)
+        self.assertEqual(core.VarDesc.VarType.READER, var.type())
+
     def test_dtype(self):
         program_desc = core.ProgramDesc()
         block = program_desc.block(0)
@@ -124,6 +135,28 @@ class TestVarDesc(unittest.TestCase):
         self.assertEqual(core.DataType.INT32, var.dtype())
         self.assertEqual(core.VarDesc.VarType.LOD_TENSOR, var.type())
 
+    def test_multiple_dtype(self):
+        program_desc = core.ProgramDesc()
+        block = program_desc.block(0)
+        var = block.var('my_reader')
+        var.set_type(core.VarDesc.VarType.READER)
+        src_types = [
+            core.DataType.INT32, core.DataType.FP64, core.DataType.FP32
+        ]
+        var.set_dtypes(src_types)
+        self.assertEqual(src_types, var.dtypes())
+        self.assertEqual(core.VarDesc.VarType.READER, var.type())
+
+    def test_multiple_lod_level(self):
+        program_desc = core.ProgramDesc()
+        block = program_desc.block(0)
+        var = block.var('my_reader')
+        var.set_type(core.VarDesc.VarType.READER)
+        src_types = [3, 1, 2]
+        var.set_lod_levels(src_types)
+        self.assertEqual(src_types, var.lod_levels())
+        self.assertEqual(core.VarDesc.VarType.READER, var.type())
+
 
 class TestBlockDesc(unittest.TestCase):
     def test_add_var(self):

python/paddle/v2/fluid/tests/test_switch.py

+#   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.
+
+import unittest
+
+import paddle.v2.fluid.core as core
+import paddle.v2.fluid.layers as layers
+import paddle.v2.fluid.framework as framework
+from paddle.v2.fluid.executor import Executor
+from paddle.v2.fluid.framework import default_startup_program
+
+
+class TestSwitch(unittest.TestCase):
+    def check_switch(self, value):
+        x = layers.fill_constant(shape=[1], dtype='float32', value=value)
+
+        zero_var = layers.fill_constant(shape=[1], dtype='float32', value=0.0)
+        one_var = layers.fill_constant(shape=[1], dtype='float32', value=1.0)
+        two_var = layers.fill_constant(shape=[1], dtype='float32', value=2.0)
+        three_var = layers.fill_constant(shape=[1], dtype='float32', value=3.0)
+
+        result = layers.create_global_var(
+            shape=[1], value=-1.0, dtype='float32', persistable=True)
+
+        with layers.Switch() as switch:
+            with switch.case(layers.less_than(x, zero_var)):
+                layers.assign(zero_var, result)
+            with switch.case(layers.less_than(x, one_var)):
+                layers.assign(one_var, result)
+            with switch.case(layers.less_than(x, two_var)):
+                layers.assign(two_var, result)
+            with switch.default():
+                layers.assign(three_var, result)
+
+        cpu = core.CPUPlace()
+        exe = Executor(cpu)
+        exe.run(default_startup_program())
+
+        out = exe.run(feed={}, fetch_list=[result])[0][0]
+        return out
+
+    def test_switch(self):
+        test_data = {(-0.1, 0), (0.1, 1), (1.1, 2), (2.1, 3)}
+        for x, expected_result in test_data:
+            main_program = framework.Program()
+            startup_program = framework.Program()
+            with framework.program_guard(main_program, startup_program):
+                result = self.check_switch(x)
+                self.assertEqual(result, expected_result)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/fluid/tests/test_target_assign_op.py

