Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

提交 632108dc 编写于 作者: F fengjiayi
浏览文件
操作

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

上级 20c8ff0f 9d723b8c
展开全部 隐藏空白更改
内联 并排
Showing with 4198 addition and 2507 deletion
.pre-commit-config.yaml
浏览文件 @ 632108dc
......@@ -34,6 +34,14 @@ repos:
entry: bash ./tools/codestyle/cpplint_pre_commit.hook
language: system
files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx)$
- repo: local
hooks:
- id: pylint-doc-string
name: pylint
description: Check python docstring style using docstring_checker.
entry: bash ./tools/codestyle/pylint_pre_commit.hook
language: system
files: \.(py)$
- repo: https://github.com/PaddlePaddle/pre-commit-golang
sha: 8337620115c25ff8333f1b1a493bd031049bd7c0
hooks:
......
.travis.yml
浏览文件 @ 632108dc
......@@ -18,6 +18,8 @@ env:
addons:
ssh_known_hosts: 13.229.163.131
before_install:
# For pylint dockstring checker
- sudo pip install pylint pytest astroid isort
- |
function timeout() { perl -e 'alarm shift; exec @ARGV' "$@"; }
script:
......
CMakeLists.txt
浏览文件 @ 632108dc
......@@ -57,7 +57,10 @@ option(GLIDE_INSTALL "Download and install go dependencies " ON)
option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF)
option(WITH_DISTRIBUTE "Compile with grpc distributed support" OFF)
option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF)
option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF)
option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF)
option(WITH_FAST_BUNDLE_TEST "Bundle tests that can be run in a single process together to reduce launch overhead" OFF)
option(WITH_CONTRIB "Compile the third-party contributation" OFF)
# CMAKE_BUILD_TYPE
if(NOT CMAKE_BUILD_TYPE)
......@@ -202,7 +205,7 @@ endif(USE_NNPACK)
add_subdirectory(proto)
if(NOT MOBILE_INFERENCE)
if(NOT MOBILE_INFERENCE AND NOT WITH_FLUID_ONLY)
# "add_subdirectory(go)" should be placed after the following loine,
# because it depends on paddle/optimizer.
add_subdirectory(paddle/optimizer)
......@@ -230,3 +233,7 @@ if(WITH_DOC)
find_python_module(recommonmark REQUIRED)
add_subdirectory(doc)
endif()
if (WITH_CONTRIB)
add_subdirectory(paddle/contrib)
endif()
Dockerfile
浏览文件 @ 632108dc
......@@ -79,6 +79,9 @@ RUN pip install pre-commit 'ipython==5.3.0' && \
pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
pip install opencv-python
#For docstring checker
RUN pip install pylint pytest astroid isort
COPY ./python/requirements.txt /root/
RUN pip install -r /root/requirements.txt
......@@ -101,6 +104,3 @@ RUN echo 'root:root' | chpasswd
RUN sed -ri 's/^PermitRootLogin\s+.*/PermitRootLogin yes/' /etc/ssh/sshd_config
RUN sed -ri 's/UsePAM yes/#UsePAM yes/g' /etc/ssh/sshd_config
EXPOSE 22
# development image default do build work
CMD ["bash", "/paddle/paddle/scripts/docker/build.sh"]
Dockerfile.android
浏览文件 @ 632108dc
......@@ -40,5 +40,3 @@ RUN mkdir -p ${ANDROID_TOOLCHAINS_DIR} && \
unzip -q android-ndk-r14b-linux-x86_64.zip && \
mv android-ndk-r14b ${ANDROID_NDK_HOME} && \
rm -rf /opt/android-ndk-tmp
CMD ["bash", "/paddle/paddle/scripts/docker/build_android.sh"]
benchmark/cluster/README.md 已删除 100644 → 0
浏览文件 @ 20c8ff0f
# Cluster Training Benchmark
## Setup
- Platform
- Kubernetes: v1.6.2
- Linux Kernel: v3.10.0
- Resource
- CPU: 10 Cores per Pod
- Memory: 5GB per Pod
- Docker Image
We use different base Docker Image to run the benchmark on Kubernetes:
- PaddlePaddle v2: paddlepaddle/paddle:0.11.0
- PaddlePaddle Fluid: paddlepaddle/paddle:[commit-id]
- TensorFlow: tensorflow/tensorflow:1.5.0-rc0
- Model
vgg16 is used in this benchmark.
## Cases
- Variable
- Batch Size of training data.
- PServer count of the training job.
- The number of trainers.
- Invariant
- The resource of trainer/pserver Pod.
### Measure the Performance for Different Batch Size
- PServer Count: 40
- Trainer Count: 100
- Metrics: mini-batch / sec
<table>
<thead>
<tr>
<th>Batch Size </th>
<th> 32</th>
<th>64</th>
<th>128 </th>
<th>256</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>TensorFlow </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
</tbody>
</table>
### Measure the Performance for Different PServer Count
- Trainer Count: 100
- Batch Size: 64
- Metrics: mini-batch / sec
<table>
<thead>
<tr>
<th>PServer Count </th>
<th>10</th>
<th>20</th>
<th>40 </th>
<th>60</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>TensorFlow </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
</tbody>
</table>
### Measure Parallel Efficiency By Increasing Trainer Count
- PServer Count: 20
- Batch Size: 64
- Metrics:
$S = \div(T1, TN)$
which S is the ratio of T1 over TN, training time of 1 and N trainers.
The parallel efficiency is:
$E = \div(S, N)$
<table>
<thead>
<tr>
<th>Trainer Counter </th>
<th>1</th>
<th>10</th>
<th>20 </th>
<th>30</th>
<th>40</th>
<th>50</th>
<th>60 </th>
<th>70</th>
<th>80</th>
<th>90</th>
<th>100 </th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>TensorFlow </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
</tr>
</tbody>
</table>
## Reproduce the benchmark
TODO
benchmark/cluster/vgg16/Dockerfile 已删除 100644 → 0
浏览文件 @ 20c8ff0f
FROM nvidia/cuda:8.0-cudnn5-runtime-ubuntu16.04
# you can get mirror list here:
# https://launchpad.net/ubuntu/+archivemirrors
ARG UBUNTU_MIRROR
RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi'
RUN apt-get update && apt-get install -y python python-dev python-pip iputils-ping libgtk2.0-dev
RUN pip install -U kubernetes opencv-python
RUN pip install paddlepaddle
# if network is slowly, you may need to add proxy here.
# ENV https_proxy=
RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.cifar.train10()" | python'
RUN pip uninstall -y paddlepaddle
# unset proxy if it is setted.
# ENV https_proxy=""
# NOTE: By default CI built wheel packages turn WITH_DISTRIBUTE=OFF,
# so we must build one with distribute support to install in this image.
ADD *.whl /
RUN pip install /*.whl && rm -f /*.whl
ENV LD_LIBRARY_PATH=/usr/local/lib
# tf k8s
RUN pip install tensorflow==1.4.0
ADD tf_k8s /usr/bin
RUN chmod +x /usr/bin/tf_k8s
ADD vgg16_tf.py /workspace/
# below lines may change a lot for debugging
ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/paddle_k8s /usr/bin
ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/k8s_tools.py /root
RUN chmod +x /usr/bin/paddle_k8s
ADD vgg16_fluid.py vgg16_v2.py /workspace/
benchmark/cluster/vgg16/README.md 已删除 100644 → 0
浏览文件 @ 20c8ff0f
# 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
### Blas settings
Setting environment variable: `MKL_NUM_THREADS=1`.
### Single Node Single Thread
- Metrics: samples / sec
<table>
<thead>
<tr>
<th>Batch Size </th>
<th> 32</th>
<th>64</th>
<th>128 </th>
<th>256</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td> 15.44 </td>
<td> 16.32 </td>
<td> 16.74 </td>
<td> 16.79 </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td> 15.97 </td>
<td> 17.04 </td>
<td> 17.60 </td>
<td> 17.83 </td>
</tr>
<tr>
<td>TensorFlow </td>
<td> 9.09 </td>
<td> 9.10 </td>
<td> 9.24 </td>
<td> 8.66 </td>
</tr>
</tbody>
</table>
### Different Batch Size
- PServer Count: 10
- Trainer Count: 20
- Metrics: samples / sec
<table>
<thead>
<tr>
<th>Batch Size </th>
<th> 32</th>
<th>64</th>
<th>128 </th>
<th>256</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td> 190.20 </td>
<td> 222.15 </td>
<td> 247.40 </td>
<td> 258.18 </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td> 170.96 </td>
<td> 233.71 </td>
<td> 256.14 </td>
<td> 329.23 </td>
</tr>
<tr>
<td>TensorFlow </td>
<td> - </td>
<td> - </td>
<td> - </td>
<td> - </td>
</tr>
</tbody>
</table>
### Accelerate Rate
- Pserver Count: 20
- Batch Size: 128
- Metrics: samples / sec
<table>
<thead>
<tr>
<th>Trainer Count </th>
<th>20</th>
<th>40</th>
<th>80</th>
<th>100</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td> 263.29 (78.64%) </td>
<td> 518.80 (77.47%) </td>
<td> 836.26 (62.44%) </td>
<td> 1019.29 (60.89%) </td>
</tr>
<tr>
<td>PaddlePaddle v2 (need more tests) </td>
<td> 326.85 (92.85%) </td>
<td> 534.58 (75.93%) </td>
<td> 853.30 (60.60%) </td>
<td> 1041.99 (59.20%) </td>
</tr>
<tr>
<td>TensorFlow </td>
<td> - </td>
<td> - </td>
<td> - </td>
<td> - </td>
</tr>
</tbody>
</table>
### Different Pserver Count
- Trainer Count: 60
- Batch Size: 128
- Metrics: samples/ sec
<table>
<thead>
<tr>
<th>PServer Count </th>
<th>3</th>
<th>6</th>
<th>10</th>
<th>20</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid(should fix in next PR) </td>
<td> 589.1 </td>
<td> 592.6 </td>
<td> 656.4 </td>
<td> 655.8 </td>
</tr>
<tr>
<td>PaddlePaddle v2 (need more tests) </td>
<td> 593.4 </td>
<td> 791.3 </td>
<td> 729.7 </td>
<td> 821.7 </td>
</tr>
<tr>
<td>TensorFlow </td>
<td> - </td>
<td> - </td>
<td> - </td>
<td> - </td>
</tr>
</tbody>
</table>
*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 已删除 100644 → 0
浏览文件 @ 20c8ff0f
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 已删除 100644 → 0
浏览文件 @ 20c8ff0f
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/run_vgg_dist.sh 已删除 100644 → 0
浏览文件 @ 20c8ff0f
#!/bin/bash
# Update to point to the source file.
VGG_SRC="vgg16_fluid.py"
export TRAINING_ROLE=PSERVER
export TRAINERS=2
export POD_IP=127.0.0.1
export PADDLE_INIT_PORT=6174
MKL_NUM_THREADS=1 python -u ${VGG_SRC} --local 0 --ps_host=127.0.0.1:6174 --trainer_hosts=127.0.0.1:6174 &
# Need to wait for the ps to start first.
sleep 10
echo "done start ps"
export TRAINING_ROLE=TRAINER
export TRAINERS=2
export POD_IP=127.0.0.1
export PADDLE_INIT_PORT=6174
CUDA_VISIBLE_DEVICES=4 MKL_NUM_THREADS=1 python -u ${VGG_SRC} --local 0 --ps_host=127.0.0.1:6174 --trainer_hosts=127.0.0.1:6174 --device=GPU --task_index=0 &
CUDA_VISIBLE_DEVICES=5 MKL_NUM_THREADS=1 python -u ${VGG_SRC} --local 0 --ps_host=127.0.0.1:6174 --trainer_hosts=127.0.0.1:6174 --device=GPU --task_index=1 &
benchmark/cluster/vgg16/tf_k8s 已删除 100644 → 0
浏览文件 @ 20c8ff0f
#!/bin/bash
check_trainer_ret() {
ret=$1
stdbuf -oL echo "job returned $ret...setting pod return message..."
stdbuf -oL echo "==============================="
if [ $ret -eq 136 ] ; then
echo "Error Arithmetic Operation(Floating Point Exception)" > /dev/termination-log
elif [ $ret -eq 139 ] ; then
echo "Segmentation Fault" > /dev/termination-log
elif [ $ret -eq 1 ] ; then
echo "General Error" > /dev/termination-log
elif [ $ret -eq 134 ] ; then
echo "Program Abort" > /dev/termination-log
fi
stdbuf -oL echo "termination log wroted..."
exit $ret
}
g_pservers=""
g_trainers=""
wait_running_pods(){
pserver_label="tf-job-pserver=${JOB_NAME}"
trainer_label="tf-job-trainer=${JOB_NAME}"
stdbuf -oL python /root/k8s_tools.py wait_pods_running ${pserver_label} ${PSERVERS_NUM}
stdbuf -oL python /root/k8s_tools.py wait_pods_running ${trainer_label} ${TRAINERS_NUM}
g_pservers=$(python /root/k8s_tools.py fetch_endpoints ${pserver_label} ${PORT})
g_trainers=$(python /root/k8s_tools.py fetch_endpoints ${trainer_label} ${PORT})
}
start_tf_pserver(){
wait_running_pods
label="tf-job-pserver=${JOB_NAME}"
pserver_id=$(python /root/k8s_tools.py fetch_id ${label})
cmd="${ENTRY} --ps_hosts=${g_pservers} --worker_hosts=${g_trainers} \
--job_name=${TF_JOB_NAME} --task_index=${pserver_id}"
stdbuf -oL sh -c "cd ${TRAINER_PACKAGE} && ${cmd}"
}
start_tf_trainer(){
wait_running_pods
label="tf-job-trainer=${JOB_NAME}"
trainer_id=$(python /root/k8s_tools.py fetch_id ${label})
cmd="${ENTRY} --ps_hosts=${g_pservers} --worker_hosts=${g_trainers} \
--job_name=${TF_JOB_NAME} --task_index=${trainer_id} --batch_size=${BATCH_SIZE}"
stdbuf -oL sh -c "cd ${TRAINER_PACKAGE} && ${cmd}"
check_trainer_ret $?
}
start_tf(){
if [[ "${TF_JOB_NAME}" == "worker" ]]; then
start_tf_trainer
else
start_tf_pserver
fi
}
usage() {
echo "usage: tf_k8s [<args>]:"
echo " start_tf Start tensorflow jobs"
}
case "$1" in
start_tf)
start_tf
;;
--help)
usage
;;
*)
usage
;;
esac
benchmark/cluster/vgg16/tf_pserver.yaml 已删除 100644 → 0
浏览文件 @ 20c8ff0f
apiVersion: extensions/v1beta1
kind: ReplicaSet
metadata:
name: vgg16job-tf-pserver
spec:
replicas: 10
template:
metadata:
labels:
tf-job-pserver: vgg16job-tf
spec:
hostNetwork: true
imagePullSecrets:
- name: job-registry-secret
containers:
- name: pserver
image: "registry.baidu.com/paddlepaddle/fluid_benchmark_tf:vgg16"
imagePullPolicy: Always
command: ["tf_k8s", "start_tf"]
ports:
- name: jobport-30236
containerPort: 30236
env:
- name: PORT
value: "32036"
- name: ENTRY
value: "python vgg16_tf.py"
- name: JOB_NAME
value: vgg16job-tf
- name: PSERVERS_NUM
value: "10"
- name: TF_JOB_NAME
value: "ps"
- name: TRAINERS_NUM
value: "20"
- name: BATCH_SIZE
value: "128"
- name: TRAINER_PACKAGE
value: "/workspace"
- name: NUM_PASSES
value: "1"
- name: NAMESPACE
valueFrom:
fieldRef:
fieldPath: "metadata.namespace"
- name: POD_IP
valueFrom:
fieldRef:
fieldPath: "status.podIP"
resources:
requests:
memory: 10Gi
cpu: 4
limits:
memory: 10Gi
cpu: 4
benchmark/cluster/vgg16/tf_trainer.yaml 已删除 100644 → 0
浏览文件 @ 20c8ff0f
apiVersion: batch/v1
kind: Job
metadata:
name: vgg16job-tf-trainer
spec:
parallelism: 20
completions: 20
template:
metadata:
labels:
tf-job-trainer: vgg16job-tf
spec:
imagePullSecrets:
- name: job-registry-secret
hostNetwork: true
containers:
- name: trainer
image: "registry.baidu.com/paddlepaddle/fluid_benchmark_tf:vgg16"
imagePullPolicy: Always
command: ["tf_k8s", "start_tf"]
ports:
- name: jobport-30236
containerPort: 30236
env:
- name: PORT
value: "32036"
- name: JOB_NAME
value: vgg16job-tf
- name: TF_JOB_NAME
value: "worker"
- name: ENTRY
value: "python vgg16_tf.py"
- name: PSERVERS_NUM
value: "10"
- name: BATCH_SIZE
value: "128"
- name: TRAINERS_NUM
value: "20"
- name: TRAINER_PACKAGE
value: "/workspace"
- name: NUM_PASSES
value: "1"
- name: NAMESPACE
valueFrom:
fieldRef:
fieldPath: "metadata.namespace"
- name: POD_IP
valueFrom:
fieldRef:
fieldPath: "status.podIP"
resources:
requests:
memory: 40Gi
cpu: 2
limits:
memory: 40Gi
cpu: 2
restartPolicy: Never
benchmark/cluster/vgg16/v2_pserver.yaml 已删除 100644 → 0
浏览文件 @ 20c8ff0f
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 已删除 100644 → 0
浏览文件 @ 20c8ff0f
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 已删除 100644 → 0
浏览文件 @ 20c8ff0f
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""VGG16 benchmark in Fluid"""
from __future__ import print_function
import sys
import time
import numpy as np
import paddle.v2 as paddle
import paddle.fluid as fluid
import paddle.fluid.core as core
import paddle.fluid.profiler as profiler
import argparse
import functools
import os
from paddle.fluid import debuger
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=16, 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='flowers',
choices=['cifar10', 'flowers'],
help='Optional dataset for benchmark.')
parser.add_argument(
'--local',
type=str2bool,
default=True,
help='Whether to run as local mode.')
parser.add_argument(
"--ps_hosts",
type=str,
default="",
help="Comma-separated list of hostname:port pairs")
parser.add_argument(
"--trainer_hosts",
type=str,
default="",
help="Comma-separated list of hostname:port pairs")
parser.add_argument(
"--profile", action='store_true', help="If set, profile a few steps.")
# Flags for defining the tf.train.Server
parser.add_argument(
"--task_index", type=int, default=0, help="Index of task within the job")
args = parser.parse_args()
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=4096, 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=4096, 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
batch_size = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(
input=predict, label=label, total=batch_size)
# inference program
inference_program = fluid.default_main_program().clone()
with fluid.program_guard(inference_program):
inference_program = fluid.io.get_inference_program(batch_acc)
# 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):
test_pass_acc = fluid.average.WeightedAverage()
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])
outs = exe.run(inference_program,
feed={"pixel": img_data,
"label": y_data},
fetch_list=[batch_acc, batch_size])
test_pass_acc.add(value=np.array(outs[0]), weight=np.array(outs[1]))
return test_pass_acc.eval()
def train_loop(exe, trainer_prog):
iters = 0
ts = time.time()
train_pass_acc = fluid.average.WeightedAverage()
for pass_id in range(args.num_passes):
# train
start_time = time.time()
num_samples = 0
train_pass_acc.reset()
def run_step(batch_id, data):
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, b_size = exe.run(
trainer_prog,
feed={"pixel": img_data,
"label": y_data},
fetch_list=[avg_cost, batch_acc, batch_size])
return loss, acc, b_size
if args.profile:
with profiler.profiler('All', 'total',
'/tmp/profile_vgg_%d' % args.task_index):
for batch_id, data in enumerate(train_reader()):
if batch_id > 5: break
run_step(batch_id, data)
total_time = 0.0
count = 0
for batch_id, data in enumerate(train_reader()):
ts = time.time()
loss, acc, b_size = run_step(batch_id, data)
iters += 1
num_samples += len(data)
train_pass_acc.add(value=acc, weight=b_size)
duration = time.time() - ts
total_time += duration
count += len(data)
print(
"Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, "
"Speed = %.2f (%.2f) img/s" % (pass_id, iters, loss, acc,
len(data) / duration,
count / total_time)
) # The accuracy is the accumulation of batches, but not the current batch.
pass_elapsed = time.time() - start_time
pass_train_acc = train_pass_acc.eval()
pass_test_acc = test(exe)
print("Task:%d Pass = %d, Training performance = %f imgs/s, "
"Train accuracy = %f, Test accuracy = %f\n" %
(args.task_index, 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:
trainers = int(os.getenv("TRAINERS")) # total trainer count
print("trainers total: ", trainers)
training_role = os.getenv(
"TRAINING_ROLE",
"TRAINER") # get the training role: trainer/pserver
t = fluid.DistributeTranspiler()
t.transpile(
trainer_id=args.task_index,
pservers=args.ps_hosts,
trainers=trainers)
if training_role == "PSERVER":
current_endpoint = os.getenv("POD_IP") + ":" + os.getenv(
"PADDLE_INIT_PORT")
if not current_endpoint:
print("need env SERVER_ENDPOINT")
exit(1)
pserver_prog = t.get_pserver_program(current_endpoint)
pserver_startup = t.get_startup_program(current_endpoint,
pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog)
elif training_role == "TRAINER":
# 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_tf.py 已删除 100644 → 0
浏览文件 @ 20c8ff0f
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""VGG16 benchmark in TensorFlow
You can get distribution example template structure here:
https://medium.com/clusterone/how-to-write-distributed-tensorflow-code-with-an-example-on-tensorport-70bf3306adcb
https://www.tensorflow.org/deploy/distributed
"""
import tensorflow as tf
import paddle.v2 as paddle
import numpy as np
import argparse
import time
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
'--batch_size', type=int, default=128, help="Batch size for training.")
parser.add_argument(
'--learning_rate',
type=float,
default=1e-3,
help="Learning rate for training.")
parser.add_argument('--num_passes', type=int, default=50, help="No. of passes.")
parser.add_argument(
'--device',
type=str,
default='CPU',
choices=['CPU', 'GPU'],
help="The device type.")
parser.add_argument(
'--data_format',
type=str,
default='NHWC',
choices=['NCHW', 'NHWC'],
help='The data order, NCHW=[batch, channels, height, width].'
'Only support NHWC right now.')
parser.add_argument(
'--data_set',
type=str,
default='cifar10',
choices=['cifar10', 'flowers'],
help='Optional dataset for benchmark.')
parser.add_argument(
"--ps_hosts",
type=str,
default="",
help="Comma-separated list of hostname:port pairs")
parser.add_argument(
"--worker_hosts",
type=str,
default="",
help="Comma-separated list of hostname:port pairs")
parser.add_argument(
"--job_name", type=str, default="", help="One of 'worker', 'ps'")
# Flags for defining the tf.train.Server
parser.add_argument(
"--task_index", type=int, default=0, help="Index of task within the job")
args = parser.parse_args()
class VGG16Model(object):
def __init__(self):
self.parameters = []
def batch_norm_relu(self, inputs, is_training):
"""Performs a batch normalization followed by a ReLU."""
# We set fused=True for a significant speed boost. See
# https://www.tensorflow.org/speed/speed_guide#common_fused_ops
inputs = tf.layers.batch_normalization(
inputs=inputs,
axis=1 if args.data_format == 'NCHW' else -1,
momentum=0.9,
epsilon=1e-05,
center=True,
scale=True,
training=is_training,
fused=True)
inputs = tf.nn.relu(inputs)
return inputs
def conv_bn_layer(self,
name,
images,
kernel_shape,
is_training,
drop_rate=0.0):
with tf.name_scope(name) as scope:
kernel = tf.Variable(
tf.truncated_normal(
kernel_shape, dtype=tf.float32, stddev=1e-1),
name='weights')
conv = tf.nn.conv2d(
images,
kernel, [1, 1, 1, 1],
data_format=args.data_format,
padding='SAME')
biases = tf.Variable(
tf.constant(
0.0, shape=[kernel_shape[-1]], dtype=tf.float32),
trainable=True,
name='biases')
out = tf.nn.bias_add(conv, biases)
out = self.batch_norm_relu(out, is_training)
out = tf.layers.dropout(out, rate=drop_rate, training=is_training)
return out
def fc_layer(self, name, inputs, shape):
with tf.name_scope(name) as scope:
fc_w = tf.Variable(
tf.truncated_normal(
shape, dtype=tf.float32, stddev=1e-1),
name='weights')
fc_b = tf.Variable(
tf.constant(
0.0, shape=[shape[-1]], dtype=tf.float32),
trainable=True,
name='biases')
out = tf.nn.bias_add(tf.matmul(inputs, fc_w), fc_b)
return out
def network(self, images, class_dim, is_training):
""" VGG16 model structure.
TODO(kuke): enable this network to support the 'NCHW' data format
"""
# conv1
conv1_1 = self.conv_bn_layer(
'conv1_1', images, [3, 3, 3, 64], is_training, drop_rate=0.3)
conv1_2 = self.conv_bn_layer(
'conv1_2', conv1_1, [3, 3, 64, 64], is_training, drop_rate=0.0)
# pool1
pool1 = tf.nn.max_pool(
conv1_2,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool1')
# conv2
conv2_1 = self.conv_bn_layer(
'conv2_1', pool1, [3, 3, 64, 128], is_training, drop_rate=0.4)
conv2_2 = self.conv_bn_layer(
'conv2_2', conv2_1, [3, 3, 128, 128], is_training, drop_rate=0.0)
# pool2
pool2 = tf.nn.max_pool(
conv2_2,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool2')
# conv3
conv3_1 = self.conv_bn_layer(
'conv3_1', pool2, [3, 3, 128, 256], is_training, drop_rate=0.4)
conv3_2 = self.conv_bn_layer(
'conv3_2', conv3_1, [3, 3, 256, 256], is_training, drop_rate=0.4)
conv3_3 = self.conv_bn_layer(
'conv3_3', conv3_2, [3, 3, 256, 256], is_training, drop_rate=0.0)
# pool3
pool3 = tf.nn.max_pool(
conv3_3,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool3')
# conv4
conv4_1 = self.conv_bn_layer(
'conv4_1', pool3, [3, 3, 256, 512], is_training, drop_rate=0.4)
conv4_2 = self.conv_bn_layer(
'conv4_2', conv4_1, [3, 3, 512, 512], is_training, drop_rate=0.4)
conv4_3 = self.conv_bn_layer(
'conv4_3', conv4_2, [3, 3, 512, 512], is_training, drop_rate=0.0)
# pool4
pool4 = tf.nn.max_pool(
conv4_3,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool4')
# conv5
conv5_1 = self.conv_bn_layer(
'conv5_1', pool4, [3, 3, 512, 512], is_training, drop_rate=0.4)
conv5_2 = self.conv_bn_layer(
'conv5_2', conv5_1, [3, 3, 512, 512], is_training, drop_rate=0.4)
conv5_3 = self.conv_bn_layer(
'conv5_3', conv5_2, [3, 3, 512, 512], is_training, drop_rate=0.0)
# pool5
pool5 = tf.nn.max_pool(
conv5_3,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool4')
# flatten
shape = int(np.prod(pool5.get_shape()[1:]))
pool5_flat = tf.reshape(pool5, [-1, shape])
# fc1
drop = tf.layers.dropout(pool5_flat, rate=0.5, training=is_training)
fc1 = self.fc_layer('fc1', drop, [shape, 512])
# fc2
bn = self.batch_norm_relu(fc1, is_training)
drop = tf.layers.dropout(bn, rate=0.5, training=is_training)
fc2 = self.fc_layer('fc2', drop, [512, 512])
fc3 = self.fc_layer('fc3', fc2, [512, class_dim])
return fc3
def run_benchmark(cluster_spec, server):
"""Run benchmark on cifar10 or flowers."""
if args.data_set == "cifar10":
class_dim = 10
raw_shape = (3, 32, 32)
dat_shape = (None, 32, 32, 3) if args.data_format == 'NHWC' else (
None, 3, 32, 32)
else:
class_dim = 102
raw_shape = (3, 224, 224)
dat_shape = (None, 224, 224, 3) if args.data_format == 'NHWC' else (
None, 3, 224, 224)
device = tf.train.replica_device_setter(
worker_device="/job:worker/task:{}".format(args.task_index),
cluster=cluster_spec)
with tf.device(device):
images = tf.placeholder(tf.float32, shape=dat_shape)
labels = tf.placeholder(tf.int64, shape=(None, ))
is_training = tf.placeholder('bool')
onehot_labels = tf.one_hot(labels, depth=class_dim)
vgg16 = VGG16Model()
logits = vgg16.network(images, class_dim, is_training)
loss = tf.losses.softmax_cross_entropy(
onehot_labels=onehot_labels, logits=logits)
avg_loss = tf.reduce_mean(loss)
correct = tf.equal(tf.argmax(logits, 1), labels)
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
global_step = tf.Variable(0, name='global_step', trainable=False)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(avg_loss, global_step=global_step)
summary_op = tf.summary.merge_all()
init_op = tf.global_variables_initializer()
# data reader
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.cifar.train10()
if args.data_set == 'cifar10' else paddle.dataset.flowers.train(),
buf_size=5120),
batch_size=args.batch_size)
test_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.cifar.test10()
if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
buf_size=5120),
batch_size=args.batch_size)
# test
def test():
test_accs = []
for batch_id, data in enumerate(test_reader()):
test_images = np.array(
map(lambda x: np.transpose(x[0].reshape(raw_shape),
axes=[1, 2, 0]) if args.data_format == 'NHWC' else x[0], data)).astype("float32")
test_labels = np.array(map(lambda x: x[1], data)).astype('int64')
test_accs.append(
accuracy.eval(feed_dict={
images: test_images,
labels: test_labels,
is_training: False
}))
return np.mean(test_accs)
config = tf.ConfigProto(
intra_op_parallelism_threads=1,
inter_op_parallelism_threads=1,
log_device_placement=True)
config.gpu_options.allow_growth = True
hooks = [tf.train.StopAtStepHook(last_step=1000000)]
with tf.train.MonitoredTrainingSession(
master=server.target,
is_chief=(args.task_index == 0),
hooks=hooks,
config=config) as sess:
iters, num_samples, start_time = 0, 0, 0.0
for pass_id in range(args.num_passes):
# train
num_samples = 0
start_time = time.time()
for batch_id, data in enumerate(train_reader()):
train_images = np.array(
map(lambda x: np.transpose(x[0].reshape(raw_shape),
axes=[1, 2, 0]) if args.data_format == 'NHWC' else x[0], data)).astype("float32")
train_labels = np.array(map(lambda x: x[1], data)).astype(
'int64')
iter_begin_time = time.time()
_, loss, acc = sess.run([train_op, avg_loss, accuracy],
feed_dict={
images: train_images,
labels: train_labels,
is_training: True
})
iters += 1
print(
"Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, Speed=%.2f imgs/sec"
% (pass_id, iters, loss, acc,
len(data) / (time.time() - iter_begin_time)))
num_samples += len(data)
train_elapsed = time.time() - start_time
# test
pass_test_acc = test()
print("Pass = %d, Train speed = %f imgs/s, Test accuracy = %f\n" %
(pass_id, num_samples / train_elapsed, pass_test_acc))
def print_arguments():
print('----------- Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
if __name__ == '__main__':
print_arguments()
ps_hosts = args.ps_hosts.split(",")
worker_hosts = args.worker_hosts.split(",")
