未验证 提交 75303664 编写于 作者: W Wu Yi 提交者: GitHub

Merge pull request #10919 from typhoonzero/remove_fluid_old_cluster_benchmark_script

remove old fluid cluster benchmark scripts
# 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
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/
# 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.
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
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
#!/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 &
#!/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
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
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
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
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
# 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()
# 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)
# 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()
......@@ -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
......@@ -271,6 +300,9 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
num_samples = 0
if iters == args.iterations:
break
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()
......
......@@ -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({
......
......@@ -54,5 +54,13 @@ envs = [
"fieldPath": "status.podIP"
}
}
},
{
"name": "PADDLE_CURRENT_IP",
"valueFrom": {
"fieldRef": {
"fieldPath": "status.podIP"
}
}
}
]
Markdown is supported
0% .
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