Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix_docker_build
guochaorong
7 years ago
commit
ee65272933
@ -1,196 +0,0 @@
|
||||
# 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
|
||||
@ -1,35 +0,0 @@
|
||||
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/
|
||||
@ -1,195 +0,0 @@
|
||||
# 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.
|
||||
@ -1,72 +0,0 @@
|
||||
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
|
||||
@ -1,69 +0,0 @@
|
||||
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
|
||||
@ -1,21 +0,0 @@
|
||||
#!/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 &
|
||||
@ -1,82 +0,0 @@
|
||||
#!/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
|
||||
@ -1,56 +0,0 @@
|
||||
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
|
||||
@ -1,58 +0,0 @@
|
||||
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
|
||||
@ -1,64 +0,0 @@
|
||||
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
|
||||
@ -1,65 +0,0 @@
|
||||
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
|
||||
File diff suppressed because it is too large
Load Diff
File diff suppressed because it is too large
Load Diff
@ -1,154 +0,0 @@
|
||||
# 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()
|
||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue