Skip to content

P
Paddle

Crayon鑫 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 fa002cd6 编写于 作者: S sneaxiy
浏览文件
操作

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

上级 e0f883e6 9169b3b8
展开全部 隐藏空白更改
内联 并排
Showing with 3134 addition and 934 deletion
CMakeLists.txt
浏览文件 @ fa002cd6
......@@ -61,6 +61,7 @@ option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF)
option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF)
option(WITH_FAST_BUNDLE_TEST "Bundle tests that can be run in a single process together to reduce launch overhead" OFF)
option(WITH_CONTRIB "Compile the third-party contributation" OFF)
option(WITH_ANAKIN "Compile with Anakin library" OFF)
option(WITH_GRPC "Use grpc as the default rpc framework" ${WITH_DISTRIBUTE})
# CMAKE_BUILD_TYPE
......@@ -193,7 +194,10 @@ set(EXTERNAL_LIBS
if(WITH_GPU)
include(cuda)
include(tensorrt)
endif(WITH_GPU)
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)
......
benchmark/fluid/fluid_benchmark.py
浏览文件 @ fa002cd6
......@@ -188,7 +188,7 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
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:
if not args.no_test and batch_acc and not args.use_reader_op:
pass_test_acc = test(exe, infer_prog, test_reader, feeder,
batch_acc)
print(", Test Accuracy: %f" % pass_test_acc)
......@@ -285,11 +285,12 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
batch_id += 1
print_train_time(start_time, time.time(), num_samples)
if not args.no_test and batch_acc:
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))
exit(0)
def print_arguments(args):
......
benchmark/fluid/models/resnet.py
浏览文件 @ fa002cd6
......@@ -199,7 +199,10 @@ def get_model(args):
batched_train_reader = paddle.batch(
paddle.reader.shuffle(
train_reader, buf_size=5120),
batch_size=args.batch_size * args.gpus)
batched_test_reader = paddle.batch(train_reader, batch_size=args.batch_size)
batch_size=args.batch_size * args.gpus,
drop_last=True)
batched_test_reader = paddle.batch(
train_reader, batch_size=args.batch_size, drop_last=True)
return avg_cost, inference_program, optimizer, batched_train_reader, batched_test_reader, batch_acc
return avg_cost, inference_program, optimizer, batched_train_reader,\
batched_test_reader, batch_acc
cmake/configure.cmake
浏览文件 @ fa002cd6
......@@ -118,6 +118,10 @@ endif()
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SIMD_FLAG}")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SIMD_FLAG}")
if(WITH_DISTRIBUTE)
add_definitions(-DPADDLE_WITH_DISTRIBUTE)
endif()
if(WITH_GOLANG)
# we need to symlink Paddle directory into GOPATH. If we
# don't do it and we have code that depends on Paddle, go
......
cmake/external/anakin.cmake 0 → 100644
浏览文件 @ fa002cd6
if (NOT WITH_ANAKIN)
return()
endif()
set(ANAKIN_INSTALL_DIR "${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()
# 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 ${ANAKIN_INSTALL_DIR}; tar xzf anakin_release_simple.tar.gz")
if (WITH_ANAKIN)
message(STATUS "Anakin for inference is enabled")
message(STATUS "Anakin is set INCLUDE:${ANAKIN_INCLUDE} LIBRARY:${ANAKIN_LIBRARY}")
fetch_include_recursively(${ANAKIN_INCLUDE})
link_directories(${ANAKIN_LIBRARY})
endif()
cmake/external/openblas.cmake
浏览文件 @ fa002cd6
......@@ -29,6 +29,8 @@ IF(NOT ${CBLAS_FOUND})
"${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
CACHE FILEPATH "openblas library." FORCE)
ADD_DEFINITIONS(-DPADDLE_USE_OPENBLAS)
SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable")
SET(OPENBLAS_COMMIT "v0.2.20")
......
doc/fluid/api/gen_doc.sh
浏览文件 @ fa002cd6
#!/bin/bash
python gen_doc.py layers --submodules control_flow device io nn ops tensor > layers.rst
python gen_doc.py layers --submodules control_flow device io nn ops tensor detection learning_rate_scheduler > layers.rst
for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer
do
......
doc/fluid/api/io.rst
浏览文件 @ fa002cd6
......@@ -59,21 +59,3 @@ get_inference_program
.. autofunction:: paddle.fluid.io.get_inference_program
:noindex:
save_checkpoint
---------------
.. autofunction:: paddle.fluid.io.save_checkpoint
:noindex:
load_checkpoint
---------------
.. autofunction:: paddle.fluid.io.load_checkpoint
:noindex:
clean_checkpoint
----------------
.. autofunction:: paddle.fluid.io.clean_checkpoint
:noindex:
doc/fluid/api/layers.rst
浏览文件 @ fa002cd6
......@@ -181,12 +181,6 @@ Print
.. autofunction:: paddle.fluid.layers.Print
:noindex:
is_empty
--------
.. autofunction:: paddle.fluid.layers.is_empty
:noindex:
device
======
......@@ -261,19 +255,6 @@ double_buffer
.. autofunction:: paddle.fluid.layers.double_buffer
:noindex:
random_data_generator
---------------------
.. autofunction:: paddle.fluid.layers.random_data_generator
:noindex:
Preprocessor
------------
.. autoclass:: paddle.fluid.layers.Preprocessor
:members:
:noindex:
nn
==
......@@ -613,30 +594,6 @@ roi_pool
.. autofunction:: paddle.fluid.layers.roi_pool
:noindex:
dice_loss
---------
.. autofunction:: paddle.fluid.layers.dice_loss
:noindex:
resize_bilinear
---------------
.. autofunction:: paddle.fluid.layers.resize_bilinear
:noindex:
gather
------
.. autofunction:: paddle.fluid.layers.gather
:noindex:
random_crop
-----------
.. autofunction:: paddle.fluid.layers.random_crop
:noindex:
ops
===
......@@ -784,12 +741,6 @@ sum
.. autofunction:: paddle.fluid.layers.sum
:noindex:
shape
-----
.. autofunction:: paddle.fluid.layers.shape
:noindex:
sigmoid
-------
......@@ -1039,3 +990,93 @@ zeros
.. autofunction:: paddle.fluid.layers.zeros
:noindex:
detection
=========
multi_box_head
--------------
.. autofunction:: paddle.fluid.layers.multi_box_head
:noindex:
bipartite_match
---------------
.. autofunction:: paddle.fluid.layers.bipartite_match
:noindex:
target_assign
-------------
.. autofunction:: paddle.fluid.layers.target_assign
:noindex:
detection_output
----------------
.. autofunction:: paddle.fluid.layers.detection_output
:noindex:
ssd_loss
--------
.. autofunction:: paddle.fluid.layers.ssd_loss
:noindex:
detection_map
-------------
.. autofunction:: paddle.fluid.layers.detection_map
:noindex:
iou_similarity
--------------
.. autofunction:: paddle.fluid.layers.iou_similarity
:noindex:
box_coder
---------
.. autofunction:: paddle.fluid.layers.box_coder
:noindex:
learning_rate_scheduler
=======================
exponential_decay
-----------------
.. autofunction:: paddle.fluid.layers.exponential_decay
:noindex:
natural_exp_decay
-----------------
.. autofunction:: paddle.fluid.layers.natural_exp_decay
:noindex:
inverse_time_decay
------------------
.. autofunction:: paddle.fluid.layers.inverse_time_decay
:noindex:
polynomial_decay
----------------
.. autofunction:: paddle.fluid.layers.polynomial_decay
:noindex:
piecewise_decay
---------------
.. autofunction:: paddle.fluid.layers.piecewise_decay
:noindex:
noam_decay
----------
.. autofunction:: paddle.fluid.layers.noam_decay
:noindex:
doc/fluid/api/optimizer.rst
浏览文件 @ fa002cd6
......@@ -89,13 +89,6 @@ DecayedAdagradOptimizer
:members:
:noindex:
RMSPropOptimizer
----------------
.. autoclass:: paddle.fluid.optimizer.RMSPropOptimizer
:members:
:noindex:
Adadelta
--------
......
doc/fluid/api/profiler.rst
浏览文件 @ fa002cd6
......@@ -23,15 +23,3 @@ profiler
.. autofunction:: paddle.fluid.profiler.profiler
:noindex:
start_profiler
--------------
.. autofunction:: paddle.fluid.profiler.start_profiler
:noindex:
stop_profiler
-------------
.. autofunction:: paddle.fluid.profiler.stop_profiler
:noindex:
doc/survey/dynamic_graph.md
浏览文件 @ fa002cd6
......@@ -171,7 +171,7 @@ Pytorch chooses immediate evaluation. It avoids ever materializing a "forward gr
## What can fluid learn from them?
TBD
Please refer to `paddle/contrib/dynamic/`.
# Appendix
......
doc/v2/api/config/evaluators.rst
浏览文件 @ fa002cd6
......@@ -101,7 +101,7 @@ value_printer
:noindex:
Detection
=====
==========
detection_map
-------------
......
doc/v2/api/config/layer.rst
浏览文件 @ fa002cd6
......@@ -11,7 +11,7 @@ Data layer
data
----
.. autoclass:: paddle.v2.layer.data
.. autofunction:: paddle.v2.layer.data
:noindex:
Fully Connected Layers
......@@ -21,12 +21,12 @@ Fully Connected Layers
fc
--
.. autoclass:: paddle.v2.layer.fc
.. autofunction:: paddle.v2.layer.fc
:noindex:
selective_fc
------------
.. autoclass:: paddle.v2.layer.selective_fc
.. autofunction:: paddle.v2.layer.selective_fc
:noindex:
Conv Layers
......@@ -34,34 +34,34 @@ Conv Layers
conv_operator
-------------
.. autoclass:: paddle.v2.layer.conv_operator
.. autofunction:: paddle.v2.layer.conv_operator
:noindex:
conv_projection
---------------
.. autoclass:: paddle.v2.layer.conv_projection
.. autofunction:: paddle.v2.layer.conv_projection
:noindex:
conv_shift
----------
.. autoclass:: paddle.v2.layer.conv_shift
.. autofunction:: paddle.v2.layer.conv_shift
:noindex:
img_conv
--------
.. autoclass:: paddle.v2.layer.img_conv
.. autofunction:: paddle.v2.layer.img_conv
:noindex:
.. _api_v2.layer_context_projection:
context_projection
------------------
.. autoclass:: paddle.v2.layer.context_projection
.. autofunction:: paddle.v2.layer.context_projection
:noindex:
row_conv
--------
.. autoclass:: paddle.v2.layer.row_conv
.. autofunction:: paddle.v2.layer.row_conv
:noindex:
Image Pooling Layer
......@@ -69,27 +69,27 @@ Image Pooling Layer
img_pool
--------
.. autoclass:: paddle.v2.layer.img_pool
.. autofunction:: paddle.v2.layer.img_pool
:noindex:
spp
---
.. autoclass:: paddle.v2.layer.spp
.. autofunction:: paddle.v2.layer.spp
:noindex:
maxout
------
.. autoclass:: paddle.v2.layer.maxout
.. autofunction:: paddle.v2.layer.maxout
:noindex:
roi_pool
--------
.. autoclass:: paddle.v2.layer.roi_pool
.. autofunction:: paddle.v2.layer.roi_pool
:noindex:
pad
----
.. autoclass:: paddle.v2.layer.pad
.. autofunction:: paddle.v2.layer.pad
:noindex:
Norm Layer
......@@ -97,27 +97,27 @@ Norm Layer
img_cmrnorm
-----------
.. autoclass:: paddle.v2.layer.img_cmrnorm
.. autofunction:: paddle.v2.layer.img_cmrnorm
:noindex:
batch_norm
----------
.. autoclass:: paddle.v2.layer.batch_norm
.. autofunction:: paddle.v2.layer.batch_norm
:noindex:
sum_to_one_norm
---------------
.. autoclass:: paddle.v2.layer.sum_to_one_norm
.. autofunction:: paddle.v2.layer.sum_to_one_norm
:noindex:
cross_channel_norm
------------------
.. autoclass:: paddle.v2.layer.cross_channel_norm
.. autofunction:: paddle.v2.layer.cross_channel_norm
:noindex:
row_l2_norm
-----------
.. autoclass:: paddle.v2.layer.row_l2_norm
.. autofunction:: paddle.v2.layer.row_l2_norm
:noindex:
Recurrent Layers
......@@ -125,22 +125,22 @@ Recurrent Layers
recurrent
---------
.. autoclass:: paddle.v2.layer.recurrent
.. autofunction:: paddle.v2.layer.recurrent
:noindex:
lstmemory
---------
.. autoclass:: paddle.v2.layer.lstmemory
.. autofunction:: paddle.v2.layer.lstmemory
:noindex:
grumemory
---------
.. autoclass:: paddle.v2.layer.grumemory
.. autofunction:: paddle.v2.layer.grumemory
:noindex:
gated_unit
-----------
.. autoclass:: paddle.v2.layer.gated_unit
.. autofunction:: paddle.v2.layer.gated_unit
:noindex:
Recurrent Layer Group
......@@ -148,32 +148,32 @@ Recurrent Layer Group
memory
------
.. autoclass:: paddle.v2.layer.memory
.. autofunction:: paddle.v2.layer.memory
:noindex:
recurrent_group
---------------
.. autoclass:: paddle.v2.layer.recurrent_group
.. autofunction:: paddle.v2.layer.recurrent_group
:noindex:
lstm_step
---------
.. autoclass:: paddle.v2.layer.lstm_step
.. autofunction:: paddle.v2.layer.lstm_step
:noindex:
gru_step
--------
.. autoclass:: paddle.v2.layer.gru_step
.. autofunction:: paddle.v2.layer.gru_step
:noindex:
beam_search
------------
.. autoclass:: paddle.v2.layer.beam_search
.. autofunction:: paddle.v2.layer.beam_search
:noindex:
get_output
----------
.. autoclass:: paddle.v2.layer.get_output
.. autofunction:: paddle.v2.layer.get_output
:noindex:
Mixed Layer
......@@ -183,54 +183,54 @@ Mixed Layer
mixed
-----
.. autoclass:: paddle.v2.layer.mixed
.. autofunction:: paddle.v2.layer.mixed
:noindex:
.. _api_v2.layer_embedding:
embedding
---------
.. autoclass:: paddle.v2.layer.embedding
.. autofunction:: paddle.v2.layer.embedding
:noindex:
scaling_projection
------------------
.. autoclass:: paddle.v2.layer.scaling_projection
.. autofunction:: paddle.v2.layer.scaling_projection
:noindex:
dotmul_projection
-----------------
.. autoclass:: paddle.v2.layer.dotmul_projection
.. autofunction:: paddle.v2.layer.dotmul_projection
:noindex:
dotmul_operator
---------------
.. autoclass:: paddle.v2.layer.dotmul_operator
.. autofunction:: paddle.v2.layer.dotmul_operator
:noindex:
full_matrix_projection
----------------------
.. autoclass:: paddle.v2.layer.full_matrix_projection
.. autofunction:: paddle.v2.layer.full_matrix_projection
:noindex:
identity_projection
-------------------
.. autoclass:: paddle.v2.layer.identity_projection
.. autofunction:: paddle.v2.layer.identity_projection
:noindex:
slice_projection
-------------------
.. autoclass:: paddle.v2.layer.slice_projection
.. autofunction:: paddle.v2.layer.slice_projection
:noindex:
table_projection
----------------
.. autoclass:: paddle.v2.layer.table_projection
.. autofunction:: paddle.v2.layer.table_projection
:noindex:
trans_full_matrix_projection
----------------------------
.. autoclass:: paddle.v2.layer.trans_full_matrix_projection
.. autofunction:: paddle.v2.layer.trans_full_matrix_projection
:noindex:
Aggregate Layers
......@@ -245,51 +245,46 @@ AggregateLevel
pooling
-------
.. autoclass:: paddle.v2.layer.pooling
.. autofunction:: paddle.v2.layer.pooling
:noindex:
.. _api_v2.layer_last_seq:
last_seq
--------
.. autoclass:: paddle.v2.layer.last_seq
.. autofunction:: paddle.v2.layer.last_seq
:noindex:
.. _api_v2.layer_first_seq:
first_seq
---------
.. autoclass:: paddle.v2.layer.first_seq
.. autofunction:: paddle.v2.layer.first_seq
:noindex:
sub_seq
---------
.. autoclass:: paddle.v2.layer.sub_seq
.. autofunction:: paddle.v2.layer.sub_seq
:noindex:
concat
------
.. autoclass:: paddle.v2.layer.concat
.. autofunction:: paddle.v2.layer.concat
:noindex:
seq_concat
----------
.. autoclass:: paddle.v2.layer.seq_concat
.. autofunction:: paddle.v2.layer.seq_concat
:noindex:
seq_slice
---------
.. autoclass:: paddle.v2.layer.seq_slice
:noindex:
kmax_sequence_score
-------------------
.. autoclass:: paddle.v2.layer.kmax_sequence_score
.. autofunction:: paddle.v2.layer.seq_slice
:noindex:
sub_nested_seq
--------------
.. autoclass:: paddle.v2.layer.sub_nested_seq
.. autofunction:: paddle.v2.layer.sub_nested_seq
:noindex:
Reshaping Layers
......@@ -297,7 +292,7 @@ Reshaping Layers
block_expand
------------
.. autoclass:: paddle.v2.layer.block_expand
.. autofunction:: paddle.v2.layer.block_expand
:noindex:
.. _api_v2.layer_expand:
......@@ -309,22 +304,22 @@ ExpandLevel
expand
------
.. autoclass:: paddle.v2.layer.expand
.. autofunction:: paddle.v2.layer.expand
:noindex:
repeat
------
.. autoclass:: paddle.v2.layer.repeat
.. autofunction:: paddle.v2.layer.repeat
:noindex:
rotate
------
.. autoclass:: paddle.v2.layer.rotate
.. autofunction:: paddle.v2.layer.rotate
:noindex:
seq_reshape
-----------
.. autoclass:: paddle.v2.layer.seq_reshape
.. autofunction:: paddle.v2.layer.seq_reshape
:noindex:
Math Layers
......@@ -332,94 +327,94 @@ Math Layers
addto
-----
.. autoclass:: paddle.v2.layer.addto
.. autofunction:: paddle.v2.layer.addto
:noindex:
linear_comb
-----------
.. autoclass:: paddle.v2.layer.linear_comb
.. autofunction:: paddle.v2.layer.linear_comb
:noindex:
interpolation
-------------
.. autoclass:: paddle.v2.layer.interpolation
.. autofunction:: paddle.v2.layer.interpolation
:noindex:
bilinear_interp
---------------
.. autoclass:: paddle.v2.layer.bilinear_interp
.. autofunction:: paddle.v2.layer.bilinear_interp
:noindex:
dropout
--------
.. autoclass:: paddle.v2.layer.dropout
.. autofunction:: paddle.v2.layer.dropout
:noindex:
dot_prod
---------
.. autoclass:: paddle.v2.layer.dot_prod
.. autofunction:: paddle.v2.layer.dot_prod
:noindex:
out_prod
--------
.. autoclass:: paddle.v2.layer.out_prod
.. autofunction:: paddle.v2.layer.out_prod
:noindex:
power
-----
.. autoclass:: paddle.v2.layer.power
.. autofunction:: paddle.v2.layer.power
:noindex:
scaling
-------
.. autoclass:: paddle.v2.layer.scaling
.. autofunction:: paddle.v2.layer.scaling
:noindex:
clip
----
.. autoclass:: paddle.v2.layer.clip
.. autofunction:: paddle.v2.layer.clip
:noindex:
resize
------
.. autoclass:: paddle.v2.layer.resize
.. autofunction:: paddle.v2.layer.resize
:noindex:
slope_intercept
---------------
.. autoclass:: paddle.v2.layer.slope_intercept
.. autofunction:: paddle.v2.layer.slope_intercept
:noindex:
tensor
------
.. autoclass:: paddle.v2.layer.tensor
.. autofunction:: paddle.v2.layer.tensor
:noindex:
.. _api_v2.layer_cos_sim:
cos_sim
-------
.. autoclass:: paddle.v2.layer.cos_sim
.. autofunction:: paddle.v2.layer.cos_sim
:noindex:
l2_distance
-----------
.. autoclass:: paddle.v2.layer.l2_distance
.. autofunction:: paddle.v2.layer.l2_distance
:noindex:
trans
-----
.. autoclass:: paddle.v2.layer.trans
.. autofunction:: paddle.v2.layer.trans
:noindex:
scale_shift
-----------
.. autoclass:: paddle.v2.layer.scale_shift
.. autofunction:: paddle.v2.layer.scale_shift
:noindex:
factorization_machine
---------------------
.. autoclass:: paddle.v2.layer.factorization_machine
.. autofunction:: paddle.v2.layer.factorization_machine
:noindex:
Sampling Layers
......@@ -427,17 +422,17 @@ Sampling Layers
maxid
-----
.. autoclass:: paddle.v2.layer.max_id
.. autofunction:: paddle.v2.layer.max_id