# Create a cluster from the parameter server and worker hosts.
cluster_spec = tf.train.ClusterSpec({
"ps": ps_hosts,
"worker": worker_hosts
})
# Create and start a server for the local task.
server = tf.train.Server(
cluster_spec, job_name=args.job_name, task_index=args.task_index)
if args.job_name == "ps":
print("start pserver")
server.join()
elif args.job_name == "worker":
print("start worker")
run_benchmark(cluster_spec, server)
benchmark/cluster/vgg16/vgg16_v2.py 已删除 100644 → 0
浏览文件 @ 20c8ff0f
# 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 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()
benchmark/fluid/README.md
浏览文件 @ 632108dc
......@@ -24,22 +24,22 @@ Currently supported `--model` argument include:
* Run the following command to start a benchmark job locally:
```bash
python fluid_benchmark.py --model mnist --parallel 1 --device GPU --with_test
python fluid_benchmark.py --model mnist --device GPU
```
You can choose to use GPU/CPU training. With GPU training, you can specify
`--parallel 1` to run multi GPU training.
`--gpus <gpu_num>` to run multi GPU training.
* Run distributed training with parameter servers:
* start parameter servers:
```bash
PADDLE_TRAINING_ROLE=PSERVER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=1 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --parallel 0 --device GPU --update_method pserver
PADDLE_TRAINING_ROLE=PSERVER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=1 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method pserver
```
* start trainers:
```bash
PADDLE_TRAINING_ROLE=PSERVER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=1 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --parallel 0 --device GPU --update_method pserver
PADDLE_TRAINING_ROLE=TRAINER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=1 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method pserver
```
* Run distributed training using NCCL2
```bash
PADDLE_PSERVER_PORT=7164 PADDLE_TRAINER_IPS=192.168.0.2,192.168.0.3 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --parallel 0 --device GPU --update_method nccl2
PADDLE_PSERVER_PORT=7164 PADDLE_TRAINER_IPS=192.168.0.2,192.168.0.3 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method nccl2
```
## Run Distributed Benchmark on Kubernetes Cluster
......@@ -48,7 +48,7 @@ We provide a script `kube_gen_job.py` to generate Kubernetes yaml files to submi
distributed benchmark jobs to your cluster. To generate a job yaml, just run:
```bash
python kube_gen_job.py --jobname myjob --pscpu 4 --cpu 8 --gpu 8 --psmemory 20 --memory 40 --pservers 4 --trainers 4 --entry "python fluid_benchmark.py --model mnist --parallel 1 --device GPU --update_method pserver --with_test" --disttype pserver
python kube_gen_job.py --jobname myjob --pscpu 4 --cpu 8 --gpu 8 --psmemory 20 --memory 40 --pservers 4 --trainers 4 --entry "python fluid_benchmark.py --model mnist --parallel 1 --device GPU --update_method pserver " --disttype pserver
```
Then the yaml files are generated under directory `myjob`, you can run:
......
benchmark/fluid/fluid_benchmark.py
浏览文件 @ 632108dc
......@@ -94,6 +94,10 @@ def parse_args():
'--memory_optimize',
action='store_true',
help='If set, optimize runtime memory before start.')
parser.add_argument(
'--use_fake_data',
action='store_true',
help='If set ommit the actual read data operators.')
parser.add_argument(
'--update_method',
type=str,
......@@ -198,6 +202,10 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
exe.run(train_prog)
return
if args.use_fake_data:
raise Exception(
"fake data is not supported in single GPU test for now.")
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(startup_prog)
......@@ -244,7 +252,31 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
batch_acc, args, train_prog, startup_prog, nccl_id_var,
num_trainers, trainer_id):
feed_var_list = [
var for var in train_prog.global_block().vars.itervalues()
if var.is_data
]
# generate fake:
if args.use_fake_data:
for var in feed_var_list:
v = startup_prog.global_block().clone_variable(var)
var.persistable = True
v.persistable = True
real_shape = list(var.shape)
real_shape[0] = args.batch_size / args.gpus
startup_prog.global_block().append_op(
outputs={"Out": v},
type="fill_constant",
attrs={"shape": real_shape,
"value": 1.0,
"dtype": var.dtype})
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
if nccl_id_var and trainer_id == 0:
#FIXME(wuyi): wait other trainer to start listening
time.sleep(30)
startup_exe = fluid.Executor(place)
startup_exe.run(startup_prog)
strategy = fluid.ExecutionStrategy()
......@@ -256,10 +288,7 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
exec_strategy=strategy,
num_trainers=num_trainers,
trainer_id=trainer_id)
feed_var_list = [
var for var in train_prog.global_block().vars.itervalues()
if var.is_data
]
feeder = fluid.DataFeeder(feed_var_list, place)
for pass_id in range(args.pass_num):
num_samples = 0
......@@ -271,7 +300,10 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
num_samples = 0
if iters == args.iterations:
break
loss, = exe.run([avg_loss.name], feed=feeder.feed(data))
if args.use_fake_data:
loss, = exe.run([avg_loss.name])
else:
loss, = exe.run([avg_loss.name], feed=feeder.feed(data))
if args.update_method == "pserver":
exe.bcast_params()
num_samples += len(data)
......
benchmark/fluid/kube_gen_job.py
浏览文件 @ 632108dc
......@@ -112,6 +112,7 @@ def gen_job():
envs.append({"name": "PSERVERS", "value": str(args.pservers)})
envs.append({"name": "ENTRY", "value": args.entry})
envs.append({"name": "PADDLE_INIT_PORT", "value": str(args.port)})
envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)})
# NOTE: these directories below are cluster specific, please modify
# this settings before you run on your own cluster.
envs.append({
......
benchmark/fluid/kube_templates/__init__.py
浏览文件 @ 632108dc
......@@ -54,5 +54,13 @@ envs = [
"fieldPath": "status.podIP"
}
}
},
{
"name": "PADDLE_CURRENT_IP",
"valueFrom": {
"fieldRef": {
"fieldPath": "status.podIP"
}
}
}
]
cmake/configure.cmake
浏览文件 @ 632108dc
......@@ -41,6 +41,10 @@ if(USE_EIGEN_FOR_BLAS)
add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS)
endif(USE_EIGEN_FOR_BLAS)
if(EIGEN_USE_THREADS)
add_definitions(-DEIGEN_USE_THREADS)
endif(EIGEN_USE_THREADS)
if(NOT WITH_PROFILER)
add_definitions(-DPADDLE_DISABLE_PROFILER)
endif(NOT WITH_PROFILER)
......
cmake/external/protobuf.cmake
浏览文件 @ 632108dc
......@@ -212,6 +212,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
${CMAKE_COMMAND} ${PROTOBUF_SOURCES_DIR}/src/${TARGET_NAME}/cmake
${OPTIONAL_ARGS}
-Dprotobuf_BUILD_TESTS=OFF
-DCMAKE_SKIP_RPATH=ON
-DCMAKE_POSITION_INDEPENDENT_CODE=ON
-DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
-DCMAKE_INSTALL_PREFIX=${PROTOBUF_INSTALL_DIR}
......
doc/fluid/api/layers.rst
浏览文件 @ 632108dc
......@@ -1003,9 +1003,9 @@ dice_loss
.. autofunction:: paddle.fluid.layers.dice_loss
:noindex:
bilinear_interp
upsampling_bilinear2d
____
.. autofunction:: paddle.fluid.layers.bilinear_interp
.. autofunction:: paddle.fluid.layers.upsampling_bilinear2d
:noindex:
doc/v2/build_and_install/build_from_source_cn.rst
浏览文件 @ 632108dc
......@@ -35,13 +35,11 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安
# 2. 可选步骤：源码中构建用于编译PaddlePaddle的Docker镜像
docker build -t paddle:dev .
# 3. 执行下面的命令编译CPU-Only的二进制
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x /paddle/paddle/scripts/paddle_build.sh build
docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 ./paddle/scripts/paddle_build.sh build
# 4. 或者也可以使用为上述可选步骤构建的镜像（必须先执行第2步）
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddle:dev
docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddle:dev ./paddle/scripts/paddle_build.sh build
注：上述命令把当前目录（源码树根目录）映射为 container 里的 :code:`/paddle` 目录。如果使用自行
构建的镜像（上述第4步）会执行 :code:`Dockerfile` 描述的默认入口程序 :code:`build.sh` 可以省略步骤3中
最后的执行脚本的命令。
注：上述命令把当前目录（源码树根目录）映射为 container 里的 :code:`/paddle` 目录。
编译完成后会在build/python/dist目录下生成输出的whl包，可以选在在当前机器安装也可以拷贝到目标机器安装：
......@@ -72,15 +70,15 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安
.. code-block:: bash
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x /paddle/paddle/scripts/docker/build.sh
docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 ./paddle/scripts/paddle_build.sh test
如果期望执行其中一个单元测试，（比如 :code:`test_sum_op` ）：
.. code-block:: bash
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 /bin/bash
bash /paddle/paddle/scripts/docker/build.sh
cd /paddle/build
docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 /bin/bash
./paddle/scripts/paddle_build.sh build
cd build
ctest -R test_sum_op -V
.. _faq_docker:
......
doc/v2/build_and_install/build_from_source_en.rst
浏览文件 @ 632108dc
......@@ -34,14 +34,12 @@ Or you can build your own image from source as the optional step below:
# 2. Optional: build development docker image from source
docker build -t paddle:dev .
# 3. Run the following command to build a CPU-Only binaries
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x /paddle/paddle/scripts/paddle_build.sh build
docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 ./paddle/scripts/paddle_build.sh build
# 4. Or, use your built Docker image to build PaddlePaddle (must run step 2)
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddle:dev
docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddle:dev ./paddle/scripts/paddle_build.sh build
NOTE: The above command try to mount the current working directory (root directory of source code)
into :code:`/paddle` directory inside docker container. If you are using your own image
(Step 4) it will run default entry-point :code:`build.sh` , so you could omit the last
command in step 3.
into :code:`/paddle` directory inside docker container.
When the compile finishes, you can get the output whl package under
build/python/dist, then you can choose to install the whl on local
......@@ -74,15 +72,15 @@ Set :code:`WITH_GPU=ON` Can also run tests on GPU.
.. code-block:: bash
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x paddle/paddle/scripts/docker/build.sh
docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 ./paddle/scripts/paddle_build.sh test
If you wish to run only one unit test, like :code:`test_sum_op`:
.. code-block:: bash
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 /bin/bash
bash /paddle/paddle/scripts/docker/build.sh
cd /paddle/build
docker run -it -v $PWD:/paddle -w /paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 /bin/bash
./paddle/scripts/paddle_build.sh build
cd build
ctest -R test_sum_op -V
.. _faq_docker:
......
doc/v2/build_and_install/docker_install_cn.rst
浏览文件 @ 632108dc
......@@ -98,7 +98,7 @@ PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Note
国内用户可以使用下面的镜像源来加速访问：
.. code-block: bash
.. code-block:: bash
docker run -p 8888:8888 docker.paddlepaddlehub.com/book
......
doc/v2/build_and_install/docker_install_en.rst
浏览文件 @ 632108dc
......@@ -105,7 +105,7 @@ We provide a packaged book image, simply issue the command:
For users in China, we provide a faster mirror:
.. code-block: bash
.. code-block:: bash
docker run -p 8888:8888 docker.paddlepaddlehub.com/book
......
paddle/.gitignore
浏览文件 @ 632108dc
......@@ -11,7 +11,6 @@ GTAGS
*.pb.cc
*.pb.h
*_pb2.py
paddle_*
output/
google/
Makefile
......
paddle/contrib/CMakeLists.txt 0 → 100644
浏览文件 @ 632108dc
# Copyright (c) 2016 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.
#
add_subdirectory(inference)
paddle/contrib/float16/README.md
浏览文件 @ 632108dc
......@@ -89,7 +89,7 @@ cd Paddle
# to `FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04` and similarly for other configurations
nvidia-docker build -t paddle:float16 .
# After running this, different results will be written to different log files in Paddle/contrib/float16/
nvidia-docker run -it -v $PWD:/paddle paddle:float16 /paddle/contrib/float16/run_float16_demo.sh
nvidia-docker run -it -v $PWD:/paddle paddle:float16 /paddle/paddle/contrib/float16/run_float16_demo.sh
```
#### Accuracy
......
paddle/contrib/float16/run_float16_demo.sh
浏览文件 @ 632108dc
......@@ -3,7 +3,7 @@
BUILD_PATH=/paddle/fp16_build
WHEEL_PATH=$BUILD_PATH/python/dist
INFER_PATH=$BUILD_PATH/paddle/fluid/inference/tests/book
DEMO_PATH=/paddle/contrib/float16
DEMO_PATH=/paddle/paddle/contrib/float16
# Use the single most powerful CUDA GPU on your machine
export CUDA_VISIBLE_DEVICES=0
......@@ -50,7 +50,6 @@ do
--repeat=1 \
$INFER_PATH/test_inference_image_classification_vgg \
--data_set=imagenet \
--dirname=$DEMO_PATH/image_classification_imagenet_vgg.inference.model \
--fp16_dirname=$DEMO_PATH/float16_image_classification_imagenet_vgg.inference.model \
--repeat=$REPEAT \
......@@ -68,7 +67,6 @@ do
--repeat=1 \
$INFER_PATH/test_inference_image_classification_resnet \
--data_set=imagenet \
--dirname=$DEMO_PATH/image_classification_imagenet_resnet.inference.model \
--fp16_dirname=$DEMO_PATH/float16_image_classification_imagenet_resnet.inference.model \
--repeat=$REPEAT \
......@@ -86,7 +84,6 @@ do
--repeat=1 \
$INFER_PATH/test_inference_image_classification_vgg \
--data_set=cifar10 \
--dirname=$DEMO_PATH/image_classification_cifar10_vgg.inference.model \
--fp16_dirname=$DEMO_PATH/float16_image_classification_cifar10_vgg.inference.model \
--repeat=$REPEAT \
......@@ -104,7 +101,6 @@ do
--repeat=1 \
$INFER_PATH/test_inference_image_classification_vgg \
--data_set=cifar10 \
--dirname=$DEMO_PATH/image_classification_cifar10_resnet.inference.model \
--fp16_dirname=$DEMO_PATH/float16_image_classification_cifar10_resnet.inference.model \
--repeat=$REPEAT \
......
paddle/contrib/inference/CMakeLists.txt 0 → 100644
浏览文件 @ 632108dc
# Copyright (c) 2016 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.
#
function(inference_api_test TARGET_NAME TEST_SRC DEP_TEST)
set(options "")
set(oneValueArgs "")
set(multiValueArgs ARGS)
cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
set(arg_list "")
if(inference_test_ARGS)
foreach(arg ${inference_test_ARGS})
list(APPEND arg_list "_${arg}")
endforeach()
else()
list(APPEND arg_list "_")
endif()
foreach(arg ${arg_list})
string(REGEX REPLACE "^_$" "" arg "${arg}")
cc_test(${TARGET_NAME}
SRCS ${TEST_SRC}
DEPS paddle_fluid_api paddle_inference_api paddle_inference_api_impl
ARGS --dirname=${PYTHON_TESTS_DIR}/book/)
# set_tests_properties(${TARGET_NAME}
# PROPERTIES DEPENDS ${DEP_TEST})
endforeach()
endfunction(inference_api_test)
cc_library(paddle_inference_api
SRCS paddle_inference_api.cc
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
cc_library(paddle_inference_api_impl
SRCS paddle_inference_api_impl.cc
DEPS paddle_inference_api paddle_fluid_api)
cc_test(test_paddle_inference_api
SRCS test_paddle_inference_api.cc
DEPS paddle_inference_api)
inference_api_test(test_paddle_inference_api_impl
test_paddle_inference_api_impl.cc
test_word2vec)
paddle/contrib/inference/paddle_inference_api.cc 0 → 100644
浏览文件 @ 632108dc
/* 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/contrib/inference/paddle_inference_api.h"
paddle/contrib/inference/paddle_inference_api.h
浏览文件 @ 632108dc
......@@ -12,49 +12,74 @@
See the License for the specific language governing permissions and
limitations under the License. */
/*
* This file contains the definition of a simple Inference API for Paddle.
*
* ATTENTION: It requires some C++ features, for lower version C++ or C, we
* might release another API.
*/
#pragma once
#include <memory>
#include <string>
#include <vector>
namespace paddle {
class Predictor {
public:
struct Attr;
Predictor() = default;
enum PaddleDType {
FLOAT32,
INT64,
};
// Build the network before inference.
bool Init(const Attr& attr);
struct PaddleBuf {
void* data; // pointer to the data memory.
size_t length; // number of memory bytes.
};
struct PaddleTensor {
std::string name; // variable name.
std::vector<int> shape;
PaddleBuf data; // blob of data.
PaddleDType dtype;
};
/*
* A simple Inference API for Paddle. Currently this API might just be used by
* non-sequence scenerios.
* TODO(Superjomn) Prepare another API for NLP-related usages.
*/
class PaddlePredictor {
public:
struct Config;
PaddlePredictor() = default;
PaddlePredictor(const PaddlePredictor&) = delete;
// Predict an record.
// Arguments:
// inputs: the name of the input variables.
// outputs: the name of the output varaibles.
// input_shapes: the shape of the input variables.
// output_shapes: the shape of the output variables.
// input_data: the data of the input variables.
// output_data: the data of the output variables.
bool Run(const std::vector<std::string>& inputs,
const std::vector<std::string>& outputs,
const std::vector<std::vector<int>>& input_shapes,
const std::vector<std::vector<int>>& output_shapes,
const std::vector<std::vector<float>>& input_data,
std::vector<std::vector<float>>* output_data);
// Clone a predictor that share the model weights.
Predictor* Clone();
// The caller should be responsible for allocating and releasing the memory of
// `inputs`. `inputs` should be alive until Run returns. caller should be
// responsible for releasing the memory of `output_data`.
virtual bool Run(const std::vector<PaddleTensor>& inputs,
std::vector<PaddleTensor>* output_data) = 0;
// Clone a predictor that share the model weights, the Cloned predictor should
// be thread-safe.
virtual std::unique_ptr<PaddlePredictor> Clone() = 0;
// Destroy the Predictor.
~Predictor();
virtual ~PaddlePredictor() {}
struct Attr {
friend std::unique_ptr<PaddlePredictor> CreatePaddlePredictor(
const PaddlePredictor::Config& config);
// The common configs for all the predictors.
struct Config {
enum class EngineKind;
std::string model_dir; // path to the model directory.
bool enable_engine{false}; // Enable to execute (part of) the model on
// third-party engines.
EngineKind engine_kind{Attr::EngineKind::kNone};
// third-party engines.
EngineKind engine_kind{Config::EngineKind::kNone};
enum class EngineKind {
kNone = -1, // Use the native Fluid facility.
