Skip to content

P
PaddleDetection

PaddlePaddle / PaddleDetection
大约 1 年 前同步成功

通知 695
Star 11112
Fork 2696
P
PaddleDetection
提交 7116cc12 编写于 作者: S sneaxiy
浏览文件
操作

Merge develop

上级 abf9832c 3fbfcd9c
展开全部 隐藏空白更改
内联 并排
Showing with 61751 addition and 9860 deletion

要显示的变更太多。

To preserve performance only 1000 of 1000+ files are displayed.
文本差异 电子邮件补丁
.gitignore
浏览文件 @ 7116cc12
......@@ -5,6 +5,7 @@ python/paddle/v2/fluid/tests/book/image_classification_resnet.inference.model/
python/paddle/v2/fluid/tests/book/image_classification_vgg.inference.model/
python/paddle/v2/fluid/tests/book/label_semantic_roles.inference.model/
*.DS_Store
*.vs
build/
build_doc/
*.user
......@@ -15,6 +16,7 @@ build_doc/
.cproject
.pydevproject
.settings/
CMakeSettings.json
Makefile
.test_env/
third_party/
......
.travis.yml
浏览文件 @ 7116cc12
......@@ -27,15 +27,6 @@ script:
# 43min timeout
paddle/scripts/paddle_docker_build.sh ${JOB}
if [ $? -eq 0 ] || [ $? -eq 142 ]; then true; else exit 1; fi;
- |
if [[ "$JOB" != "doc" ]]; then exit 0; fi;
# For document only
if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
if [[ "$TRAVIS_BRANCH" != "develop" && ! "$TRAVIS_BRANCH" =~ ^v[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then exit 0; fi;
export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/master/scripts/deploy/deploy_docs.sh
export DOCS_DIR=`pwd`
cd ..
curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH $DOCS_DIR $DOCS_DIR/build/doc/
notifications:
email:
on_success: change
......
AUTHORS.md
浏览文件 @ 7116cc12
......@@ -46,6 +46,7 @@
| tianbingsz | Tian-Bing Xu |
| tpatejko | Tomasz Patejko |
| typhoonzero | Yi Wu |
| velconia | Qi-Yang Min |
| wanghaoshuang | Hao-Shuang Wang |
| wangyang59 | Yang Wang |
| wangzhen-nlp | Zhen Wang |
......
CMakeLists.txt
浏览文件 @ 7116cc12
......@@ -24,6 +24,9 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
"${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION}")
message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
"${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
if(WIN32)
set(CMAKE_STATIC_LIBRARY_PREFIX lib)
endif(WIN32)
if(NOT CMAKE_CROSSCOMPILING)
find_package(CUDA QUIET)
......@@ -65,7 +68,15 @@ option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better d
option(WITH_ANAKIN "Compile with Anakin library" OFF)
option(WITH_GRPC "Use grpc as the default rpc framework" ${WITH_DISTRIBUTE})
option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF)
option(WITH_INFERENCE "Compile fluid inference library" ON)
option(WITH_SYSTEM_BLAS "Use system blas library" OFF)
option(PY_VERSION "Compile PaddlePaddle with python3 support" ${PY_VERSION})
# PY_VERSION
if(NOT PY_VERSION)
set(PY_VERSION 2.7)
endif()
set(PYBIND11_PYTHON_VERSION ${PY_VERSION})
# CMAKE_BUILD_TYPE
if(NOT CMAKE_BUILD_TYPE)
......@@ -103,6 +114,11 @@ if(ANDROID OR IOS)
add_definitions(-DPADDLE_MOBILE_INFERENCE)
endif()
if (APPLE OR WIN32)
set(WITH_MKL OFF CACHE STRING
"Disable MKL for building on mac and windows" FORCE)
endif()
set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
"A path setting third party libraries download & build directories.")
......@@ -141,6 +157,8 @@ endif()
########################################################################################
include(external/mklml) # download mklml package
include(external/xbyak) # download xbyak package
include(external/libxsmm) # download, build, install libxsmm
include(external/zlib) # download, build, install zlib
include(external/gflags) # download, build, install gflags
include(external/glog) # download, build, install glog
......@@ -150,12 +168,20 @@ include(external/python) # download, build, install python
include(external/openblas) # download, build, install openblas
include(external/mkldnn) # download, build, install mkldnn
include(external/swig) # download, build, install swig
include(external/warpctc) # download, build, install warpctc
include(external/boost) # download boost
include(external/any) # download libn::any
include(external/eigen) # download eigen3
include(external/pybind11) # download pybind11
include(external/cares)
include(external/cub)
if (NOT WIN32)
# there is no official support of snappystream, warpctc, nccl, cupti in windows
include(external/snappy) # download snappy
include(external/snappystream) # download snappystream
include(external/warpctc) # download, build, install warpctc
include(cupti)
endif (NOT WIN32)
if(WITH_DISTRIBUTE)
if(WITH_GRPC)
......@@ -178,19 +204,27 @@ if(WITH_BRPC_RDMA)
endif()
endif()
include(external/snappy) # download snappy
include(external/snappystream)
include(external/threadpool)
include(external/threadpool)
include(flags) # set paddle compile flags
include(cudnn) # set cudnn libraries, must before configure
include(cupti)
include(configure) # add paddle env configuration
if(WITH_GPU)
include(cuda)
include(tensorrt)
endif()
if(WITH_MKL OR WITH_MKLML)
include(external/anakin)
elseif()
set(WITH_ANAKIN OFF CACHE STRING "Anakin is used in MKL only now." FORCE)
endif()
include(generic) # simplify cmake module
include(package) # set paddle packages
include(ccache) # set ccache for compilation
include(util) # set unittest and link libs
include(rdma) # set rdma libraries
include(flags) # set paddle compile flags
include(version) # set PADDLE_VERSION
include(coveralls) # set code coverage
include(inference_lib) # add paddle fluid inference libraries
......@@ -210,14 +244,6 @@ set(EXTERNAL_LIBS
${PYTHON_LIBRARIES}
)
if(WITH_GPU)
include(cuda)
include(tensorrt)
include(external/anakin)
else()
set(WITH_ANAKIN OFF CACHE STRING "Anakin is valid only when GPU is set." FORCE)
endif()
if(WITH_AMD_GPU)
find_package(HIP)
include(hip)
......@@ -227,6 +253,10 @@ if(WITH_MKLML)
list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB})
endif()
if(WITH_LIBXSMM)
list(APPEND EXTERNAL_LIBS ${LIBXSMM_LIBS})
endif()
if(WITH_MKLDNN)
list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB})
endif()
......@@ -266,7 +296,3 @@ 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 @ 7116cc12
......@@ -53,7 +53,7 @@ RUN curl -s -q https://glide.sh/get | sh
# and its size is only one-third of the official one.
# 2. Manually add ~IPluginFactory() in IPluginFactory class of NvInfer.h, otherwise, it couldn't work in paddle.
# See https://github.com/PaddlePaddle/Paddle/issues/10129 for details.
RUN wget -qO- http://paddlepaddledeps.bj.bcebos.com/TensorRT-4.0.0.3.Ubuntu-16.04.4.x86_64-gnu.cuda-8.0.cudnn7.0.tar.gz | \
RUN wget -qO- http://paddlepaddledeps.cdn.bcebos.com/TensorRT-4.0.0.3.Ubuntu-16.04.4.x86_64-gnu.cuda-8.0.cudnn7.0.tar.gz | \
tar -xz -C /usr/local && \
cp -rf /usr/local/TensorRT/include /usr && \
cp -rf /usr/local/TensorRT/lib /usr
......@@ -80,7 +80,7 @@ RUN pip install pre-commit 'ipython==5.3.0' && \
pip install opencv-python
#For docstring checker
RUN pip install pylint pytest astroid isort
RUN pip install pylint pytest astroid isort LinkChecker
COPY ./python/requirements.txt /root/
RUN pip install -r /root/requirements.txt
......
README.md
浏览文件 @ 7116cc12
......@@ -18,7 +18,21 @@ learning to many products at Baidu.
Our vision is to enable deep learning for everyone via PaddlePaddle.
Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
### Lastest PaddlePaddle Version: [Fluid](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid)
### Latest PaddlePaddle Release: [Fluid 0.15.0](https://github.com/PaddlePaddle/Paddle/tree/v0.15.0)
### Install Latest Stable Release:
```
# Linux CPU
pip install paddlepaddle
# Linux GPU cuda9cudnn7
pip install paddlepaddle-gpu
# Linux GPU cuda8cudnn7
pip install paddlepaddle-gpu==0.14.0.post87
# Linux GPU cuda8cudnn5
pip install paddlepaddle-gpu==0.14.0.post85
# For installation on other platform, refer to http://paddlepaddle.org/
```
## Features
......@@ -62,33 +76,26 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl
## Installation
It is recommended to check out the
[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/docker_install_en.html)
before looking into the
[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/build_from_source_en.html).
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/install/install_doc.html) on our website.
## Documentation
We provide [English](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html) and
[Chinese](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html) documentation.
We provide [English](http://paddlepaddle.org/documentation/docs/en/0.15.0/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/index.html) documentation.
- [Deep Learning 101](http://www.paddlepaddle.org/docs/develop/book/01.fit_a_line/index.html)
- [Deep Learning 101](https://github.com/PaddlePaddle/book)
You might want to start from this online interactive book that can run in a Jupyter Notebook.
- [Distributed Training](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/cluster/index_en.html)
- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/user_guides/howto/training/cluster_howto.html)
You can run distributed training jobs on MPI clusters.
- [Distributed Training on Kubernetes](http://www.paddlepaddle.org/docs/develop/documentation/en/howto/cluster/multi_cluster/k8s_en.html)
You can also run distributed training jobs on Kubernetes clusters.
- [Python API](http://www.paddlepaddle.org/docs/develop/api/en/overview.html)
- [Python API](http://paddlepaddle.org/documentation/api/zh/0.15.0/fluid.html)
Our new API enables much shorter programs.
- [How to Contribute](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/dev/contribute_to_paddle_en.html)
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/advanced_usage/development/contribute_to_paddle.html)
We appreciate your contributions!
......
benchmark/fluid/Dockerfile
浏览文件 @ 7116cc12
......@@ -11,6 +11,7 @@ RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/lib/libcudnn.so && ln -s
# Add "ENV http_proxy=http://ip:port" if your download is slow, and don't forget to unset it at runtime.
# exmaple: unset http_proxy && unset https_proxy && python fluid_benchmark.py ...
RUN pip install -U pip
RUN pip install -U kubernetes paddlepaddle
......@@ -27,5 +28,6 @@ ADD *.whl /
RUN pip install /*.whl && rm -f /*.whl
ENV LD_LIBRARY_PATH=/usr/local/lib
ADD fluid_benchmark.py recordio_converter.py args.py recordio_converter.py run.sh run_fluid_benchmark.sh /workspace/
ADD fluid_benchmark.py recordio_converter.py args.py recordio_converter.py run.sh run_fluid_benchmark.sh imagenet_reader.py /workspace/
ADD models/ /workspace/models/
benchmark/fluid/args.py
浏览文件 @ 7116cc12
......@@ -17,7 +17,8 @@ import argparse
__all__ = ['parse_args', ]
BENCHMARK_MODELS = [
"machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm"
"machine_translation", "resnet", "se_resnext", "vgg", "mnist",
"stacked_dynamic_lstm", "resnet_with_preprocess"
]
......@@ -67,12 +68,12 @@ def parse_args():
'--cpus',
type=int,
default=1,
help='If cpus > 1, will use ParallelDo to run, else use Executor.')
help='If cpus > 1, will set ParallelExecutor to use multiple threads.')
parser.add_argument(
'--data_set',
type=str,
default='flowers',
choices=['cifar10', 'flowers'],
choices=['cifar10', 'flowers', 'imagenet'],
help='Optional dataset for benchmark.')
parser.add_argument(
'--infer_only', action='store_true', help='If set, run forward only.')
......@@ -122,6 +123,11 @@ def parse_args():
type=str,
default="",
help='Directory that contains all the training recordio files.')
parser.add_argument(
'--test_data_path',
type=str,
default="",
help='Directory that contains all the test data (NOT recordio).')
parser.add_argument(
'--use_inference_transpiler',
action='store_true',
......@@ -130,5 +136,15 @@ def parse_args():
'--no_random',
action='store_true',
help='If set, keep the random seed and do not shuffle the data.')
parser.add_argument(
'--use_lars',
action='store_true',
help='If set, use lars for optimizers, ONLY support resnet module.')
parser.add_argument(
'--reduce_strategy',
type=str,
choices=['reduce', 'all_reduce'],
default='all_reduce',
help='Specify the reduce strategy, can be reduce, all_reduce')
args = parser.parse_args()
return args
benchmark/fluid/fluid_benchmark.py
浏览文件 @ 7116cc12
......@@ -16,6 +16,7 @@ import argparse
import cProfile
import time
import os
import traceback
import numpy as np
......@@ -27,7 +28,7 @@ import paddle.fluid.transpiler.distribute_transpiler as distribute_transpiler
from args import *
def append_nccl2_prepare(trainer_id):
def append_nccl2_prepare(trainer_id, startup_prog):
if trainer_id >= 0:
# append gen_nccl_id at the end of startup program
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
......@@ -40,11 +41,11 @@ def append_nccl2_prepare(trainer_id):
current_endpoint = os.getenv("PADDLE_CURRENT_IP") + ":" + port
worker_endpoints.remove(current_endpoint)
nccl_id_var = fluid.default_startup_program().global_block().create_var(
nccl_id_var = startup_prog.global_block().create_var(
name="NCCLID",
persistable=True,
type=fluid.core.VarDesc.VarType.RAW)
fluid.default_startup_program().global_block().append_op(
startup_prog.global_block().append_op(
type="gen_nccl_id",
inputs={},
outputs={"NCCLID": nccl_id_var},
......@@ -59,7 +60,7 @@ def append_nccl2_prepare(trainer_id):
"nccl-based dist train.")
def dist_transpile(trainer_id, args):
def dist_transpile(trainer_id, args, train_prog, startup_prog):
if trainer_id < 0:
return None, None
......@@ -80,137 +81,74 @@ def dist_transpile(trainer_id, args):
# the role, should be either PSERVER or TRAINER
training_role = os.getenv("PADDLE_TRAINING_ROLE")
t = distribute_transpiler.DistributeTranspiler()
config = distribute_transpiler.DistributeTranspilerConfig()
config.slice_var_up = not args.no_split_var
t = distribute_transpiler.DistributeTranspiler(config=config)
t.transpile(
trainer_id,
# NOTE: *MUST* use train_prog, for we are using with guard to
# generate different program for train and test.
program=train_prog,
pservers=pserver_endpoints,
trainers=trainers,
sync_mode=not args.async_mode,
slice_var_up=not args.no_split_var)
startup_program=startup_prog)
if training_role == "PSERVER":
pserver_program = t.get_pserver_program(current_endpoint)
pserver_startup_program = t.get_startup_program(current_endpoint,
pserver_program)
pserver_startup_program = t.get_startup_program(
current_endpoint, pserver_program, startup_program=startup_prog)
return pserver_program, pserver_startup_program
elif training_role == "TRAINER":
train_program = t.get_trainer_program()
return train_program, fluid.default_startup_program()
return train_program, startup_prog
else:
raise ValueError(
'PADDLE_TRAINING_ROLE environment variable must be either TRAINER or PSERVER'
)
def test(exe, inference_program, test_reader, feeder, batch_acc):
accuracy_evaluator = fluid.metrics.Accuracy()
for batch_id, data in enumerate(test_reader()):
acc = exe.run(inference_program,
feed=feeder.feed(data),
fetch_list=[batch_acc])
accuracy_evaluator.update(value=np.array(acc), weight=len(data))
def test_parallel(exe, test_args, args, test_prog, feeder):
acc_evaluators = []
for i in xrange(len(test_args[2])):
acc_evaluators.append(fluid.metrics.Accuracy())
return accuracy_evaluator.eval()
# TODO(wuyi): replace train, train_parallel, test functions with new trainer
# API once it is ready.
def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
args, train_prog, startup_prog):
if os.getenv("PADDLE_TRAINING_ROLE") == "PSERVER":
place = core.CPUPlace()
exe = fluid.Executor(place)
exe.run(startup_prog)
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)
# Use inference_transpiler to speedup
if not args.use_reader_op:
feed_var_list = [
var for var in train_prog.global_block().vars.itervalues()
if var.is_data
]
feeder = fluid.DataFeeder(feed_var_list, place)
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in range(args.pass_num):
train_losses = []
if not args.use_reader_op:
reader_generator = train_reader()
batch_id = 0
data = None
to_fetch = [v.name for v in test_args[2]]
if args.use_reader_op:
test_args[4].start()
while True:
if not args.use_reader_op:
data = next(reader_generator, None)
if data == None:
break
if iters == args.iterations:
try:
acc_rets = exe.run(fetch_list=to_fetch)
for i, e in enumerate(acc_evaluators):
e.update(
value=np.array(acc_rets[i]), weight=args.batch_size)
except fluid.core.EOFException as eof:
test_args[4].reset()
break
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
else:
for batch_id, data in enumerate(test_args[3]()):
acc_rets = exe.run(feed=feeder.feed(data), fetch_list=to_fetch)
for i, e in enumerate(acc_evaluators):
e.update(value=np.array(acc_rets[i]), weight=len(data))
if args.use_reader_op:
try:
loss = exe.run(train_prog, fetch_list=[avg_loss])
except fluid.core.EnforceNotMet as ex:
break
else:
loss = exe.run(train_prog,
feed=feeder.feed(data),
fetch_list=[avg_loss])
iters += 1
batch_id += 1
# FIXME(wuyi): For use_reader_op, if the current
# pass is not the last, the last batch of this pass
# is also equal to args.batch_size.
if args.use_reader_op:
num_samples += args.batch_size * args.gpus
else:
num_samples += len(data)
train_losses.append(loss)
print("Pass: %d, Iter: %d, Loss: %f\n" %
(pass_id, iters, np.mean(train_losses)))
print_train_time(start_time, time.time(), num_samples)
print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses))),
# evaluation
if not args.no_test and batch_acc and not args.use_reader_op:
if args.use_inference_transpiler:
t = fluid.InferenceTranspiler()
t.transpile(infer_prog, place)
pass_test_acc = test(exe, infer_prog, test_reader, feeder,
batch_acc)
print(", Test Accuracy: %f" % pass_test_acc)
print("\n")
# TODO(wuyi): add warmup passes to get better perf data.
exit(0)
return [e.eval() for e in acc_evaluators]
# TODO(wuyi): replace train, train_parallel, test functions with new trainer
# API once it is ready.
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):
# NOTE: only need to benchmark using parallelexe
def train_parallel(train_args, test_args, args, train_prog, test_prog,
startup_prog, nccl_id_var, num_trainers, trainer_id):
over_all_start = time.time()
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
feeder = None
if not args.use_reader_op:
feed_var_list = [
var for var in train_prog.global_block().vars.itervalues()
if var.is_data
]
feeder = fluid.DataFeeder(feed_var_list, place)
# generate fake:
if args.use_fake_data:
for var in feed_var_list:
v = startup_prog.global_block().clone_variable(var)
v = startup_prog.global_block()._clone_variable(var)
var.persistable = True
v.persistable = True
......@@ -230,62 +168,119 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
startup_exe = fluid.Executor(place)
startup_exe.run(startup_prog)
strategy = fluid.ExecutionStrategy()
strategy.num_threads = 1
strategy.num_threads = args.cpus
strategy.allow_op_delay = False
build_strategy = fluid.BuildStrategy()
if args.reduce_strategy == "reduce":
build_strategy.reduce_strategy = fluid.BuildStrategy(
).ReduceStrategy.Reduce
else:
build_strategy.reduce_strategy = fluid.BuildStrategy(
).ReduceStrategy.AllReduce
avg_loss = train_args[0]
if args.update_method == "pserver":
# parameter server mode distributed training, merge
# gradients on local server, do not initialize
# ParallelExecutor with multi server all-reduce mode.
num_trainers = 1
trainer_id = 0
exe = fluid.ParallelExecutor(
True,
avg_loss.name,
main_program=train_prog,
exec_strategy=strategy,
build_strategy=build_strategy,
num_trainers=num_trainers,
trainer_id=trainer_id)
if not args.no_test:
if args.update_method == "pserver":
test_scope = None
else:
# NOTE: use an empty scope to avoid test exe using NCCLID
test_scope = fluid.Scope()
test_exe = fluid.ParallelExecutor(
True, main_program=test_prog, share_vars_from=exe)
for pass_id in range(args.pass_num):
num_samples = 0
iters = 0
start_time = time.time()
if not args.use_reader_op:
reader_generator = train_reader()
reader_generator = train_args[3]() #train_reader
batch_id = 0
data = None
if args.use_reader_op:
train_args[4].start()
while True:
if not args.use_reader_op:
data = next(reader_generator, None)
if data == None:
break
if args.profile and batch_id == 5:
profiler.start_profiler("All")
profiler.reset_profiler()
elif args.profile and batch_id == 10:
print("profiling total time: ", time.time() - start_time)
profiler.stop_profiler("total", "/tmp/profile_%d_pass%d" %
(trainer_id, pass_id))
if iters == args.iterations:
reader_generator.close()
break
if args.profile and pass_id == 0 and batch_id == 5:
profiler.start_profiler("All")
elif args.profile and pass_id == 0 and batch_id == 10:
profiler.stop_profiler("total", "/tmp/profile_%d" % trainer_id)
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
fetch_list = [avg_loss.name]
acc_name_list = [v.name for v in train_args[2]]
fetch_list.extend(acc_name_list)
if args.use_fake_data or args.use_reader_op:
try:
loss, = exe.run([avg_loss.name])
fetch_ret = exe.run(fetch_list)
except fluid.core.EOFException as eof:
break
except fluid.core.EnforceNotMet as ex:
traceback.print_exc()
break
else:
loss, = exe.run([avg_loss.name], feed=feeder.feed(data))
fetch_ret = exe.run(fetch_list, feed=feeder.feed(data))
if args.use_reader_op:
num_samples += args.batch_size * args.gpus
else:
num_samples += len(data)
iters += 1
if batch_id % 1 == 0:
print("Pass %d, batch %d, loss %s" %
(pass_id, batch_id, np.array(loss)))
fetched_data = [np.mean(np.array(d)) for d in fetch_ret]
print("Pass %d, batch %d, loss %s, accucacys: %s" %
(pass_id, batch_id, fetched_data[0], fetched_data[1:]))
batch_id += 1
print_train_time(start_time, time.time(), num_samples)
if not args.no_test and batch_acc and not args.use_reader_op:
# we have not implement record io for test
# skip test when use args.use_reader_op
test_acc = test(startup_exe, infer_prog, test_reader, feeder,
batch_acc)
print("Pass: %d, Test Accuracy: %f\n" % (pass_id, test_acc))
if args.use_reader_op:
train_args[4].reset() # reset reader handle
else:
del reader_generator
if not args.no_test and test_args[2]:
test_feeder = None
if not args.use_reader_op:
test_feed_var_list = [
var for var in test_prog.global_block().vars.itervalues()
if var.is_data
]
test_feeder = fluid.DataFeeder(test_feed_var_list, place)
test_ret = test_parallel(test_exe, test_args, args, test_prog,
test_feeder)
print("Pass: %d, Test Accuracy: %s\n" %
(pass_id, [np.mean(np.array(v)) for v in test_ret]))
print("total train time: ", time.time() - over_all_start)
def print_arguments(args):
......@@ -327,44 +322,46 @@ def main():
if args.use_cprof:
pr = cProfile.Profile()
pr.enable()
model_def = __import__("models.%s" % args.model, fromlist=["models"])
train_args = list(model_def.get_model(args))
train_args.append(args)
# Run optimizer.minimize(avg_loss)
train_args[2].minimize(train_args[0])
if args.memory_optimize:
fluid.memory_optimize(fluid.default_main_program())
train_prog = fluid.Program()
test_prog = fluid.Program()
startup_prog = fluid.Program()
train_args = list(model_def.get_model(args, True, train_prog, startup_prog))
test_args = list(model_def.get_model(args, False, test_prog, startup_prog))
all_args = [train_args, test_args, args]
if args.update_method == "pserver":
train_prog, startup_prog = dist_transpile(trainer_id, args)
train_prog, startup_prog = dist_transpile(trainer_id, args, train_prog,
startup_prog)
if not train_prog:
raise Exception(
"Must configure correct environments to run dist train.")
train_args.extend([train_prog, startup_prog])
all_args.extend([train_prog, test_prog, startup_prog])
if args.gpus > 1 and os.getenv("PADDLE_TRAINING_ROLE") == "TRAINER":
train_args.extend([nccl_id_var, num_trainers, trainer_id])
train_parallel(*train_args)
train(*train_args)
all_args.extend([nccl_id_var, num_trainers, trainer_id])
train_parallel(*all_args)
elif os.getenv("PADDLE_TRAINING_ROLE") == "PSERVER":
# start pserver with Executor
server_exe = fluid.Executor(fluid.CPUPlace())
server_exe.run(startup_prog)
server_exe.run(train_prog)
exit(0)
# for other update methods, use default programs
train_args.append(fluid.default_main_program())
train_args.append(fluid.default_startup_program())
all_args.extend([train_prog, test_prog, startup_prog])
if args.update_method == "nccl2":
nccl_id_var, num_trainers, trainer_id = append_nccl2_prepare(trainer_id)
if args.gpus == 1:
# NOTE: parallel executor use profiler interanlly
if args.use_nvprof and args.device == 'GPU':
with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof:
train(*train_args)
else:
train(*train_args)
else:
if args.device == "CPU":
raise Exception("Only support GPU perf with parallel exe")
train_args.extend([nccl_id_var, num_trainers, trainer_id])
train_parallel(*train_args)
nccl_id_var, num_trainers, trainer_id = append_nccl2_prepare(
trainer_id, startup_prog)
if args.device == "CPU":
raise Exception("Only support GPU perf with parallel exe")
all_args.extend([nccl_id_var, num_trainers, trainer_id])
train_parallel(*all_args)
if __name__ == "__main__":
......
benchmark/fluid/imagenet_reader.py 0 → 100644
浏览文件 @ 7116cc12
# 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 os
import math
import random
import functools
import numpy as np
from threading import Thread
import subprocess
import time
from Queue import Queue
import paddle
from PIL import Image, ImageEnhance
random.seed(0)
DATA_DIM = 224
THREAD = int(os.getenv("PREPROCESS_THREADS", "10"))
BUF_SIZE = 5120
DATA_DIR = '/mnt/ImageNet'
TRAIN_LIST = '/mnt/ImageNet/train.txt'
TEST_LIST = '/mnt/ImageNet/val.txt'
img_mean = np.array([0.485, 0.456, 0.406]).reshape((3, 1, 1))
img_std = np.array([0.229, 0.224, 0.225]).reshape((3, 1, 1))
def resize_short(img, target_size):
percent = float(target_size) / min(img.size[0], img.size[1])
resized_width = int(round(img.size[0] * percent))
resized_height = int(round(img.size[1] * percent))
img = img.resize((resized_width, resized_height), Image.LANCZOS)
return img
def crop_image(img, target_size, center):
width, height = img.size
size = target_size
if center == True:
w_start = (width - size) / 2
h_start = (height - size) / 2
else:
w_start = random.randint(0, width - size)
h_start = random.randint(0, height - size)
w_end = w_start + size
h_end = h_start + size
img = img.crop((w_start, h_start, w_end, h_end))
return img
def random_crop(img, size, scale=[0.08, 1.0], ratio=[3. / 4., 4. / 3.]):
aspect_ratio = math.sqrt(random.uniform(*ratio))
w = 1. * aspect_ratio
h = 1. / aspect_ratio
bound = min((float(img.size[0]) / img.size[1]) / (w**2),
(float(img.size[1]) / img.size[0]) / (h**2))
scale_max = min(scale[1], bound)
scale_min = min(scale[0], bound)
target_area = img.size[0] * img.size[1] * random.uniform(scale_min,
scale_max)
target_size = math.sqrt(target_area)
w = int(target_size * w)
h = int(target_size * h)
i = random.randint(0, img.size[0] - w)
j = random.randint(0, img.size[1] - h)
img = img.crop((i, j, i + w, j + h))
img = img.resize((size, size), Image.LANCZOS)
return img
def rotate_image(img):
angle = random.randint(-10, 10)
img = img.rotate(angle)
return img
def distort_color(img):
def random_brightness(img, lower=0.5, upper=1.5):
e = random.uniform(lower, upper)
return ImageEnhance.Brightness(img).enhance(e)
def random_contrast(img, lower=0.5, upper=1.5):
e = random.uniform(lower, upper)
return ImageEnhance.Contrast(img).enhance(e)
def random_color(img, lower=0.5, upper=1.5):
e = random.uniform(lower, upper)
return ImageEnhance.Color(img).enhance(e)
ops = [random_brightness, random_contrast, random_color]
random.shuffle(ops)
img = ops[0](img)
img = ops[1](img)
img = ops[2](img)
return img
def process_image(sample, mode, color_jitter, rotate):
img_path = sample[0]
img = Image.open(img_path)
if mode == 'train':
if rotate: img = rotate_image(img)
img = random_crop(img, DATA_DIM)
else:
img = resize_short(img, target_size=256)
img = crop_image(img, target_size=DATA_DIM, center=True)
if mode == 'train':
if color_jitter:
img = distort_color(img)
if random.randint(0, 1) == 1:
img = img.transpose(Image.FLIP_LEFT_RIGHT)
if img.mode != 'RGB':
img = img.convert('RGB')
img = np.array(img).astype('float32').transpose((2, 0, 1)) / 255
img -= img_mean
img /= img_std
if mode == 'train' or mode == 'val':
return img, sample[1]
elif mode == 'test':
return [img]
class XmapEndSignal():
pass
def xmap_readers(mapper,
reader,
process_num,
buffer_size,
order=False,
print_queue_state=True):
end = XmapEndSignal()
# define a worker to read samples from reader to in_queue
def read_worker(reader, in_queue):
for i in reader():
in_queue.put(i)
in_queue.put(end)
# define a worker to read samples from reader to in_queue with order flag
def order_read_worker(reader, in_queue, file_queue):
in_order = 0
for i in reader():
in_queue.put((in_order, i))
in_order += 1
in_queue.put(end)
# define a worker to handle samples from in_queue by mapper
# and put mapped samples into out_queue
def handle_worker(in_queue, out_queue, mapper):
sample = in_queue.get()
while not isinstance(sample, XmapEndSignal):
r = mapper(sample)
out_queue.put(r)
sample = in_queue.get()
in_queue.put(end)
out_queue.put(end)
# define a worker to handle samples from in_queue by mapper
# and put mapped samples into out_queue by order
def order_handle_worker(in_queue, out_queue, mapper, out_order):
ins = in_queue.get()
while not isinstance(ins, XmapEndSignal):
order, sample = ins
r = mapper(sample)
while order != out_order[0]:
pass
out_queue.put(r)
out_order[0] += 1
ins = in_queue.get()
in_queue.put(end)
out_queue.put(end)
def xreader():
file_queue = Queue()
in_queue = Queue(buffer_size)
out_queue = Queue(buffer_size)
out_order = [0]
# start a read worker in a thread
target = order_read_worker if order else read_worker
t = Thread(target=target, args=(reader, in_queue))
t.daemon = True
t.start()
# start several handle_workers
target = order_handle_worker if order else handle_worker
args = (in_queue, out_queue, mapper, out_order) if order else (
in_queue, out_queue, mapper)
workers = []
for i in xrange(process_num):
worker = Thread(target=target, args=args)
worker.daemon = True
workers.append(worker)
for w in workers:
w.start()
sample = out_queue.get()
start_t = time.time()
while not isinstance(sample, XmapEndSignal):
yield sample
sample = out_queue.get()
if time.time() - start_t > 3:
if print_queue_state:
print("queue sizes: ", in_queue.qsize(), out_queue.qsize())
start_t = time.time()
finish = 1
while finish < process_num:
sample = out_queue.get()
if isinstance(sample, XmapEndSignal):
finish += 1
else:
yield sample
return xreader
def _reader_creator(file_list,
mode,
shuffle=False,
color_jitter=False,
rotate=False,
xmap=True):
def reader():
with open(file_list) as flist:
full_lines = [line.strip() for line in flist]
if shuffle:
random.shuffle(full_lines)
if mode == 'train':
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
trainer_count = int(os.getenv("PADDLE_TRAINERS"))
per_node_lines = len(full_lines) / trainer_count
lines = full_lines[trainer_id * per_node_lines:(trainer_id + 1)
* per_node_lines]
print(
"read images from %d, length: %d, lines length: %d, total: %d"
% (trainer_id * per_node_lines, per_node_lines, len(lines),
len(full_lines)))
else:
lines = full_lines
for line in lines:
if mode == 'train':
img_path, label = line.split()
img_path = img_path.replace("JPEG", "jpeg")
img_path = os.path.join(DATA_DIR, "train", img_path)
yield (img_path, int(label))
elif mode == 'val':
img_path, label = line.split()
img_path = img_path.replace("JPEG", "jpeg")
img_path = os.path.join(DATA_DIR, "val", img_path)
yield (img_path, int(label))
elif mode == 'test':
img_path = os.path.join(DATA_DIR, line)
yield [img_path]
mapper = functools.partial(
process_image, mode=mode, color_jitter=color_jitter, rotate=rotate)
return paddle.reader.xmap_readers(mapper, reader, THREAD, BUF_SIZE)
def load_raw_image_uint8(sample):
img_arr = np.array(Image.open(sample[0])).astype('int64')
return img_arr, int(sample[1])
def train_raw(file_list=TRAIN_LIST, shuffle=True):
def reader():
with open(file_list) as flist:
full_lines = [line.strip() for line in flist]
if shuffle:
random.shuffle(full_lines)
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
trainer_count = int(os.getenv("PADDLE_TRAINERS"))
per_node_lines = len(full_lines) / trainer_count
lines = full_lines[trainer_id * per_node_lines:(trainer_id + 1) *
per_node_lines]
print("read images from %d, length: %d, lines length: %d, total: %d"
% (trainer_id * per_node_lines, per_node_lines, len(lines),
len(full_lines)))
for line in lines:
img_path, label = line.split()
img_path = img_path.replace("JPEG", "jpeg")
img_path = os.path.join(DATA_DIR, "train", img_path)
yield (img_path, int(label))
return paddle.reader.xmap_readers(load_raw_image_uint8, reader, THREAD,
BUF_SIZE)
def train(file_list=TRAIN_LIST, xmap=True):
return _reader_creator(
file_list,
'train',
shuffle=True,
color_jitter=False,
rotate=False,
xmap=xmap)
def val(file_list=TEST_LIST, xmap=True):
return _reader_creator(file_list, 'val', shuffle=False, xmap=xmap)
def test(file_list=TEST_LIST):
return _reader_creator(file_list, 'test', shuffle=False)
if __name__ == "__main__":
c = 0
start_t = time.time()
for d in train()():
c += 1
if c >= 10000:
break
spent = time.time() - start_t
print("read 10000 speed: ", 10000 / spent, spent)
benchmark/fluid/kube_gen_job.py
浏览文件 @ 7116cc12
......@@ -163,6 +163,19 @@ def gen_job():
volumes.append({"name": "dshm", "emptyDir": {"medium": "Memory"}})
volumeMounts.append({"mountPath": "/dev/shm", "name": "dshm"})
# add ceph volumes
volumes.append({
"name": "ceph-data",
"cephfs": {
"monitors": ["192.168.16.23:6789"],
"secretRef": {
"name": "ceph-secret"
},
"user": "admin",
}
})
volumeMounts.append({"mountPath": "/mnt/data", "name": "ceph-data"})
tn["spec"]["template"]["spec"]["volumes"] = volumes
tn_container["volumeMounts"] = volumeMounts
......
benchmark/fluid/models/__init__.py
浏览文件 @ 7116cc12
......@@ -13,5 +13,6 @@
# limitations under the License.
__all__ = [
"machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm"
"machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm",
"resnet_with_preprocess"
]
benchmark/fluid/models/machine_translation.py
浏览文件 @ 7116cc12
......@@ -12,6 +12,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""seq2seq model for fluid."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
......@@ -181,7 +182,7 @@ def lodtensor_to_ndarray(lod_tensor):
return ndarray
def get_model(args):
def get_model(args, is_train, main_prog, startup_prog):
if args.use_reader_op:
raise Exception("machine_translation do not support reader op for now.")
embedding_dim = 512
......@@ -190,30 +191,27 @@ def get_model(args):
dict_size = 30000
beam_size = 3
max_length = 250
avg_cost, feeding_list = seq_to_seq_net(
embedding_dim,
encoder_size,
decoder_size,
dict_size,
dict_size,
False,
beam_size=beam_size,
max_length=max_length)
# clone from default main program
inference_program = fluid.default_main_program().clone()
optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate)
train_batch_generator = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.wmt14.train(dict_size), buf_size=1000),
batch_size=args.batch_size * args.gpus)
test_batch_generator = paddle.batch(
with fluid.program_guard(main_prog, startup_prog):
with fluid.unique_name.guard():
avg_cost, feeding_list = seq_to_seq_net(
embedding_dim,
encoder_size,
decoder_size,
dict_size,
dict_size,
False,
beam_size=beam_size,
max_length=max_length)
if is_train:
optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate)
optimizer.minimize(avg_cost)
batch_generator = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.wmt14.test(dict_size), buf_size=1000),
batch_size=args.batch_size)
paddle.dataset.wmt14.train(dict_size)
if is_train else paddle.dataset.wmt14.test(dict_size),
buf_size=1000),
batch_size=args.batch_size * args.gpus)
return avg_cost, inference_program, optimizer, train_batch_generator, \
test_batch_generator, None
return avg_cost, optimizer, [], batch_generator, None
benchmark/fluid/models/mnist.py
浏览文件 @ 7116cc12
......@@ -65,61 +65,53 @@ def cnn_model(data):
return predict
def get_model(args):
if args.use_reader_op:
filelist = [
os.path.join(args.data_path, f) for f in os.listdir(args.data_path)
]
data_file = fluid.layers.open_files(
filenames=filelist,
shapes=[[-1, 1, 28, 28], (-1, 1)],
lod_levels=[0, 0],
dtypes=["float32", "int64"],
thread_num=args.gpus,
pass_num=args.pass_num)
data_file = fluid.layers.double_buffer(
fluid.layers.batch(
data_file, batch_size=args.batch_size))
images, label = fluid.layers.read_file(data_file)
else:
images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE)
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
if args.device == 'CPU' and args.cpus > 1:
places = fluid.layers.get_places(args.cpus)
pd = fluid.layers.ParallelDo(places)
with pd.do():
predict = cnn_model(pd.read_input(images))
label = pd.read_input(label)
def get_model(args, is_train, main_prog, startup_prog):
# NOTE: mnist is small, we don't implement data sharding yet.
opt = None
data_file_handle = None
with fluid.program_guard(main_prog, startup_prog):
if args.use_reader_op:
filelist = [
os.path.join(args.data_path, f)
for f in os.listdir(args.data_path)
]
data_file_handle = fluid.layers.open_files(
filenames=filelist,
shapes=[[-1, 1, 28, 28], (-1, 1)],
lod_levels=[0, 0],
dtypes=["float32", "int64"],
thread_num=1,
pass_num=1)
data_file = fluid.layers.double_buffer(
fluid.layers.batch(
data_file_handle, batch_size=args.batch_size))
with fluid.unique_name.guard():
if args.use_reader_op:
input, label = fluid.layers.read_file(data_file)
else:
images = fluid.layers.data(
name='pixel', shape=[1, 28, 28], dtype='float32')
label = fluid.layers.data(
name='label', shape=[1], dtype='int64')
predict = cnn_model(images)
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
# Evaluator
batch_acc = fluid.layers.accuracy(input=predict, label=label)
pd.write_output(avg_cost)
pd.write_output(batch_acc)
avg_cost, batch_acc = pd()
avg_cost = fluid.layers.mean(avg_cost)
batch_acc = fluid.layers.mean(batch_acc)
else:
# Train program
predict = cnn_model(images)
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
# Evaluator
batch_acc = fluid.layers.accuracy(input=predict, label=label)
# inference program
inference_program = fluid.default_main_program().clone()
# Optimization
opt = fluid.optimizer.AdamOptimizer(
learning_rate=0.001, beta1=0.9, beta2=0.999)
# Optimization
if is_train:
opt = fluid.optimizer.AdamOptimizer(
learning_rate=0.001, beta1=0.9, beta2=0.999)
opt.minimize(avg_cost)
if args.memory_optimize:
fluid.memory_optimize(main_prog)
# Reader
train_reader = paddle.batch(
paddle.dataset.mnist.train(), batch_size=args.batch_size * args.gpus)
test_reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=args.batch_size)
return avg_cost, inference_program, opt, train_reader, test_reader, batch_acc
if is_train:
reader = paddle.dataset.mnist.train()
else:
reader = paddle.dataset.mnist.test()
batched_reader = paddle.batch(
reader, batch_size=args.batch_size * args.gpus)
return avg_cost, opt, [batch_acc], batched_reader, data_file_handle
benchmark/fluid/models/resnet.py
浏览文件 @ 7116cc12
......@@ -20,6 +20,7 @@ import functools
import numpy as np
import time
import os
import math
import cProfile, pstats, StringIO
......@@ -27,182 +28,215 @@ import paddle
import paddle.fluid as fluid
import paddle.fluid.core as core
import paddle.fluid.profiler as profiler
from recordio_converter import imagenet_train, imagenet_test
def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'):
conv1 = fluid.layers.conv2d(
input=input,
filter_size=filter_size,
num_filters=ch_out,
stride=stride,
padding=padding,
act=None,
bias_attr=False)
return fluid.layers.batch_norm(input=conv1, act=act)
def shortcut(input, ch_out, stride):
ch_in = input.shape[1] # if args.data_format == 'NCHW' else input.shape[-1]
if ch_in != ch_out:
return conv_bn_layer(input, ch_out, 1, stride, 0, None)
else:
return input
def basicblock(input, ch_out, stride):
short = shortcut(input, ch_out, stride)
conv1 = conv_bn_layer(input, ch_out, 3, stride, 1)
conv2 = conv_bn_layer(conv1, ch_out, 3, 1, 1, act=None)
return fluid.layers.elementwise_add(x=short, y=conv2, act='relu')
def bottleneck(input, ch_out, stride):
short = shortcut(input, ch_out * 4, stride)
conv1 = conv_bn_layer(input, ch_out, 1, stride, 0)
conv2 = conv_bn_layer(conv1, ch_out, 3, 1, 1)
conv3 = conv_bn_layer(conv2, ch_out * 4, 1, 1, 0, act=None)
return fluid.layers.elementwise_add(x=short, y=conv3, act='relu')
from imagenet_reader import train, val
train_parameters = {
"input_size": [3, 224, 224],
"input_mean": [0.485, 0.456, 0.406],
"input_std": [0.229, 0.224, 0.225],
"learning_strategy": {
"name": "piecewise_decay",
"batch_size": 256,
"epochs": [30, 60, 90],
"steps": [0.1, 0.01, 0.001, 0.0001]
}
}
class ResNet():
def __init__(self, layers=50, is_train=True):
self.params = train_parameters
self.layers = layers
self.is_train = is_train
def net(self, input, class_dim=1000):
layers = self.layers
supported_layers = [50, 101, 152]
assert layers in supported_layers, \
"supported layers are {} but input layer is {}".format(supported_layers, layers)
if layers == 50:
depth = [3, 4, 6, 3]
elif layers == 101:
depth = [3, 4, 23, 3]
elif layers == 152:
depth = [3, 8, 36, 3]
num_filters = [64, 128, 256, 512]
conv = self.conv_bn_layer(
input=input, num_filters=64, filter_size=7, stride=2, act='relu')
conv = fluid.layers.pool2d(
input=conv,
pool_size=3,
pool_stride=2,
pool_padding=1,
pool_type='max')
for block in range(len(depth)):
for i in range(depth[block]):
conv = self.bottleneck_block(
input=conv,
num_filters=num_filters[block],
stride=2 if i == 0 and block != 0 else 1)
pool = fluid.layers.pool2d(
input=conv, pool_size=7, pool_type='avg', global_pooling=True)
stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
out = fluid.layers.fc(input=pool,
size=class_dim,
act='softmax',
param_attr=fluid.param_attr.ParamAttr(
initializer=fluid.initializer.Uniform(-stdv,
stdv)))
return out
def conv_bn_layer(self,
input,
num_filters,
filter_size,
stride=1,
groups=1,
act=None):
conv = fluid.layers.conv2d(
input=input,
num_filters=num_filters,
filter_size=filter_size,
stride=stride,
padding=(filter_size - 1) // 2,
groups=groups,
act=None,
bias_attr=False)
return fluid.layers.batch_norm(
input=conv, act=act, is_test=not self.is_train)
def shortcut(self, input, ch_out, stride):
ch_in = input.shape[1]
if ch_in != ch_out or stride != 1:
return self.conv_bn_layer(input, ch_out, 1, stride)
else:
return input
def bottleneck_block(self, input, num_filters, stride):
conv0 = self.conv_bn_layer(
input=input, num_filters=num_filters, filter_size=1, act='relu')
conv1 = self.conv_bn_layer(
input=conv0,
num_filters=num_filters,
filter_size=3,
stride=stride,
act='relu')
conv2 = self.conv_bn_layer(
input=conv1, num_filters=num_filters * 4, filter_size=1, act=None)
def layer_warp(block_func, input, ch_out, count, stride):
res_out = block_func(input, ch_out, stride)
for i in range(1, count):
res_out = block_func(res_out, ch_out, 1)
return res_out
short = self.shortcut(input, num_filters * 4, stride)
return fluid.layers.elementwise_add(x=short, y=conv2, act='relu')
def resnet_imagenet(input, class_dim, depth=50, data_format='NCHW'):
cfg = {
18: ([2, 2, 2, 1], basicblock),
34: ([3, 4, 6, 3], basicblock),
50: ([3, 4, 6, 3], bottleneck),
101: ([3, 4, 23, 3], bottleneck),
152: ([3, 8, 36, 3], bottleneck)
}
stages, block_func = cfg[depth]
conv1 = conv_bn_layer(input, ch_out=64, filter_size=7, stride=2, padding=3)
pool1 = fluid.layers.pool2d(
input=conv1, pool_type='avg', pool_size=3, pool_stride=2)
res1 = layer_warp(block_func, pool1, 64, stages[0], 1)
res2 = layer_warp(block_func, res1, 128, stages[1], 2)
res3 = layer_warp(block_func, res2, 256, stages[2], 2)
res4 = layer_warp(block_func, res3, 512, stages[3], 2)
pool2 = fluid.layers.pool2d(
input=res4,
pool_size=7,
pool_type='avg',
pool_stride=1,
global_pooling=True)
out = fluid.layers.fc(input=pool2, size=class_dim, act='softmax')
return out
def resnet_cifar10(input, class_dim, depth=32, data_format='NCHW'):
assert (depth - 2) % 6 == 0
n = (depth - 2) // 6
conv1 = conv_bn_layer(
input=input, ch_out=16, filter_size=3, stride=1, padding=1)
res1 = layer_warp(basicblock, conv1, 16, n, 1)
res2 = layer_warp(basicblock, res1, 32, n, 2)
res3 = layer_warp(basicblock, res2, 64, n, 2)
pool = fluid.layers.pool2d(
input=res3, pool_size=8, pool_type='avg', pool_stride=1)
out = fluid.layers.fc(input=pool, size=class_dim, act='softmax')
return out
def get_model(args):
model = resnet_cifar10
if args.data_set == "cifar10":
class_dim = 10
if args.data_format == 'NCHW':
dshape = [3, 32, 32]
else:
dshape = [32, 32, 3]
model = resnet_cifar10
train_reader = paddle.dataset.cifar.train10()
test_reader = paddle.dataset.cifar.test10()
elif args.data_set == "flowers":
def _model_reader_dshape_classdim(args, is_train):
model = None
reader = None
if args.data_set == "flowers":
class_dim = 102
if args.data_format == 'NCHW':
dshape = [3, 224, 224]
else:
dshape = [224, 224, 3]
model = resnet_imagenet
train_reader = paddle.dataset.flowers.train()
test_reader = paddle.dataset.flowers.test()
if is_train:
reader = paddle.dataset.flowers.train()
else:
reader = paddle.dataset.flowers.test()
elif args.data_set == "imagenet":
class_dim = 1000
if args.data_format == 'NCHW':
dshape = [3, 224, 224]
else:
dshape = [224, 224, 3]
model = resnet_imagenet
if not args.data_path:
raise Exception(
"Must specify --data_path when training with imagenet")
train_reader = imagenet_train(args.data_path)
test_reader = imagenet_test(args.data_path)
if args.use_reader_op:
filelist = [
os.path.join(args.data_path, f) for f in os.listdir(args.data_path)
]
data_file = fluid.layers.open_files(
filenames=filelist,
shapes=[[-1] + dshape, (-1, 1)],
lod_levels=[0, 0],
dtypes=["float32", "int64"],
thread_num=args.gpus,
pass_num=args.pass_num)
data_file = fluid.layers.double_buffer(
fluid.layers.batch(
data_file, batch_size=args.batch_size))
input, label = fluid.layers.read_file(data_file)
else:
input = fluid.layers.data(name='data', shape=dshape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
if args.device == 'CPU' and args.cpus > 1:
places = fluid.layers.get_places(args.cpus)
pd = fluid.layers.ParallelDo(places)
with pd.do():
predict = model(pd.read_input(input), class_dim)
label = pd.read_input(label)
if not args.use_reader_op:
if is_train:
reader = train()
else:
reader = val()
else:
if is_train:
reader = train(xmap=False)
else:
reader = val(xmap=False)
return reader, dshape, class_dim
def get_model(args, is_train, main_prog, startup_prog):
reader, dshape, class_dim = _model_reader_dshape_classdim(args, is_train)
pyreader = None
trainer_count = int(os.getenv("PADDLE_TRAINERS"))
with fluid.program_guard(main_prog, startup_prog):
with fluid.unique_name.guard():
if args.use_reader_op:
pyreader = fluid.layers.py_reader(
capacity=args.batch_size * args.gpus,
shapes=([-1] + dshape, (-1, 1)),
dtypes=('float32', 'int64'),
name="train_reader" if is_train else "test_reader",
use_double_buffer=True)
input, label = fluid.layers.read_file(pyreader)
else:
input = fluid.layers.data(
name='data', shape=dshape, dtype='float32')
label = fluid.layers.data(
name='label', shape=[1], dtype='int64')
model = ResNet(is_train=is_train)
predict = model.net(input, class_dim=class_dim)
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
batch_acc = fluid.layers.accuracy(input=predict, label=label)
pd.write_output(avg_cost)
pd.write_output(batch_acc)
avg_cost, batch_acc = pd()
avg_cost = fluid.layers.mean(avg_cost)
batch_acc = fluid.layers.mean(batch_acc)
batch_acc1 = fluid.layers.accuracy(input=predict, label=label, k=1)
batch_acc5 = fluid.layers.accuracy(input=predict, label=label, k=5)
# configure optimize
optimizer = None
if is_train:
if args.use_lars:
lars_decay = 1.0
else:
lars_decay = 0.0
total_images = 1281167 / trainer_count
step = int(total_images / (args.batch_size * args.gpus) + 1)
epochs = [30, 60, 90]
bd = [step * e for e in epochs]
base_lr = args.learning_rate
lr = []
lr = [base_lr * (0.1**i) for i in range(len(bd) + 1)]
optimizer = fluid.optimizer.Momentum(
learning_rate=fluid.layers.piecewise_decay(
boundaries=bd, values=lr),
momentum=0.9,
regularization=fluid.regularizer.L2Decay(1e-4))
optimizer.minimize(avg_cost)
if args.memory_optimize:
fluid.memory_optimize(main_prog)
# config readers
if not args.use_reader_op:
batched_reader = paddle.batch(
reader if args.no_random else paddle.reader.shuffle(
reader, buf_size=5120),
batch_size=args.batch_size * args.gpus,
drop_last=True)
else:
predict = model(input, class_dim)
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
batch_acc = fluid.layers.accuracy(input=predict, label=label)
inference_program = fluid.default_main_program().clone()
with fluid.program_guard(inference_program):
inference_program = fluid.io.get_inference_program(
target_vars=[batch_acc])
optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
batched_train_reader = paddle.batch(
train_reader if args.no_random else paddle.reader.shuffle(
train_reader, buf_size=5120),
batch_size=args.batch_size * args.gpus,
drop_last=True)
batched_test_reader = paddle.batch(
test_reader, batch_size=args.batch_size, drop_last=True)
return avg_cost, inference_program, optimizer, batched_train_reader,\
batched_test_reader, batch_acc
batched_reader = None
pyreader.decorate_paddle_reader(
paddle.batch(
reader if args.no_random else paddle.reader.shuffle(
reader, buf_size=5120),
batch_size=args.batch_size))
return avg_cost, optimizer, [batch_acc1,
batch_acc5], batched_reader, pyreader
benchmark/fluid/models/resnet_with_preprocess.py 0 → 100644
浏览文件 @ 7116cc12
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import functools
import numpy as np
import time
import os
import cProfile, pstats, StringIO
import paddle
import paddle.fluid as fluid
import paddle.fluid.core as core
import paddle.fluid.profiler as profiler
# from recordio_converter import imagenet_train, imagenet_test
from imagenet_reader import train_raw, val
def conv_bn_layer(input,
ch_out,
filter_size,
stride,
padding,
act='relu',
is_train=True):
conv1 = fluid.layers.conv2d(
input=input,
filter_size=filter_size,
num_filters=ch_out,
stride=stride,
padding=padding,
act=None,
bias_attr=False)
return fluid.layers.batch_norm(input=conv1, act=act, is_test=not is_train)
def shortcut(input, ch_out, stride, is_train=True):
ch_in = input.shape[1] # if args.data_format == 'NCHW' else input.shape[-1]
if ch_in != ch_out:
return conv_bn_layer(
input, ch_out, 1, stride, 0, None, is_train=is_train)
else:
return input
def basicblock(input, ch_out, stride, is_train=True):
short = shortcut(input, ch_out, stride, is_train=is_train)
conv1 = conv_bn_layer(input, ch_out, 3, stride, 1, is_train=is_train)
conv2 = conv_bn_layer(conv1, ch_out, 3, 1, 1, act=None, is_train=is_train)
return fluid.layers.elementwise_add(x=short, y=conv2, act='relu')
def bottleneck(input, ch_out, stride, is_train=True):
short = shortcut(input, ch_out * 4, stride, is_train=is_train)
conv1 = conv_bn_layer(input, ch_out, 1, stride, 0, is_train=is_train)
conv2 = conv_bn_layer(conv1, ch_out, 3, 1, 1, is_train=is_train)
conv3 = conv_bn_layer(
conv2, ch_out * 4, 1, 1, 0, act=None, is_train=is_train)
return fluid.layers.elementwise_add(x=short, y=conv3, act='relu')
def layer_warp(block_func, input, ch_out, count, stride):
res_out = block_func(input, ch_out, stride)
for i in range(1, count):
res_out = block_func(res_out, ch_out, 1)
return res_out
def resnet_imagenet(input,
class_dim,
depth=50,
data_format='NCHW',
is_train=True):
cfg = {
18: ([2, 2, 2, 1], basicblock),
34: ([3, 4, 6, 3], basicblock),
50: ([3, 4, 6, 3], bottleneck),
101: ([3, 4, 23, 3], bottleneck),
152: ([3, 8, 36, 3], bottleneck)
}
stages, block_func = cfg[depth]
conv1 = conv_bn_layer(input, ch_out=64, filter_size=7, stride=2, padding=3)
pool1 = fluid.layers.pool2d(
input=conv1, pool_type='avg', pool_size=3, pool_stride=2)
res1 = layer_warp(block_func, pool1, 64, stages[0], 1)
res2 = layer_warp(block_func, res1, 128, stages[1], 2)
res3 = layer_warp(block_func, res2, 256, stages[2], 2)
res4 = layer_warp(block_func, res3, 512, stages[3], 2)
pool2 = fluid.layers.pool2d(
input=res4,
pool_size=7,
pool_type='avg',
pool_stride=1,
global_pooling=True)
out = fluid.layers.fc(input=pool2, size=class_dim, act='softmax')
return out
def resnet_cifar10(input, class_dim, depth=32, data_format='NCHW'):
assert (depth - 2) % 6 == 0
n = (depth - 2) // 6
conv1 = conv_bn_layer(
input=input, ch_out=16, filter_size=3, stride=1, padding=1)
res1 = layer_warp(basicblock, conv1, 16, n, 1)
res2 = layer_warp(basicblock, res1, 32, n, 2)
res3 = layer_warp(basicblock, res2, 64, n, 2)
pool = fluid.layers.pool2d(
input=res3, pool_size=8, pool_type='avg', pool_stride=1)
out = fluid.layers.fc(input=pool, size=class_dim, act='softmax')
return out
def _model_reader_dshape_classdim(args, is_train):
model = resnet_cifar10
reader = None
if args.data_set == "cifar10":
class_dim = 10
if args.data_format == 'NCHW':
dshape = [3, 32, 32]
else:
dshape = [32, 32, 3]
model = resnet_cifar10
if is_train:
reader = paddle.dataset.cifar.train10()
else:
reader = paddle.dataset.cifar.test10()
elif args.data_set == "flowers":
class_dim = 102
if args.data_format == 'NCHW':
dshape = [3, 224, 224]
else:
dshape = [224, 224, 3]
model = resnet_imagenet
if is_train:
reader = paddle.dataset.flowers.train()
else:
reader = paddle.dataset.flowers.test()
elif args.data_set == "imagenet":
class_dim = 1000
if args.data_format == 'NCHW':
dshape = [3, 224, 224]
else:
dshape = [224, 224, 3]
model = resnet_imagenet
if not args.data_path:
raise Exception(
"Must specify --data_path when training with imagenet")
if not args.use_reader_op:
if is_train:
reader = train_raw()
else:
reader = val()
else:
if is_train:
reader = train_raw()
else:
reader = val(xmap=False)
return model, reader, dshape, class_dim
def get_model(args, is_train, main_prog, startup_prog):
model, reader, dshape, class_dim = _model_reader_dshape_classdim(args,
is_train)
pyreader = None
trainer_count = int(os.getenv("PADDLE_TRAINERS"))
with fluid.program_guard(main_prog, startup_prog):
with fluid.unique_name.guard():
if args.use_reader_op:
pyreader = fluid.layers.py_reader(
capacity=args.batch_size * args.gpus,
shapes=([-1] + dshape, (-1, 1)),
dtypes=('uint8', 'int64'),
name="train_reader" if is_train else "test_reader",
use_double_buffer=True)
input, label = fluid.layers.read_file(pyreader)
else:
input = fluid.layers.data(
name='data', shape=dshape, dtype='uint8')
label = fluid.layers.data(
name='label', shape=[1], dtype='int64')
# add imagenet preprocessors
random_crop = fluid.layers.random_crop(input, dshape)
casted = fluid.layers.cast(random_crop, 'float32')
# input is HWC
trans = fluid.layers.transpose(casted, [0, 3, 1, 2]) / 255.0
img_mean = fluid.layers.tensor.assign(
np.array([0.485, 0.456, 0.406]).astype('float32').reshape((3, 1,
1)))
img_std = fluid.layers.tensor.assign(
np.array([0.229, 0.224, 0.225]).astype('float32').reshape((3, 1,
1)))
h1 = fluid.layers.elementwise_sub(trans, img_mean, axis=1)
h2 = fluid.layers.elementwise_div(h1, img_std, axis=1)
# pre_out = (trans - img_mean) / img_std
predict = model(h2, class_dim, is_train=is_train)
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
batch_acc1 = fluid.layers.accuracy(input=predict, label=label, k=1)
batch_acc5 = fluid.layers.accuracy(input=predict, label=label, k=5)
# configure optimize
optimizer = None
if is_train:
if args.use_lars:
lars_decay = 1.0
else:
lars_decay = 0.0
total_images = 1281167 / trainer_count
step = int(total_images / args.batch_size + 1)
epochs = [30, 60, 80, 90]
bd = [step * e for e in epochs]
base_lr = args.learning_rate
lr = []
lr = [base_lr * (0.1**i) for i in range(len(bd) + 1)]
optimizer = fluid.optimizer.Momentum(
learning_rate=base_lr,
#learning_rate=fluid.layers.piecewise_decay(
# boundaries=bd, values=lr),
momentum=0.9,
regularization=fluid.regularizer.L2Decay(1e-4))
optimizer.minimize(avg_cost)
if args.memory_optimize:
fluid.memory_optimize(main_prog)
# config readers
if not args.use_reader_op:
batched_reader = paddle.batch(
reader if args.no_random else paddle.reader.shuffle(
reader, buf_size=5120),
batch_size=args.batch_size * args.gpus,
drop_last=True)
else:
batched_reader = None
pyreader.decorate_paddle_reader(
paddle.batch(
# reader if args.no_random else paddle.reader.shuffle(
# reader, buf_size=5120),
reader,
batch_size=args.batch_size))
return avg_cost, optimizer, [batch_acc1,
batch_acc5], batched_reader, pyreader
benchmark/fluid/models/se_resnext.py 0 → 100644
浏览文件 @ 7116cc12
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle
import paddle.fluid as fluid
import math
import os
from imagenet_reader import train, val
__all__ = [
"SE_ResNeXt", "SE_ResNeXt50_32x4d", "SE_ResNeXt101_32x4d",
"SE_ResNeXt152_32x4d", "get_model"
]
train_parameters = {
"input_size": [3, 224, 224],
"input_mean": [0.485, 0.456, 0.406],
"input_std": [0.229, 0.224, 0.225],
"learning_strategy": {
"name": "piecewise_decay",
"batch_size": 256,
"epochs": [30, 60, 90],
"steps": [0.1, 0.01, 0.001, 0.0001]
}
}
class SE_ResNeXt():
def __init__(self, layers=50, is_train=True):
self.params = train_parameters
self.layers = layers
self.is_train = is_train
def net(self, input, class_dim=1000):
layers = self.layers
supported_layers = [50, 101, 152]
assert layers in supported_layers, \
"supported layers are {} but input layer is {}".format(supported_layers, layers)
if layers == 50:
cardinality = 32
reduction_ratio = 16
depth = [3, 4, 6, 3]
num_filters = [128, 256, 512, 1024]
conv = self.conv_bn_layer(
input=input,
num_filters=64,
filter_size=7,
stride=2,
act='relu')
conv = fluid.layers.pool2d(
input=conv,
pool_size=3,
pool_stride=2,
pool_padding=1,
pool_type='max')
elif layers == 101:
cardinality = 32
reduction_ratio = 16
depth = [3, 4, 23, 3]
num_filters = [128, 256, 512, 1024]
conv = self.conv_bn_layer(
input=input,
num_filters=64,
filter_size=7,
stride=2,
act='relu')
conv = fluid.layers.pool2d(
input=conv,
pool_size=3,
pool_stride=2,
pool_padding=1,
pool_type='max')
elif layers == 152:
cardinality = 64
reduction_ratio = 16
depth = [3, 8, 36, 3]
num_filters = [128, 256, 512, 1024]
conv = self.conv_bn_layer(
input=input,
num_filters=64,
filter_size=3,
stride=2,
act='relu')
conv = self.conv_bn_layer(
input=conv, num_filters=64, filter_size=3, stride=1, act='relu')
conv = self.conv_bn_layer(
input=conv,
num_filters=128,
filter_size=3,
stride=1,
act='relu')
conv = fluid.layers.pool2d(
input=conv, pool_size=3, pool_stride=2, pool_padding=1, \
pool_type='max')
for block in range(len(depth)):
for i in range(depth[block]):
conv = self.bottleneck_block(
input=conv,
num_filters=num_filters[block],
stride=2 if i == 0 and block != 0 else 1,
cardinality=cardinality,
reduction_ratio=reduction_ratio)
pool = fluid.layers.pool2d(
input=conv, pool_size=7, pool_type='avg', global_pooling=True)
drop = fluid.layers.dropout(x=pool, dropout_prob=0.5)
stdv = 1.0 / math.sqrt(drop.shape[1] * 1.0)
out = fluid.layers.fc(input=drop,
size=class_dim,
act='softmax',
param_attr=fluid.param_attr.ParamAttr(
initializer=fluid.initializer.Uniform(-stdv,
stdv)))
return out
def shortcut(self, input, ch_out, stride):
ch_in = input.shape[1]
if ch_in != ch_out or stride != 1:
filter_size = 1
return self.conv_bn_layer(input, ch_out, filter_size, stride)
else:
return input
def bottleneck_block(self, input, num_filters, stride, cardinality,
reduction_ratio):
conv0 = self.conv_bn_layer(
input=input, num_filters=num_filters, filter_size=1, act='relu')
conv1 = self.conv_bn_layer(
input=conv0,
num_filters=num_filters,
filter_size=3,
stride=stride,
groups=cardinality,
act='relu')
conv2 = self.conv_bn_layer(
input=conv1, num_filters=num_filters * 2, filter_size=1, act=None)
scale = self.squeeze_excitation(
input=conv2,
num_channels=num_filters * 2,
reduction_ratio=reduction_ratio)
short = self.shortcut(input, num_filters * 2, stride)
return fluid.layers.elementwise_add(x=short, y=scale, act='relu')
def conv_bn_layer(self,
input,
num_filters,
filter_size,
stride=1,
groups=1,
act=None):
conv = fluid.layers.conv2d(
input=input,
num_filters=num_filters,
filter_size=filter_size,
stride=stride,
padding=(filter_size - 1) / 2,
groups=groups,
act=None,
bias_attr=False)
return fluid.layers.batch_norm(
input=conv, act=act, is_test=not self.is_train)
def squeeze_excitation(self, input, num_channels, reduction_ratio):
pool = fluid.layers.pool2d(
input=input, pool_size=0, pool_type='avg', global_pooling=True)
stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
squeeze = fluid.layers.fc(input=pool,
size=num_channels / reduction_ratio,
act='relu',
param_attr=fluid.param_attr.ParamAttr(
initializer=fluid.initializer.Uniform(
-stdv, stdv)))
stdv = 1.0 / math.sqrt(squeeze.shape[1] * 1.0)
excitation = fluid.layers.fc(input=squeeze,
size=num_channels,
act='sigmoid',
param_attr=fluid.param_attr.ParamAttr(
initializer=fluid.initializer.Uniform(
-stdv, stdv)))
scale = fluid.layers.elementwise_mul(x=input, y=excitation, axis=0)
return scale
def SE_ResNeXt50_32x4d():
model = SE_ResNeXt(layers=50)
return model
def SE_ResNeXt101_32x4d():
model = SE_ResNeXt(layers=101)
return model
def SE_ResNeXt152_32x4d():
model = SE_ResNeXt(layers=152)
return model
def get_model(args, is_train, main_prog, startup_prog):
model = SE_ResNeXt(layers=50)
batched_reader = None
pyreader = None
trainer_count = int(os.getenv("PADDLE_TRAINERS"))
dshape = train_parameters["input_size"]
with fluid.program_guard(main_prog, startup_prog):
with fluid.unique_name.guard():
if args.use_reader_op:
pyreader = fluid.layers.py_reader(
capacity=10,
shapes=([-1] + dshape, (-1, 1)),
dtypes=('float32', 'int64'),
name="train_reader" if is_train else "test_reader",
use_double_buffer=True)
input, label = fluid.layers.read_file(pyreader)
else:
input = fluid.layers.data(
name='data', shape=dshape, dtype='float32')
label = fluid.layers.data(
name='label', shape=[1], dtype='int64')
out = model.net(input=input)
cost = fluid.layers.cross_entropy(input=out, label=label)
avg_cost = fluid.layers.mean(x=cost)
acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
optimizer = None
if is_train:
if args.use_lars:
lars_decay = 1.0
else:
lars_decay = 0.0
total_images = 1281167 / trainer_count
step = int(total_images / args.batch_size + 1)
epochs = [40, 80, 100]
bd = [step * e for e in epochs]
base_lr = args.learning_rate
lr = []
lr = [base_lr * (0.1**i) for i in range(len(bd) + 1)]
optimizer = fluid.optimizer.Momentum(
# learning_rate=base_lr,
learning_rate=fluid.layers.piecewise_decay(
boundaries=bd, values=lr),
momentum=0.9,
regularization=fluid.regularizer.L2Decay(1e-4),
LARS_weight_decay=lars_decay)
optimizer.minimize(avg_cost)
if args.memory_optimize:
fluid.memory_optimize(main_prog)
# config readers
if is_train:
reader = train()
else:
reader = val()
if not args.use_reader_op:
batched_reader = paddle.batch(
reader, batch_size=args.batch_size * args.gpus, drop_last=True)
else:
pyreader.decorate_paddle_reader(
paddle.batch(
reader, batch_size=args.batch_size))
return avg_cost, optimizer, [acc_top1, acc_top5], batched_reader, pyreader
benchmark/fluid/models/stacked_dynamic_lstm.py
浏览文件 @ 7116cc12
......@@ -26,7 +26,6 @@ import numpy
import paddle
import paddle.dataset.imdb as imdb
import paddle.fluid as fluid
import paddle.batch as batch
import paddle.fluid.profiler as profiler
word_dict = imdb.word_dict()
......@@ -43,19 +42,7 @@ def crop_sentence(reader, crop_size):
return __impl__
def get_model(args):
if args.use_reader_op:
raise Exception(
"stacked_dynamic_lstm do not support reader op for now.")
lstm_size = 512
emb_dim = 512
crop_size = 1500
data = fluid.layers.data(
name="words", shape=[1], lod_level=1, dtype='int64')
sentence = fluid.layers.embedding(
input=data, size=[len(word_dict), emb_dim])
def lstm_net(sentence, lstm_size):
sentence = fluid.layers.fc(input=sentence, size=lstm_size, act='tanh')
rnn = fluid.layers.DynamicRNN()
......@@ -97,31 +84,47 @@ def get_model(args):
last = fluid.layers.sequence_pool(rnn(), 'last')
logit = fluid.layers.fc(input=last, size=2, act='softmax')
loss = fluid.layers.cross_entropy(
input=logit,
label=fluid.layers.data(
name='label', shape=[1], dtype='int64'))
loss = fluid.layers.mean(x=loss)
return logit
# add acc
batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(input=logit, label=fluid.layers.data(name='label', \
shape=[1], dtype='int64'), total=batch_size_tensor)
inference_program = fluid.default_main_program().clone()
with fluid.program_guard(inference_program):
inference_program = fluid.io.get_inference_program(
target_vars=[batch_acc, batch_size_tensor])
adam = fluid.optimizer.Adam()
def get_model(args, is_train, main_prog, startup_prog):
if args.use_reader_op:
raise Exception(
"stacked_dynamic_lstm do not support reader op for now.")
lstm_size = 512
emb_dim = 512
crop_size = 1500
train_reader = batch(
with fluid.program_guard(main_prog, startup_prog):
with fluid.unique_name.guard():
data = fluid.layers.data(
name="words", shape=[1], lod_level=1, dtype='int64')
sentence = fluid.layers.embedding(
input=data, size=[len(word_dict), emb_dim])
logit = lstm_net(sentence, lstm_size)
loss = fluid.layers.cross_entropy(
input=logit,
label=fluid.layers.data(
name='label', shape=[1], dtype='int64'))
loss = fluid.layers.mean(x=loss)
# add acc
batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(input=logit, label=fluid.layers.data(name='label', \
shape=[1], dtype='int64'), total=batch_size_tensor)
if is_train:
adam = fluid.optimizer.Adam()
adam.minimize(loss)
if is_train:
reader = crop_sentence(imdb.train(word_dict), crop_size)
else:
reader = crop_sentence(imdb.test(word_dict), crop_size)
batched_reader = paddle.batch(
paddle.reader.shuffle(
crop_sentence(imdb.train(word_dict), crop_size), buf_size=25000),
reader, buf_size=25000),
batch_size=args.batch_size * args.gpus)
test_reader = batch(
paddle.reader.shuffle(
crop_sentence(imdb.test(word_dict), crop_size), buf_size=25000),
batch_size=args.batch_size)
return loss, inference_program, adam, train_reader, test_reader, batch_acc
return loss, adam, [batch_acc], batched_reader, None
benchmark/fluid/models/vgg.py
浏览文件 @ 7116cc12
......@@ -25,7 +25,7 @@ import functools
import os
def vgg16_bn_drop(input):
def vgg16_bn_drop(input, is_train=True):
def conv_block(input, num_filter, groups, dropouts):
return fluid.nets.img_conv_group(
input=input,
......@@ -46,13 +46,13 @@ def vgg16_bn_drop(input):
drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5)
fc1 = fluid.layers.fc(input=drop, size=512, act=None)
bn = fluid.layers.batch_norm(input=fc1, act='relu')
bn = fluid.layers.batch_norm(input=fc1, act='relu', is_test=not is_train)
drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5)
fc2 = fluid.layers.fc(input=drop2, size=512, act=None)
return fc2
def get_model(args):
def get_model(args, is_train, main_prog, startup_prog):
if args.data_set == "cifar10":
classdim = 10
if args.data_format == 'NCHW':
......@@ -65,57 +65,56 @@ def get_model(args):
data_shape = [3, 224, 224]
else:
data_shape = [224, 224, 3]
filelist = [
os.path.join(args.data_path, f) for f in os.listdir(args.data_path)
]
with fluid.program_guard(main_prog, startup_prog):
if args.use_reader_op:
data_file_handle = fluid.layers.open_files(
filenames=filelist,
shapes=[[-1] + data_shape, (-1, 1)],
lod_levels=[0, 0],
dtypes=["float32", "int64"],
thread_num=1,
pass_num=1)
data_file = fluid.layers.double_buffer(
fluid.layers.batch(
data_file_handle, batch_size=args.batch_size))
with fluid.unique_name.guard():
if args.use_reader_op:
images, label = fluid.layers.read_file(data_file)
else:
images = fluid.layers.data(
name='data', shape=data_shape, dtype='float32')
label = fluid.layers.data(
name='label', shape=[1], dtype='int64')
# Train program
net = vgg16_bn_drop(images, is_train=is_train)
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)
if args.use_reader_op:
filelist = [
os.path.join(args.data_path, f) for f in os.listdir(args.data_path)
]
data_file = fluid.layers.open_files(
filenames=filelist,
shapes=[[-1] + data_shape, (-1, 1)],
lod_levels=[0, 0],
dtypes=["float32", "int64"],
thread_num=args.gpus,
pass_num=args.pass_num)
data_file = fluid.layers.double_buffer(
fluid.layers.batch(
data_file, batch_size=args.batch_size))
images, label = fluid.layers.read_file(data_file)
else:
images = fluid.layers.data(
name='data', 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_tensor = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(
input=predict, label=label, total=batch_size_tensor)
# inference program
inference_program = fluid.default_main_program().clone()
with fluid.program_guard(inference_program):
inference_program = fluid.io.get_inference_program(
target_vars=[batch_acc, batch_size_tensor])
# Optimization
optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate)
# Evaluator
batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(
input=predict, label=label, total=batch_size_tensor)
# Optimization
if is_train:
optimizer = fluid.optimizer.Adam(
learning_rate=args.learning_rate)
optimizer.minimize(avg_cost)
# data reader
train_reader = paddle.batch(
if is_train:
reader = paddle.dataset.cifar.train10() \
if args.data_set == 'cifar10' else paddle.dataset.flowers.train()
else:
reader = paddle.dataset.cifar.test10() \
if args.data_set == 'cifar10' else paddle.dataset.flowers.test()
batched_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.cifar.train10()
if args.data_set == 'cifar10' else paddle.dataset.flowers.train(),
buf_size=5120),
reader, buf_size=5120),
batch_size=args.batch_size * args.gpus)
test_reader = paddle.batch(
paddle.dataset.cifar.test10()
if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
batch_size=args.batch_size)
return avg_cost, inference_program, optimizer, train_reader, test_reader, batch_acc
return avg_cost, optimizer, [batch_acc], batched_reader, data_file_handle
benchmark/paddle/image/run.sh
浏览文件 @ 7116cc12
#!/bin/bash
set -e
function train() {
......
benchmark/paddle/image/run_mkl_infer.sh
浏览文件 @ 7116cc12
#!/bin/bash
set -e
function clock_to_seconds() {
......
benchmark/paddle/image/run_mkl_train.sh
浏览文件 @ 7116cc12
#!/bin/bash
set -e
function train() {
......
benchmark/paddle/image/run_openblas_infer.sh
浏览文件 @ 7116cc12
#!/bin/bash
set -e
function clock_to_seconds() {
......
benchmark/paddle/rnn/run.sh
浏览文件 @ 7116cc12
#!/bin/bash
set -e
function train() {
......
benchmark/tensorflow/image/run.sh
浏览文件 @ 7116cc12
#!/bin/bash
set -e
function test() {
......
benchmark/tensorflow/image/run_multi.sh
浏览文件 @ 7116cc12
#!/bin/bash
set -e
function test() {
......
benchmark/tensorflow/rnn/run.sh
浏览文件 @ 7116cc12
#!/bin/bash
set -e
function test() {
......
benchmark/tensorflow/rnn/run_multi.sh
浏览文件 @ 7116cc12
#!/bin/bash
set -e
function test() {
......
cmake/configure.cmake
浏览文件 @ 7116cc12
......@@ -50,13 +50,22 @@ if(NOT WITH_PROFILER)
endif(NOT WITH_PROFILER)
if(NOT CMAKE_CROSSCOMPILING)
if(WITH_AVX AND AVX_FOUND)
if(WITH_AVX AND AVX512F_FOUND)
set(SIMD_FLAG ${AVX512F_FLAG})
elseif(WITH_AVX AND AVX2_FOUND)
set(SIMD_FLAG ${AVX2_FLAG})
elseif(WITH_AVX AND AVX_FOUND)
set(SIMD_FLAG ${AVX_FLAG})
elseif(SSE3_FOUND)
set(SIMD_FLAG ${SSE3_FLAG})
endif()
endif()
if(WIN32)
# windows stupid compile option for all targets.
add_definitions(-D_XKEYCHECK_H)
endif(WIN32)
if(NOT WITH_GOLANG)
add_definitions(-DPADDLE_WITHOUT_GOLANG)
endif(NOT WITH_GOLANG)
......@@ -97,6 +106,23 @@ if(WITH_GPU)
endif()
include_directories(${TENSORRT_INCLUDE_DIR})
endif()
if(WITH_ANAKIN)
if(${CUDA_VERSION_MAJOR} VERSION_LESS 8)
message(WARNING "Anakin needs CUDA >= 8.0 to compile. Force WITH_ANAKIN=OFF")
set(WITH_ANAKIN OFF CACHE STRING "Anakin is valid only when CUDA >= 8.0." FORCE)
endif()
if(${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
message(WARNING "Anakin needs CUDNN >= 7.0 to compile. Force WITH_ANAKIN=OFF")
set(WITH_ANAKIN OFF CACHE STRING "Anakin is valid only when CUDNN >= 7.0." FORCE)
endif()
endif()
if(WITH_ANAKIN)
# NOTICE(minqiyang): the end slash is important because $CUDNN_INCLUDE_DIR
# is a softlink to real cudnn.h directory
set(ENV{CUDNN_INCLUDE_DIR} "${CUDNN_INCLUDE_DIR}/")
get_filename_component(CUDNN_LIBRARY_DIR ${CUDNN_LIBRARY} DIRECTORY)
set(ENV{CUDNN_LIBRARY} ${CUDNN_LIBRARY_DIR})
endif()
elseif(WITH_AMD_GPU)
add_definitions(-DPADDLE_WITH_HIP)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -D__HIP_PLATFORM_HCC__")
......
cmake/cuda.cmake
浏览文件 @ 7116cc12
......@@ -169,14 +169,19 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF)
# Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc.
# So, don't set these flags here.
if (NOT WIN32) # windows msvc2015 support c++11 natively.
# -std=c++11 -fPIC not recoginize by msvc, -Xcompiler will be added by cmake.
list(APPEND CUDA_NVCC_FLAGS "-std=c++11")
list(APPEND CUDA_NVCC_FLAGS "--use_fast_math")
list(APPEND CUDA_NVCC_FLAGS "-Xcompiler -fPIC")
endif(NOT WIN32)
list(APPEND CUDA_NVCC_FLAGS "--use_fast_math")
# in cuda9, suppress cuda warning on eigen
list(APPEND CUDA_NVCC_FLAGS "-w")
# Set :expt-relaxed-constexpr to suppress Eigen warnings
list(APPEND CUDA_NVCC_FLAGS "--expt-relaxed-constexpr")
if (NOT WIN32)
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_DEBUG})
elseif(CMAKE_BUILD_TYPE STREQUAL "Release")
......@@ -187,6 +192,13 @@ elseif(CMAKE_BUILD_TYPE STREQUAL "MinSizeRel")
# nvcc 9 does not support -Os. Use Release flags instead
list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE})
endif()
else(NOT WIN32)
if(CMAKE_BUILD_TYPE STREQUAL "Release")
list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG")
else()
message(FATAL "Windows only support Release build now. Please set visual studio build type to Release, x64 build.")
endif()
endif(NOT WIN32)
mark_as_advanced(CUDA_BUILD_CUBIN CUDA_BUILD_EMULATION CUDA_VERBOSE_BUILD)
mark_as_advanced(CUDA_SDK_ROOT_DIR CUDA_SEPARABLE_COMPILATION)
cmake/cudnn.cmake
浏览文件 @ 7116cc12
......@@ -21,11 +21,29 @@ list(APPEND CUDNN_CHECK_LIBRARY_DIRS
${CUDNN_ROOT}/lib64
${CUDNN_ROOT}/lib
${CUDNN_ROOT}/lib/${TARGET_ARCH}-linux-gnu
${CUDNN_ROOT}/local/cuda-${CUDA_VERSION}/targets/${TARGET_ARCH}-linux/lib/
$ENV{CUDNN_ROOT}
$ENV{CUDNN_ROOT}/lib64
$ENV{CUDNN_ROOT}/lib
/usr/lib)
find_library(CUDNN_LIBRARY NAMES libcudnn.so libcudnn.dylib # libcudnn_static.a
/usr/lib
${CUDA_TOOLKIT_ROOT_DIR}
${CUDA_TOOLKIT_ROOT_DIR}/lib/x64
)
set(CUDNN_LIB_NAME "")
if (LINUX)
set(CUDNN_LIB_NAME "libcudnn.so")
endif(LINUX)
if(WIN32)
# only support cudnn7
set(CUDNN_LIB_NAME "cudnn.lib" "cudnn64_7.dll")
endif(WIN32)
if(Apple)
set(CUDNN_LIB_NAME "libcudnn.dylib" "libcudnn.so")
endif(Apple)
find_library(CUDNN_LIBRARY NAMES ${CUDNN_LIB_NAME} # libcudnn_static.a
PATHS ${CUDNN_CHECK_LIBRARY_DIRS} ${CUDNN_INCLUDE_DIR} ${__libpath_hist}
NO_DEFAULT_PATH
DOC "Path to cuDNN library.")
......
cmake/external/anakin.cmake
浏览文件 @ 7116cc12
......@@ -2,43 +2,74 @@ if (NOT WITH_ANAKIN)
return()
endif()
set(ANAKIN_INSTALL_DIR "${THIRD_PARTY_PATH}/install/anakin" CACHE PATH
"Anakin install path." FORCE)
set(ANAKIN_INCLUDE "${ANAKIN_INSTALL_DIR}" CACHE STRING "root of Anakin header files")
set(ANAKIN_LIBRARY "${ANAKIN_INSTALL_DIR}" CACHE STRING "path of Anakin library")
set(ANAKIN_COMPILE_EXTRA_FLAGS -Wno-error=unused-variable -Wno-error=format-extra-args -Wno-error=comment -Wno-error=format -Wno-error=switch -Wno-error=return-type -Wno-error=non-virtual-dtor -Wno-reorder -Wno-error=cpp)
set(ANAKIN_LIBRARY_URL "https://github.com/pangge/Anakin/releases/download/3.0/anakin_release_simple.tar.gz")
# A helper function used in Anakin, currently, to use it, one need to recursively include
# nearly all the header files.
function(fetch_include_recursively root_dir)
if (IS_DIRECTORY ${root_dir})
include_directories(${root_dir})
endif()
file(GLOB ALL_SUB RELATIVE ${root_dir} ${root_dir}/*)
foreach(sub ${ALL_SUB})
if (IS_DIRECTORY ${root_dir}/${sub})
fetch_include_recursively(${root_dir}/${sub})
endif()
endforeach()
endfunction()
if (NOT EXISTS "${ANAKIN_INSTALL_DIR}")
# download library
message(STATUS "Download Anakin library from ${ANAKIN_LIBRARY_URL}")
execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}")
execute_process(COMMAND bash -c "rm -rf ${ANAKIN_INSTALL_DIR}/*")
execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; wget -q ${ANAKIN_LIBRARY_URL}")
execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}")
execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; tar xzf anakin_release_simple.tar.gz")
option(ANAKIN_ENABLE_OP_TIMER "Get more detailed information with Anakin op time" OFF)
if(ANAKIN_ENABLE_OP_TIMER)
add_definitions(-DPADDLE_ANAKIN_ENABLE_OP_TIMER)
endif()
if (WITH_ANAKIN)
message(STATUS "Anakin for inference is enabled")
message(STATUS "Anakin is set INCLUDE:${ANAKIN_INCLUDE} LIBRARY:${ANAKIN_LIBRARY}")
fetch_include_recursively(${ANAKIN_INCLUDE})
link_directories(${ANAKIN_LIBRARY})
INCLUDE(ExternalProject)
set(ANAKIN_SOURCE_DIR ${THIRD_PARTY_PATH}/anakin)
# the anakin install dir is only default one now
set(ANAKIN_INSTALL_DIR ${THIRD_PARTY_PATH}/anakin/src/extern_anakin/output)
set(ANAKIN_INCLUDE ${ANAKIN_INSTALL_DIR})
set(ANAKIN_LIBRARY ${ANAKIN_INSTALL_DIR})
set(ANAKIN_SHARED_LIB ${ANAKIN_LIBRARY}/libanakin.so)
set(ANAKIN_SABER_LIB ${ANAKIN_LIBRARY}/libanakin_saber_common.so)
include_directories(${ANAKIN_INCLUDE})
include_directories(${ANAKIN_INCLUDE}/saber/)
include_directories(${ANAKIN_INCLUDE}/saber/core/)
include_directories(${ANAKIN_INCLUDE}/saber/funcs/impl/x86/)
include_directories(${ANAKIN_INCLUDE}/saber/funcs/impl/cuda/base/cuda_c/)
set(ANAKIN_COMPILE_EXTRA_FLAGS
-Wno-error=unused-but-set-variable -Wno-unused-but-set-variable
-Wno-error=unused-variable -Wno-unused-variable
-Wno-error=format-extra-args -Wno-format-extra-args
-Wno-error=comment -Wno-comment
-Wno-error=format -Wno-format
-Wno-error=maybe-uninitialized -Wno-maybe-uninitialized
-Wno-error=switch -Wno-switch
-Wno-error=return-type -Wno-return-type
-Wno-error=non-virtual-dtor -Wno-non-virtual-dtor
-Wno-error=ignored-qualifiers
-Wno-ignored-qualifiers
-Wno-sign-compare
-Wno-reorder
-Wno-error=cpp)
if(WITH_GPU)
set(CMAKE_ARGS_PREFIX -DUSE_GPU_PLACE=YES -DCUDNN_ROOT=${CUDNN_ROOT} -DCUDNN_INCLUDE_DIR=${CUDNN_INCLUDE_DIR})
else()
set(CMAKE_ARGS_PREFIX -DUSE_GPU_PLACE=NO)
endif()
ExternalProject_Add(
extern_anakin
${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS ${MKLML_PROJECT}
GIT_REPOSITORY "https://github.com/PaddlePaddle/Anakin"
GIT_TAG "3c8554f4978628183566ab7dd6c1e7e66493c7cd"
PREFIX ${ANAKIN_SOURCE_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS ${CMAKE_ARGS_PREFIX}
-DUSE_X86_PLACE=YES
-DBUILD_WITH_UNIT_TEST=NO
-DPROTOBUF_ROOT=${THIRD_PARTY_PATH}/install/protobuf
-DMKLML_ROOT=${THIRD_PARTY_PATH}/install/mklml
-DENABLE_OP_TIMER=${ANAKIN_ENABLE_OP_TIMER}
${EXTERNAL_OPTIONAL_ARGS}
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${ANAKIN_INSTALL_DIR}
)
message(STATUS "Anakin for inference is enabled")
message(STATUS "Anakin is set INCLUDE:${ANAKIN_INCLUDE} LIBRARY:${ANAKIN_LIBRARY}")
add_library(anakin_shared SHARED IMPORTED GLOBAL)
set_property(TARGET anakin_shared PROPERTY IMPORTED_LOCATION ${ANAKIN_SHARED_LIB})
add_dependencies(anakin_shared extern_anakin protobuf mklml)
add_library(anakin_saber SHARED IMPORTED GLOBAL)
set_property(TARGET anakin_saber PROPERTY IMPORTED_LOCATION ${ANAKIN_SABER_LIB})
add_dependencies(anakin_saber extern_anakin protobuf mklml)
list(APPEND external_project_dependencies anakin_shared anakin_saber)
cmake/external/boost.cmake
浏览文件 @ 7116cc12
......@@ -28,7 +28,12 @@ if((NOT DEFINED BOOST_TAR) OR (NOT DEFINED BOOST_URL))
set(BOOST_TAR "boost_1_41_0" CACHE STRING "" FORCE)
set(BOOST_URL "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
endif()
MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
IF (WIN32)
MESSAGE(WARNING, "In windows, boost can not be downloaded automaticlly, please build it manually and put it at " ${THIRD_PARTY_PATH}install/boost)
else()
MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
ENDIF(WIN32)
set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost)
set(BOOST_DOWNLOAD_DIR "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}")
set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)
......@@ -36,12 +41,13 @@ set_directory_properties(PROPERTIES CLEAN_NO_CUSTOM 1)
include_directories(${BOOST_INCLUDE_DIR})
if (NOT WIN32)
ExternalProject_Add(
${BOOST_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
DOWNLOAD_DIR ${BOOST_DOWNLOAD_DIR}
DOWNLOAD_COMMAND wget --no-check-certificate ${BOOST_URL} -c -q -O ${BOOST_TAR}.tar.gz
&& tar zxf ${BOOST_TAR}.tar.gz
&& tar zxf ${BOOST_TAR}.tar.gz
DOWNLOAD_NO_PROGRESS 1
PREFIX ${BOOST_SOURCES_DIR}
CONFIGURE_COMMAND ""
......@@ -49,8 +55,9 @@ ExternalProject_Add(
INSTALL_COMMAND ""
UPDATE_COMMAND ""
)
endif(NOT WIN32)
if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
if (${CMAKE_VERSION} VERSION_LESS "3.3.0" OR NOT WIN32)
set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/boost_dummy.c)
file(WRITE ${dummyfile} "const char *dummy = \"${dummyfile}\";")
add_library(boost STATIC ${dummyfile})
......
cmake/external/cub.cmake 0 → 100644
浏览文件 @ 7116cc12
if(NOT WITH_GPU)
return()
endif()
include(ExternalProject)
set(CUB_SOURCE_DIR ${THIRD_PARTY_PATH}/cub)
set(CUB_INCLUDE_DIR ${CUB_SOURCE_DIR}/src/extern_cub)
include_directories(${CUB_INCLUDE_DIR})
ExternalProject_Add(
extern_cub
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/NVlabs/cub.git"
GIT_TAG "v1.8.0"
PREFIX ${CUB_SOURCE_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
INSTALL_COMMAND ""
TEST_COMMAND ""
)
if(${CMAKE_VERSION} VERSION_LESS "3.3.0")
set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cub_dummy.c)
file(WRITE ${dummyfile} "const char *dummy = \"${dummyfile}\";")
add_library(cub STATIC ${dummyfile})
else()
add_library(cub INTERFACE)
endif()
add_dependencies(cub extern_cub)
LIST(APPEND externl_project_dependencies cub)
cmake/external/gflags.cmake
浏览文件 @ 7116cc12
......@@ -18,7 +18,7 @@ SET(GFLAGS_SOURCES_DIR ${THIRD_PARTY_PATH}/gflags)
SET(GFLAGS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/gflags)
SET(GFLAGS_INCLUDE_DIR "${GFLAGS_INSTALL_DIR}/include" CACHE PATH "gflags include directory." FORCE)
IF(WIN32)
set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/gflags.lib" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/libgflags.lib" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
ELSE(WIN32)
set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/libgflags.a" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
ENDIF(WIN32)
......@@ -45,7 +45,13 @@ ExternalProject_Add(
-DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
-DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
)
IF(WIN32)
IF(NOT EXISTS "${GFLAGS_INSTALL_DIR}/lib/libgflags.lib")
add_custom_command(TARGET extern_gflags POST_BUILD
COMMAND cmake -E rename ${GFLAGS_INSTALL_DIR}/lib/gflags_static.lib ${GFLAGS_INSTALL_DIR}/lib/libgflags.lib
)
ENDIF()
ENDIF(WIN32)
ADD_LIBRARY(gflags STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET gflags PROPERTY IMPORTED_LOCATION ${GFLAGS_LIBRARIES})
ADD_DEPENDENCIES(gflags extern_gflags)
......@@ -60,3 +66,4 @@ IF(WITH_C_API)
INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib)
ENDIF()
ENDIF()
cmake/external/glog.cmake
浏览文件 @ 7116cc12
......@@ -60,6 +60,13 @@ ExternalProject_Add(
-DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
-DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
)
IF(WIN32)
IF(NOT EXISTS "${GLOG_INSTALL_DIR}/lib/libglog.lib")
add_custom_command(TARGET extern_glog POST_BUILD
COMMAND cmake -E rename ${GLOG_INSTALL_DIR}/lib/glog.lib ${GLOG_INSTALL_DIR}/lib/libglog.lib
)
ENDIF()
ENDIF(WIN32)
ADD_LIBRARY(glog STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET glog PROPERTY IMPORTED_LOCATION ${GLOG_LIBRARIES})
......
cmake/external/grpc.cmake
浏览文件 @ 7116cc12
......@@ -44,12 +44,13 @@ ExternalProject_Add(
# 3. keep only zlib, cares, protobuf, boringssl under "third_party",
# checkout and clean other dirs under third_party
# 4. remove .git, and package the directory.
URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.10.x.tar.gz"
URL "http://paddlepaddledeps.cdn.bcebos.com/grpc-v1.10.x.tar.gz"
URL_MD5 "1f268a2aff6759839dccd256adcc91cf"
PREFIX ${GRPC_SOURCES_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_IN_SOURCE 1
PATCH_COMMAND cp ${PADDLE_SOURCE_DIR}/patches/grpc/grpc_library.h ${GRPC_SOURCES_DIR}/src/extern_grpc/include/grpcpp/impl/codegen/grpc_library.h && cp ${PADDLE_SOURCE_DIR}/patches/grpc/completion_queue.h ${GRPC_SOURCES_DIR}/src/extern_grpc/include/grpcpp/impl/codegen/completion_queue.h
# NOTE(yuyang18):
# Disable -Werror, otherwise the compile will fail in MacOS.
# It seems that we cannot configure that by make command.
......
cmake/external/libxsmm.cmake 0 → 100644
浏览文件 @ 7116cc12
# 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.
#
OPTION(WITH_LIBXSMM "Compile with libxsmm" OFF)
IF(NOT WITH_LIBXSMM)
return()
ENDIF()
IF(WIN32 OR APPLE OR ANDROID OR IOS)
MESSAGE(WARNING "Windows, Mac or Mobile are not supported with libxsmm in Paddle yet.")
SET(WITH_LIBXSMM OFF CACHE STRING "Disable LIBXSMM" FORCE)
return()
ENDIF()
INCLUDE (ExternalProject)
SET(LIBXSMM_SOURCES_DIR ${THIRD_PARTY_PATH}/libxsmm)
SET(LIBXSMM_INSTALL_DIR ${THIRD_PARTY_PATH}/install/libxsmm)
SET(LIBXSMM_INCLUDE_DIR "${LIBXSMM_INSTALL_DIR}/include" CACHE PATH "LIBXSMM include directory." FORCE)
SET(LIBXSMM_LIBRARY_DIR "${LIBXSMM_INSTALL_DIR}/lib" CACHE PATH "LIBXSMM library directory." FORCE)
SET(LIBXSMM_LIBS "${LIBXSMM_LIBRARY_DIR}/libxsmm.a"
"${LIBXSMM_LIBRARY_DIR}/libxsmmnoblas.a")
ExternalProject_Add(
extern_libxsmm
GIT_REPOSITORY "https://github.com/hfp/libxsmm.git"
GIT_TAG "7cc03b5b342fdbc6b6d990b190671c5dbb8489a2"
PREFIX ${LIBXSMM_SOURCES_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_IN_SOURCE 1
BUILD_COMMAND $(MAKE) --silent PREFIX=${LIBXSMM_INSTALL_DIR} CXX=g++ CC=gcc WARP=0 install
INSTALL_COMMAND ""
)
ADD_LIBRARY(libxsmm STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET libxsmm PROPERTY IMPORTED_LOCATION "${LIBXSMM_LIBRARY_DIR}/libxsmm.a")
SET_PROPERTY(TARGET libxsmm PROPERTY IMPORTED_LOCATION "${LIBXSMM_LIBRARY_DIR}/libxsmmnoblas.a")
MESSAGE(STATUS "Libxsmm library: ${LIBXSMM_LIBS}")
include_directories(${LIBXSMM_INCLUDE_DIR})
ADD_DEFINITIONS(-DPADDLE_WITH_LIBXSMM)
ADD_DEPENDENCIES(libxsmm extern_libxsmm)
LIST(APPEND external_project_dependencies libxsmm)
cmake/external/mkldnn.cmake
浏览文件 @ 7116cc12
......@@ -24,7 +24,7 @@ SET(MKLDNN_INSTALL_DIR ${THIRD_PARTY_PATH}/install/mkldnn)
SET(MKLDNN_INC_DIR "${MKLDNN_INSTALL_DIR}/include" CACHE PATH "mkldnn include directory." FORCE)
IF(WIN32 OR APPLE)
MESSAGE(WARNING
MESSAGE(WARNING
"Windows or Mac is not supported with MKLDNN in Paddle yet."
"Force WITH_MKLDNN=OFF")
SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in Windows and MacOS" FORCE)
......@@ -54,11 +54,13 @@ ExternalProject_Add(
${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS ${MKLDNN_DEPENDS}
GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git"
GIT_TAG "a29d8487a63afca3d5b8c5bbdbb473cf8ccc6e51"
GIT_TAG "64e03a1939e0d526aa8e9f2e3f7dc0ad8d372944"
PREFIX ${MKLDNN_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
CMAKE_ARGS -DMKLROOT=${MKLML_ROOT}
CMAKE_ARGS -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
CMAKE_ARGS -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
......
cmake/external/openblas.cmake
浏览文件 @ 7116cc12
......@@ -17,20 +17,29 @@ IF(USE_EIGEN_FOR_BLAS)
ENDIF(USE_EIGEN_FOR_BLAS)
INCLUDE(cblas)
# IF(WIN32 AND NOT ${CBLAS_FOUND})
IF(NOT ${CBLAS_FOUND})
INCLUDE(ExternalProject)
SET(CBLAS_SOURCES_DIR ${THIRD_PARTY_PATH}/openblas)
SET(CBLAS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/openblas)
SET(CBLAS_INC_DIR "${CBLAS_INSTALL_DIR}/include" CACHE PATH "openblas include directory." FORCE)
SET(CBLAS_INCLUDE_DIR "${CBLAS_INSTALL_DIR}/include" CACHE PATH "openblas include directory." FORCE)
SET(CBLAS_LIBRARIES
"${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
CACHE FILEPATH "openblas library." FORCE)
ADD_DEFINITIONS(-DPADDLE_USE_OPENBLAS)
IF (WIN32)
SET(CBLAS_FOUND true)
MESSAGE(WARNING, "In windows, openblas only support msvc build, please build it manually and put it at " ${CBLAS_INSTALL_DIR})
ENDIF(WIN32)
IF (NOT WIN32)
SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable")
SET(OPENBLAS_COMMIT "v0.2.20")
......@@ -69,7 +78,6 @@ IF(NOT ${CBLAS_FOUND})
ENDIF()
SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs)
ExternalProject_Add(
extern_openblas
${EXTERNAL_PROJECT_LOG_ARGS}
......@@ -84,9 +92,11 @@ IF(NOT ${CBLAS_FOUND})
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
)
ELSE()
ENDIF(NOT WIN32)
SET(CBLAS_PROVIDER openblas)
IF(WITH_C_API)
INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
INSTALL(DIRECTORY ${CBLAS_INCLUDE_DIR} DESTINATION third_party/openblas)
# Because libopenblas.a is a symbolic link of another library, thus need to
# install the whole directory.
IF(ANDROID)
......@@ -107,7 +117,8 @@ IF(NOT ${CBLAS_FOUND})
ENDIF(NOT ${CBLAS_FOUND})
MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}")
INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
MESSAGE(STATUS "BLAS Include: ${CBLAS_INCLUDE_DIR}")
INCLUDE_DIRECTORIES(${CBLAS_INCLUDE_DIR})
# FIXME(gangliao): generate cblas target to track all high performance
# linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
......@@ -121,6 +132,11 @@ ELSE()
TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES})
ENDIF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
IF(WITH_LIBXSMM)
TARGET_LINK_LIBRARIES(cblas ${LIBXSMM_LIBS})
ADD_DEPENDENCIES(cblas extern_libxsmm)
ENDIF()
IF(NOT ${CBLAS_FOUND})
ADD_DEPENDENCIES(cblas extern_openblas)
LIST(APPEND external_project_dependencies cblas)
......
cmake/external/protobuf.cmake
浏览文件 @ 7116cc12
......@@ -14,11 +14,14 @@
INCLUDE(ExternalProject)
# Always invoke `FIND_PACKAGE(Protobuf)` for importing function protobuf_generate_cpp
IF(NOT WIN32)
FIND_PACKAGE(Protobuf QUIET)
ENDIF(NOT WIN32)
macro(UNSET_VAR VAR_NAME)
UNSET(${VAR_NAME} CACHE)
UNSET(${VAR_NAME})
endmacro()
UNSET_VAR(PROTOBUF_INCLUDE_DIR)
UNSET_VAR(PROTOBUF_FOUND)
UNSET_VAR(PROTOBUF_PROTOC_EXECUTABLE)
......@@ -94,12 +97,14 @@ macro(PROMPT_PROTOBUF_LIB)
SET(protobuf_DEPS ${ARGN})
MESSAGE(STATUS "Protobuf protoc executable: ${PROTOBUF_PROTOC_EXECUTABLE}")
MESSAGE(STATUS "Protobuf-lite library: ${PROTOBUF_LITE_LIBRARY}")
MESSAGE(STATUS "Protobuf library: ${PROTOBUF_LIBRARY}")
MESSAGE(STATUS "Protoc library: ${PROTOBUF_PROTOC_LIBRARY}")
MESSAGE(STATUS "Protobuf version: ${PROTOBUF_VERSION}")
INCLUDE_DIRECTORIES(${PROTOBUF_INCLUDE_DIR})
# Assuming that all the protobuf libraries are of the same type.
IF(${PROTOBUF_LIBRARY} MATCHES "${CMAKE_STATIC_LIBRARY_SUFFIX}$")
IF(${PROTOBUF_LIBRARY} MATCHES ${CMAKE_STATIC_LIBRARY_SUFFIX})
SET(protobuf_LIBTYPE STATIC)
ELSEIF(${PROTOBUF_LIBRARY} MATCHES "${CMAKE_SHARED_LIBRARY_SUFFIX}$")
SET(protobuf_LIBTYPE SHARED)
......@@ -137,18 +142,25 @@ macro(SET_PROTOBUF_VERSION)
endmacro()
set(PROTOBUF_ROOT "" CACHE PATH "Folder contains protobuf")
IF (WIN32)
SET(PROTOBUF_ROOT ${THIRD_PARTY_PATH}/install/protobuf)
MESSAGE(WARNING, "In windows, protobuf only support msvc build, please build it manually and put it at " ${PROTOBUF_ROOT})
ENDIF(WIN32)
if (NOT "${PROTOBUF_ROOT}" STREQUAL "")
find_path(PROTOBUF_INCLUDE_DIR google/protobuf/message.h PATHS ${PROTOBUF_ROOT}/include NO_DEFAULT_PATH)
find_library(PROTOBUF_LIBRARY protobuf PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
find_library(PROTOBUF_LITE_LIBRARY protobuf-lite PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
find_library(PROTOBUF_PROTOC_LIBRARY protoc PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
find_library(PROTOBUF_LIBRARY protobuf libprotobuf.lib PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
find_library(PROTOBUF_LITE_LIBRARY protobuf-lite libprotobuf-lite.lib PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
find_library(PROTOBUF_PROTOC_LIBRARY protoc libprotoc.lib PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
find_program(PROTOBUF_PROTOC_EXECUTABLE protoc PATHS ${PROTOBUF_ROOT}/bin NO_DEFAULT_PATH)
if (PROTOBUF_INCLUDE_DIR AND PROTOBUF_LIBRARY AND PROTOBUF_LITE_LIBRARY AND PROTOBUF_PROTOC_LIBRARY AND PROTOBUF_PROTOC_EXECUTABLE)
message(STATUS "Using custom protobuf library in ${PROTOBUF_ROOT}.")
SET(PROTOBUF_FOUND true)
SET_PROTOBUF_VERSION()
PROMPT_PROTOBUF_LIB()
else()
message(WARNING "Cannot find protobuf library in ${PROTOBUF_ROOT}.")
message(WARNING "Cannot find protobuf library in ${PROTOBUF_ROOT}")
endif()
endif()
......@@ -239,6 +251,7 @@ IF(CMAKE_CROSSCOMPILING)
CACHE FILEPATH "protobuf executable." FORCE)
ENDIF()
IF(NOT PROTOBUF_FOUND)
build_protobuf(extern_protobuf FALSE)
......
cmake/external/python.cmake
浏览文件 @ 7116cc12
......@@ -18,8 +18,9 @@ ENDIF()
INCLUDE(python_module)
FIND_PACKAGE(PythonInterp 2.7)
FIND_PACKAGE(PythonLibs 2.7)
FIND_PACKAGE(PythonInterp ${PY_VERSION})
FIND_PACKAGE(PythonLibs ${PY_VERSION})
# Fixme: Maybe find a static library. Get SHARED/STATIC by FIND_PACKAGE.
ADD_LIBRARY(python SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET python PROPERTY IMPORTED_LOCATION ${PYTHON_LIBRARIES})
......
cmake/external/xbyak.cmake 0 → 100644
浏览文件 @ 7116cc12
# 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.
set(WITH_XBYAK ON)
if(WIN32 OR APPLE)
SET(WITH_XBYAK OFF CACHE STRING "Disable XBYAK in Windows and MacOS" FORCE)
return()
endif()
include(ExternalProject)
set(XBYAK_PROJECT extern_xbyak)
set(XBYAK_PREFIX_DIR ${THIRD_PARTY_PATH}/xbyak)
set(XBYAK_INSTALL_ROOT ${THIRD_PARTY_PATH}/install/xbyak)
set(XBYAK_INC_DIR ${XBYAK_INSTALL_ROOT}/include)
include_directories(${XBYAK_INC_DIR})
include_directories(${XBYAK_INC_DIR}/xbyak)
add_definitions(-DPADDLE_WITH_XBYAK)
# xbyak options
add_definitions(-DXBYAK64)
add_definitions(-DXBYAK_NO_OP_NAMES)
ExternalProject_Add(
${XBYAK_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS ""
GIT_REPOSITORY "https://github.com/herumi/xbyak.git"
GIT_TAG "v5.661" # Jul 26th
PREFIX ${XBYAK_PREFIX_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${XBYAK_INSTALL_ROOT}
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${XBYAK_INSTALL_ROOT}
)
if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/xbyak_dummy.c)
file(WRITE ${dummyfile} "const char *dummy_xbyak = \"${dummyfile}\";")
add_library(xbyak STATIC ${dummyfile})
else()
add_library(xbyak INTERFACE)
endif()
add_dependencies(xbyak ${XBYAK_PROJECT})
list(APPEND external_project_dependencies xbyak)
cmake/flags.cmake
浏览文件 @ 7116cc12
......@@ -102,7 +102,6 @@ set(COMMON_FLAGS
-fno-omit-frame-pointer
-Wall
-Wextra
-Werror
-Wnon-virtual-dtor
-Wdelete-non-virtual-dtor
-Wno-unused-parameter
......@@ -115,6 +114,11 @@ set(COMMON_FLAGS
-Wno-error=terminate # Warning in PADDLE_ENFORCE
)
# https://github.com/PaddlePaddle/Paddle/issues/12773
if (NOT WIN32)
list(APPEND COMMON_FLAGS -Werror)
endif()
set(GPU_COMMON_FLAGS
-fPIC
-fno-omit-frame-pointer
......@@ -142,6 +146,11 @@ else()
${GPU_COMMON_FLAGS})
endif()
if(UNIX AND NOT APPLE)
# except apple from nix*Os family
set(LINUX TRUE)
endif(UNIX AND NOT APPLE)
foreach(flag ${COMMON_FLAGS})
safe_set_cflag(CMAKE_C_FLAGS ${flag})
......
cmake/generic.cmake
浏览文件 @ 7116cc12
......@@ -148,7 +148,8 @@ function(merge_static_libs TARGET_NAME)
COMMAND rm "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a"
COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles}
)
else() # general UNIX: use "ar" to extract objects and re-add to a common lib
endif(APPLE)
if(LINUX) # general UNIX: use "ar" to extract objects and re-add to a common lib
set(target_DIR ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}.dir)
foreach(lib ${libs})
......@@ -187,7 +188,36 @@ function(merge_static_libs TARGET_NAME)
COMMAND ${CMAKE_AR} crs ${target_LIBNAME} `find ${target_DIR} -name '*.o'`
COMMAND ${CMAKE_RANLIB} ${target_LIBNAME}
WORKING_DIRECTORY ${target_DIR})
endif()
endif(LINUX)
if(WIN32) # windows do not support gcc/nvcc combined compiling. Use msvc lib.exe to merge libs.
# Make the generated dummy source file depended on all static input
# libs. If input lib changes,the source file is touched
# which causes the desired effect (relink).
add_custom_command(OUTPUT ${target_SRCS}
COMMAND ${CMAKE_COMMAND} -E touch ${target_SRCS}
DEPENDS ${libs})
# Generate dummy staic lib
file(WRITE ${target_SRCS} "const char *dummy_${TARGET_NAME} = \"${target_SRCS}\";")
add_library(${TARGET_NAME} STATIC ${target_SRCS})
target_link_libraries(${TARGET_NAME} ${libs_deps})
foreach(lib ${libs})
# Get the file names of the libraries to be merged
#if(NOT $<TARGET_FILE:${lib}> MATCHES "lib.*\\.lib")
# message("library" ${lib})
# set(libfiles ${libfiles} lib$<TARGET_FILE:${lib}>)
#else()
set(libfiles ${libfiles} $<TARGET_FILE:${lib}>)
#endif()
endforeach()
# windows cmd return error in clean env.
# COMMAND del "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}/${TARGET_NAME}.lib"
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
COMMAND lib /OUT:${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.lib ${libfiles}
)
endif(WIN32)
endfunction(merge_static_libs)
function(cc_library TARGET_NAME)
......@@ -195,6 +225,10 @@ function(cc_library TARGET_NAME)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(cc_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
if(WIN32)
# add libxxx.lib prefix in windows
set(${TARGET_NAME}_LIB_NAME "${CMAKE_STATIC_LIBRARY_PREFIX}${TARGET_NAME}${CMAKE_STATIC_LIBRARY_SUFFIX}" CACHE STRING "output library name for target ${TARGET_NAME}")
endif(WIN32)
if(cc_library_SRCS)
if(cc_library_SHARED OR cc_library_shared) # build *.so
add_library(${TARGET_NAME} SHARED ${cc_library_SRCS})
......@@ -263,8 +297,11 @@ function(cc_test TARGET_NAME)
COMMAND ${TARGET_NAME} ${cc_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
if (${cc_test_SERIAL})
set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
endif()
endif()
endfunction(cc_test)
......@@ -328,8 +365,11 @@ function(nv_test TARGET_NAME)
add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
add_test(${TARGET_NAME} ${TARGET_NAME})
if (nv_test_SERIAL)
set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
endif()
endif()
endfunction(nv_test)
......@@ -577,7 +617,9 @@ function(py_test TARGET_NAME)
set(multiValueArgs SRCS DEPS ARGS ENVS)
cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
add_test(NAME ${TARGET_NAME}
COMMAND env FLAGS_init_allocated_mem=true PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
COMMAND env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
FLAGS_cpu_deterministic=true
PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
endif()
......
cmake/inference_lib.cmake
浏览文件 @ 7116cc12
......@@ -101,6 +101,7 @@ if(WITH_MKLDNN)
)
endif()
if (NOT WIN32)
if(NOT MOBILE_INFERENCE AND NOT RPI)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
copy(snappy_lib
......@@ -120,14 +121,19 @@ if(NOT MOBILE_INFERENCE AND NOT RPI)
DSTS ${dst_dir} ${dst_dir}/lib
DEPS zlib)
endif()
endif(NOT WIN32)
# paddle fluid module
set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
set(dst_dir "${FLUID_INSTALL_DIR}/paddle/fluid")
set(module "framework")
copy(framework_lib DEPS framework_py_proto
if (NOT WIN32)
set(framework_lib_deps framework_py_proto)
endif(NOT WIN32)
copy(framework_lib DEPS ${framework_lib_deps}
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module}
${src_dir}/${module}/ir/*.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module} ${dst_dir}/${module}/ir
)
set(module "memory")
......@@ -138,29 +144,22 @@ copy(memory_lib
set(inference_deps paddle_fluid_shared paddle_fluid)
if(WITH_CONTRIB)
message(STATUS "installing contrib")
set(contrib_dst_dir "${FLUID_INSTALL_DIR}/contrib/inference")
if (WITH_ANAKIN AND WITH_GPU)
copy(contrib_anakin_inference_lib DEPS paddle_inference_api inference_anakin_api
SRCS
${PADDLE_BINARY_DIR}/paddle/contrib/inference/libinference_anakin_api* # compiled anakin api
${PADDLE_BINARY_DIR}/third_party/install/anakin/*.tar.gz # anakin release
DSTS ${contrib_dst_dir}/anakin ${contrib_dst_dir}/anakin)
list(APPEND inference_deps contrib_anakin_inference_lib)
endif()
copy(contrib_inference_lib DEPS paddle_inference_api paddle_inference_api_shared
SRCS ${PADDLE_SOURCE_DIR}/paddle/contrib/inference/paddle_inference_api.h
${PADDLE_BINARY_DIR}/paddle/contrib/inference/libpaddle_inference_api*
DSTS ${contrib_dst_dir} ${contrib_dst_dir})
list(APPEND inference_deps contrib_inference_lib)
set(module "inference/api")
if (WITH_ANAKIN AND WITH_MKL)
copy(anakin_inference_lib DEPS paddle_inference_api inference_anakin_api
SRCS
${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libinference_anakin_api* # compiled anakin api
${ANAKIN_INSTALL_DIR} # anakin release
DSTS ${dst_dir}/inference/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin)
list(APPEND inference_deps anakin_inference_lib)
endif()
set(module "inference")
copy(inference_lib DEPS ${inference_deps}
SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
DSTS ${dst_dir}/${module} ${dst_dir}/${module}
${src_dir}/${module}/api/paddle_inference_api.h ${src_dir}/${module}/api/demo_ci
${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
)
set(module "platform")
......
cmake/simd.cmake
浏览文件 @ 7116cc12
......@@ -10,6 +10,7 @@ if(CMAKE_COMPILER_IS_GNUCC OR CMAKE_COMPILER_IS_GNUCXX OR CMAKE_CXX_COMPILER_ID
set(SSE3_FLAG "-msse3")
set(AVX_FLAG "-mavx")
set(AVX2_FLAG "-mavx2")
set(AVX512F_FLAG "-mavx512f")
elseif(MSVC)
set(MMX_FLAG "/arch:MMX")
set(SSE2_FLAG "/arch:SSE2")
......@@ -81,5 +82,16 @@ int main()
return 0;
}" AVX2_FOUND)
# Check AVX512F
set(CMAKE_REQUIRED_FLAGS ${AVX512F_FLAG})
set(AVX512F_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
CHECK_CXX_SOURCE_RUNS("
#include <immintrin.h>
int main()
{
__m512i a = _mm512_undefined_epi32();
return 0;
}" AVX512F_FOUND)
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_RETAINED})
mark_as_advanced(MMX_FOUND SSE2_FOUND SSE3_FOUND AVX_FOUND AVX2_FOUND)
mark_as_advanced(MMX_FOUND SSE2_FOUND SSE3_FOUND AVX_FOUND AVX2_FOUND AVX512F_FOUND)
cmake/tensorrt.cmake
浏览文件 @ 7116cc12
......@@ -16,7 +16,9 @@ find_library(TENSORRT_LIBRARY NAMES libnvinfer.so libnvinfer.a
DOC "Path to TensorRT library.")
if(TENSORRT_INCLUDE_DIR AND TENSORRT_LIBRARY)
if(WITH_DSO)
set(TENSORRT_FOUND ON)
endif(WITH DSO)
else()
set(TENSORRT_FOUND OFF)
endif()
......
cmake/version.cmake
浏览文件 @ 7116cc12
# Get the latest git tag.
set(PADDLE_VERSION $ENV{PADDLE_VERSION})
set(tmp_version "HEAD")
set(TAG_VERSION_REGEX "[0-9]+\\.[0-9]+\\.[0-9]+(\\.(a|b|rc)\\.[0-9]+)?")
set(COMMIT_VERSION_REGEX "[0-9a-f]+[0-9a-f]+[0-9a-f]+[0-9a-f]+[0-9a-f]+")
while ("${PADDLE_VERSION}" STREQUAL "")
# Check current branch name
execute_process(
COMMAND ${GIT_EXECUTABLE} describe --tags --abbrev=0 ${tmp_version}
COMMAND ${GIT_EXECUTABLE} rev-parse --abbrev-ref ${tmp_version}
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
OUTPUT_VARIABLE GIT_TAG_NAME
RESULT_VARIABLE GIT_RESULT
OUTPUT_VARIABLE GIT_BRANCH_NAME
RESULT_VARIABLE GIT_BRANCH_RESULT
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
if (NOT ${GIT_RESULT})
# Check the tag is a correct version
if (${GIT_TAG_NAME} MATCHES "v[0-9]+\\.[0-9]+\\.[0-9]+(\\.(a|b|rc)\\.[0-9]+)?")
string(REPLACE "v" "" PADDLE_VERSION ${GIT_TAG_NAME})
else() # otherwise, get the previous git tag name.
set(tmp_version "${GIT_TAG_NAME}~1")
if (NOT ${GIT_BRANCH_RESULT})
execute_process(
COMMAND ${GIT_EXECUTABLE} describe --tags --abbrev=0 --always ${tmp_version}
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
OUTPUT_VARIABLE GIT_TAG_NAME
RESULT_VARIABLE GIT_RESULT
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
if (NOT ${GIT_RESULT})
# Check if current branch is release branch
if (${GIT_BRANCH_NAME} MATCHES "release/${TAG_VERSION_REGEX}")
# Check the tag is a correct version
if (${GIT_TAG_NAME} MATCHES "${COMMIT_VERSION_REGEX}")
# if no tag was found, set PADDLE_VERSION to 0.0.0 to represent latest
set(PADDLE_VERSION "0.0.0")
elseif (${GIT_TAG_NAME} MATCHES "v${TAG_VERSION_REGEX}")
string(REPLACE "v" "" PADDLE_VERSION ${GIT_TAG_NAME})
else() # otherwise, get the previous git tag name.
set(tmp_version "${GIT_TAG_NAME}~1")
endif()
else()
# otherwise, we always set PADDLE_VERSION to 0.0.0 to represent latest
set(PADDLE_VERSION "0.0.0")
endif()
else()
set(PADDLE_VERSION "0.0.0")
message(WARNING "Cannot add paddle version from git tag")
endif()
else()
set(PADDLE_VERSION "0.0.0")
message(WARNING "Cannot add paddle version from git tag")
message(WARNING "Cannot add paddle version for wrong git branch result")
endif()
endwhile()
......
doc/fluid/api/executor.rst
浏览文件 @ 7116cc12
......@@ -38,11 +38,3 @@ _switch_scope
.. autofunction:: paddle.fluid.executor._switch_scope
:noindex:
.. _api_fluid_executor_fetch_var:
fetch_var
---------
.. autofunction:: paddle.fluid.executor.fetch_var
:noindex:
doc/fluid/api/fluid.rst
浏览文件 @ 7116cc12
......@@ -106,22 +106,6 @@ _switch_scope
.. autofunction:: paddle.fluid._switch_scope
:noindex:
.. _api_fluid_fetch_var:
fetch_var
---------
.. autofunction:: paddle.fluid.fetch_var
:noindex:
.. _api_fluid_Go:
Go
--
.. autoclass:: paddle.fluid.Go
:members:
:noindex:
.. _api_fluid_make_channel:
......
doc/fluid/api/layers.rst
浏览文件 @ 7116cc12
......@@ -822,6 +822,14 @@ pad
.. autofunction:: paddle.fluid.layers.pad
:noindex:
.. _api_fluid_layers_pad_constant_like:
pad_constant_like
---
.. autofunction:: paddle.fluid.layers.pad_constant_like
:noindex:
.. _api_fluid_layers_label_smooth:
label_smooth
......@@ -1145,6 +1153,14 @@ sigmoid
.. autofunction:: paddle.fluid.layers.sigmoid
:noindex:
.. _api_fluid_layers_hsigmoid:
hsigmoid
-------
.. autofunction:: paddle.fluid.layers.hsigmoid
:noindex:
.. _api_fluid_layers_logsigmoid:
logsigmoid
......@@ -1768,3 +1784,11 @@ reverse
.. autofunction:: paddle.fluid.layers.reverse
:noindex:
.. _api_fluid_layers_rank_loss:
rank_loss
-------
.. autofunction:: paddle.fluid.layers.rank_loss
:noindex:
doc/fluid/design/dist_train/dist_train_nccl2.md
浏览文件 @ 7116cc12
# Distributed Training with NCCL2
We design a pattern that can enable training with `ParallelExecutor` and
using [NCCL2](https://developer.nvidia.com/nccl) as it's collective
use [NCCL2](https://developer.nvidia.com/nccl) as it's collective
communication library.
In `ParallelExecutor` we can use `AllReduce` or `Reduce` and `Broadcast`
......@@ -9,14 +9,14 @@ to do multi GPU training. And if we initialize NCCL2 communicators as
ranks in a distributed environment, we can simply run the `ParallelExecutor`
as a distributed program! The only thing that may be different than in
the single node version is that we need to broadcast the NCCL unique ID
to all the nodes, and initialize communicators using that ID, so NCCL2
will know each other as ranks.
to all the nodes and initialize communicators using that ID, so NCCL2
can know each other as ranks.
To achieve this feature, we introduce a new operator: `gen_nccl_id` op,
so we are ***not*** "bind to" running NCCL2 with MPI, we can run it in
what ever platform you like.
whatever platform you like.
It have two running modes:
It has two running modes:
1. Generate and broadcast mode, which should be used on trainer 0;
1. Listen and fetch mode, which should be used on trainers other than 0.
......@@ -29,7 +29,7 @@ initialize NCCL communicator objects.
<img src="src/ncc2_design.png">
The above figure indicates the general process when training with NCCL2
distributed. Each trainer have the number of communicators equal to the
distributed. Each trainer has the number of communicators equal to the
number of GPUs, but the ranks should match the global ranks number: here
we have total 8 GPUs, so `nranks==8`, for each trainer, the ranks should
be from 0 ~ 3 on trainer 0 and 4 ~ 7 on trainer 1.
doc/fluid/design/dist_train/distributed_lookup_table_design.md
浏览文件 @ 7116cc12
# Design Doc: Distributed Lookup Table Operator
A lookup table operator in PaddlePaddle where the table could be out
A distribute lookup table operator in PaddlePaddle where the table could be out
of the memory of a computer.
## Background
......@@ -24,14 +24,14 @@ memory, so we'd need a distributed storage service, which supports the
lookup of rows.
The following figure illustrates the multiplication of x with two
non-zero elements, or say, two symbols, and a lookup table W:
non-zero elements, or say two symbols, and a lookup table W:
![lookup table](./src/lookup_table.png)
### The Backward Algorithm
The backward algorithm computes W'(x) using W(x). W'(x) has the same
scale of size as W(x) and is much smaller than W.
the scale of size as W(x) and is much smaller than W.
To optimize W given W', we can do simple SGD update:
......@@ -44,111 +44,46 @@ $$W = f(W, W')$$
The following figure illustrates the backward pass of the lookup
operator: ![lookup table training](./src/lookup_table_training.png)
## Distributed Storage Service
The forward algorithm requires a distributed storage service for W.
The backward algorithm prefers that the storage system can apply the
optimization algorithm on W. The following two sections describe two
solutions -- the former doesn't require that the storage service can
do optimization, the latter does.
### Storage Service Doesn't Optimize
In this design, we use highly-optimized distributed storage, e.g.,
memcached, as the storage service, and we run the optimization
algorithm on parameter servers of PaddlePaddle. The following figure
illustrates the training process.
<!--
Note: please update the following URL when update this digraph.
<img src='https://g.gravizo.com/svg?
digraph G {
rankdir="LR";
subgraph cluster1 {
P1 [label="pserver 1"];
P2 [label="pserver 2"];
T1 [label="trainer 1"];
T2 [label="trainer 2"];
T3 [label="trainer 3"];
}
KV [label="memcached"];
T1 -> P1;
T1 -> P2;
T2 -> P1;
T2 -> P2;
T3 -> P1;
T3 -> P2;
P1 -> KV [color=gray, weight=0.1];
KV -> P1 [color=gray, weight=0.1];
P2 -> KV [color=gray, weight=0.1];
KV -> P2 [color=gray, weight=0.1];
KV -> T1 [color=gray, weight=0.1];
KV -> T2 [color=gray, weight=0.1];
KV -> T3 [color=gray, weight=0.1];
}
)
'/>
-->
<img src='https://g.gravizo.com/svg?%20digraph%20G%20{%20rankdir=%22LR%22;%20subgraph%20cluster1%20{%20P1%20[label=%22pserver%201%22];%20P2%20[label=%22pserver%202%22];%20T1%20[label=%22trainer%201%22];%20T2%20[label=%22trainer%202%22];%20T3%20[label=%22trainer%203%22];%20}%20KV%20[label=%22memcached%22];%20T1%20-%3E%20P1;%20T1%20-%3E%20P2;%20T2%20-%3E%20P1;%20T2%20-%3E%20P2;%20T3%20-%3E%20P1;%20T3%20-%3E%20P2;%20P1%20-%3E%20KV%20[color=gray,%20weight=0.1];%20KV%20-%3E%20P1%20[color=gray,%20weight=0.1];%20P2%20-%3E%20KV%20[color=gray,%20weight=0.1];%20KV%20-%3E%20P2%20[color=gray,%20weight=0.1];%20KV%20-%3E%20T1%20[color=gray,%20weight=0.1];%20KV%20-%3E%20T2%20[color=gray,%20weight=0.1];%20KV%20-%3E%20T3%20[color=gray,%20weight=0.1];%20}'/>
Each trainer runs the forward and backward passes using their local
data:
1. In the forward pass, when a trainer runs the forward algorithm of a
lookup operator, it retrieves W(x) from the storage service.