:noindex:
sampling_id
-----------
.. autoclass:: paddle.v2.layer.sampling_id
.. autofunction:: paddle.v2.layer.sampling_id
:noindex:
multiplex
---------
.. autoclass:: paddle.v2.layer.multiplex
.. autofunction:: paddle.v2.layer.multiplex
:noindex:
.. _api_v2.layer_costs:
......@@ -447,97 +442,97 @@ Cost Layers
cross_entropy_cost
------------------
.. autoclass:: paddle.v2.layer.cross_entropy_cost
.. autofunction:: paddle.v2.layer.cross_entropy_cost
:noindex:
cross_entropy_with_selfnorm_cost
--------------------------------
.. autoclass:: paddle.v2.layer.cross_entropy_with_selfnorm_cost
.. autofunction:: paddle.v2.layer.cross_entropy_with_selfnorm_cost
:noindex:
multi_binary_label_cross_entropy_cost
-------------------------------------
.. autoclass:: paddle.v2.layer.multi_binary_label_cross_entropy_cost
.. autofunction:: paddle.v2.layer.multi_binary_label_cross_entropy_cost
:noindex:
classification_cost
-------------------
.. autoclass:: paddle.v2.layer.classification_cost
.. autofunction:: paddle.v2.layer.classification_cost
:noindex:
huber_regression_cost
-------------------------
.. autoclass:: paddle.v2.layer.huber_regression_cost
.. autofunction:: paddle.v2.layer.huber_regression_cost
:noindex:
huber_classification_cost
-------------------------
.. autoclass:: paddle.v2.layer.huber_classification_cost
.. autofunction:: paddle.v2.layer.huber_classification_cost
:noindex:
lambda_cost
-----------
.. autoclass:: paddle.v2.layer.lambda_cost
.. autofunction:: paddle.v2.layer.lambda_cost
:noindex:
square_error_cost
-----------------
.. autoclass:: paddle.v2.layer.square_error_cost
.. autofunction:: paddle.v2.layer.square_error_cost
:noindex:
rank_cost
---------
.. autoclass:: paddle.v2.layer.rank_cost
.. autofunction:: paddle.v2.layer.rank_cost
:noindex:
sum_cost
---------
.. autoclass:: paddle.v2.layer.sum_cost
.. autofunction:: paddle.v2.layer.sum_cost
:noindex:
crf
---
.. autoclass:: paddle.v2.layer.crf
.. autofunction:: paddle.v2.layer.crf
:noindex:
crf_decoding
------------
.. autoclass:: paddle.v2.layer.crf_decoding
.. autofunction:: paddle.v2.layer.crf_decoding
:noindex:
ctc
---
.. autoclass:: paddle.v2.layer.ctc
.. autofunction:: paddle.v2.layer.ctc
:noindex:
warp_ctc
--------
.. autoclass:: paddle.v2.layer.warp_ctc
.. autofunction:: paddle.v2.layer.warp_ctc
:noindex:
nce
---
.. autoclass:: paddle.v2.layer.nce
.. autofunction:: paddle.v2.layer.nce
:noindex:
hsigmoid
---------
.. autoclass:: paddle.v2.layer.hsigmoid
.. autofunction:: paddle.v2.layer.hsigmoid
:noindex:
smooth_l1_cost
--------------
.. autoclass:: paddle.v2.layer.smooth_l1_cost
.. autofunction:: paddle.v2.layer.smooth_l1_cost
:noindex:
multibox_loss
--------------
.. autoclass:: paddle.v2.layer.multibox_loss
.. autofunction:: paddle.v2.layer.multibox_loss
:noindex:
detection_output
----------------
.. autoclass:: paddle.v2.layer.detection_output
.. autofunction:: paddle.v2.layer.detection_output
:noindex:
Check Layer
......@@ -545,7 +540,7 @@ Check Layer
eos
---
.. autoclass:: paddle.v2.layer.eos
.. autofunction:: paddle.v2.layer.eos
:noindex:
Activation
......@@ -553,5 +548,5 @@ Activation
prelu
--------
.. autoclass:: paddle.v2.layer.prelu
.. autofunction:: paddle.v2.layer.prelu
:noindex:
doc/v2/api/index_en.rst
浏览文件 @ fa002cd6
......@@ -8,4 +8,3 @@ API
model_configs.rst
data.rst
run_logic.rst
fluid/index.rst
doc/v2/build_and_install/pip_install_cn.rst
浏览文件 @ fa002cd6
......@@ -60,6 +60,7 @@ paddlepaddle-gpu==0.11.0 使用CUDA 7.5和cuDNN 5编译的0.11.0版
"cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`_"
"cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
"cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
"cuda9.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
.. _pip_dependency:
......
doc/v2/build_and_install/pip_install_en.rst
浏览文件 @ fa002cd6
......@@ -63,6 +63,7 @@ If the links below shows up the login form, just click "Log in as guest" to star
"cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`__"
"cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
"cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
"cuda9.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda90cudnn7avxMkl/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`__"
.. _pip_dependency:
......
paddle/contrib/CMakeLists.txt
浏览文件 @ fa002cd6
......@@ -14,3 +14,4 @@
#
add_subdirectory(inference)
add_subdirectory(tape)
paddle/contrib/inference/CMakeLists.txt
浏览文件 @ fa002cd6
......@@ -17,48 +17,9 @@ if(APPLE)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move")
endif(APPLE)
set(ANAKIN_INCLUDE "" CACHE STRING "root of Anakin header files")
set(ANAKIN_LIBRARY "" CACHE STRING "path of Anakin library")
set(inference_deps paddle_inference_api paddle_fluid_api)
# if anakin is set enable anakin api implementation
if(ANAKIN_INCLUDE AND ANAKIN_LIBRARY)
set(ANAKIN_FOUND ON)
else()
set(ANAKIN_FOUND OFF)
endif()
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 (ANAKIN_FOUND)
# Anakin's code style doesn't follow google c style.
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=unused-variable -Wno-error=format-extra-args -Wno-error=comment -Wno-error=format -Wno-error=switch -Wno-error=return-type -Wno-error=non-virtual-dtor -Wno-reorder -Wno-error=cpp")
message(STATUS "Anakin for inference is enabled")
message(STATUS "Anakin is set INCLUDE:${ANAKIN_INCLUDE} LIBRARY:${ANAKIN_LIBRARY}")
fetch_include_recursively(${ANAKIN_INCLUDE})
link_directories(${ANAKIN_LIBRARY})
nv_library(inference_anakin_api SHARED SRCS paddle_inference_api.cc paddle_inference_api_anakin_engine.cc)
target_link_libraries(inference_anakin_api anakin anakin_saber_common)
list(APPEND inference_deps inference_anakin_api)
endif()
function(inference_api_test TARGET_NAME)
if (WITH_TESTING)
set(options "")
......@@ -79,7 +40,7 @@ function(inference_api_test TARGET_NAME)
endfunction(inference_api_test)
cc_library(paddle_inference_api
SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
cc_test(test_paddle_inference_api
......@@ -89,9 +50,17 @@ cc_test(test_paddle_inference_api
inference_api_test(test_paddle_inference_api_impl
ARGS test_word2vec test_image_classification)
if (ANAKIN_FOUND)
if (WITH_ANAKIN)
# Due to Anakin do not have official library releases and the versions of protobuf and cuda do not match Paddle's,
# so anakin library will not be merged to our official inference library. To use anakin prediction API, one need to
# compile the libinference_anakin_api.a and compile with anakin.so.
nv_library(inference_anakin_api SHARED SRCS paddle_inference_api.cc paddle_inference_api_anakin_engine.cc)
target_compile_options(inference_anakin_api BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
target_link_libraries(inference_anakin_api anakin anakin_saber_common)
cc_test(inference_anakin_test SRCS paddle_inference_api_anakin_engine_tester.cc
DEPS ${inference_deps})
ARGS --model=${ANAKIN_INSTALL_DIR}/mobilenet_v2.anakin.bin
DEPS inference_anakin_api)
target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
endif()
if(WITH_TESTING)
......
paddle/contrib/inference/paddle_inference_api_anakin_engine.cc
浏览文件 @ fa002cd6
......@@ -12,9 +12,8 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include <cuda.h>
#include "paddle/contrib/inference/paddle_inference_api_anakin_engine.h"
#include <cuda.h>
namespace paddle {
......
paddle/contrib/inference/paddle_inference_api_anakin_engine.h
浏览文件 @ fa002cd6
......@@ -19,10 +19,9 @@ limitations under the License. */
#pragma once
// NOTE This header file do not have namespace.
//#include <test/framework/net/paddle_api.h>
#include "paddle/contrib/inference/paddle_inference_api.h"
// from anakin
#include "framework/core/net/net.h"
#include "saber/saber_types.h"
......
paddle/contrib/inference/paddle_inference_api_anakin_engine_tester.cc
浏览文件 @ fa002cd6
......@@ -12,17 +12,19 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <gflags/gflags.h>
#include <glog/logging.h>
#include <gtest/gtest.h>
#include "gflags/gflags.h"
#include "paddle/contrib/inference/paddle_inference_api.h"
DEFINE_string(model, "", "Directory of the inference model.");
namespace paddle {
AnakinConfig GetConfig() {
AnakinConfig config;
config.model_file = "./mobilenet_v2.anakin.bin";
config.model_file = FLAGS_model;
config.device = 0;
config.max_batch_size = 1;
return config;
......
paddle/contrib/tape/CMakeLists.txt 0 → 100644
浏览文件 @ fa002cd6
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
if(APPLE)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move")
endif(APPLE)
cc_library(tape_variable SRCS variable.cc DEPS ${FLUID_CORE_MODULES})
cc_library(tape SRCS tape.cc DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB} tape_variable)
cc_test(test_tape
SRCS test_tape.cc
DEPS tape tape_variable)
paddle/contrib/tape/README.md 0 → 100644
浏览文件 @ fa002cd6
# Dynamic Graph on Fluid
PaddlePaddle Fluid is targeting the autodiff without tape, which, however, is very
challenging and we are still way from there. DyNet and PyTorch provide a good design
idea, the *tape*, that significantly eases the challenge. Also, DyNet provides
a C++ API that is as convenient as Python but with higher efficiency and could
conveniently integrate with industrial/production systems. This package, `tape`,
combines the good of
1. tape from PyTorch and DyNet
2. C++ API and core from DyNet
3. rich set of operators from PaddlePaddle
## Overview
We can implement Dynet-like Tape(See this [survey](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/survey/dynamic_graph.md))
by wrapping Paddle Fluid's `Operator` and `Variable`.