......@@ -66,4 +91,8 @@ public:
};
};
// A factory to help create difference predictor.
template <typename ConfigT>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor(const ConfigT& config);
} // namespace paddle
paddle/contrib/inference/paddle_inference_api_impl.cc 0 → 100644
浏览文件 @ 632108dc
/* 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 <sys/time.h>
#include <algorithm>
#include <map>
#include <set>
#include <sstream>
#include <string>
#include <utility>
#include <vector>
#include "paddle/contrib/inference/paddle_inference_api_impl.h"
namespace paddle {
namespace {
// Timer for timer
class Timer {
public:
double start;
double startu;
void tic() {
struct timeval tp;
gettimeofday(&tp, NULL);
start = tp.tv_sec;
startu = tp.tv_usec;
}
double toc() {
struct timeval tp;
gettimeofday(&tp, NULL);
double used_time_ms =
(tp.tv_sec - start) * 1000.0 + (tp.tv_usec - startu) / 1000.0;
return used_time_ms;
}
};
template <class T>
std::string num2str(T a) {
std::stringstream istr;
istr << a;
return istr.str();
}
} // namespace
bool PaddlePredictorImpl::Init() {
VLOG(3) << "Predictor::init()";
// TODO(panyx0718): Should CPU vs GPU device be decided by id?
if (config_.device >= 0) {
place_ = paddle::platform::CUDAPlace(config_.device);
} else {
place_ = paddle::platform::CPUPlace();
}
paddle::framework::InitDevices(false);
executor_.reset(new paddle::framework::Executor(place_));
scope_.reset(new paddle::framework::Scope());
// Initialize the inference program
if (!config_.model_dir.empty()) {
// Parameters are saved in separate files sited in
// the specified `dirname`.
inference_program_ = paddle::inference::Load(
executor_.get(), scope_.get(), config_.model_dir);
} else if (!config_.prog_file.empty() && !config_.param_file.empty()) {
// All parameters are saved in a single file.
// The file names should be consistent with that used
// in Python API `fluid.io.save_inference_model`.
inference_program_ = paddle::inference::Load(
executor_.get(), scope_.get(), config_.prog_file, config_.param_file);
} else {
LOG(ERROR) << "fail to load inference model.";
return false;
}
ctx_ = executor_->Prepare(*inference_program_, 0);
// Create variables
// TODO(panyx0718): Why need to test share_variables here?
if (config_.share_variables) {
executor_->CreateVariables(*inference_program_, scope_.get(), 0);
}
// Get the feed_target_names and fetch_target_names
feed_target_names_ = inference_program_->GetFeedTargetNames();
fetch_target_names_ = inference_program_->GetFetchTargetNames();
return true;
}
bool PaddlePredictorImpl::Run(const std::vector<PaddleTensor> &inputs,
std::vector<PaddleTensor> *output_data) {
VLOG(3) << "Predictor::predict";
Timer timer;
timer.tic();
// set feed variable
std::map<std::string, const paddle::framework::LoDTensor *> feed_targets;
std::vector<paddle::framework::LoDTensor> feeds;
if (!SetFeed(inputs, &feeds)) {
LOG(ERROR) << "fail to set feed";
return false;
}
for (size_t i = 0; i < feed_target_names_.size(); ++i) {
feed_targets[feed_target_names_[i]] = &feeds[i];
}
// get fetch variable
std::map<std::string, paddle::framework::LoDTensor *> fetch_targets;
std::vector<paddle::framework::LoDTensor> fetchs;
fetchs.resize(fetch_target_names_.size());
for (size_t i = 0; i < fetch_target_names_.size(); ++i) {
fetch_targets[fetch_target_names_[i]] = &fetchs[i];
}
// Run the inference program
// if share variables, we need not create variables
executor_->RunPreparedContext(ctx_.get(),
scope_.get(),
&feed_targets,
&fetch_targets,
!config_.share_variables);
if (!GetFetch(fetchs, output_data)) {
LOG(ERROR) << "fail to get fetchs";
return false;
}
VLOG(3) << "predict cost: " << timer.toc() << "ms";
return true;
}
std::unique_ptr<PaddlePredictor> PaddlePredictorImpl::Clone() {
VLOG(3) << "Predictor::clone";
std::unique_ptr<PaddlePredictorImpl> cls(new PaddlePredictorImpl(config_));
if (!cls->InitShared(this)) {
LOG(ERROR) << "fail to call InitShared";
return nullptr;
}
return cls;
}
// TODO(panyx0718): Consider merge with Init()?
bool PaddlePredictorImpl::InitShared(PaddlePredictorImpl *cls) {
VLOG(3) << "Predictor::init_shared";
// 1. Define place, executor, scope
if (this->config_.device >= 0) {
place_ = paddle::platform::CUDAPlace();
} else {
place_ = paddle::platform::CPUPlace();
}
this->executor_.reset(new paddle::framework::Executor(this->place_));
this->scope_.reset(new paddle::framework::Scope());
// Initialize the inference program
if (!this->config_.model_dir.empty()) {
// Parameters are saved in separate files sited in
// the specified `dirname`.
this->inference_program_ = paddle::inference::Load(
this->executor_.get(), this->scope_.get(), this->config_.model_dir);
} else if (!this->config_.prog_file.empty() &&
!this->config_.param_file.empty()) {
// All parameters are saved in a single file.
// The file names should be consistent with that used
// in Python API `fluid.io.save_inference_model`.
this->inference_program_ =
paddle::inference::Load(this->executor_.get(),
this->scope_.get(),
this->config_.prog_file,
this->config_.param_file);
}
this->ctx_ = this->executor_->Prepare(*this->inference_program_, 0);
// 3. create variables
// TODO(panyx0718): why test share_variables.
if (config_.share_variables) {
this->executor_->CreateVariables(
*this->inference_program_, this->scope_.get(), 0);
}
// 4. Get the feed_target_names and fetch_target_names
this->feed_target_names_ = this->inference_program_->GetFeedTargetNames();
this->fetch_target_names_ = this->inference_program_->GetFetchTargetNames();
return true;
}
bool PaddlePredictorImpl::SetFeed(
const std::vector<PaddleTensor> &inputs,
std::vector<paddle::framework::LoDTensor> *feeds) {
VLOG(3) << "Predictor::set_feed";
if (inputs.size() != feed_target_names_.size()) {
LOG(ERROR) << "wrong feed input size.";
return false;
}
for (size_t i = 0; i < feed_target_names_.size(); ++i) {
paddle::framework::LoDTensor input;
paddle::framework::DDim ddim =
paddle::framework::make_ddim(inputs[i].shape);
void *input_ptr;
if (inputs[i].dtype == PaddleDType::INT64) {
input_ptr =
input.mutable_data<int64_t>(ddim, paddle::platform::CPUPlace());
} else if (inputs[i].dtype == PaddleDType::FLOAT32) {
input_ptr = input.mutable_data<float>(ddim, paddle::platform::CPUPlace());
} else {
LOG(ERROR) << "unsupported feed type " << inputs[i].dtype;
return false;
}
// TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
std::memcpy(static_cast<void *>(input_ptr),
inputs[i].data.data,
inputs[i].data.length);
feeds->push_back(input);
LOG(ERROR) << "Actual feed type " << feeds->back().type().name();
}
return true;
}
bool PaddlePredictorImpl::GetFetch(
const std::vector<paddle::framework::LoDTensor> &fetchs,
std::vector<PaddleTensor> *outputs) {
VLOG(3) << "Predictor::get_fetch";
outputs->resize(fetchs.size());
for (size_t i = 0; i < fetchs.size(); ++i) {
// TODO(panyx0718): Support fetch of other types.
if (fetchs[i].type() != typeid(float)) {
LOG(ERROR) << "only support fetching float now.";
return false;
}
std::vector<int> shape;
auto dims_i = fetchs[i].dims();
auto lod = fetchs[i].lod();
const float *output_ptr = fetchs[i].data<float>();
// const int64_t* output_ptr = fetchs[i].data<int64_t>();
auto num = fetchs[i].numel();
std::vector<float> data;
if (0 == lod.size()) {
std::copy(output_ptr, output_ptr + num, std::back_inserter(data));
for (int j = 0; j < dims_i.size(); ++j) {
shape.push_back(dims_i[j]);
}
} else {
// for batch detection
// image[0] -> output[0] shape {145, 6}
// image[1] -> output[1] shape {176, 6}
// then,
// the batch output shape {321, 6}
// the lod {{0, 145, 321}}
// so we should append output[0] to {176, 6}
size_t max_dim = 0;
for (size_t j = 1; j < lod[0].size(); j++) {
max_dim = std::max(max_dim, lod[0][j] - lod[0][j - 1]);
}
size_t common_dim = lod[0].back() == 0 ? 0 : num / lod[0].back();
if (max_dim > 0) {
data.resize((lod[0].size() - 1) * max_dim * common_dim, 0);
}
for (size_t j = 1; j < lod[0].size(); j++) {
size_t start = lod[0][j - 1] * common_dim;
size_t end = lod[0][j] * common_dim;
if (end > start) {
std::copy(output_ptr + start,
output_ptr + end,
data.begin() + (j - 1) * max_dim * common_dim);
}
}
shape.push_back(lod[0].size() - 1);
shape.push_back(max_dim);
for (int j = 1; j < dims_i.size(); ++j) {
shape.push_back(dims_i[j]);
}
}
outputs->at(i).shape = shape;
outputs->at(i).data.length = sizeof(float) * data.size();
outputs->at(i).data.data = malloc(outputs->at(i).data.length);
std::memcpy(
outputs->at(i).data.data, data.data(), outputs->at(i).data.length);
outputs->at(i).dtype = PaddleDType::FLOAT32;
// TODO(panyx0718): support other types? fill tensor name? avoid a copy.