1. The trainer computes W'(x) in the backward pass using W(x).
During the global update process:
1. Each trainer uploads its W'(x) to parameter servers.
1. The parameter server runs the optimization algorithm, e.g., the
Adam optimization algorithm, which requires that
1. The parameter server retrieves W(x) from memcached, and
1. The parameter server pushes $\Delta W(x)=f(W(x), lambda \sum_j
W'(x))$ to memcached, where $f$ denotes the optimization
algorithm.
### Storage Service Does Optimize
This design is very similar to the above one, except that the
optimization algorithm $f$ runs on the storage service.
- Pro: parameter servers do not retrieve W(x) from the storage
service, thus saves half network communication.
- Con: the storage service needs to be able to run the optimization
algorithm.
## Distributed Sparse Table in Fluid
For another design, we can implement a distributed sparse table in Fluid,
and don't need to maintain an external storage component while training.
You may need to read Fluid [Distributed Training Architecture](./distributed_architecture.md)
and [Parameter Server](./parameter_server.md) before going on.
![fluid lookup remote table](./src/fluid_lookup_remote_table.png)
Partition a large table into multiple pserver instances
1. `DistributeTranspiler` would split the table partitioned into some small
table blocks with some partitioned algorithms such as
[RoundRobin](https://en.wikipedia.org/wiki/Round-robin_scheduling),
[Hash](https://en.wikipedia.org/wiki/Hash) and etc...
1. For some cases, the range of input `Ids` is very wide and unpredictable, so the sparse
table would be able to fill a new value for the id that didn't appear before with
zero, uniform random or Gaussian distribution.
For each Trainer's training process:
1. In the forward pass, we use `pre-fetch` op to pre-fetch parameter blocks according to the
input `Ids` from PServers instead of the local `lookup_table` op, and then merge the blocks
into a parameter `W`.
1. Compute `GRAD@W'` in the backward pass using the pre-fetched `W` and send it to PServer to
execute the optimize pass.
## Conclusion
Let us do the "storage service does not optimize" solution first, as a
baseline at least, because it is easier to use a well-optimized
distributed storage service like memcached. We can do the "storage
service does optimize" solution later or at the same time, which, if
implemented carefully, should have better performance than the former.
## Distributed Lookup Table
### Problem 1: The lookup table may be very large.
In the condition like the search engine and recommendation system, the number of feature Id may be very large, say 100,000,000,000, then for a float value lookup table of size 8, the total size of the table is:
```
100,000,000,000 * 8 * 4(Bytes) = 2980.23 GB
```
### Solution: Distributed storage
1. Paddle use [SelectedRows](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/selected_rows.md) as the storage format for the lookup table, the lookup table parameter will be split to multi-machine according to the hash of the feature ID, and data will also be split and send to the same machine to prefetch the parameter.
1. For common parameters, the trainer will get the whole parameter for training, but for the big lookup table, the trainer can not store the whole parameter. Because the input data feature is very sparse, every time we only need a few parameters for training, so we use `prefetch_op` to only prefetch the parameter needed to trainer.
### Problem 2. The Id in the lookup table is not sure before training.
The feature Id is calculated by the hash function because the feature data source is so large, we can not get all the Id before training. So we can not initialize the table before training.
### Solution: Id auto growth
At the beginning of training, paddle only malloc the memory for the lookup table at parameter server side, the Id and it's value will not be initialized. During training, when a parameter server received an Id, if it is already in the lookup table, it will return the existing parameter, if the Id does not exist, paddle will add it into the lookup table and initialize the value for it.
### Problem 3: parameter load and save
For common parameters, paddle use trainer to save and load them. But for distributed lookup table, trainer cannot do this because it's large size.
### Solution: Parameter server side save and load
Paddle support parameter server side save and load for distribute lookup table. Each machine of parameter servers will only save and load part of the whole table.
## Architecture
The whole architecture of the distribute lookup table is as below:
### Training steps:
1. Read a batch of data, the data is feature ids.
1. The input ids will be split by `split_ids_op` with the same hash function of the lookup table.
1. The `prefetch_op` use the split result to prefetch parameters back from the lookup table.
1. Run forward-backward to get the gradient of the lookup table.
1. `split_ids_op` split the gradient and then use `send_op` to the parameter server.
1. parameter server update the table with the received gradient.
![distribute lookup table](./src/distributed_lookup_table.jpeg)
doc/fluid/design/ir/overview.md 0 → 100644
浏览文件 @ 7116cc12
## Motivation
There is a `gap` between the `Program` defined by
user and the `Executable` that can be scheduled
efficiently on heterogeneous hardware, either locally
or distributedly.
Usually, the `gap` is bridged by
* A serious transformations with defined order.
* These transformations usually involve
`insert, delete, clustering, split, dependency analysis`.
* Has a simple way to verify and debug each transformation.
* Flexible to add, remove or customize transformations to fit
the requirements of various algorithms (models) and hardware secenarios.
Some other events also push us to a better unified pattern.
* The deep learning framework is built around the concepts of graphs.
To leverage tools such as compilation (e.g. TVM and nGraph) or
cross-framework conversion (e.g. ONNX), we also need a intermediate
representation that can be connected to the rest of the ecosystem.
We need a unified pattern to naturally support the requirements
described above. The pattern should fit both training, inference
and other offline serielized model transformations.
Learned from LLVM and other deep learning framework, we draft the
design below.
## Design
### Major Concepts
#### Node
`Node` represents an operation that performs some computation or
a variable that is input or output of operation.
`Node`s are connected to other `Node`s via inputs and outputs.
Other properties (maybe device placement information) can be added
to `Node` in the future if it's a
common requirement of many other `Pass`es. Otherwise, it should live
in a `Node` wrapper class that is private to some `Pass` or be
a local member of a `Pass`.
#### Graph
`Graph` contains a list of `Node`s, which are connected to
each other via inputs and outputs.
TODO: Better definitions for the graph.
`Graph` can also contain `Attribute`s. `Attribute`s
can be `any` thing. For example, it can be a list of "wraper"
nodes. The `wrapper` nodes compose `Node`s and provide
helper method for execution or transformation. `Attribute`
can also contain other things that describe some properties of
the `Graph` or `Graph` nodes. `Attribute` can be passed
across `Pass`. However, it should be used with care.
```cpp
class Graph {
public:
explicit Graph(const ProgramDesc &program);
bool Has(const std::string &attr_name) const;
template <typename AttrType>
AttrType &Get(const std::string &attr_name) const;
template <typename AttrType>
void Set(const std::string &attr_name, AttrType *attr);
const std::unordered_set<ir::Node *> &Nodes() const;
// Create a normal variable with non-null VarDesc.
ir::Node *CreateVarNode(VarDesc *var_desc);
// Create a normal runnable operator with OpDesc.
ir::Node *CreateOpNode(OpDesc *op_desc);
// Create a control dependency var that connects 2 operations. The
// var doesn't hold any data. Other than that, it's no different from
// other var, considering dependency analysis.
ir::Node *CreateControlDepVar();
// A more free style way of creating a graph node. Mostly use for test
// or "copy" from another node. Avoid using it if possible.
ir::Node *CreateEmptyNode(const std::string &name, ir::Node::Type type);
// Clear all node information of the graph and return the ownership of the
// nodes.
std::vector<std::unique_ptr<ir::Node>> ReleaseNodes();
};
```
#### Pass
`Pass` represents a transformation of `Graph`. Its input
is a `Graph` and its output is also a `Graph`. For example,
a `Pass` can simply print out the `Graph`. A `Pass`
can also fuse some `Graph`'s `Node`s.
```cpp
class Pass {
public:
std::unique_ptr<Graph> Apply(std::unique_ptr<Graph> graph) const {
// Some correctness check.
auto new_graph = ApplyImpl(std::move(graph));
// Some correctness check.
return new_graph;
}
// Get a reference to the attributed previously set.
template <typename AttrType>
AttrType &Get(const std::string &attr_name) const;
// Set a pointer to the attribute. Pass takes ownership of the attribute.
template <typename AttrType>
void Set(const std::string &attr_name, AttrType *attr) ;
// Set a pointer to the attribute. Pass doesn't take ownership. Caller
// should delete the attribute.
template <typename AttrType>
void SetNotOwned(const std::string &attr_name, AttrType *attr);
protected:
virtual std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const = 0;
};
// In my_pass.cc
class MyPass : public Pass {
protected:
std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const override {
// do something.
return graph;
}
}
REGISTER_PASS(my_pass, MyPass)
.RequirePassAttr("places")
.RequireGraphAttr("dep_vars");
// To use the pass.
auto my_pass = ir::PassRegistry::Instance().Get("my_pass");
graph = my_pass->Apply(std::move(graph));
// Note: to force link my_pass.cc, in the code:
USE_PASS(my_pass);
```
#### Optimize
`Optimize` contains a series of `Pass` with defined order.
`Optimize` transforms a `Graph` that only contains raw
modeling logic to a `Graph` that can be run efficiently while
maintaining the original modeling logic.
### Optimize Process
* Program is first converted to Graph.
* Graph goes through a series of Pass
* Graph is transformed from raw model logic to a
form that is efficient to execute.
```
// Program->ProgramToGraph->Graph->Pass1->Graph->Pass2->Graph->Pass3->Graph->Executor
auto graph = Graph(program);
graph = PassRegistry::Instance().Get("op_fuse_pass").Apply(std::move(grah));
// For more complex Pass, Optimize Process can provide Pass attributes.
auto mem_opt_pass = PassRegistry::Instance().Get("memory_optimization_pass");
mem_opt_pass.SetNotOwned<int>("optimize_level", 1);
mem_opt_pass->Apply(std::move(graph));
graph = PassRegistry::Instance().Get("multi_devices_pass").Apply(std::move(grah));
graph = PassRegistry::Instance().Get("multi_devices_check_pass").Apply(std::move(grah));
Executor exe;
exe.Run(graph);
```
doc/fluid/design/modules/python_api.md
浏览文件 @ 7116cc12
......@@ -98,13 +98,13 @@ class Block(objects):
def append_operator(self, ...):
self.ops.append(Operator(self, ...))
def prepend_operator(self, ...): # Parameter's ctor prepands initialize operators.
def _prepend_operator(self, ...): # Parameter's ctor prepands initialize operators.
self.ops.prepend(Operator(self, ...))
```
`create_parameter` is necessary because parameters are global variables, defined in the global block, but can be created in some sub-blocks. For example, an FC layer in the step block of an RNN operator.
`prepend_operator` is necessary because the constructor of `Parameter` needs to create the initialize (or load) operator of the parameter, and would like to put it in the *preamble* of the global block.
`_prepend_operator` is necessary because the constructor of `Parameter` needs to create the initialize (or load) operator of the parameter, and would like to put it in the *preamble* of the global block.
### Operator
......
doc/fluid/design/others/graph_survey.md
浏览文件 @ 7116cc12
......@@ -28,7 +28,7 @@ def get_symbol(num_classes=10, **kwargs):
Varible here is actually a Symbol. Every basic Symbol will correspond to one Node, and every Node has its own NodeAttr. There is a op field in NodeAttr class, when a Symbol represents Variable(often input data), the op field is null.
Varible here is actually a Symbol. Every basic Symbol will correspond to one Node, and every Node has its own AnyAttr. There is a op field in AnyAttr class, when a Symbol represents Variable(often input data), the op field is null.
Symbol contains a data member, std::vector<NodeEntry> outputs, and NodeEntry cantains a poniter to Node. We can follow the Node pointer to get all the Graph.
......
doc/fluid/design/quantization/fixed_point_quantization.md 0 → 100644
浏览文件 @ 7116cc12
Fixed-point quantization uses lower bits, for example, 2-bit, 3-bit or 8-bit fixed point to represent weights and activations, which usually are in singe-precision float-point with 32 bits. The fixed-point representation has advantages in reducing memory bandwidth, lowering power consumption and computational resources as well as the model storage requirements. It is especially important for the inference in embedded-device deployment.
According to some experiments, the apporach to quantize the model trained in float point directly works effectively on the large models, like the VGG model having many parameters. But the accuracy drops a lot for the small model. In order to improve the tradeoff between accuracy and latency, many quantized training apporaches are proposed.
This document is to design a quantized training framework on Fluid. The first part will introduce how to quantize, The second part will describe the quantized training framework. The last part will illustrate how to calculate the quantization scale.
### How to quantize
There are many ways to quantize the float value to fixed-point value. For example:
$$ r = min(max(x, a), b)$$
$$ s = \frac{b - a}{n - 1} $$
$$ q = \left \lfloor \frac{r - a}{s} \right \rceil $$
where, $x$ is the float value to be quantized, $[a, b]$ is the quantization range, $a$ is the minimum value and $b$ is the maximal value. $\left \lfloor \right \rceil$ denotes rounding to the nearest integer. If the quantization level is $k$, $n$ is $2^k$, for example, $k$ is 8 and $n$ is 256. $q$ is the quantized integer.
The quantization we applied is parameterized by the number of quantization levels and maximum absolute value:
$$ M = max(abs(x)) $$
$$ q = \left \lfloor \frac{x}{M} * (n - 1) \right \rceil $$
where, $x$ is the float value to be quantized, $M$ is maximum absolute value. $\left \lfloor \right \rceil$ denotes rounding to the nearest integer. For 8 bit quantization, $n=2^{8}=256$. $q$ is the quantized integer.
Wether the *min-max* quantization or *max-abs* quantization, they also can be represent:
$q = scale * r + b$
We call *min-max*, *max-abs* as the quantization arguments, also call them quantization scale or quantization range.
How to calculate the quantization scale (or maximum absolute value) for inference will be described in the last part.
### Training Framework
#### Forward pass
The forward pass is simulated quantization, see Figure 1.
The training framework is as following figure.
<p align="center">
<img src="quantization_forward.png" width="300" height="340"><br/>
Figure 1. Forward in training with simulated quantization.
</p>
- Firstly, both input and weight will be quantized to 8-bit integers.
- Second, do the multiplication (or convolution) operation with integers.
- Third, dequantize the multiplication (or convolution) results to 32-bit float point.
- Finally, do bias-addition in float type of 32 bit. Here, the bias is not quantized.
For general matrix multiplication (GEMM), quantize for $X$ and $W$:
$$ X_q = \left \lfloor \frac{X}{X_m} * (n - 1) \right \rceil $$
$$ W_q = \left \lfloor \frac{W}{W_m} * (n - 1) \right \rceil $$
Do GEMM:
$$ Y = X_q * W_q $$
Dequantize $Y$:
$$
\begin{align}
Y_{dq} &=\frac{Y}{(n - 1) * (n - 1)} * X_m * W_m \\\
&=\frac{X_q * W_q}{(n - 1) * (n - 1)} * X_m * W_m \\\
&=(\frac{X_q}{n - 1} * X_m) * (\frac{W_q}{n - 1} * W_m)
\end{align}
$$
From these formulas, dequantization also can be moved before GEMM, do dequantization for $Xq$ and $Wq$ at first, then do GEMM. The forward workflow in training is equivalent to following framework.
<p align="center">
<img src="quantization_equivalent_forward.png" width="300" height="330"><br/>
Figure 2. Equivalent forward in training with simulated quantization.
</p>
We use this equivalent workflow in the training. In our desigin, there is a quantization transpiler to insert the quantization operator and the de-quantization operator in the Fluid `ProgramDesc`. Since the outputs of quantization and de-quantization operator are still in floating point, they are called faked quantization and de-quantization operator. And the training framework is called simulated quantization.
#### Backward pass
See Figure 3. The gradients are calculated by dequantized weights and activations. All inputs and outputs are float point with 32-bit. And in the weight updating process, the gradients will be added to the original weight, not the quantized or dequantized weights.
<p align="center">
<img src="quantization_backward_and_optimization.png"><br/>
Figure 3. Backward and weight updating in training with simulated quantization.
</p>
So the quantization transipler will change some inputs of the corresponding backward operators.
### How to calculate quantization scale
There are two strategies to calculate quantization scale, we call them dynamic and static strategy. The dynamic strategy calculates the quantization scale value each iteration. The static strategy keeps the quantization scale for different inputs.
For weights, we apply the dynamic strategy in the training, that is to say, the quantization scale will be recalculated during each iteration until the traning is finished.
For activations, the quantization scales are estimated during training, then used in inference. There are several different ways to estimate them:
1. Calculate the mean of maximum absolute during a window.
2. Calculate the max of maximum absolute during a window.
3. Calculate the running mean of maximum absolute during a window, as follows:
$$ Vt = (1 - k) * V + k * V_{t-1} $$
where, $V$ is the maximum absolute value of current batch, $Vt$ is the running mean value. $k$ is a factor, such as 0.9.
doc/fluid/dev/new_op_cn.md
浏览文件 @ 7116cc12
......@@ -36,19 +36,19 @@
<tbody>
<tr>
<td>OpProtoMake定义 </td>
<td>`.cc`文件，Backward Op不需要定义OpProtoMake </td>
<td>.cc 文件，Backward Op不需要定义OpProtoMake </td>
</tr>
<tr>
<td>Op定义 </td>
<td> `.cc`文件</td>
<td> .cc 文件</td>
</tr>
<tr>
<td>Kernel实现 </td>
<td> CPU、CUDA共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，CUDA 实现在`.cu`文件中。</td>
<td> CPU、CUDA共享Kernel实现在.h 文件中，否则，CPU 实现在.cc 文件中，CUDA 实现在.cu 文件中。</td>
</tr>
<tr>
<td>注册Op </td>
<td> Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，CUDA实现在`.cu`文件中</td>
<td> Op注册实现在.cc 文件；Kernel注册CPU实现在.cc 文件中，CUDA实现在.cu 文件中</td>
</tr>
</tbody>
</table>
......@@ -119,10 +119,29 @@ $$Out = scale*X$$
这个例子有`AddAttr<AttrType>("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。
### 定义GradProtoMaker类
每个Op的必须有一个对应的GraProtoMaker，若未定制对应前向Op的GradProtoMaker，fluid提供了DefaultGradProtoMaker，默认注册会使用全部输入输出，包括Input, Output, Output@Grad等，使用不需要的变量的会造成显存浪费。
下面示例定义了ScaleOp的GradProtoMaker。
```cpp
class ScaleGradMaker : public framework::SingleGradOpDescMaker {
public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
std::unique_ptr<framework::OpDesc> Apply() const override {
auto *grad_op = new framework::OpDesc();
grad_op->SetType("scale");
grad_op->SetInput("X", OutputGrad("Out"));
grad_op->SetOutput("Out", InputGrad("X"));
grad_op->SetAttr("scale", GetAttr("scale"));
return std::unique_ptr<framework::OpDesc>(grad_op);
}
};
```
### 定义Operator类
下面的点实现了MulOp的定义：
下面实现了MulOp的定义：
```cpp
class MulOp : public framework::OperatorWithKernel {
......@@ -334,3 +353,83 @@ ctest -R test_mul_op
- 注册Op时的类型名，需要和该Op的名字一样。即不允许在`A_op.cc`里面，注册`REGISTER_OPERATOR(B, ...)`等，这将会导致单元测试出错。
- 如果Op没有实现CUDA Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
- 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。
### PADDLE_ENFORCE使用注意
实现Op时检查数据的合法性需要使用PADDLE_ENFORCE以及PADDLE_ENFORCE_EQ等宏定义，基本格式如下：
```
PADDLE_ENFORCE(表达式, 错误提示信息)
PADDLE_ENFORCE_EQ(比较对象A, 比较对象B, 错误提示信息)
```
如果表达式为真，或者比较对象A=B，则检查通过，否则会终止程序运行，向用户反馈相应的错误提示信息。
为了确保提示友好易懂，开发者需要注意其使用方法。
#### 总体原则
任何使用了PADDLE_ENFORCE与PADDLE_ENFORCE_**检查的地方，必须有详略得当的备注解释！**错误提示信息**不能为空！
#### 提示信息书写标准
1. [required] 哪里错了？为什么错了？
- 例如：`ValueError: Mismatched label shape`
2. [optional] 期望的输入是什么样的？实际的输入是怎样的？
- 例如：`Expected labels dimension=1. Received 4.`
3. [optional] 能否给出修改意见？
- 例如：`Suggested Fix:If your classifier expects one-hot encoding label,check your n_classes argument to the estimatorand/or the shape of your label.Otherwise, check the shape of your label.`
如果并非必要或者简洁的描述即可表达清楚以上要点，根据情况书写亦可。
##### FAQ 典型问题
1. 无报错信息或报错信息过于简单，不能给用户提供有效的提示！
问题示例1 ：未写提示信息
```
PADDLE_ENFORCE(ctx->HasInput("X"), "");
```
问题示例2 ：提示信息过于简单
```
PADDLE_ENFORCE(i != nullptr, "i must be set"); // i是什么？
```
2. 在报错信息中使用开发人员定义的变量缩写，不易理解！
问题示例：
```
PADDLE_ENFORCE(forward_pd != nullptr,
"Fail to find eltwise_fwd_pd in device context"); //eltwise_fwd_pd用户可能看不懂
```
3. OP内部调用非法接口：Op内部如果出现Output = ShareDataWith(Input)
问题示例：
```cpp
auto *out = ctx.Output<framework::LoDTensor>("Out");
auto *in = ctx.Input<framework::LoDTensor>("X");
out->ShareDataWith(*in);
```
Op内部如果出现Output = ShareDataWith(Input)，相当于operator图的中有一条隐藏边，连接了Input和Output，这条边无法在图分析中表达，引发基于图优化的错误。
4. OP实现的性能实践
调用了eigen的broadcast, chop等操作，性能会比手写cuda kernel差几倍以上。此时cpu的实现可以复用eigen，gpu实现可以实现cuda kernel.
#### OP InferShape检查提示信息特别说明
- 检查输入输出变量，请统一遵循以下格式
`Input(变量名) of OP名 operator should not be null.`
正确示例：
```
PADDLE_ENFORCE(ctx->HasInput("Input"),
"Input(Input) of LSTMP operator should not be null.");
```
- 反向Op的输入输出检查，要写明反向Op的名字
正确示例：
```
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of LoDResetGrad opreator should not be null.");
```
doc/fluid/dev/releasing_process_cn.md
浏览文件 @ 7116cc12
# PaddlePaddle发行规范
PaddlePaddle使用git-flow branching model做分支管理，使用[Semantic Versioning](http://semver.org/)标准表示PaddlePaddle版本号。
PaddlePaddle使用Trunk Based Development，使用[Semantic Versioning](http://semver.org/)标准表示PaddlePaddle版本号。
PaddlePaddle每次发新的版本，遵循以下流程:
1. 从`develop`分支派生出新的分支，分支名为`release/版本号`。例如，`release/0.10.0`
1. 将新分支的版本打上tag，tag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`，第二个为`0.10.0rc2`，依次类推。
1. 对这个版本的提交，做如下几个操作:
* 使用Regression Test List作为检查列表，测试本次release的正确性。
* 如果失败，记录下所有失败的例子，在这个`release/版本号`分支中，修复所有bug后，Patch号加一，到第二步
* 修改`python/setup.py.in`中的版本信息,并将`istaged`字段设为`True`。
* 将这个版本的python wheel包发布到pypi。
* 更新Docker镜像（参考后面的操作细节）。
1. 第三步完成后，将`release/版本号`分支合入master分支，将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。
1. 协同完成Release Note的书写。
2. 将新分支的版本打上tag，tag为`版本号rc-Patch号`。例如，第一个tag为`0.10.0-rc0`。
3. 新分支一般不接受新的feature和优化。QA在release分支上进行测试。研发基于最新的develop开发。
4. QA和研发发现的bug，在develop上修复验证后，cherry-pick修复到release分支。直到release分支相对稳定。
5. 如果有需要，在release分支最新代码上打上新的tag，比如`0.10.0-rc1`，让更多的用户加入测试。重复3-4步。
6. release分支稳定后，打上正式的release tag，比如`0.10.0`。
7. 将这个版本的python wheel包发布到pypi。
8. 更新Docker镜像（参考后面的操作细节）。
需要注意的是:
* `release/版本号`分支一旦建立，一般不允许再从`develop`分支合入`release/版本号`。这样保证`release/版本号`分支功能的封闭，方便测试人员测试PaddlePaddle的行为。
* 在`release/版本号`分支存在的时候，如果有bugfix的行为，需要将bugfix的分支同时merge到`master`, `develop`和`release/版本号`这三个分支。
* bug修复需要先在develop上进行，然后进入release分支。而不是直接在release分支上开发。
* release分支原则上只接受修复类的修改，不接受新feature。
## 发布wheel包到pypi
......@@ -61,24 +60,21 @@ docker push [镜像]:[version]
## PaddlePaddle 分支规范
PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，并适应github的特性做了一些区别。
* PaddlePaddle的主版本库遵循[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范。其中:
* `master`分支为稳定(stable branch)版本分支。每一个`master`分支的版本都是经过单元测试和回归测试的版本。
* `develop`分支为开发(develop branch)版本分支。每一个`develop`分支的版本都经过单元测试，但并没有经过回归测试。
* `release/版本号`分支为每一次Release时建立的临时分支。在这个阶段的代码正在经历回归测试。
PaddlePaddle开发过程使用[Trunk Based Development](https://trunkbaseddevelopment.com/) 开发规范。
* 其他用户的fork版本库并不需要严格遵守[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，但所有fork的版本库的所有分支都相当于特性分支。
* 建议，开发者fork的版本库使用`develop`分支同步主版本库的`develop`分支
* 建议，开发者fork的版本库中，再基于`develop`版本fork出自己的功能分支。
* 当功能分支开发完毕后，向PaddlePaddle的主版本库提交`Pull Reuqest`，进而进行代码评审。
* 在评审过程中，开发者修改自己的代码，可以继续在自己的功能分支提交代码。
* `develop`分支为开发(develop branch)版本分支。每一个`develop`分支的版本都经过单元测试。并且会经过模型回归测试。
* `release/版本号`分支为每一次Release时建立的临时分支。release分支主要用于测试，bug修复和最终发版。
* `master`分支因为历史原因，已经废弃。
* BugFix分支也是在开发者自己的fork版本库维护，与功能分支不同的是，BugFix分支需要分别给主版本库的`master`、`develop`与可能有的`release/版本号`分支，同时提起`Pull Request`。
* 其他开发者fork的feature branch。
* 建议，开发者的feature branch需要同步主版本库的`develop`分支。
* 建议，开发者的feature branch需要基于主版本库中的`develop`分支。
* 当feature branch开发完毕后，向PaddlePaddle的主版本库提交`Pull Reuqest`，进而进行代码评审。
* 在评审过程中，开发者修改自己的代码，可以继续在自己的feature branch提交代码。
## PaddlePaddle回归测试列表
本列表说明PaddlePaddle发版之前需要测试的功能点。
TODO
### PaddlePaddle Book中所有章节
......
doc/fluid/dev/releasing_process_en.md
浏览文件 @ 7116cc12
......@@ -4,26 +4,21 @@ PaddlePaddle manages its branches using "git-flow branching model", and [Semanti
Each time we release a new PaddlePaddle version, we should follow the below steps:
1. Fork a new branch from `develop` named `release/[version]`, e.g. `release/0.10.0`.
1. Push a new tag on the release branch, the tag name should be like `[version]rc.patch`. The
first tag should be `0.10.0rc1`, and the second should be `0.10.0.rc2` and so on.
1. After that, we should do:
* Run all regression test on the Regression Test List (see PaddlePaddle TeamCity CI), to confirm
that this release has no major bugs.
* If regression test fails, we must fix those bugs and create a new `release/[version]`
branch from previous release branch.
* Modify `python/setup.py.in`, change the version number and change `ISTAGED` to `True`.
* Publish PaddlePaddle release wheel packages to pypi (see below instructions for detail).
* Update the Docker images (see below instructions for detail).
1. After above step, merge `release/[version]` branch to master and push a tag on the master commit,
then merge `master` to `develop`.
1. Update the Release Note.
***NOTE:***
* Do ***NOT*** merge commits from develop branch to release branches to keep the release branch contain
features only for current release, so that we can test on that version.
* If we want to fix bugs on release branches, we must merge the fix to master, develop and release branch.
1. Create a new release branch from `develop`，named `release/[version]`. E.g.，`release/0.10.0`
2. Create a new tag for the release branch, tag format: `version-rc.Patch`. E.g. the first tag is `0.10.0-rc0`。
3. New release branch normally doesn't accept new features or optimizations. QA will test on the release branch. Developer should develop based on `develop` branch.
4. If QA or Developer find bugs. They should first fix and verify on `develop` branch. Then cherry-pick the fix to the release branch. Wait until the release branch is stable.
5. If necessary, create a new tag on the relese branch, e.g. `0.10.0-rc1`. Involve more users to try it and repeat step 3-4.
6. After release branch is stable，Create the official release tag，such as `0.10.0`.
7. Release the python wheel package to pypi.
8. Update the docker image (More details below).
NOTE:
* bug fix should happen on `develop` branch, then cherry-pick to relese branch. Avoid developing directly on release branch.
* release normally only accept bug fixes. Don't add new features.
## Publish Wheel Packages to pypi
......@@ -50,6 +45,33 @@ pop-up box, choose the current release branch and click "Run Build" button. You
* pypi does not allow overwrite the already uploaded version of wheel package, even if you delete the
old version. you must change the version number before upload a new one.
### Publish wheel Packages for MacOS
You need to build the binary wheel package for MacOS before publishing, to
make sure that the package can be used by many versions of MacOS
(10.11, 10.12, 10.13) and different python installs (python.org, homebrew, etc.),
you must build the package ***exactly*** following below steps:
Build steps:
1. install python from python.org downloads, and make sure it's currently in use
in your system.
1. `export MACOSX_DEPLOYMENT_TARGET=10.11`, use `10.11` is enough for recent versions.
1. `git clone https://github.com/PaddlePaddle/Paddle.git && cd Paddle && mkdir build && cd build`
1. `cmake -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_SYSTEM_BLAS=OFF ..`, make sure the output of `cmake` command is using the correct python interpreter installed from python.org
1. `make -j`
1. `pip install delocate`
1. `mkdir fixed_wheel && delocate-wheel -w fixed_wheel python/dist/*.whl`
Then the whl under `fixed_wheel` is ready to upload.
Install steps:
1. run `pip install paddlepaddle...whl`
1. find the `libpython.dylib` that are currently in use:
- for python.org package installs, do nothing.
- for other python installs, find the path of `libpython*.dylib` and `export LD_LIBRARY_PATH=you path && DYLD_LIBRARY_PATH=your path`
## Publish Docker Images
Our CI tool will push latest images to DockerHub, so we only need to push a version tag like:
......@@ -70,26 +92,22 @@ You can then checkout the latest pushed tags at https://hub.docker.com/r/paddlep
## Branching Model
We use [git-flow](http://nvie.com/posts/a-successful-git-branching-model/) as our branching model,
with some modifications:
* `master` branch is the stable branch. Each version on the master branch is tested and guaranteed.
* `develop` branch is for development. Each commit on develop branch has passed CI unit test, but no
regression tests are run.
* `release/[version]` branch is used to publish each release. Latest release version branches have
bugfix only for that version, but no feature updates.
* Developer forks are not required to follow
[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)
branching model, all forks is like a feature branch.
* Advise: developer fork's develop branch is used to sync up with main repo's develop branch.
* Advise: developer use it's fork's develop branch to for new branch to start developing.
* Use that branch on developer's fork to create pull requests and start reviews.
* developer can push new commits to that branch when the pull request is open.
* Bug fixes are also started from developers forked repo. And, bug fixes branch can merge to
`master`, `develop` and `releases`.
PaddlePaddle uses [Trunk Based Development](https://trunkbaseddevelopment.com/) as our branching model.
* `develop` branch is used for development. Each comment to `develop` branc goes through unit tests and model regression tests.
* `release/[version]` branch is used for each release. Release branch is used for tests, bug fix and evetual release.
* `master` branch as been deprecated for historical reasons
* Developer's feature branch。
* Developer's feature branch should sync with upstream `develop` branch.
* Developer's feature branch should be forked from upstream `develop` branch.
* After feature branch is ready, create a `Pull Request` against the Paddle repo and go through code review.
* In the review process, develop modify codes and push to their own feature branch.
## PaddlePaddle Regression Test List
TODO
### All Chapters of PaddlePaddle Book
We need to guarantee that all the chapters of PaddlePaddle Book can run correctly. Including
......
doc/fluid/dev/use_eigen_cn.md
浏览文件 @ 7116cc12
......@@ -7,7 +7,7 @@
Eigen Tensor模块对element-wise计算提供了强大的支持，并且书写一份代码，可以同时在CPU、GPU执行。但Eigen Tensor是一个正在开发中的模块，因此可能测试不够完备，文档较少。
关于Eigen Tensor模块的详细介绍请参考[文档1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md) 和[文档2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md)
关于Eigen Tensor模块的详细介绍请参考[Eigen文档](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md)
## paddle::framework::Tensor
......
doc/fluid/howto/cluster/nccl2_rdma_training.md
浏览文件 @ 7116cc12
# Distributed Training with NCCL2 and RDMA
When doing distributed multi-GPU training, network bandwith often becomes the
bottle neck. We introduce a way to use NCCL2 to do such training job to
achieve best performace.
When doing distributed multi-GPU training, network bandwidth often becomes the
bottleneck. We introduce a way to use NCCL2 to do such training job to
achieve best performance.
## Prepare Hardwares with RDMA and Multiple GPUs
## Prepare Hardware with RDMA and Multiple GPUs
I'm using two Linux servers each of them is installed with 8 GPUs and
I'm using two Linux servers each of them installed with 8 GPUs and
one 100Gb RDMA card.
Base environment is:
......@@ -25,7 +25,7 @@ In general, the steps including:
1. Use docker to run tests and make sure GPUs and RDMA can work inside
the container.
I'll ommit section "Install GPU drivers" because we can find it easily
I'll omit the section "Install GPU drivers" because we can find it easily
somewhere else.
### Install RDMA drivers
......@@ -33,7 +33,7 @@ somewhere else.
For my case, I've got two machines with device
"Mellanox Technologies MT27700 Family [ConnectX-4]" installed. The OS was
"CentOS 7.4" and I updated the kernel to version 4.4 so that docker can
work with latest overlay2 filesystem.
work with the latest overlay2 filesystem.
***NOTE: before you start, make sure you have a way to get a console
of the server other than ssh because we may need to re-configure the
......@@ -45,14 +45,14 @@ network device.***
1. Run `./mlnxofedinstall --add-kernel-support` in the software package.
1. Run `/etc/init.d/openibd restart` to make everything work, note that
this operation may cause the network goes down if you are using this
RDMA device as default network device and use ssh to login the server.
RDMA device as default network device and use ssh to log in the server.
1. Re-configure the network interface, for example:
`ifconfig eth2 192.168.16.30/20 up`, then add routes if needed:
`ip route add default via 192.168.16.1 dev eth2`.
1. Do the same thing on the other node.
1. Use `ping` to test if the two nodes have typical ICMP connection.
1. Use either `udaddy` or `ib_write_bw` to test the network connection is
ready and have the desired bandwith.
ready and have the desired bandwidth.
### Prepare Docker Image to Run RDMA Programs
......@@ -60,7 +60,7 @@ network device.***
package in it.
1. Start a docker container and mount GPU driver libs into it (you can
skip this step if you are using nvidia-docker).
1. Mount RDMA dirvers and libs into the docker image (see below section),
1. Mount RDMA drivers and libs into the docker image (see below section),
also `udaddy` and `ib_write_bw` if needed.
1. Mount GPU devices and RDMA devices into the container using `--device`
or just use privileged mode `--privileged`.
......
doc/fluid/howto/inference/build_and_install_lib_cn.rst
浏览文件 @ 7116cc12
......@@ -7,13 +7,13 @@
====================== ========================================
版本说明 C++预测库
====================== ========================================
cpu_avx_mkl `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/fluid.tgz>`_
cpu_avx_openblas `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/fluid.tgz>`_
cpu_noavx_openblas `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/fluid.tgz>`_
cuda7.5_cudnn5_avx_mkl `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/fluid.tgz>`_
cuda8.0_cudnn5_avx_mkl `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/fluid.tgz>`_
cuda8.0_cudnn7_avx_mkl `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/fluid.tgz>`_
cuda9.0_cudnn7_avx_mkl `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/.lastSuccessful/fluid.tgz>`_
cpu_avx_mkl `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/fluid.tgz/?branch=0.14.0>`_
cpu_avx_openblas `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/fluid.tgz/?branch=0.14.0>`_
cpu_noavx_openblas `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/fluid.tgz/?branch=0.14.0>`_
cuda7.5_cudnn5_avx_mkl `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/fluid.tgz/?branch=0.14.0>`_
cuda8.0_cudnn5_avx_mkl `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/fluid.tgz/?branch=0.14.0>`_
cuda8.0_cudnn7_avx_mkl `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/fluid.tgz/?branch=0.14.0>`_
cuda9.0_cudnn7_avx_mkl `fluid.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/.lastSuccessful/fluid.tgz/?branch=0.14.0>`_
====================== ========================================
从源码编译
......
doc/fluid/howto/optimization/timeline_cn.md
浏览文件 @ 7116cc12
# 如何使用timeline工具做性能分析
1. 在训练的主循环外加上`with profiler.profiler(...)`。运行之后，代码会在`/tmp/profile`目录下生成一个profile的记录文件。
1. 在训练的主循环外加上`profiler.start_profiler(...)`和`profiler.stop_profiler(...)`。运行之后，代码会在`/tmp/profile`目录下生成一个profile的记录文件。
**提示：**
请不要在timeline记录信息时运行太多次迭代，因为timeline中的记录数量和迭代次数是成正比的。
```python
with profiler.profiler('All', 'total', '/tmp/profile') as prof:
for pass_id in range(pass_num):
for batch_id, data in enumerate(train_reader()):
exe.run(fluid.default_main_program(),
feed=feeder.feed(data),
fetch_list=[])
for pass_id in range(pass_num):
for batch_id, data in enumerate(train_reader()):
if pass_id == 0 and batch_id == 5:
profiler.start_profiler("All")
elif pass_id == 0 and batch_id == 10:
profiler.stop_profiler("total", "/tmp/profile")
exe.run(fluid.default_main_program(),
feed=feeder.feed(data),
fetch_list=[])
...
```
1. 运行`python paddle/tools/timeline.py`来处理`/tmp/profile`，这个程序默认会生成一个`/tmp/timeline`文件，你也可以用命令行参数来修改这个路径，请参考[timeline.py](https://github.com/PaddlePaddle/Paddle/blob/develop/tools/timeline.py)。
```python
python Paddle/tools/timeline.py --profile_path=/tmp/profile --timeline_path=timeline
```
1. 打开chrome浏览器，访问<chrome://tracing/>，用`load`按钮来加载生成的`timeline`文件。
......
doc/fluid/howto/optimization/timeline_en.md
浏览文件 @ 7116cc12
# how to use timeline tool to do profile
1. Add `with profiler.profiler(...)` to the main training loop. After run, the code will generate a profile record file `/tmp/profile`. **Warning**: Please do not run too many batches when use profiler to record timeline information, for the profile record will grow with the batch number.
1. Add `profiler.start_profiler(...)`和`profiler.stop_profiler(...)` to the main training loop. After run, the code will generate a profile record file `/tmp/profile`. **Warning**: Please do not run too many batches when use profiler to record timeline information, for the profile record will grow with the batch number.
```python
with profiler.profiler('All', 'total', '/tmp/profile') as prof:
for pass_id in range(pass_num):
for batch_id, data in enumerate(train_reader()):
exe.run(fluid.default_main_program(),
feed=feeder.feed(data),
fetch_list=[],
use_program_cache=True)
for pass_id in range(pass_num):
for batch_id, data in enumerate(train_reader()):
if pass_id == 0 and batch_id == 5:
profiler.start_profiler("All")
elif pass_id == 0 and batch_id == 10:
profiler.stop_profiler("total", "/tmp/profile")
exe.run(fluid.default_main_program(),
feed=feeder.feed(data),
fetch_list=[])
...
```
......@@ -17,6 +19,10 @@
file `/tmp/timeline` by default. You can change the path by cmd parameter, please take a look at
[timeline.py](https://github.com/PaddlePaddle/Paddle/blob/develop/tools/timeline.py) for details.
```python
python Paddle/tools/timeline.py --profile_path=/tmp/profile --timeline_path=timeline
```
1. Open chrome and visit <chrome://tracing/>, use `load` button to load the generated `timeline` file.
![chrome tracing](./tracing.jpeg)
......
doc/fluid/howto/performance/error_clip.md
浏览文件 @ 7116cc12
......@@ -78,7 +78,7 @@ def error_clip_callback(block, context):
op_desc = block.desc.op(block.desc.op_size() - 1)
for grad_n in filter(lambda n: grad_to_var.has_key(n),
op_desc.output_arg_names()):
fwd_var = block.var_recursive(grad_to_var[grad_n])
fwd_var = block.__var_recursive(grad_to_var[grad_n])
error_clip = getattr(fwd_var, "error_clip", None)
if not (error_clip is None or isinstance(error_clip,
BaseErrorClipAttr)):
......
doc/fluid/index_cn.rst
浏览文件 @ 7116cc12
PaddlePaddle Fluid
==========================
.. PaddlePaddle Fluid documentation master file, created by
sphinx-quickstart on Thu Jun 7 17:04:53 2018.
You can adapt this file completely to your liking, but it should at least
contain the root `toctree` directive.
##############
欢迎使用 Fluid
##############
.. toctree::
:maxdepth: 1
:maxdepth: 1
getstarted/index_cn.rst
build_and_install/index_cn.rst
design/index_cn.rst
howto/index_cn.rst
dev/index_cn.rst
faq/index_cn.rst
new_docs/beginners_guide/index.rst
new_docs/user_guides/index.rst
new_docs/advanced_usage/index.rst
new_docs/faq/index_cn.rst
doc/fluid/new_docs/advanced_usage/benchmark.rst 0 → 100644
浏览文件 @ 7116cc12
#################
如何进行基准测试
#################
本文介绍如何给深度学习框架做基准测试。基准测试主要包含验证模型的精度和性能两方面，下文包含搭建测试环境，选择基准测试模型，验证测试结果等几方面内容。
验证深度学习框架，可分为训练和测试两个阶段， 验证指标略有不同，本文只介绍训练阶段的指标验证。训练阶段关注的是模型训练集上的精度，训练集是完备的，因此关注大batch\_size下的训练速度,关注吞吐量，例如图像模型常用的batch\_size=128, 多卡情况下会加大；预测阶段关注的是在测试集上的精度，线上服务测试数据不能提前收集，因此关注小batch\_size下的预测速度，关注延迟，例如预测服务常用的batch\_size=1, 4等。
`Fluid <https://github.com/PaddlePaddle/Paddle>`__ 是PaddlePaddle从0.11.0版本开始引入的设计，本文的基准测试在该版本上完成。
环境搭建
""""""""""""
基准测试中模型精度和硬件、框架无关，由模型结构和数据共同决定；性能方面由测试硬件和框架性能决定。框架基准测试为了对比框架之间的差异，控制硬件环境，系统库等版本一致。下文中的对比实验都在相同的硬件条件和系统环境条件下进行.
不同架构的GPU卡性能差异巨大，在验证模型在GPU上训练性能时，可使用NVIDIA提供的工具:code `nvidia-smi` 检验当前使用的GPU型号，如果测试多卡训练性能，需确认硬件连接是 `nvlink <https://zh.wikipedia.org/zh/NVLink>`__ 或 `PCIe <https://zh.wikipedia.org/zh-hans/PCI_Express>`__ 。 同样地，CPU型号会极大影响模型在CPU上的训练性能。可读取`/proc/cpuinfo`中的参数，确认当前正在使用的CPU型号。
下载GPU对应的Cuda Tool Kit和 Cudnn，或者使用NVIDIA官方发布的nvidia-docker镜像 `nvidia-docker <https://github.com/NVIDIA/nvidia-docker>`__, 镜像内包含了Cuda和Cudnn，本文采用这种方式。 Cuda Tool Kit包含了GPU代码使用到的基础库，影响在此基础上编译出的Fluid二进制运行性能。
准备好Cuda环境后，从github上的下载Paddle并源码编译，会生成对应的最适合当前GPU的sm\_arch二进制\ `sm\_arch <https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html>`__\ 。另外，cudnn对卷积类任务影响巨大，在基准测试中需要小版本一致，例如Cudnn7.0.2与Cudnn7.1.4在Resnet上有5%以上差异。
选择基准模型
""""""""""""
对框架做基准测试，需要覆盖不同训练任务和不同大小的模型，本文中选取了图像和NLP的最为常用的5个模型。
============ ============ ================= ============
任务种类 模型名称 网络结构 数据集
============ ============ ================= ============
图像分类 mnist Lenet mnist
图像分类 VGG VGG-16 Flowers102
图像分类 Resnet Resnet-50 Flowers102
文本分类 Stacked-LSTM Stacked-LSTM IMDB
机器翻译 seq-seq Stacked-LSTM wmt14
============ ============ ================= ============
其中mnist, VGG, Resnet属于CNN模型, stacked-lstm, seq2seq代表RNN模型。
`benchmark <https://github.com/PaddlePaddle/Paddle/tree/develop/benchmark/fluid>`__
基准模型测试脚本中，均跳过了前几个batch的训练过程，原因是加载数据和分配显存受系统当前运行情况影响，会导致统计性能不准确。运行完若干个轮次后，统计对应指标。
基准模型的数据的选择方面，数据量大且验证效果多的公开数据集为首选。图像模型VGG和resnet, 本文选择了 `flowers102 <http://www.robots.ox.ac.uk/~vgg/data/flowers/102/>`__ ，图像大小预处理为和Imagenet相同大小，因此性能可直接对比
NLP模型的公开且影响力大数据集较少，seq2seq模型选择了wmt14数据，stacked-lstm模型中选择了 `imdb <https://www.imdb.com/interfaces/>`__ 数据。
注意，图像模型每条样本大小相同，图像经过变换后大小一致，因此经过的计算路径基本相同，计算速度和显存占用波动较小，可以从若干个batch的数据中采样得到当前的训练性能数据。而NLP模型由于样本长度不定，计算路径和显存占用也不相同，因此只能完整运行若干个轮次后，统计速度和显存消耗。
显存分配是特别耗时的操作，因此Fluid默认会占用所有可用显存空间形成显存池，用以加速计算过程中的显存分配。如果需要统计模型真实显存消耗，可设置环境变量`FLAGS_fraction_of_gpu_memory_to_use=0.0`，观察最大显存开销。
测试过程
""""""""""""
- CPU 单机单线程测试
测试CPU上单线程的性能，先设置CUDA的环境变量为空，``CUDA_VISIBLE_DEVICES=``，并通过环境变量关闭OpenMP和MKL的多线程 ``OMP_NUM_THREADS=1``， ``MKL_NUM_THREADS=1;``。
然后代码中设置为使用CPUPlace，如果使用Paddle代码库中的脚本，只需要命令行参数传入 use_gpu=False即可。
.. code-block:: python
>>> import paddle.fluid as fluid
>>> place = fluid.CPUPlace()
.. code:: bash
docker run -it --name CASE_NAME --security-opt seccomp=unconfined -v $PWD/benchmark:/benchmark paddlepaddle/paddle:latest-dev /bin/bash
- GPU 单机单卡测试
本教程使用了Cuda8, Cudnn7.0.1。来源为:code `nvidia/cuda:8.0-cudnn7-devel-ubuntu16.04`
.. code:: bash
nvidia-docker run -it --name CASE_NAME --security-opt seccomp=unconfined -v $PWD/benchmark:/benchmark -v /usr/lib/x86_64-linux-gnu:/usr/lib/x86_64-linux-gnu paddlepaddle/paddle:latest-dev /bin/bash
在单卡上测试，设置CUDA的环境变量使用一块GPU，``CUDA_VISIBLE_DEVICES=0``
然后代码中设置为使用CUDAPlace，如果使用Paddle代码库中的脚本，只需要命令行参数传入 use_gpu=True即可。
.. code-block:: python
>>> import paddle.fluid as fluid
>>> place = fluid.CUDAPlace(0) // 0 指第0块GPU
测试结果
""""""""""""
本教程对比相同环境下的Fluid0.12.0和TensorFlow1.4.0的性能表现。
硬件环境为 CPU: Intel(R) Xeon(R) CPU E5-2660 v4 @ 2.00GHz, GPU: TITAN X(Pascal) 12G x 1, Nvidia-Driver 384.90。
系统环境为Ubuntu 16.04.3 LTS, 本文中采用了docker环境，系统版本为nvidia-docker17.05.0-ce。
测试的Fluid版本为\ `v.0.12.0 <https://github.com/PaddlePaddle/Paddle/releases/tag/v.0.12.0>`__ 。
TensorFlow版本为\ `v.1.4.0-rc1 <https://github.com/tensorflow/tensorflow/tree/v1.4.0-rc1>`__ 。
使用的脚本和配置见\ `benchmark <https://github.com/PaddlePaddle/Paddle/tree/develop/benchmark/fluid>`__ 。
图表中统计单位为samples/秒。
- CPU 单机单线程测试结果
================ ==================== ===================
Speed Fluid CPU TensorFlow CPU
================ ==================== ===================
mnist 1298.75 samples/s 637.57 samples/s
VGG-16 0.4147 images/s 0.1229 images/s
Resnet-50 1.6935 images/s 0.3657 images/s
Stacked-LSTM 472.3225 words/s 48.2293words/s
Seq2Seq 217.1655 words/s 28.6164 words/s
================ ==================== ===================
- GPU 单机单卡测试结果
=============== ===================== =================
Speed Fluid GPU TensorFlow GPU
=============== ===================== =================
mnist 19710.90 samples/s 15576.3 samples/s
VGG-16 59.83327 images/s 40.9967 images/s
Resnet-50 105.84412 97.8923 images/s
Stacked-LSTM 1319.99315 1608.2526 words/s
Seq2Seq 7147.89081 6845.1161 words/s
=============== ===================== =================
doc/fluid/new_docs/advanced_usage/deploy/anakin_arm_benchmark.md 0 → 100644
浏览文件 @ 7116cc12
# Anakin ARM 性能测试
## 测试环境和参数:
+ 测试模型Mobilenetv1, mobilenetv2, mobilenet-ssd
+ 采用android ndk交叉编译，gcc 4.9，enable neon， ABI： armveabi-v7a with neon -mfloat-abi=softfp
+ 测试平台
- 荣耀v9(root): 处理器:麒麟960, 4 big cores in 2.36GHz, 4 little cores in 1.8GHz
- nubia z17:处理器:高通835, 4 big cores in 2.36GHz, 4 little cores in 1.9GHz
- 360 N5:处理器:高通653, 4 big cores in 1.8GHz, 4 little cores in 1.4GHz
+ 多线程：openmp
+ 时间：warmup10次，运行10次取均值
+ ncnn版本：来源于github的master branch中commits ID：307a77f04be29875f40d337cfff6df747df09de6（msg:convert LogisticRegressionOutput)版本
+ TFlite版本：来源于github的master branch中commits ID：65c05bc2ac19f51f7027e66350bc71652662125c（msg:Removed unneeded file copy that was causing failure in Pi builds)版本
在BenchMark中本文将使用**`ncnn`**、**`TFlite`**和**`Anakin`**进行性能对比分析
## BenchMark model
> 注意在性能测试之前，请先将测试model通过[External Converter](#10003)转换为Anakin model
> 对这些model，本文在ARM上进行多线程的单batch size测试。
- [Mobilenet v1](#11) *caffe model 可以在[这儿](https://github.com/shicai/MobileNet-Caffe)下载*
- [Mobilenet v2](#22) *caffe model 可以在[这儿](https://github.com/shicai/MobileNet-Caffe)下载*
- [mobilenet-ssd](#33) *caffe model 可以在[这儿](https://github.com/chuanqi305/MobileNet-SSD)下载*
### <span id = '11'> mobilenetv1 </span>
|platform | Anakin (1) | Anakin (2) | Anakin (4) | ncnn (1) | ncnn (2) | ncnn (4) | TFlite (1) | TFlite (2) | TFlite (4)|
|:---: | :---: | :---: | :---:| :---:| :---:| :---:| :---:| :---:| :---:|
|麒麟960|107.7ms|61.1ms|38.2ms|152.8ms|85.2ms|51.9ms|152.6ms|nan|nan|
|高通835|105.7ms|63.1ms|~~46.8ms~~|152.7ms|87.0ms|~~92.7ms~~|146.9ms|nan|nan|
|高通653|120.3ms|64.2ms|46.6ms|202.5ms|117.6ms|84.8ms|158.6ms|nan|nan|
### <span id = '22'> mobilenetv2 </span>
|platform | Anakin (1) | Anakin (2) | Anakin (4) | ncnn (1) | ncnn (2) | ncnn (4) | TFlite (1) | TFlite (2) | TFlite (4)|
|:---: | :---: | :---: | :---:| :---:| :---:| :---:| :---:| :---:| :---:|
|麒麟960|93.1ms|53.9ms|34.8ms|144.4ms|84.3ms|55.3ms|100.6ms|nan|nan|
|高通835|93.0ms|55.6ms|41.1ms|139.1ms|88.4ms|58.1ms|95.2ms|nan|nan|
|高通653|106.6ms|64.2ms|48.0ms|199.9ms|125.1ms|98.9ms|108.5ms|nan|nan|
### <span id = '33'> mobilenet-ssd </span>
|platform | Anakin (1) | Anakin (2) | Anakin (4) | ncnn (1) | ncnn (2) | ncnn (4) | TFlite (1) | TFlite (2) | TFlite (4)|