The user API is straight forward since
1. it is imperative. And it uses host language's control flow logic.
1. it avoids extra concepts such as `Scope` and `Executor`.
All of these benefits come at the cost of just adding one line `reset_global_tape`
at every iteration.
## Code Structure
In short, the `Tape` contains a vector of `OpHandle`s. And an `OpHandle` contains its
`type`, the pointers to the `Variable`s, and necessary attributes.
```c++
class Variable {
public:
VriableHandle Grad(); // returns its gradient variable
private:
framework::VarDesc desc_; // compile time infershape, necessary for lazy execution
framework::Variable var_; // run time variable, holds data memory
};
using VariableHandle = shared_ptr<Variable>;
struct OpHandle {
string type_;
map<string, vector<VariableHandle>> inputs_;
map<string, vector<VariableHandle>> outputs_;
AttributeMap attrs_;
};
class Tape {
public:
void AddOp(OpHandle); // add op
void Forward(); // execute the tape_
void Backward(); // execute the backward of the tape_
private:
vector<OpHandle> tape_;
};
```
We uses `Function` to indicate layers. It takes care of parameter
initialization and `AddOp` to the Tape when it is called.
```c++
class Linear {
public:
Linear(int in_dim, int out_dim, const std::string &act)
: w_(new Variable("LinearWeight")),
b_(new Variable("LinearBias")),
act_(act) {
Tape init_tape;
std::string initializer = "fill_constant";
framework::AttributeMap attrs;
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{in_dim, out_dim};
attrs["value"] = 1.0f;
init_tape.AddOp(initializer, {}, {{"Out", {w_}}}, attrs);
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{out_dim};
attrs["value"] = 1.0f;
init_tape.AddOp(initializer, {}, {{"Out", {b_}}}, attrs);
init_tape.Forward();
}
VariableHandle operator()(VariableHandle input) {
VariableHandle pre_bias(new Variable("linear"));
get_global_tape().AddOp("mul",
{{"X", {input}}, {"Y", {w_}}},
{{"Out", {pre_bias}}},
{{"x_num_col_dims", 1}, {"y_num_col_dims", 1}});
VariableHandle pre_act(new Variable("linear"));
get_global_tape().AddOp("elementwise_add",
{{"X", {pre_bias}}, {"Y", {b_}}},
{{"Out", {pre_act}}},
{{"axis", 1}});
VariableHandle post_act(new Variable("linear"));
get_global_tape().AddOp(act_,
{{"X", {pre_act}}},
{{"Out", {post_act}}},
{});
return post_act;
}
std::vector<VariableHandle> Params() { return {w_, b_}; }
private:
VariableHandle w_;
VariableHandle b_;
std::string act_;
};
```
## User API
```c++
// Model function
paddle::tape::Linear linear1(3, 3, "relu"); // init weight and bias
paddle::tape::Linear linear2(3, 3, "relu"); // init weight and bias
paddle::tape::Mean mean;
// Optimizer
paddle::tape::SGD sgd(0.001);
// Data Feeder
paddle::tape::Fill data_feeder(...);
VariableHandle input(new paddle::tape::Variable("input"));
VariableHandle label(new paddle::tape::Variable("label"));
for (int i = 0; i < 2; ++i) {
reset_global_tape();
data_feeder(input, label);
auto loss = softmax(linear2(linear1(input)), label); // compile time InferShape & InferVarType
LOG(INFO) << loss.value(); // Run forward up to loss
// Run backward, store gradient of w at w->Grad()
get_global_tape.Backward(loss);
// Update w
sgd(linear1.Params());
sgd(linear2.Params());
}
```
<details>
<summary></summary>
digraph G {
subgraph cluster_0 {
node [shape=record,style=filled];
style=filled;
color=lightgrey;
linear1 [label="{type: mul | {input | {<before_mul1>X: before_mul1 |<weight1> Y: weight1}} | {output |<before_bias1> Out: before_bias1}}"];
elementwise_add1 [label="{type: elementwise_add | {input | {<before_bias1>X: before_bias1 |<bias1> Y: bias1}} | {output |<before_act1> Out: before_act1}}"];
relu1 [label="{type: relu | {input | {<before_act1>X: before_act1 }} | {output |<after_act1> Out: after_act1}}"];
linear1 -> elementwise_add1->relu1;
label = "forward tape";
}
linear1:before_mul1->before_mul1
linear1:weight1->weight1
linear1:before_bias1->before_bias1
elementwise_add1:bias1->bias1
elementwise_add1:before_bias1->before_bias1
elementwise_add1:before_act1->before_act1
relu1:before_act1->before_act1
relu1:after_act1->after_act1
subgraph cluster_1 {
node [shape=record,style=filled];
style=filled;
color=lightgrey;
linear1_grad [label="{type: mul_grad | {input | {<before_mul1>X: before_mul1 |<weight1> Y: weight1|<before_bias1_grad> Out_grad: before_bias1_grad}} | {output |{<before_mul1_grad>X_grad: before_mul1_grad |<weight1_grad> Y_grad: weight1_grad}}}"];
elementwise_add1_grad [label="{type: elementwise_add_grad | {input | <before_act1_grad> Out_grad: before_act1_grad} | {output |{<before_bias1_grad>X_grad: before_bias1_grad |<bias1_grad> Y_grad: bias1_grad}}}"];
relu1_grad [label="{type: relu_grad | {input |<after_act1_grad> Out_grad: after_act1_grad} | {ouput | {<before_act1_grad>X_grad: before_act1_grad }}}"];
linear1_grad -> elementwise_add1_grad ->relu1_grad [dir=back];
label = "backward tape";
}
relu1_grad:after_act1_grad->after_act1_grad
relu1_grad:before_act1_grad->before_act1_grad
elementwise_add1_grad:before_act1_grad->before_act1_grad
elementwise_add1_grad:before_bias1_grad->before_bias1_grad
elementwise_add1_grad:bias1_grad->bias1_grad
linear1_grad:before_mul1->before_mul1
linear1_grad:weight1->weight1
linear1_grad:before_bias1_grad->before_bias1_grad
linear1_grad:before_mul1_grad->before_mul1_grad
linear1_grad:weight1_grad->weight1_grad
subgraph cluster_2 {
node [shape=record];
label = "Linear1";
weight1
bias1
}
weight1 -> weight1_grad [ label="Grad()", style="dashed" ];
bias1 -> bias1_grad [ label="Grad()", style="dashed"];
}
</details>
![Image](https://github.com/tonyyang-svail/Paddle/blob/cpp_tap/paddle/contrib/tape/computation_graph.png)
## Code Reuse
We want to stay close to Paddle Fluid as much as possible.
### Reuse All Operators
As all Ops are registered at `OpInfoMap`, the effort of adding a new `Function`
is about 10 lines of code, similar to expose an operator to Python.
### Reuse Compile Time InferShape and InferVarType
Note that all the symbolic information is stored at `tape::Varaible::desc_`, instead
of `ProgramDesc.block.vars`, we create a temporary `BlockDesc` to do `InferShape` and
`InferVarType` every time we `AddOp` to the tape.
### Reuse Operator::Run
We use smart pointer, instead of `Scope`, to manage memory. So we create a temporary
`Scope` for every `Operator::Run()`.
## Possible Feature
### Release Memory on Backward
We can release memory aggressively. During backward, we can delete the OpHandle once
we have finished its backward. Since all the variable is managed by smart pointer, the
memory is automatically released when its `ref_count` goes to 0.
### Kernel Fusion
As a symbolic representation of the Tape is constructed first before the actual
execution, it would be possible to perform graph optimization. One use case is kernel
fusion.