+#   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.
+
+import unittest
+import numpy as np
+import random
+from op_test import OpTest
+
+
+def gen_match_and_neg_indices(num_prior, gt_lod, neg_lod):
+    if len(gt_lod) != len(neg_lod):
+        raise AssertionError("The input arguments are illegal.")
+
+    batch_size = len(gt_lod) - 1
+
+    match_indices = -1 * np.ones((batch_size, num_prior)).astype('int32')
+    neg_indices = np.zeros((neg_lod[-1], 1)).astype('int32')
+
+    for n in range(batch_size):
+        gt_num = gt_lod[n + 1] - gt_lod[n]
+        ids = random.sample([i for i in range(num_prior)], gt_num)
+        match_indices[n, ids] = [i for i in range(gt_num)]
+
+        ret_ids = set([i for i in range(num_prior)]) - set(ids)
+        s = neg_lod[n]
+        e = neg_lod[n + 1]
+        l = e - s
+        neg_ids = random.sample(ret_ids, l)
+        neg_indices[s:e, :] = np.array(neg_ids).astype('int32').reshape(l, 1)
+
+    return match_indices, neg_indices
+
+
+def target_assign(encoded_box, gt_label, match_indices, neg_indices, gt_lod,
+                  neg_lod, background_label):
+    batch_size, num_prior = match_indices.shape
+
+    # init target bbox
+    trg_box = np.zeros((batch_size, num_prior, 4)).astype('float32')
+    # init weight for target bbox
+    trg_box_wt = np.zeros((batch_size, num_prior, 1)).astype('float32')
+    # init target label
+    trg_label = np.ones((batch_size, num_prior, 1)).astype('int32')
+    trg_label = trg_label * background_label
+    # init weight for target label
+    trg_label_wt = np.zeros((batch_size, num_prior, 1)).astype('float32')
+
+    for i in range(batch_size):
+        cur_indices = match_indices[i]
+        col_ids = np.where(cur_indices > -1)
+        col_val = cur_indices[col_ids]
+
+        gt_start = gt_lod[i]
+        # target bbox
+        for v, c in zip(col_val + gt_start, col_ids[0].tolist()):
+            trg_box[i][c][:] = encoded_box[v][c][:]
+
+        # weight for target bbox
+        trg_box_wt[i][col_ids] = 1.0
+
+        trg_label[i][col_ids] = gt_label[col_val + gt_start]
+
+        trg_label_wt[i][col_ids] = 1.0
+        # set target label weight to 1.0 for the negative samples
+        neg_ids = neg_indices[neg_lod[i]:neg_lod[i + 1]]
+        trg_label_wt[i][neg_ids] = 1.0
+
+    return trg_box, trg_box_wt, trg_label, trg_label_wt
+
+
+class TestTargetAssginOp(OpTest):
+    def setUp(self):
+        self.op_type = "target_assign"
+
+        num_prior = 120
+        num_class = 21
+        gt_lod = [0, 5, 11, 23]
+        neg_lod = [0, 4, 7, 13]
+        batch_size = len(gt_lod) - 1
+        num_gt = gt_lod[-1]
+        background_label = 0
+
+        encoded_box = np.random.random((num_gt, num_prior, 4)).astype('float32')
+        gt_label = np.random.randint(
+            num_class, size=(num_gt, 1)).astype('int32')
+        match_indices, neg_indices = gen_match_and_neg_indices(num_prior,
+                                                               gt_lod, neg_lod)
+        trg_box, trg_box_wt, trg_label, trg_label_wt = target_assign(
+            encoded_box, gt_label, match_indices, neg_indices, gt_lod, neg_lod,
+            background_label)
+
+        self.inputs = {
+            'EncodedGTBBox': (encoded_box, [gt_lod]),
+            'GTScoreLabel': (gt_label, [gt_lod]),
+            'MatchIndices': (match_indices),
+            'NegIndices': (neg_indices, [neg_lod]),
+        }
+        self.attrs = {'background_label': background_label}
+        self.outputs = {
+            'PredBBoxLabel': (trg_box),
+            'PredBBoxWeight': (trg_box_wt),
+            'PredScoreLabel': (trg_label),
+            'PredScoreWeight': (trg_label_wt),
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == '__main__':
+    unittest.main()