|:---: | :---: | :---: | :---:| :---:| :---:| :---:| :---:| :---:| :---:|
|麒麟960|213.9ms|120.5ms|74.5ms|307.9ms|166.5ms|104.2ms|nan|nan|nan|
|高通835|213.0ms|125.7ms|~~98.4ms~~|292.9ms|177.9ms|~~167.8ms~~|nan|nan|nan|
|高通653|236.0ms|129.6ms|96.0ms|377.7ms|228.9ms|165.0ms|nan|nan|nan
## How to run those Benchmark models?
1. 首先, 使用[External Converter](../docs/Manual/Converter_en.md)对caffe model 进行转换
2. 然后将转换后的Anakin model和编译好的benchmark_arm 二进制文件通过'adb push'命令上传至测试机
3. 接着在测试机含有Anakin model的目录中运行'./benchmark_arm ./ anakin_model.anakin.bin 1 10 10 1' 命令
4. 最后，终端显示器上将会打印该模型的运行时间
5. 其中运行命令的参数个数和含义可以通过运行'./benchmark_arm'看到
doc/fluid/new_docs/advanced_usage/deploy/anakin_example.md 0 → 100644
浏览文件 @ 7116cc12
# Example
Anakin目前只支持NCHW的格式
示例文件在test/framework/net下
## 在NV的GPU上运行CNN模型
示例文件为打开example_nv_cnn_net.cpp，整体流程如下：
- 将模型的的path设置为anakin模型的路径，初始化NV平台的图对象。 anakin模型可以通过转换器转化caffe或fluid的模型得到
- 根据模型设置网络图的输入尺寸，进行图优化
- 根据优化后的网络图初始化网络执行器
- 取出网络的输入tensor，将数据拷贝到输入tensor
- 运行推导
- 取出网络的输出tensor
以NV平台为例演示Anakin框架的使用方法，注意编译时需要打开GPU编译开关
## 在X86上运行RNN模型
示例文件为example_x86_rnn_net.cpp
整体流程与在NV的GPU上运行CNN模型相似，不同之处如下：
- 使用X86标识初始化图对象和网络执行器对象
- rnn模型的输入尺寸是可变的，初始化图时的输入维度是维度的最大值，输入维度N代表总的词的个数。还需要设置输入tensor的seq_offset来标示这些词是如何划分为句子的,如{0,5,12}表示共有12个词，其中第0到第4个词是第一句话，第5到第11个词是第二句话
以X86平台为例演示Anakin框架的使用方法，注意编译时需要打开X86编译开关
## 在NV的GPU上使用Anakin的线程池运行CNN模型
示例文件为example_nv_cnn_net_multi_thread.cpp ，示例使用worker的同步预测接口
整体流程与在NV的GPU上运行CNN模型相似，不同之处如下：
- 用模型地址和线程池大小初始化worker对象
- 将输入tensor注入任务队列,获得输出tensor
doc/fluid/new_docs/advanced_usage/deploy/anakin_gpu_benchmark.md 0 → 100644
浏览文件 @ 7116cc12
# Anakin GPU Benchmark
## Machine:
> CPU: `12-core Intel(R) Xeon(R) CPU E5-2620 v2 @2.10GHz`
> GPU: `Tesla P4`
> cuDNN: `v7`
## Counterpart of anakin :
The counterpart of **`Anakin`** is the acknowledged high performance inference engine **`NVIDIA TensorRT 3`** , The models which TensorRT 3 doesn't support we use the custom plugins to support.
## Benchmark Model
The following convolutional neural networks are tested with both `Anakin` and `TenorRT3`.
You can use pretrained caffe model or the model trained by youself.
> Please note that you should transform caffe model or others into anakin model with the help of [`external converter ->`](../docs/Manual/Converter_en.md)
- [Vgg16](#1) *caffe model can be found [here->](https://gist.github.com/jimmie33/27c1c0a7736ba66c2395)*
- [Yolo](#2) *caffe model can be found [here->](https://github.com/hojel/caffe-yolo-model)*
- [Resnet50](#3) *caffe model can be found [here->](https://github.com/KaimingHe/deep-residual-networks#models)*
- [Resnet101](#4) *caffe model can be found [here->](https://github.com/KaimingHe/deep-residual-networks#models)*
- [Mobilenet v1](#5) *caffe model can be found [here->](https://github.com/shicai/MobileNet-Caffe)*
- [Mobilenet v2](#6) *caffe model can be found [here->](https://github.com/shicai/MobileNet-Caffe)*
- [RNN](#7) *not support yet*
We tested them on single-GPU with single-thread.
### <span id = '1'>VGG16 </span>
- Latency (`ms`) of different batch
| BatchSize | TensorRT | Anakin |
| --- | --- | --- |
| 1 | 8.8690 | 8.2815 |
| 2 | 15.5344 | 13.9116 |
| 4 | 26.6000 | 21.8747 |
| 8 | 49.8279 | 40.4076 |
| 32 | 188.6270 | 163.7660 |
- GPU Memory Used (`MB`)
| BatchSize | TensorRT | Anakin |
| --- | --- | --- |
| 1 | 963 | 997 |
| 2 | 965 | 1039 |
| 4 | 991 | 1115 |
| 8 | 1067 | 1269 |
| 32 | 1715 | 2193 |
### <span id = '2'>Yolo </span>
- Latency (`ms`) of different batch
| BatchSize | TensorRT | Anakin |
| --- | --- | --- |
| 1 | 16.4596| 15.2124 |
| 2 | 26.6347| 25.0442 |
| 4 | 43.3695| 43.5017 |
| 8 | 80.9139 | 80.9880 |
| 32 | 293.8080| 310.8810 |
- GPU Memory Used (`MB`)
| BatchSize | TensorRT | Anakin |
| --- | --- | --- |
| 1 | 1569 | 1775 |
| 2 | 1649 | 1815 |
| 4 | 1709 | 1887 |
| 8 | 1731 | 2031 |
| 32 | 2253 | 2907 |
### <span id = '3'> Resnet50 </span>
- Latency (`ms`) of different batch
| BatchSize | TensorRT | Anakin |
| --- | --- | --- |
| 1 | 4.2459 | 4.1061 |
| 2 | 6.2627 | 6.5159 |
| 4 | 10.1277 | 11.3327 |
| 8 | 17.8209 | 20.6680 |
| 32 | 65.8582 | 77.8858 |
- GPU Memory Used (`MB`)
| BatchSize | TensorRT | Anakin |
| --- | --- | --- |
| 1 | 531 | 503 |
| 2 | 543 | 517 |
| 4 | 583 | 541 |
| 8 | 611 | 589 |
| 32 | 809 | 879 |
### <span id = '4'> Resnet101 </span>
- Latency (`ms`) of different batch
| BatchSize | TensorRT | Anakin |
| --- | --- | --- |
| 1 | 7.5562 | 7.0837 |
| 2 | 11.6023 | 11.4079 |
| 4 | 18.3650 | 20.0493 |
| 8 | 32.7632 | 36.0648 |
| 32 | 123.2550 | 135.4880 |
- GPU Memory Used (`MB)`
| BatchSize | TensorRT | Anakin |
| --- | --- | --- |
| 1 | 701 | 683 |
| 2 | 713 | 697 |
| 4 | 793 | 721 |
| 8 | 819 | 769 |
| 32 | 1043 | 1059 |
### <span id = '5'> MobileNet V1 </span>
- Latency (`ms`) of different batch
| BatchSize | TensorRT | Anakin |
| --- | --- | --- |
| 1 | 45.5156 | 1.3947 |
| 2 | 46.5585 | 2.5483 |
| 4 | 48.4242 | 4.3404 |
| 8 | 52.7957 | 8.1513 |
| 32 | 83.2519 | 31.3178 |
- GPU Memory Used (`MB`)
| BatchSize | TensorRT | Anakin |
| --- | --- | --- |
| 1 | 329 | 283 |
| 2 | 345 | 289 |
| 4 | 371 | 299 |
| 8 | 393 | 319 |
| 32 | 531 | 433 |
### <span id = '6'> MobileNet V2</span>
- Latency (`ms`) of different batch
| BatchSize | TensorRT | Anakin |
| --- | --- | --- |
| 1 | 65.6861 | 2.9842 |
| 2 | 66.6814 | 4.7472 |
| 4 | 69.7114 | 7.4163 |
| 8 | 76.1092 | 12.8779 |
| 32 | 124.9810 | 47.2142 |
- GPU Memory Used (`MB`)
| BatchSize | TensorRT | Anakin |
| --- | --- | --- |
| 1 | 341 | 293 |
| 2 | 353 | 301 |
| 4 | 385 | 319 |
| 8 | 421 | 351 |
| 32 | 637 | 551 |
## How to run those Benchmark models?
> 1. At first, you should parse the caffe model with [`external converter`](https://github.com/PaddlePaddle/Anakin/blob/b95f31e19993a192e7428b4fcf852b9fe9860e5f/docs/Manual/Converter_en.md).
> 2. Switch to *source_root/benchmark/CNN* directory. Use 'mkdir ./models' to create ./models and put anakin models into this file.
> 3. Use command 'sh run.sh', we will create files in logs to save model log with different batch size. Finally, model latency summary will be displayed on the screen.
> 4. If you want to get more detailed information with op time, you can modify CMakeLists.txt with setting `ENABLE_OP_TIMER` to `YES`, then recompile and run. You will find detailed information in model log file.
doc/fluid/new_docs/advanced_usage/deploy/convert_paddle_to_anakin.md 0 → 100644
浏览文件 @ 7116cc12
# 模型转换指南
Anakin 支持不同框架的模型预测。但由于格式的差别，Anakin 需要您预先转换模型。本文档介绍如何转换模型。
## 简介
Anakin 模型转换器输入支持 Caffe 和 Fluid 两种格式的预测模型，模型包含网络结构（model 或 prototxt）和权重参数（param 或 caffemodel）。
模型转换的输出是一个 bin 文件，它作为 Anakin 框架的 graph 参数导入。
您还可以使用模型转换器的 launch board 功能生成网络结构的 HTML 预览。
## 系统要求
- python 2.7+
- pyyaml
- flask
- protobuf 3.5+
## 用法
### 1、环境
转换器所需的依赖标注于 *系统要求* 一节。
### 2、配置
您需要对 *config.yaml* 文件进行修改以告知您的需求。工程中给出了 *config.yaml* 示例，下面作进一步说明。
#### config.yaml
```bash
OPTIONS:
Framework: CAFFE # 依框架类型填写 CAFFE 或 FLUID
SavePath: ./output # 转换结束后模型的保存位置
ResultName: googlenet # 输出模型的名字
Config:
LaunchBoard: ON # 是否生成网络结构预览页面
Server:
ip: 0.0.0.0
port: 8888 # 从一个可用端口访问预览页面
OptimizedGraph: # 当您使用了 Anakin 框架的 Optimized 功能时，才应该打开此项
enable: OFF
path: /path/to/anakin_optimized_anakin_model/googlenet.anakin.bin.saved
LOGGER:
LogToPath: ./log/ # 生成日志的路径
WithColor: ON
TARGET:
CAFFE:
# 当 Framework 为 CAFFE 时需填写
ProtoPaths:
- /path/to/caffe/src/caffe/proto/caffe.proto
PrototxtPath: /path/to/your/googlenet.prototxt
ModelPath: /path/to/your/googlenet.caffemodel
FLUID:
# 当 Framework 为 FLUID 时需填写
Debug: NULL
ProtoPaths:
- /
PrototxtPath: /path/to/fluid/inference_model
ModelPath: /path/to/fluid/inference_model
# ...
```
### 3、转换
在完成配置文件的修改后，您只需执行 ```python converter.py``` 就可以进行模型转换了。
### 4、预览
最后一步，就是在浏览器中查看令人振奋的转换结果！网址是在 *config.yaml* 中配置的，例如 http://0.0.0.0:8888 。
> 注意：若您使用了默认的 IP 地址 0.0.0.0，请在预览时使用真实的服务器地址 real_ip:port 替代它。
doc/fluid/new_docs/advanced_usage/deploy/index_anakin.rst 0 → 100644
浏览文件 @ 7116cc12
Anakin - 服务器端加速引擎
#######################
使用文档
~~~~~~~
.. toctree::
:maxdepth: 1
install_anakin.md
convert_paddle_to_anakin.md
run_anakin_on_arm.md
anakin_tutorial.md
anakin_example.md
anakin_gpu_benchmark.md
anakin_arm_benchmark.md
开发文档
~~~~~~~
.. toctree::
:maxdepth: 1
how_to_add_anakin_op.md
how_to_support_new_device_in_anakin.md
doc/fluid/new_docs/advanced_usage/deploy/index_mobile.rst 0 → 100644
浏览文件 @ 7116cc12
移动端部署
##########
.. toctree::
:maxdepth: 2
mobile_build.md
mobile_dev.md
doc/fluid/new_docs/advanced_usage/deploy/mobile_dev.md 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/fluid/new_docs/advanced_usage/development/contribute_to_paddle.md 0 → 120000
浏览文件 @ 7116cc12
../../../dev/contribute_to_paddle_cn.md
\ No newline at end of file
doc/fluid/new_docs/advanced_usage/development/cpu_profiling_cn.md 0 → 120000
浏览文件 @ 7116cc12
../../../howto/optimization/cpu_profiling_cn.md
\ No newline at end of file
doc/fluid/new_docs/advanced_usage/development/host_memory_profiling_cn.md 0 → 120000
浏览文件 @ 7116cc12
../../../howto/optimization/host_memory_profiling_cn.md
\ No newline at end of file
doc/fluid/new_docs/advanced_usage/development/new_op.md 0 → 120000
浏览文件 @ 7116cc12
../../../dev/new_op_cn.md
\ No newline at end of file
doc/fluid/new_docs/advanced_usage/index.rst 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/fluid/new_docs/advanced_usage/pics/anakin_fm_ch.png 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/fluid/new_docs/beginners_guide/basics/index.rst 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/fluid/new_docs/beginners_guide/index.rst 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/fluid/new_docs/beginners_guide/install/paddleci.png 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/fluid/new_docs/faq/faq.rst 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/fluid/new_docs/faq/index_cn.rst 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/fluid/new_docs/user_guides/howto/debug/index.rst 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/fluid/new_docs/user_guides/howto/debug/visualdl.md 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/fluid/new_docs/user_guides/howto/training/index.rst 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/fluid/new_docs/user_guides/index.rst 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/fluid/new_docs/user_guides/models/index.rst 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/survey/dynamic_graph.md
浏览文件 @ 7116cc12
此差异已折叠。
doc/survey/op_fusion_design.md 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
doc/v2/api/index_en.rst
浏览文件 @ 7116cc12
此差异已折叠。
doc/v2/faq/parameter/index_en.rst
浏览文件 @ 7116cc12
此差异已折叠。
paddle/contrib/inference/demo/CMakeLists.txt 已删除 100644 → 0
浏览文件 @ abf9832c
此差异已折叠。
paddle/contrib/inference/demo/README.md 已删除 100644 → 0
浏览文件 @ abf9832c
此差异已折叠。
paddle/fluid/API.spec 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/CMakeLists.txt
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/.gitignore 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/array.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/attribute.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/block_desc.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/data_type.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/data_type.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/data_type_test.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/executor.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/executor.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/CMakeLists.txt 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/attention_lstm_fuse_pass.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/fc_fuse_pass.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/fc_fuse_pass.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/fc_fuse_pass_tester.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/fc_gru_fuse_pass.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/fc_gru_fuse_pass.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/fc_lstm_fuse_pass.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/fc_lstm_fuse_pass.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/fuse_pass_base.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_helper.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_helper.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_helper_test.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_pattern_detector.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_pattern_detector.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_test.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_to_program_pass.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_to_program_pass.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_to_program_pass_test.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_traits.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_traits.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_viz_pass.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/graph_viz_pass.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/infer_clean_graph_pass.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/node.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/node.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/pass.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/pass.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/pass_test.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/lod_tensor.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/mixed_vector_test.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/op_desc.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/op_desc.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/op_info.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/operator.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/reader.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/reader.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/reader_test.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/rw_lock.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/rw_lock_test.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/scope.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/scope.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/tensor.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/tensor.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/var_type.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/version.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/version.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/framework/version_test.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/analysis/analyzer_main.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/analysis/flags.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/analysis/fluid_to_ir_pass.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/analysis/fluid_to_ir_pass.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/analysis/ir_pass_manager.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/analysis/ir_pass_manager.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/analysis/model_store_pass.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/analysis/model_store_pass.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/api/analysis_predictor.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/api/analysis_predictor.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/api/api_anakin_engine.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/api/demo_ci/.gitignore 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/api/demo_ci/CMakeLists.txt 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/api/demo_ci/README.md 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/api/demo_ci/clean.sh 0 → 100755
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/api/demo_ci/run.sh 0 → 100755
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/api/helper.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/api/helper.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/api/timer.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/check_symbol.sh 0 → 100755
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/io.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/io.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/paddle_fluid.sym 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/concat_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/pool2d_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/tensorrt/convert/softmax_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/tests/api/CMakeLists.txt 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/tests/api/analyzer_lac_tester.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/tests/api/analyzer_ner_tester.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/tests/api/analyzer_vis_tester.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/inference/tests/api/tester_helper.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/memory/malloc.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/adam_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/attention_lstm_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/attention_lstm_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/auc_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/auc_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/cast_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/concat_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/conv_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/crop_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/crop_op.cu
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/crop_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/detection/bbox_util.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/distributed/grpc_serde.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/distributed/grpc_serde.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/distributed/varhandle_test.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/extract_rows_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fake_quantize_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fake_quantize_op.cu 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fake_quantize_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fc_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/feed_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fetch_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fill_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/flatten_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fused_elemwise_activation_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fused_elemwise_activation_op.cu 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fused_elemwise_activation_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fusion_gru_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fusion_gru_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fusion_lstm_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fusion_lstm_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fusion_seqexpand_concat_fc_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/fusion_seqexpand_concat_fc_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/gather_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/gru_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/gru_op.cu.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/gru_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/hierarchical_sigmoid_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/hierarchical_sigmoid_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/load_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/math/blas.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/math/compound_functors.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/math/cpu_lstm_compute.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/math/cpu_lstm_compute.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/math/cpu_vec.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/math/cpu_vec_test.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/math/fc_compute.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/math/functors.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/math/im2col_cfo_cpu.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/math/matrix_bit_code.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/math/matrix_bit_code.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/math/padding.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/maxout_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/mul_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/mul_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/norm_op.cu
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/norm_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/one_hot_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/pad2d_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/pad2d_op.cu 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/pad_constant_like_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/pad_constant_like_op.cu 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/pad_constant_like_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/pad_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/prelu_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/prelu_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/print_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/read_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/reader/buffered_reader.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/reader/buffered_reader.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/recv_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/rmsprop_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sampling_id_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sampling_id_op.cu 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sampling_id_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/save_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/scale_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/scale_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/scatter_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/send_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sequence_enumerate_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sequence_enumerate_op.cu 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sequence_enumerate_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sequence_mask_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sequence_mask_op.cu 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sequence_mask_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sequence_pad_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sequence_pad_op.cu 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sequence_pad_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sgd_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/shape_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/shape_op.cu
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/shape_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/slice_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/slice_op.cu
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/slice_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/softmax_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/squeeze_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/stack_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/stack_op.cu 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/stack_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/sum_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/top_k_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/unsqueeze_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/unstack_op.cc 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/unstack_op.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/warpctc_op.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/operators/while_op.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/cpu_helper.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/cpu_info.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/cpu_info.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/cuda_helper_test.cu 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/enforce.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/float16.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/gpu_info.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/init.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/macros.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/nccl_helper.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/port.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/profiler.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/profiler.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/platform/variant.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/pybind/CMakeLists.txt
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/pybind/const_value.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/pybind/protobuf.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/pybind/recordio.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/pybind/tensor_py.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/recordio/scanner.cc
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/string/CMakeLists.txt
浏览文件 @ 7116cc12
此差异已折叠。
paddle/fluid/string/pretty_log.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
paddle/legacy/capi/Arguments.cpp
浏览文件 @ 7116cc12
此差异已折叠。
paddle/legacy/capi/arguments.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/legacy/utils/PythonUtil.cpp
浏览文件 @ 7116cc12
此差异已折叠。
paddle/legacy/utils/PythonUtil.h
浏览文件 @ 7116cc12
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ 7116cc12
此差异已折叠。
paddle/scripts/submit_local.sh.in
浏览文件 @ 7116cc12
此差异已折叠。
patches/grpc/completion_queue.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
patches/grpc/grpc_library.h 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/CMakeLists.txt
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/__init__.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/batch.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/compat.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/__init__.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/cifar.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/common.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/conll05.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/flowers.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/image.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/imdb.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/imikolov.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/mnist.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/movielens.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/mq2007.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/sentiment.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/voc2012.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/wmt14.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/dataset/wmt16.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/annotations.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/average.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/backward.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/clip.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/concurrency.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/contrib/__init__.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/contrib/memory_usage_calc.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/data_feeder.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/debugger.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/evaluator.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/executor.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/framework.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/graphviz.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/inferencer.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/initializer.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/io.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/layer_helper.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/layers/ops.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/layers/utils.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/lod_tensor.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/metrics.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/net_drawer.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/nets.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/op.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/optimizer.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/param_attr.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/profiler.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/regularizer.py
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/tests/demo/pyreader.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/tests/test_beam_search_decoder.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/tests/unittests/dist_mnist.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/tests/unittests/dist_se_resnext.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/tests/unittests/dist_transformer.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/tests/unittests/dist_word2vec.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/tests/unittests/test_compat.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/tests/unittests/test_desc_clone.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/tests/unittests/test_dist_base.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/tests/unittests/test_fc_op.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/tests/unittests/test_flatten_op.py 0 → 100644
浏览文件 @ 7116cc12
此差异已折叠。
python/paddle/fluid/tests/unittests/test_hsigmoid_op.py 0 → 100644
浏览文件 @ 7116cc12
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