paddle/contrib/tape/function.h 0 → 100644
浏览文件 @ fa002cd6
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <string>
#include "paddle/contrib/tape/tape.h"
#include "paddle/contrib/tape/variable.h"
#include "paddle/fluid/framework/type_defs.h"
namespace paddle {
namespace tape {
class Function {};
class Fill {
public:
Fill(const std::string &initializer, const framework::AttributeMap &attrs)
: initializer_(initializer), attrs_(attrs) {}
void operator()(VariableHandle var) {
get_global_tape().AddOp(initializer_, {}, {{"Out", {var}}}, attrs_);
}
private:
const std::string initializer_;
const framework::AttributeMap attrs_;
};
class Mean {
public:
VariableHandle operator()(VariableHandle var) {
VariableHandle out(new Variable("mean"));
get_global_tape().AddOp("mean", {{"X", {var}}}, {{"Out", {out}}}, {});
return out;
}
};
class Linear {
public:
Linear(int in_dim, int out_dim, const std::string &act)
: w_(new Variable("LinearWeight")),
b_(new Variable("LinearBias")),
act_(act) {
Tape init_tape;
std::string initializer = "fill_constant";
framework::AttributeMap attrs;
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{in_dim, out_dim};
attrs["value"] = 1.0f;
init_tape.AddOp(initializer, {}, {{"Out", {w_}}}, attrs);
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{out_dim};
attrs["value"] = 1.0f;
init_tape.AddOp(initializer, {}, {{"Out", {b_}}}, attrs);
init_tape.Forward();
}
VariableHandle operator()(VariableHandle input) {
VariableHandle pre_bias(new Variable("linear"));
get_global_tape().AddOp("mul",
{{"X", {input}}, {"Y", {w_}}},
{{"Out", {pre_bias}}},
{{"x_num_col_dims", 1}, {"y_num_col_dims", 1}});
VariableHandle pre_act(new Variable("linear"));
get_global_tape().AddOp("elementwise_add",
{{"X", {pre_bias}}, {"Y", {b_}}},
{{"Out", {pre_act}}},
{{"axis", 1}});
VariableHandle post_act(new Variable("linear"));
get_global_tape().AddOp(
act_, {{"X", {pre_act}}}, {{"Out", {post_act}}}, {});
return post_act;
}
std::vector<VariableHandle> Params() { return {w_, b_}; }
private:
VariableHandle w_;
VariableHandle b_;
std::string act_;
};
class SGD {
public:
SGD(float learning_rate) : learning_rate_(new Variable("sgd")) {
Tape init_tape;
std::string initializer = "fill_constant";
framework::AttributeMap attrs;
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{1};
attrs["value"] = learning_rate;
init_tape.AddOp(initializer, {}, {{"Out", {learning_rate_}}}, attrs);
init_tape.Forward();
}
void operator()(VariableHandle input) {
PADDLE_ENFORCE(get_global_tape().HasBeenBackwarded(),
"optimization must happen after the backward");
Tape temp_tape;
temp_tape.AddOp("sgd",
{{"Param", {input}},
{"LearningRate", {learning_rate_}},
{"Grad", {input->Grad()}}},
{{"ParamOut", {input}}},
{});
temp_tape.Forward();
}
private:
VariableHandle learning_rate_;
};
}
}
paddle/contrib/tape/tape.cc 0 → 100644
浏览文件 @ fa002cd6
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/contrib/tape/tape.h"
#include <list>
#include <map>
#include <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/dim.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/pybind/pybind.h"
namespace paddle {
namespace tape {
// borrowed from
// https://stackoverflow.com/questions/874134/find-if-string-ends-with-another-string-in-c
inline bool ends_with(std::string const &value, std::string const &ending) {
if (ending.size() > value.size()) return false;
return std::equal(ending.rbegin(), ending.rend(), value.rbegin());
}
std::ostream &operator<<(std::ostream &os, const framework::VarDesc &var_desc) {
os << var_desc.Name();
os << "[" << var_desc.GetType() << "]";
os << "[" << var_desc.GetDataType() << "]";
os << "{";
for (auto &i : var_desc.GetShape()) {
os << i << ",";
}
os << "}";
return os;
}
std::string to_string(const std::string &type,
const VariableHandleMap &in_vars,
const VariableHandleMap &out_vars,
const framework::AttributeMap &attrs) {
std::stringstream ss;
ss << type << " ";
for (auto &param_name : in_vars) {
for (auto &var : param_name.second) {
ss << param_name.first << ":(" << var->Desc() << ") ";
}
}
for (auto &param_name : out_vars) {
for (auto &var : param_name.second) {
ss << param_name.first << ":(" << var->Desc() << ") ";
}
}
return ss.str();
}
framework::OpDesc CreateOpDesc(const std::string &type,
const VariableHandleMap &in_vars,
const VariableHandleMap &out_vars,
const framework::AttributeMap &attrs) {
framework::VariableNameMap inputs;
for (auto &param_name : in_vars) {
for (auto &var : param_name.second) {
inputs[param_name.first].emplace_back(var->Name());
}
}
framework::VariableNameMap outputs;
for (auto &param_name : out_vars) {
for (auto &var : param_name.second) {
outputs[param_name.first].emplace_back(var->Name());
}
}
return framework::OpDesc(type, inputs, outputs, attrs);
}
void InferShapeAndVarType(const std::string &type,
const VariableHandleMap &in_vars,
VariableHandleMap *out_vars,
const framework::AttributeMap &attrs) {
framework::OpDesc op_desc = CreateOpDesc(type, in_vars, *out_vars, attrs);
// Create a temporary block for compile-time
framework::ProgramDesc program_desc;
framework::BlockDesc *block_desc = program_desc.MutableBlock(0);
PADDLE_ENFORCE(block_desc);
for (auto &param_name : in_vars) {
for (auto &var : param_name.second) {
*block_desc->Var(var->Name())->Proto() = *var->MutableDesc()->Proto();
}
}
for (auto &param_name : *out_vars) {
for (auto &var : param_name.second) {
*block_desc->Var(var->Name())->Proto() = *var->MutableDesc()->Proto();
}
}
LOG(INFO) << "- " << to_string(type, in_vars, *out_vars, attrs);
op_desc.InferShape(*block_desc);
op_desc.InferVarType(block_desc);
for (auto &param_name : *out_vars) {
for (auto &var : param_name.second) {
*var->MutableDesc()->Proto() = *block_desc->Var(var->Name())->Proto();
}
}
LOG(INFO) << "+ " << to_string(type, in_vars, *out_vars, attrs);
}
void Tape::AddOp(const std::string &type,
const VariableHandleMap &in_vars,
VariableHandleMap out_vars,
const framework::AttributeMap &attrs) {
InferShapeAndVarType(type, in_vars, &out_vars, attrs);
tape_.emplace_back(type, in_vars, out_vars, attrs);
}
// Temporary Scope for Operator::Run()
class ScopeWrapper : public framework::Scope {
public:
ScopeWrapper(const VariableHandleMap &in_vars,
const VariableHandleMap &out_vars) {
for (auto &v : in_vars) {
for (auto &vv : v.second) {
if (!vars_.count(vv->Name())) {
vars_[vv->Name()].reset(vv->Var());
}
}
}
for (auto &v : out_vars) {
for (auto &vv : v.second) {
if (!vars_.count(vv->Name())) {
vars_[vv->Name()].reset(vv->Var());
}
}
}
}
~ScopeWrapper() {
for (auto &pair : vars_) {
pair.second.release();
}
}
};
void Tape::Forward() {
LOG(INFO) << "Starting forward -------------------------";
PADDLE_ENFORCE(!has_been_backwarded_);
while (current_position_ < tape_.size()) {
OpHandle &op = tape_[current_position_];
// Create Output Tensor, this is only necessary for OpWithKernel
for (auto &param2var : op.outputs_) {
for (auto &var : param2var.second) {
var->InitializeVariable();
}
}
framework::OpDesc op_desc =
CreateOpDesc(op.type_, op.inputs_, op.outputs_, op.attrs_);
ScopeWrapper scope(op.inputs_, op.outputs_);
framework::OpRegistry::CreateOp(op_desc)->Run(scope, platform::CPUPlace());
current_position_++;
}
LOG(INFO) << "Finishing forward -------------------------";
}
void Tape::Backward(VariableHandle target) {
PADDLE_ENFORCE(!has_been_backwarded_);
Forward();
// TODO(tonyyang-svail): check output of last op is target
backward_tape_.reset(new Tape());
framework::AttributeMap attrs;
// FIXME(tonyyang-svail): Need to infer_data_type
attrs["dtype"] = framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{1};
attrs["value"] = 1.0f;
backward_tape_->AddOp(
"fill_constant", {}, {{"Out", {target->Grad()}}}, attrs);
for (auto it = tape_.rbegin(); it != tape_.rend(); ++it) {
framework::OpDesc op_desc =
CreateOpDesc(it->type_, it->inputs_, it->outputs_, it->attrs_);
std::unordered_map<std::string, std::string> grad_to_var;
std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs =
framework::OpInfoMap::Instance()
.Get(op_desc.Type())
.GradOpMaker()(op_desc, {}, &grad_to_var, {});
for (auto &op_desc : grad_op_descs) {
std::unordered_map<std::string, VariableHandle> name2var;
for (auto &param2vars : it->inputs_) {
for (auto &a : param2vars.second) {
name2var[a->Name()] = a;
}
}
for (auto &param2vars : it->outputs_) {
for (auto &a : param2vars.second) {
name2var[a->Name()] = a;
}
}
VariableHandleMap in_vars;
VariableHandleMap out_vars;
std::map<const framework::VariableNameMap *, VariableHandleMap *>
loop_over{{&op_desc->Inputs(), &in_vars},
{&op_desc->Outputs(), &out_vars}};
for (auto &each : loop_over) {
auto &vmp = *each.first;
auto &vhm = *each.second;
for (auto &p2a : vmp) {
for (auto &argu : p2a.second) {
if (name2var.count(argu)) {
vhm[p2a.first].push_back(name2var[argu]);
} else {
PADDLE_ENFORCE(ends_with(argu, framework::kGradVarSuffix),
argu.c_str());
std::string name = argu.substr(
0, argu.size() - std::strlen(framework::kGradVarSuffix));
PADDLE_ENFORCE(name2var.count(name), name.c_str());
vhm[p2a.first].push_back(name2var[name]->Grad());
}
}
}
}
backward_tape_->AddOp(
op_desc->Type(), in_vars, out_vars, op_desc->GetAttrMap());
}
// TODO(tonyyang-svail): how to fill empty grad?
// TODO(tonyyang-svail): Sum var grad is necessary
}
backward_tape_->Forward();
has_been_backwarded_ = true;
}
Tape &get_global_tape() {
static Tape T;
return T;
}
void reset_global_tape() { get_global_tape() = Tape(); }
}
}
paddle/contrib/tape/tape.h 0 → 100644
浏览文件 @ fa002cd6
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <map>
#include <memory>
#include <string>
#include <vector>
#include "paddle/contrib/tape/variable.h"
namespace paddle {
namespace tape {
using VariableHandleMap = std::map<std::string, std::vector<VariableHandle>>;
struct OpHandle {
OpHandle(const std::string &type,
const VariableHandleMap &in_vars,
const VariableHandleMap &out_vars,
const framework::AttributeMap &attrs)
: type_(type), inputs_(in_vars), outputs_(out_vars), attrs_(attrs) {}
std::string type_;
VariableHandleMap inputs_;
VariableHandleMap outputs_;
framework::AttributeMap attrs_;
};
class Tape {
public:
void AddOp(const std::string &type,
const VariableHandleMap &in_vars,
VariableHandleMap out_vars,
const framework::AttributeMap &attrs);
void Forward();
void Backward(VariableHandle target);
bool HasBeenBackwarded() { return has_been_backwarded_; }
private:
bool has_been_backwarded_ = false;
size_t current_position_ = 0;
std::vector<OpHandle> tape_;
std::shared_ptr<Tape> backward_tape_;
};
Tape &get_global_tape();
void reset_global_tape();
}
}
paddle/contrib/tape/test_tape.cc 0 → 100644
浏览文件 @ fa002cd6
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "gtest/gtest.h"
#include "paddle/contrib/tape/function.h"
using namespace paddle::tape;
TEST(Tape, TestMLP) {
LOG(INFO) << "TestMLP";
Linear linear1(3, 3, "relu");
Linear linear2(3, 3, "relu");
Mean mean;
SGD sgd(0.001);
std::string initializer = "fill_constant";
paddle::framework::AttributeMap attrs;
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{3, 3};
attrs["value"] = 1.0f;
Fill filler(initializer, attrs);
for (int i = 0; i < 2; ++i) {
reset_global_tape();
VariableHandle input(new Variable("input"));
filler(input);
auto loss = mean(linear2(linear1(input)));
get_global_tape().Backward(loss);
for (auto w : linear1.Params()) {
sgd(w);
}
for (auto w : linear2.Params()) {
sgd(w);
}
}
}
int main(int argc, char** argv) {
std::vector<paddle::platform::Place> places;
places.emplace_back(paddle::platform::CPUPlace());
paddle::platform::DeviceContextPool::Init(places);
testing::InitGoogleTest(&argc, argv);
return RUN_ALL_TESTS();
}
paddle/contrib/tape/variable.cc 0 → 100644
浏览文件 @ fa002cd6
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/contrib/tape/variable.h"
namespace paddle {
namespace tape {
void Variable::InitializeVariable() {
LOG(INFO) << "Initialzing " << desc_.Name() << " as " << desc_.GetType();
framework::proto::VarType::Type var_type = desc_.GetType();
if (var_type == framework::proto::VarType::LOD_TENSOR) {
var_.GetMutable<framework::LoDTensor>();
} else if (var_type == framework::proto::VarType::SELECTED_ROWS) {
var_.GetMutable<framework::SelectedRows>();
} else {
PADDLE_THROW("Variable type %d is not in [LOD_TENSOR, SELECTED_ROWS]",
var_type);
}
}
}
}
paddle/contrib/tape/variable.h 0 → 100644
浏览文件 @ fa002cd6
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <memory>
#include "paddle/fluid/framework/operator.h" // framework::kGradVarSuffix
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/variable.h"
namespace paddle {
namespace tape {
class Variable;
using VariableHandle = std::shared_ptr<Variable>;
/*
* Combination of
* framework::VarDesc desc_;
* framework::Variable var_;
*/
class Variable {
public:
Variable(const std::string pre_fix)
: desc_(pre_fix + std::to_string(count())) {}
Variable(const std::string pre_fix, bool is_grad)
: desc_(pre_fix + (is_grad ? framework::kGradVarSuffix
: std::to_string(count()))) {}
~Variable() { LOG(INFO) << "Deleting " << Name(); }
// Instantiate LoDTensor/SelectedRow
void InitializeVariable();
VariableHandle Grad() {
if (grad_.expired()) {
VariableHandle new_grad(new Variable(desc_.Name(), true));
grad_ = new_grad;
return new_grad;
} else {
return VariableHandle(grad_);
}
}
// Stochastic Gradient Descent with Momentum
// VariableHandle Momentum ();
// void init(const std::string& initializer,
// const framework::AttributeMap& attrs);
// void value() {};
const framework::VarDesc& Desc() const { return desc_; }
framework::VarDesc* MutableDesc() { return &desc_; }
// TODO(tonyyang-svail): No need to expose name
std::string Name() const { return desc_.Name(); }
framework::Variable* Var() { return &var_; }
private:
int count() {
static int counter = 0;
return counter++;
}
framework::VarDesc desc_;
framework::Variable var_;
std::weak_ptr<Variable> grad_;
};
}
}
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ fa002cd6
......@@ -83,8 +83,13 @@ cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor)
cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope
framework_proto glog lod_rank_table feed_fetch_method)
if(WITH_DISTRIBUTE)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
else()
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method)
endif()
cc_library(parallel_executor SRCS parallel_executor.cc DEPS ssa_graph_builder_factory threaded_ssa_graph_executor scope_buffered_ssa_graph_executor)
......
paddle/fluid/framework/details/ssa_graph_checker.h
浏览文件 @ fa002cd6
......@@ -19,7 +19,7 @@
namespace paddle {
namespace framework {
namespace details {
class SSAGraph;
struct SSAGraph;
class SSAGraghBuilderWithChecker : public SSAGraphBuilder {
public:
......
paddle/fluid/framework/executor.cc
浏览文件 @ fa002cd6
......@@ -20,6 +20,9 @@ limitations under the License. */
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/reader.h"
#ifdef PADDLE_WITH_DISTRIBUTE
#include "paddle/fluid/operators/detail/grpc_client.h"
#endif
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/platform/profiler.h"
......@@ -44,6 +47,14 @@ ExecutorPrepareContext::~ExecutorPrepareContext() {
Executor::Executor(const platform::Place& place) : place_(place) {}
#ifdef PADDLE_WITH_DISTRIBUTE
void Executor::Complete() {
::paddle::operators::detail::RPCClient::GetInstance<
::paddle::operators::detail::GRPCClient>()
->SendComplete();
}
#endif
void InitializeVariable(Variable* var, proto::VarType::Type var_type) {
if (var_type == proto::VarType::LOD_TENSOR) {
var->GetMutable<LoDTensor>();
......@@ -319,8 +330,12 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
}
for (auto& op : ctx->ops_) {
VLOG(3) << place_ << " " << op->DebugStringEx(local_scope);
VLOG(4) << place_ << " " << op->DebugStringEx(local_scope);
op->Run(*local_scope, place_);
// NOTE! Please do not delete this line, it's usefull because the debug
// string before and after op.run are different, after run the output
// will have right shape which is usefull for debug.
VLOG(3) << place_ << " " << op->DebugStringEx(local_scope);
if (FLAGS_benchmark) {
VLOG(2) << "Memory used after operator " + op->Type() + " running: "
......
paddle/fluid/framework/executor.h
浏览文件 @ fa002cd6
......@@ -44,6 +44,13 @@ class Executor {
explicit Executor(const platform::Place& place);
#ifdef PADDLE_WITH_DISTRIBUTE
/*
* Sending signal to pserver to mark current trainer stop.
*/
void Complete();
#endif
/* @Brief
* Runtime evaluation of the given ProgramDesc under certain Scope
*
......
paddle/fluid/framework/operator.cc
浏览文件 @ fa002cd6
......@@ -69,6 +69,19 @@ static DDim GetDims(const Scope& scope, const std::string& name,
}
}
static int GetRowSize(const Scope& scope, const std::string& name) {
Variable* var = scope.FindVar(name);
if (var == nullptr) {
return -1;
}
if (var->IsType<SelectedRows>()) {
return var->Get<SelectedRows>().rows().size();
}
return -1;
}
static LoD GetLoD(const Scope& scope, const std::string& name) {
Variable* var = scope.FindVar(name);
auto default_lod = LoD({{}});
......@@ -85,6 +98,7 @@ static LoD GetLoD(const Scope& scope, const std::string& name) {
}
void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
VLOG(10) << "- " << DebugStringEx(&scope);
if (platform::is_gpu_place(place)) {
#ifndef PADDLE_WITH_CUDA
PADDLE_THROW("Cannot run operator on place %s", place);
......@@ -94,6 +108,7 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
#endif
}
RunImpl(scope, place);
VLOG(10) << "+ " << DebugStringEx(&scope);
}
bool OperatorBase::HasInputs(const std::string& name) const {
......@@ -153,6 +168,10 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
for (size_t i = 0; i < input.second.size(); ++i) {
ss << input.second[i];
if (scope) {
int row_size = GetRowSize(*scope, input.second[i]);
if (row_size >= 0) {
ss << "[row_size=" << row_size << "]";
}
ss << "[" << GetDims(*scope, input.second[i], true) << "]";
ss << "(" << GetLoD(*scope, input.second[i]) << ")";
}
......@@ -173,6 +192,10 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
for (size_t i = 0; i < output.second.size(); ++i) {
ss << output.second[i];
if (scope) {
int row_size = GetRowSize(*scope, output.second[i]);
if (row_size >= 0) {
ss << "[row_size=" << row_size << "]";
}
ss << "[" << GetDims(*scope, output.second[i], true) << "]";
ss << "(" << GetLoD(*scope, output.second[i]) << ")";
}
......
paddle/fluid/framework/parallel_executor.cc
浏览文件 @ fa002cd6
......@@ -145,9 +145,9 @@ void ParallelExecutor::BCastParamsToGPUs(
auto &dims = main_tensor.dims();
if (paddle::platform::is_gpu_place(main_tensor.place())) {
#ifdef PADDLE_WITH_CUDA
std::vector<void *> buffers;
size_t numel = main_tensor.numel();
ncclDataType_t data_type = platform::ToNCCLDataType(main_tensor.type());
platform::NCCLGroupGuard guard;
for (size_t i = 0; i < member_->places_.size(); ++i) {
auto place = member_->places_[i];
void *buffer;
......@@ -159,11 +159,21 @@ void ParallelExecutor::BCastParamsToGPUs(
t->Resize(dims);
buffer = t->mutable_data(place, main_tensor.type());
}
auto &nccl_ctx = member_->nccl_ctxs_->at(place);
platform::dynload::ncclBcast(buffer, numel, data_type, 0,
nccl_ctx.comm_, nccl_ctx.stream());
buffers.push_back(buffer);
}
member_->nccl_ctxs_->WaitAll();
PADDLE_ENFORCE_EQ(member_->places_.size(), buffers.size(),
"variables' buffer size to bcast NOT equal to places");
{
platform::NCCLGroupGuard guard;
for (size_t i = 0; i < member_->places_.size(); ++i) {
auto &nccl_ctx = member_->nccl_ctxs_->at(member_->places_[i]);
platform::dynload::ncclBcast(buffers[i], numel, data_type, 0,
nccl_ctx.comm_, nccl_ctx.stream());
}
member_->nccl_ctxs_->WaitAll();
}
#else
PADDLE_THROW("Not compiled with CUDA");
#endif
......
paddle/fluid/framework/reader.h
浏览文件 @ fa002cd6
......@@ -35,14 +35,15 @@ class ReaderBase {
class DecoratedReader : public ReaderBase {
public:
explicit DecoratedReader(ReaderBase* reader) : ReaderBase(), reader_(reader) {
explicit DecoratedReader(const std::shared_ptr<ReaderBase>& reader)
: ReaderBase(), reader_(reader) {
PADDLE_ENFORCE_NOT_NULL(reader_);
}
void ReInit() override { reader_->ReInit(); }
protected:
ReaderBase* reader_;
std::shared_ptr<ReaderBase> reader_;
};
class FileReader : public ReaderBase {
......@@ -64,7 +65,7 @@ class ReaderHolder {
public:
void Reset(ReaderBase* reader) { reader_.reset(reader); }
ReaderBase* Get() const { return reader_.get(); }
std::shared_ptr<ReaderBase> Get() const { return reader_; }
void ReadNext(std::vector<LoDTensor>* out) {
PADDLE_ENFORCE_NOT_NULL(reader_);
......@@ -76,7 +77,7 @@ class ReaderHolder {
}
private:
std::unique_ptr<ReaderBase> reader_;
std::shared_ptr<ReaderBase> reader_;
};
} // namespace framework
......
paddle/fluid/framework/scope.h
浏览文件 @ fa002cd6
......@@ -81,6 +81,9 @@ class Scope {
// Rename variable to a new name and return the new name
std::string Rename(const std::string& origin_name) const;
protected:
mutable std::unordered_map<std::string, std::unique_ptr<Variable>> vars_;
private:
// Call Scope::NewScope for a sub-scope.
explicit Scope(Scope const* parent) : parent_(parent) {}
......@@ -93,8 +96,6 @@ class Scope {
// Caller doesn't own the returned Variable.
Variable* FindVarLocally(const std::string& name) const;
mutable std::unordered_map<std::string, std::unique_ptr<Variable>> vars_;
// Scope in `kids_` are owned by this class.
mutable std::list<Scope*> kids_;
Scope const* parent_{nullptr};
......
paddle/fluid/inference/io.cc
浏览文件 @ fa002cd6
......@@ -20,16 +20,20 @@ limitations under the License. */
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/pybind/pybind.h"
DEFINE_string(devices, "", "The devices to be used which is joined by comma.");
DEFINE_bool(init_p2p, false, "Whether to init p2p.");
DEFINE_int32(math_num_threads, 1,
"Number of threads used to run math functions.");
namespace paddle {
namespace inference {
void Init(const std::vector<std::string> argv) {
framework::InitGflags(argv);
operators::math::SetNumThreads(FLAGS_math_num_threads);
// init devices
std::vector<int> devices;
std::string token;
......
paddle/fluid/operators/batch_norm_mkldnn_op.cc
浏览文件 @ fa002cd6
......@@ -19,10 +19,17 @@ limitations under the License. */
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
using batch_norm_bwd = mkldnn::batch_normalization_backward;
using batch_norm_fwd = mkldnn::batch_normalization_forward;
using framework::DataLayout;
using framework::Tensor;
using mkldnn::memory;
using mkldnn::primitive;
using mkldnn::reorder;
using mkldnn::stream;
using paddle::platform::MKLDNNDeviceContext;
using paddle::platform::MKLDNNMemDesc;
using mkldnn::memory;
using platform::to_void_cast;
template <typename T>
using EigenArrayMap =
......@@ -64,21 +71,12 @@ void run_batch_norm_op(Args &&... args) {
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
}
template <typename T>
inline void *cast_const_to_void(const T *t) {
return static_cast<void *>(const_cast<T *>(t));
}
} // namespace
template <typename T>
class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
auto data_layout_str = ctx.Attr<std::string>("data_layout");
auto data_layout = framework::StringToDataLayout(data_layout_str);
PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW,
"MKLDNN batch normalization handles only NCHW data layout");
const float epsilon = ctx.Attr<float>("epsilon");
const float momentum = ctx.Attr<float>("momentum");
const bool is_test = ctx.Attr<bool>("is_test");
......@@ -99,41 +97,53 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
const auto *scale = ctx.Input<Tensor>("Scale");
const auto *shift = ctx.Input<Tensor>("Bias");
y->mutable_data<T>(ctx.GetPlace());
mean_out->mutable_data<T>(ctx.GetPlace());
variance_out->mutable_data<T>(ctx.GetPlace());
PADDLE_ENFORCE(x->layout() == DataLayout::kMKLDNN &&
x->format() != memory::format::format_undef,
"Wrong layout/format set for Input x tensor");
const T *x_data = x->data<T>();
const T *mean_data = mean->data<T>();
const T *variance_data = variance->data<T>();
T *y_data = y->mutable_data<T>(ctx.GetPlace());
T *mean_out_data = mean_out->mutable_data<T>(ctx.GetPlace());
T *variance_out_data = variance_out->mutable_data<T>(ctx.GetPlace());
T *batch_mean_data = nullptr;
T *batch_variance_data = nullptr;
if (!is_test) {
batch_mean->mutable_data<T>(ctx.GetPlace());
batch_variance->mutable_data<T>(ctx.GetPlace());
batch_mean_data = batch_mean->mutable_data<T>(ctx.GetPlace());
batch_variance_data = batch_variance->mutable_data<T>(ctx.GetPlace());
}
auto propagation = is_test == true ? mkldnn::prop_kind::forward_scoring
: mkldnn::prop_kind::forward_training;
auto dims = paddle::framework::vectorize2int(x->dims());
auto src_md =
MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
auto dst_md =
MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
auto src_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine};
auto dst_pd = mkldnn::memory::primitive_desc{dst_md, mkldnn_engine};
auto src = mkldnn::memory{src_pd, cast_const_to_void(x->data<T>())};
auto dst = mkldnn::memory{dst_pd, y->data<T>()};
auto src_tz = paddle::framework::vectorize2int(x->dims());
auto scale_tz = paddle::framework::vectorize2int(scale->dims());
PADDLE_ENFORCE(scale_tz.size() == 1, "Dims of scale tensor is NOT 1");
const unsigned int ic = scale_tz[0];
unsigned flags = mkldnn::use_scale_shift;
if (is_test) flags |= mkldnn::use_global_stats;
// create mkldnn memory from input x tensor
auto src_memory =
memory({{{src_tz}, memory::data_type::f32, x->format()}, mkldnn_engine},
to_void_cast(x_data));
// create primitive descriptor for batch norm forward
using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>;
auto batch_norm_fwd_desc =
bn_fwd_types::op_desc{propagation, src_md, epsilon, flags};
auto batch_norm_fwd_pd =
bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine};
auto batch_norm_fwd_desc = bn_fwd_types::op_desc{
propagation, src_memory.get_primitive_desc().desc(), epsilon, flags};
std::shared_ptr<batch_norm_fwd::primitive_desc> batch_norm_fwd_pd =
std::shared_ptr<batch_norm_fwd::primitive_desc>(
new batch_norm_fwd::primitive_desc(batch_norm_fwd_desc,
mkldnn_engine));
const unsigned int ic = dims[1];
// Save the pd to be used in backward pass
const std::string key = ctx.op().Output("SavedMean");
const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd";
dev_ctx.SetBlob(key_batch_norm_fwd_pd, batch_norm_fwd_pd);
// MKLDNN requires a single piece of memory for scale and shift/bias data
const size_t scaleshift_size = 2 * ic;
......@@ -143,73 +153,58 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
copy_to_weights(scale->data<T>(), scale->data<T>() + ic, shift->data<T>(),
shift->data<T>() + ic, &scaleshift_data);
auto scaleshift_memory = mkldnn::memory{
batch_norm_fwd_pd.weights_primitive_desc(), scaleshift_data.data()};
// crate mkldnn memory for weights(scale/shift)
auto scaleshift_memory = memory(batch_norm_fwd_pd->weights_primitive_desc(),
scaleshift_data.data());
if (is_test) {
auto mean_memory = mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(),
cast_const_to_void(mean->data<T>())};
// create mkldnn memory for output y tensor
auto dst_memory = memory(batch_norm_fwd_pd->dst_primitive_desc(), y_data);
if (is_test) {
// create mkldnn memory for stats (as input)
auto mean_memory = memory(batch_norm_fwd_pd->mean_primitive_desc(),
to_void_cast(mean_data));
auto variance_memory =
mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(),
cast_const_to_void(variance->data<T>())};
memory(batch_norm_fwd_pd->variance_primitive_desc(),
to_void_cast(variance_data));
run_batch_norm_op<typename bn_fwd_types::op_type>(
batch_norm_fwd_pd, src, (const mkldnn::primitive::at &)mean_memory,
*batch_norm_fwd_pd, src_memory,
(const mkldnn::primitive::at &)mean_memory,
(const mkldnn::primitive::at &)variance_memory, scaleshift_memory,
dst);
dst_memory);
} else {
// create mkldnn memory for stats (as output)
auto mean_memory =
mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(),
cast_const_to_void(batch_mean->data<T>())};
auto variance_memory =
mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(),
cast_const_to_void(batch_variance->data<T>())};
memory(batch_norm_fwd_pd->mean_primitive_desc(), batch_mean_data);
auto variance_memory = memory(
batch_norm_fwd_pd->variance_primitive_desc(), batch_variance_data);
run_batch_norm_op<bn_fwd_types::op_type>(batch_norm_fwd_pd, src,
scaleshift_memory, dst,
run_batch_norm_op<bn_fwd_types::op_type>(*batch_norm_fwd_pd, src_memory,
scaleshift_memory, dst_memory,
mean_memory, variance_memory);
}
if (!is_test) {
const unsigned int in = dims[0];
const unsigned int sample_size = x->numel() / in / ic;
// saved_xx is use just in this batch of data
EigenVectorArrayMap<T> saved_mean_e(
batch_mean->mutable_data<T>(ctx.GetPlace()), ic);
EigenVectorArrayMap<T> saved_variance_e(
batch_variance->mutable_data<T>(ctx.GetPlace()), ic);
saved_mean_e.setZero();
saved_variance_e.setZero();
const unsigned int x_arr_size = in * ic;
ConstEigenArrayMap<T> x_arr(x->data<T>(), sample_size, x_arr_size);
for (unsigned int nc = 0; nc < x_arr_size; ++nc) {
saved_mean_e(nc % ic) += x_arr.col(nc).sum();
}
saved_mean_e /= in * sample_size;
for (unsigned int nc = 0; nc < x_arr_size; ++nc) {
saved_variance_e(nc % ic) +=
(x_arr.col(nc) - saved_mean_e(nc % ic)).matrix().squaredNorm();
}
saved_variance_e /= in * sample_size;
ConstEigenVectorArrayMap<T> mean_arr{mean->data<T>(), ic};
ConstEigenVectorArrayMap<T> variance_arr{variance->data<T>(), ic};
EigenVectorArrayMap<T> running_mean_arr(
mean_out->mutable_data<T>(ctx.GetPlace()), ic);
EigenVectorArrayMap<T> running_var_arr(
variance_out->mutable_data<T>(ctx.GetPlace()), ic);
// mkldnn only compute stats for current batch
// so we need compute momentum stats via Eigen lib
EigenVectorArrayMap<T> batch_mean_e(batch_mean_data, ic);
EigenVectorArrayMap<T> batch_variance_e(batch_variance_data, ic);
ConstEigenVectorArrayMap<T> mean_e(mean_data, ic);
ConstEigenVectorArrayMap<T> variance_e{variance_data, ic};
EigenVectorArrayMap<T> running_mean_e(mean_out_data, ic);
EigenVectorArrayMap<T> running_variance_e(variance_out_data, ic);
auto one_minus_momentum = 1. - momentum;
running_mean_arr =
mean_arr * momentum + saved_mean_e * one_minus_momentum;
running_var_arr =
variance_arr * momentum + saved_variance_e * one_minus_momentum;
running_mean_e = mean_e * momentum + batch_mean_e * one_minus_momentum;
running_variance_e =
variance_e * momentum + batch_variance_e * one_minus_momentum;
}
y->set_layout(DataLayout::kMKLDNN);
y->set_format(
(memory::format)dst_memory.get_primitive_desc().desc().data.format);
}
};
......@@ -217,11 +212,6 @@ template <typename T>