}
return true;
}
std::unique_ptr<PaddlePredictorImpl> CreatePaddlePredictorImpl(
const VisConfig &config) {
VLOG(3) << "create PaddlePredictorImpl";
// 1. GPU memeroy
std::vector<std::string> flags;
if (config.fraction_of_gpu_memory >= 0.0f ||
config.fraction_of_gpu_memory <= 0.95f) {
flags.push_back("dummpy");
std::string flag = "--fraction_of_gpu_memory_to_use=" +
num2str<float>(config.fraction_of_gpu_memory);
flags.push_back(flag);
VLOG(3) << "set flag: " << flag;
framework::InitGflags(flags);
}
std::unique_ptr<PaddlePredictorImpl> predictor(
new PaddlePredictorImpl(config));
if (!predictor->Init()) {
return nullptr;
}
return predictor;
}
} // namespace paddle
paddle/contrib/inference/paddle_inference_api_impl.h 0 → 100644
浏览文件 @ 632108dc
/* 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 <glog/logging.h>
#include <memory>
#include <string>
#include <vector>
#include "paddle/contrib/inference/paddle_inference_api.h"
#include "paddle/fluid/framework/ddim.h"
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/inference/io.h"
#include "paddle/fluid/platform/profiler.h"
namespace paddle {
struct VisConfig : public PaddlePredictor::Config {
int device;
float fraction_of_gpu_memory;
std::string prog_file;
std::string param_file;
bool share_variables;
};
/*
* Do not use this, just a demo indicating how to customize a Predictor.
*/
class PaddlePredictorImpl : public PaddlePredictor {
public:
explicit PaddlePredictorImpl(const VisConfig &config) : config_(config) {}
bool Init();
bool Run(const std::vector<PaddleTensor> &inputs,
std::vector<PaddleTensor> *output_data) override;
std::unique_ptr<PaddlePredictor> Clone() override;
~PaddlePredictorImpl() override{};
private:
bool InitShared(PaddlePredictorImpl *cls);
bool SetFeed(const std::vector<PaddleTensor> &input_datas,
std::vector<paddle::framework::LoDTensor> *feeds);
bool GetFetch(const std::vector<paddle::framework::LoDTensor> &fetchs,
std::vector<PaddleTensor> *output_data);
VisConfig config_;
paddle::platform::Place place_;
std::unique_ptr<paddle::framework::Executor> executor_;
std::unique_ptr<paddle::framework::Scope> scope_;
std::unique_ptr<paddle::framework::ExecutorPrepareContext> ctx_;
std::unique_ptr<paddle::framework::ProgramDesc> inference_program_;
std::vector<std::string> feed_target_names_;
std::vector<std::string> fetch_target_names_;
};
std::unique_ptr<PaddlePredictorImpl> CreatePaddlePredictorImpl(
const VisConfig &config);
} // namespace paddle
paddle/contrib/inference/test_paddle_inference_api.cc 0 → 100644
浏览文件 @ 632108dc
/* 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/contrib/inference/paddle_inference_api.h"
#include <glog/logging.h>
#include <gtest/gtest.h>
namespace paddle {
/*
* Do not use this, just a demo indicating how to customize a config for a
* specific predictor.
*/
struct DemoConfig : public PaddlePredictor::Config {
float other_config;
};
/*
* Do not use this, just a demo indicating how to customize a Predictor.
*/
class DemoPredictor : public PaddlePredictor {
public:
explicit DemoPredictor(const DemoConfig &config) {
LOG(INFO) << "I get other_config " << config.other_config;
}
bool Run(const std::vector<PaddleTensor> &inputs,
std::vector<PaddleTensor> *output_data) override {
LOG(INFO) << "Run";
return false;
}
std::unique_ptr<PaddlePredictor> Clone() override { return nullptr; }
~DemoPredictor() override {}
};
template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<DemoConfig>(
const DemoConfig &config) {
std::unique_ptr<PaddlePredictor> x(new DemoPredictor(config));
return x;
}
TEST(paddle_inference_api, demo) {
DemoConfig config;
config.other_config = 1.7;
auto predictor = CreatePaddlePredictor(config);
std::vector<PaddleTensor> outputs;
predictor->Run({}, &outputs);
}
} // namespace paddle
paddle/contrib/inference/test_paddle_inference_api_impl.cc 0 → 100644
浏览文件 @ 632108dc
/* 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 <glog/logging.h>
#include <gtest/gtest.h>
#include "gflags/gflags.h"
#include "paddle/contrib/inference/paddle_inference_api_impl.h"
#include "paddle/fluid/inference/tests/test_helper.h"
DEFINE_string(dirname, "", "Directory of the inference model.");
namespace paddle {
PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) {
PaddleTensor pt;
pt.data.data = t->data<void>();
if (t->type() == typeid(int64_t)) {
pt.data.length = t->numel() * sizeof(int64_t);
pt.dtype = PaddleDType::INT64;
} else if (t->type() == typeid(float)) {
pt.data.length = t->numel() * sizeof(float);
pt.dtype = PaddleDType::FLOAT32;
} else {
LOG(FATAL) << "unsupported type.";
}
pt.shape = framework::vectorize2int(t->dims());
return pt;
}
TEST(paddle_inference_api_impl, word2vec) {
VisConfig config;
config.model_dir = FLAGS_dirname + "word2vec.inference.model";
LOG(INFO) << "dirname " << config.model_dir;
config.fraction_of_gpu_memory = 0.85;
config.device = 0;
config.share_variables = true;
std::unique_ptr<PaddlePredictorImpl> predictor =
CreatePaddlePredictorImpl(config);
framework::LoDTensor first_word, second_word, third_word, fourth_word;
framework::LoD lod{{0, 1}};
int64_t dict_size = 2073; // The size of dictionary
SetupLoDTensor(&first_word, lod, static_cast<int64_t>(0), dict_size - 1);
SetupLoDTensor(&second_word, lod, static_cast<int64_t>(0), dict_size - 1);
SetupLoDTensor(&third_word, lod, static_cast<int64_t>(0), dict_size - 1);
SetupLoDTensor(&fourth_word, lod, static_cast<int64_t>(0), dict_size - 1);
std::vector<PaddleTensor> cpu_feeds;
cpu_feeds.push_back(LodTensorToPaddleTensor(&first_word));
cpu_feeds.push_back(LodTensorToPaddleTensor(&second_word));
cpu_feeds.push_back(LodTensorToPaddleTensor(&third_word));
cpu_feeds.push_back(LodTensorToPaddleTensor(&fourth_word));
std::vector<PaddleTensor> outputs;
ASSERT_TRUE(predictor->Run(cpu_feeds, &outputs));
ASSERT_EQ(outputs.size(), 1);
for (size_t i = 0; i < outputs.size(); ++i) {
size_t len = outputs[i].data.length;
float* data = static_cast<float*>(outputs[i].data.data);
for (int j = 0; j < len / sizeof(float); ++j) {
ASSERT_LT(data[j], 1.0);
ASSERT_GT(data[j], -1.0);
}
free(outputs[i].data.data);
}
}
} // namespace paddle
paddle/fluid/framework/op_desc.cc
浏览文件 @ 632108dc
......@@ -243,13 +243,8 @@ const std::unordered_map<std::string, Attribute> &OpDesc::GetAttrMap() const {
}
void OpDesc::Rename(const std::string &old_name, const std::string &new_name) {
for (auto &input : inputs_) {
std::replace(input.second.begin(), input.second.end(), old_name, new_name);
}
for (auto &output : outputs_) {
std::replace(output.second.begin(), output.second.end(), old_name,
new_name);
}
RenameInput(old_name, new_name);
RenameOutput(old_name, new_name);
need_update_ = true;
}
......@@ -274,6 +269,13 @@ void OpDesc::RenameInput(const std::string &old_name,
for (auto &input : inputs_) {
std::replace(input.second.begin(), input.second.end(), old_name, new_name);
}
auto it = attrs_.find(framework::OpProtoAndCheckerMaker::OpRoleVarAttrName());
if (it != attrs_.end()) {
auto &op_vars = boost::get<std::vector<std::string>>(it->second);
std::replace(op_vars.begin(), op_vars.end(), old_name, new_name);
}
need_update_ = true;
}
......
paddle/fluid/framework/shape_inference.h
浏览文件 @ 632108dc
......@@ -63,6 +63,7 @@ class InferShapeContext {
std::vector<InferShapeVarPtr> GetInputVarPtrs(const std::string &name);
std::vector<InferShapeVarPtr> GetOutputVarPtrs(const std::string &name);
virtual InferShapeVarPtr GetVarPtr(const std::string &name) = 0;
// Note: In while op, we need this to be public
void SetDims(const std::vector<std::string> &names,
......@@ -81,8 +82,6 @@ class InferShapeContext {
const std::vector<std::string> &names) const;
virtual proto::VarType::Type GetVarType(const std::string &name) const = 0;
virtual InferShapeVarPtr GetVarPtr(const std::string &name) = 0;
};
} // namespace framework
......
paddle/fluid/inference/CMakeLists.txt
浏览文件 @ 632108dc
set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor init)
# TODO(panyx0718): Should this be called paddle_fluid_inference_api_internal?
cc_library(paddle_fluid_api
SRCS io.cc
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
......
paddle/fluid/inference/tensorrt/convert/CMakeLists.txt
浏览文件 @ 632108dc
nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES})
nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc io_converter.cc
DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine
SERIAL)
# This test is not stable
# See https://paddleci.ngrok.io/viewLog.html?tab=buildLog&buildTypeId=Paddle_PrCi2&buildId=36834&_focus=8828
#nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc io_converter.cc
# DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine
# SERIAL)
nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor)
paddle/fluid/operators/detail/sendrecvop_utils.cc
浏览文件 @ 632108dc
......@@ -149,12 +149,14 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
}
if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef PADDLE_WITH_CUDA
// GPU data is copied to CPU buffer when sending,
// free the buffer when possible.
destroy_callback = [](void* backing) {
platform::CUDAPinnedPlace cuda_pinned;
memory::Free(cuda_pinned, backing);
};
#endif
}
std::string header;
......
paddle/fluid/operators/detection/CMakeLists.txt
浏览文件 @ 632108dc
......@@ -24,6 +24,8 @@ detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc)
detection_library(prior_box_op SRCS prior_box_op.cc prior_box_op.cu)
detection_library(target_assign_op SRCS target_assign_op.cc
target_assign_op.cu)
detection_library(polygon_box_transform_op SRCS polygon_box_transform_op.cc
polygon_box_transform_op.cu)
# Export local libraries to parent
set(DETECTION_LIBRARY ${LOCAL_DETECTION_LIBS} PARENT_SCOPE)
paddle/fluid/operators/detection/polygon_box_transform_op.cc 0 → 100644
浏览文件 @ 632108dc
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename DeviceContext, typename T>
class PolygonBoxTransformCPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
"It must use CUDAPlace.");
auto* in = ctx.Input<Tensor>("Input");
auto in_dims = in->dims();
const T* in_data = in->data<T>();
auto* out = ctx.Output<Tensor>("Output");
T* out_data = out->mutable_data<T>(ctx.GetPlace());
int batch_size = in_dims[0];
int geo_channel = in_dims[1];
int height = in_dims[2];
int width = in_dims[3];
int id = 0;
for (int id_n = 0; id_n < batch_size * geo_channel; ++id_n) {
for (int id_h = 0; id_h < height; ++id_h) {
for (int id_w = 0; id_w < width; ++id_w) {
id = id_n * height * width + width * id_h + id_w;
if (id_n % 2 == 0) {
out_data[id] = id_w - in_data[id];
} else {
out_data[id] = id_h - in_data[id];
}
}
}
}
}
};
class PolygonBoxTransformOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(
ctx->HasInput("Input"),
"Input (Input) of polygon_box transform op should not be null.");
PADDLE_ENFORCE(
ctx->HasOutput("Output"),
"Output (Output) of polygon_box transform op should not be null.");
auto in_dim = ctx->GetInputDim("Input");
PADDLE_ENFORCE_EQ(in_dim.size(), 4, "input's rank must be 4.");
PADDLE_ENFORCE_EQ(in_dim[1] % 2, 0,
"input's second dimension must be even.");
ctx->SetOutputDim("Output", in_dim);
}
};
class PolygonBoxTransformOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput(
"Input",
"The input with shape [batch_size, geometry_channels, height, width]");
AddOutput("Output", "The output with the same shape as input");
AddComment(R"DOC(
PolygonBoxTransform Operator.
The input is the final geometry output in detection network.
We use 2*n numbers to denote the coordinate shift from n corner vertices of
the polygon_box to the pixel location. As each distance offset contains two numbers (xi, yi),
the geometry output contains 2*n channels.
PolygonBoxTransform Operator is used to transform the coordinate shift to the real coordinate.
)DOC");
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(polygon_box_transform, ops::PolygonBoxTransformOp,
ops::PolygonBoxTransformOpMaker,
paddle::framework::EmptyGradOpMaker);
REGISTER_OP_CPU_KERNEL(
polygon_box_transform,
ops::PolygonBoxTransformCPUKernel<paddle::platform::CPUPlace, float>,
ops::PolygonBoxTransformCPUKernel<paddle::platform::CPUPlace, double>);
paddle/fluid/operators/detection/polygon_box_transform_op.cu 0 → 100644
浏览文件 @ 632108dc
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/platform/cuda_primitives.h"
#include "paddle/fluid/platform/gpu_info.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
using platform::PADDLE_CUDA_NUM_THREADS;
#define CUDA_BLOCK_SIZE 16
template <typename T>
__global__ void PolygonBoxTransformKernel(const int n, const int h, const int w,
const T* input, T* output) {
int id_n = threadIdx.x + blockDim.x * blockIdx.x;
int id_h = threadIdx.y + blockDim.y * blockIdx.y;
int id_w = threadIdx.z + blockDim.z * blockIdx.z;
if (id_n < n && id_h < h && id_w < w) {
int id = id_n * h * w + w * id_h + id_w;
if (id_n % 2 == 0) {
output[id] = id_w - input[id];
} else {
output[id] = id_h - input[id];
}
}
}
template <typename T>
class PolygonBoxTransformOpCUDAKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
"It must use CUDAPlace.");
auto* in = ctx.Input<Tensor>("Input");
auto in_dims = in->dims();
const T* in_data = in->data<T>();
auto* out = ctx.Output<Tensor>("Output");
T* out_data = out->mutable_data<T>(ctx.GetPlace());
int batch_size = in_dims[0];
int geo_channels = in_dims[1];
int height = in_dims[2];
int width = in_dims[3];
dim3 threadsPerBlock(
PADDLE_CUDA_NUM_THREADS / (CUDA_BLOCK_SIZE * CUDA_BLOCK_SIZE),
CUDA_BLOCK_SIZE, CUDA_BLOCK_SIZE);
dim3 numBlocks((batch_size * geo_channels) / threadsPerBlock.x,
(height + threadsPerBlock.y - 1) / threadsPerBlock.y,
(width + threadsPerBlock.z - 1) / threadsPerBlock.z);
auto stream = ctx.cuda_device_context().stream();
PolygonBoxTransformKernel<T><<<numBlocks, threadsPerBlock, 0, stream>>>(
batch_size * geo_channels, height, width, in_data, out_data);
}
};
} // namespace operators
} // namespace paddle
REGISTER_OP_CUDA_KERNEL(
polygon_box_transform,
paddle::operators::PolygonBoxTransformOpCUDAKernel<float>,
paddle::operators::PolygonBoxTransformOpCUDAKernel<double>);
paddle/fluid/operators/fill_constant_batch_size_like_op.h
浏览文件 @ 632108dc
......@@ -24,6 +24,14 @@ class FillConstantBatchSizeLikeOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* out = ctx.Output<framework::Tensor>("Out");
auto* in = ctx.Input<framework::LoDTensor>("Input");
if (in->lod().size() && ctx.Attr<int>("input_dim_idx") == 0) {
// set the correct batch size for the LoDTensor.
auto odims = out->dims();
int output_dim_idx = ctx.Attr<int>("output_dim_idx");
odims[output_dim_idx] = static_cast<int>(in->lod().back().size()) - 1;
out->mutable_data<T>(odims, ctx.GetPlace());
}
out->mutable_data<T>(ctx.GetPlace());
auto value = ctx.Attr<float>("value");
......
paddle/fluid/operators/math/cross_entropy.cc
浏览文件 @ 632108dc
......@@ -46,7 +46,10 @@ class CrossEntropyFunctor<platform::CPUDeviceContext, T> {
const int64_t* label_data = labels->data<int64_t>();
for (int i = 0; i < batch_size; ++i) {
int index = i * class_num + label_data[i];
int lbl = label_data[i];
PADDLE_ENFORCE_GE(lbl, 0);
PADDLE_ENFORCE_LT(lbl, class_num);
int index = i * class_num + lbl;
loss_data[i] = -math::TolerableValue<T>()(std::log(prob_data[index]));
}
}
......
paddle/fluid/operators/reader/CMakeLists.txt
浏览文件 @ 632108dc
......@@ -23,6 +23,7 @@ reader_library(create_recordio_file_reader_op SRCS create_recordio_file_reader_o
reader_library(create_double_buffer_reader_op SRCS create_double_buffer_reader_op.cc)
reader_library(create_multi_pass_reader_op SRCS create_multi_pass_reader_op.cc)
reader_library(create_threaded_reader_op SRCS create_threaded_reader_op.cc)
reader_library(create_custom_reader_op SRCS create_custom_reader_op.cc)
cc_test(reader_blocking_queue_test SRCS reader_blocking_queue_test.cc)
# Export local libraries to parent
......
paddle/fluid/operators/reader/reader_op_registry.cc
浏览文件 @ 632108dc
......@@ -115,6 +115,7 @@ void DecoratedReaderInferShape::operator()(
boost::get<framework::VarDesc*>(ctx->GetOutputVarPtrs("Out")[0]);
out_reader->SetLoDLevels(in_reader->GetLoDLevels());
}
void DecoratedReaderInferVarType::operator()(
const framework::OpDesc& op_desc, framework::BlockDesc* block) const {
std::string in_reader_name = op_desc.Input("UnderlyingReader")[0];
......
paddle/function/EigenGemm.cpp
浏览文件 @ 632108dc
此差异已折叠。
paddle/function/EigenThreadDevice.h 0 → 100644
浏览文件 @ 632108dc
此差异已折叠。
paddle/optimizer/CMakeLists.txt
浏览文件 @ 632108dc
......@@ -7,6 +7,10 @@ set(OPITMIZER_SRCS
sgd_optimizer.cc
)
cc_library(paddle_optimizer STATIC SRCS ${OPITMIZER_SRCS} DEPS paddle_proto glog)
cc_test(serialization_test SRCS serialization_test.cc DEPS paddle_proto)
cc_test(parameter_optimizer_test SRCS parameter_optimizer_test.cc DEPS paddle_optimizer)
add_library(paddle_optimizer ${OPITMIZER_SRCS})
target_link_libraries(paddle_optimizer paddle_proto glog)
if (WITH_TESTING)
add_unittest(serialization_test serialization_test.cc)
add_unittest(parameter_optimizer_test parameter_optimizer_test.cc)
endif()
paddle/scripts/docker/build.sh 已删除 100755 → 0
浏览文件 @ 20c8ff0f
此差异已折叠。
paddle/scripts/docker/build_android.sh 已删除 100644 → 0
浏览文件 @ 20c8ff0f
此差异已折叠。
paddle/scripts/docker/entrypoint 已删除 100755 → 0
浏览文件 @ 20c8ff0f
#!/bin/bash
/usr/sbin/sshd -D &
jupyter notebook --ip=0.0.0.0 /paddle/book/
paddle/scripts/docker/test.sh 已删除 100755 → 0
浏览文件 @ 20c8ff0f
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ 632108dc
此差异已折叠。
paddle/scripts/travis/build_doc.sh 已删除 100755 → 0
浏览文件 @ 20c8ff0f
此差异已折叠。
paddle/scripts/travis/build_ios.sh 已删除 100755 → 0
浏览文件 @ 20c8ff0f
此差异已折叠。
paddle/scripts/travis/check_style.sh 已删除 100755 → 0
浏览文件 @ 20c8ff0f
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ 632108dc
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ 632108dc
此差异已折叠。
python/paddle/fluid/lod_tensor.py
浏览文件 @ 632108dc
此差异已折叠。
tools/codestyle/docstring_checker.py 0 → 100644
浏览文件 @ 632108dc
此差异已折叠。
tools/codestyle/pylint_pre_commit.hook 0 → 100755
浏览文件 @ 632108dc
此差异已折叠。
tools/codestyle/test_docstring_checker.py 0 → 100644
浏览文件 @ 632108dc
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或