class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
public:
void Compute(const paddle::framework::ExecutionContext &ctx) const override {
auto data_layout_str = ctx.Attr<std::string>("data_layout");
auto data_layout = framework::StringToDataLayout(data_layout_str);
PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW,
"MKLDNN batch normalization handles only NCHW data layout");
auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
auto mkldnn_engine = dev_ctx.GetEngine();
......@@ -238,88 +228,132 @@ class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
auto *diff_scale = ctx.Output<Tensor>(framework::GradVarName("Scale"));
auto *diff_shift = ctx.Output<Tensor>(framework::GradVarName("Bias"));
diff_x->mutable_data<T>(ctx.GetPlace());
diff_scale->mutable_data<T>(ctx.GetPlace());
diff_shift->mutable_data<T>(ctx.GetPlace());
PADDLE_ENFORCE(diff_y->layout() == DataLayout::kMKLDNN &&
diff_y->format() != memory::format::format_undef,
"Wrong layout/format set for Input diff_y tensor");
const T *x_data = x->data<T>();
const T *diff_y_data = diff_y->data<T>();
const T *batch_mean_data = batch_mean->data<T>();
const T *batch_variance_data = batch_variance->data<T>();
const T *scale_data = scale->data<T>();
const T *shift_data = shift->data<T>();
T *diff_x_data = diff_x->mutable_data<T>(ctx.GetPlace());
T *diff_scale_data = diff_scale->mutable_data<T>(ctx.GetPlace());
T *diff_shift_data = diff_shift->mutable_data<T>(ctx.GetPlace());
auto src_tz = paddle::framework::vectorize2int(x->dims());
auto diff_src_tz = src_tz;
auto dst_tz = src_tz;
auto diff_dst_tz = dst_tz;
auto scale_tz = paddle::framework::vectorize2int(scale->dims());
PADDLE_ENFORCE(scale_tz.size() == 1, "Dims of scale tensor is NOT 1");
const unsigned int ic = scale_tz[0];
// Retrieve bn_fwd_pd from device context
const std::string key = ctx.op().Input("SavedMean");
const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd";
auto batch_norm_fwd_pd =
std::static_pointer_cast<batch_norm_fwd::primitive_desc>(
dev_ctx.GetBlob(key_batch_norm_fwd_pd));
PADDLE_ENFORCE(batch_norm_fwd_pd != nullptr,
"Fail to find batch_norm_fwd_pd in device context");
auto dims = paddle::framework::vectorize2int(x->dims());
unsigned flags = mkldnn::use_scale_shift | !mkldnn::use_global_stats;
using bn_bwd_types = bn_type_traits<mkldnn::batch_normalization_backward>;
auto src_md =
MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
auto dst_md =
MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
auto diff_src_md =
MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
auto diff_dst_md =
MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw);
// create mkldnn memory from input diff_y tensor
auto user_diff_dst_memory =
memory({{{diff_dst_tz}, memory::data_type::f32, diff_y->format()},
mkldnn_engine},
to_void_cast(diff_y_data));
using bn_bwd_types = bn_type_traits<mkldnn::batch_normalization_backward>;
using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>;
// create mkldnn memory from input x tensor
auto src_memory =
memory({{{src_tz}, memory::data_type::f32, x->format()}, mkldnn_engine},
to_void_cast(x_data));
auto batch_norm_fwd_desc = bn_fwd_types::op_desc{
mkldnn::prop_kind::forward_training, src_md, epsilon, flags};
auto batch_norm_fwd_pd =
bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine};
// for diff_dst, try to use same format as dst in forward pass
auto diff_dst_pd = batch_norm_fwd_pd.get()->dst_primitive_desc();
auto diff_dst_md = diff_dst_pd.desc();
// create primitive descriptor for batch norm backward
unsigned flags = mkldnn::use_scale_shift;
auto batch_norm_bwd_desc = bn_bwd_types::op_desc{
mkldnn::prop_kind::backward, diff_dst_md, dst_md, epsilon, flags};
mkldnn::prop_kind::backward, diff_dst_md,
src_memory.get_primitive_desc().desc(), epsilon, flags};
auto batch_norm_bwd_pd = bn_bwd_types::op_prim{
batch_norm_bwd_desc, mkldnn_engine, batch_norm_fwd_pd};
auto src = mkldnn::memory{{src_md, mkldnn_engine},
cast_const_to_void(x->data<T>())};
auto mean = mkldnn::memory{batch_norm_bwd_pd.mean_primitive_desc(),
cast_const_to_void(batch_mean->data<T>())};
auto variance =
mkldnn::memory{batch_norm_bwd_pd.variance_primitive_desc(),
cast_const_to_void(batch_variance->data<T>())};
auto diff_dst = mkldnn::memory{{diff_dst_md, mkldnn_engine},
cast_const_to_void(diff_y->data<T>())};
batch_norm_bwd_desc, mkldnn_engine, *batch_norm_fwd_pd};
// reorder user_diff_dst if it's not in preferred format
auto diff_dst_memory = user_diff_dst_memory;
primitive reorder_diff_dst;
bool is_diff_dst_reordered = false;
if (diff_dst_pd != user_diff_dst_memory.get_primitive_desc()) {
diff_dst_memory = memory(diff_dst_pd);
reorder_diff_dst = reorder(user_diff_dst_memory, diff_dst_memory);
is_diff_dst_reordered = true;
}
const unsigned int ic = dims[1];
// create mkldnn memory for input tensors (src/mean/variance)
auto mean_memory = memory(batch_norm_bwd_pd.mean_primitive_desc(),
to_void_cast(batch_mean_data));
auto variance_memory = memory(batch_norm_bwd_pd.variance_primitive_desc(),
to_void_cast(batch_variance_data));
// MKLDNN requires a single piece of memory for scale and shift/bias data
const size_t scaleshift_size = 2 * ic;
std::vector<T> scaleshift_data;
scaleshift_data.reserve(scaleshift_size);
copy_to_weights(scale->data<T>(), scale->data<T>() + ic, shift->data<T>(),
shift->data<T>() + ic, &scaleshift_data);
copy_to_weights(scale_data, scale_data + ic, shift_data, shift_data + ic,
&scaleshift_data);
auto scaleshift_memory = mkldnn::memory{
batch_norm_bwd_pd.weights_primitive_desc(), scaleshift_data.data()};
// create mkldnn memory for input tensors (scale/shift)
auto scaleshift_memory = memory(batch_norm_bwd_pd.weights_primitive_desc(),
scaleshift_data.data());
// create mkldnn memory for output diff weights (combined scale/shift)
std::vector<T> diff_scaleshift_data;
diff_scaleshift_data.reserve(scaleshift_size);
copy_to_weights(diff_scale->data<T>(), diff_scale->data<T>() + ic,
diff_shift->data<T>(), diff_shift->data<T>() + ic,
&diff_scaleshift_data);
auto diff_scaleshift_memory =
mkldnn::memory{batch_norm_bwd_pd.diff_weights_primitive_desc(),
diff_scaleshift_data.data()};
auto diff_src = mkldnn::memory{{diff_src_md, mkldnn_engine},
static_cast<void *>(diff_x->data<T>())};
run_batch_norm_op<bn_bwd_types::op_type>(
batch_norm_bwd_pd, src, mean, variance, diff_dst, scaleshift_memory,
diff_src, diff_scaleshift_memory);
memory(batch_norm_bwd_pd.diff_weights_primitive_desc(),
diff_scaleshift_data.data());
// here assume diff_src is in the same format of src
auto diff_src_memory = memory(src_memory.get_primitive_desc(), diff_x_data);
// finally create batch_norm backward primitive
auto batch_norm_bwd_prim =
batch_norm_bwd(batch_norm_bwd_pd, src_memory, mean_memory,
variance_memory, diff_dst_memory, scaleshift_memory,
diff_src_memory, diff_scaleshift_memory);
// execute optional reorder and batch_norm backward primitive
std::vector<primitive> pipeline;
if (is_diff_dst_reordered) pipeline.push_back(reorder_diff_dst);
pipeline.push_back(batch_norm_bwd_prim);
stream(stream::kind::eager).submit(pipeline).wait();
// copy back diff sacle/shift to output tensors (diff scale/shift)
diff_scaleshift_data.resize(scaleshift_size);
auto it = std::begin(diff_scaleshift_data);
std::copy(it, std::next(it, ic), diff_scale->data<T>());
std::copy(it, std::next(it, ic), diff_scale_data);
std::copy(std::next(it, ic), std::end(diff_scaleshift_data),
diff_shift->data<T>());
diff_shift_data);
// set layout/format of output tensors
diff_x->set_layout(DataLayout::kMKLDNN);
diff_x->set_format((memory::format)diff_src_memory.get_primitive_desc()
.desc()
.data.format);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_KERNEL(batch_norm, MKLDNN, paddle::platform::CPUPlace,
REGISTER_OP_KERNEL(batch_norm, MKLDNN, ::paddle::platform::CPUPlace,
ops::BatchNormMKLDNNOpKernel<float>);
REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, paddle::platform::CPUPlace,
REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, ::paddle::platform::CPUPlace,
ops::BatchNormMKLDNNGradOpKernel<float>);
paddle/fluid/operators/batch_norm_op.cc
浏览文件 @ fa002cd6
......@@ -110,19 +110,19 @@ class BatchNormOp : public framework::OperatorWithKernel {
ctx.Input<Tensor>("Variance")->type()),
"Variance input should be of float type");
framework::LibraryType library_{framework::LibraryType::kPlain};
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::LibraryType library = framework::LibraryType::kPlain;
framework::DataLayout layout = framework::DataLayout::kAnyLayout;
#ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain &&
if (library == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
library = framework::LibraryType::kMKLDNN;
layout = framework::DataLayout::kMKLDNN;
}
#endif
return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout,
library_);
library);
}
};
......@@ -370,19 +370,21 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
PADDLE_THROW("can't find Y@GRAD");
}
framework::LibraryType library_{framework::LibraryType::kPlain};
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout_ = framework::DataLayout::kAnyLayout;
framework::LibraryType library = framework::LibraryType::kPlain;
framework::DataLayout layout = framework::DataLayout::kAnyLayout;
#ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain &&
if (library == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
layout_ = framework::DataLayout::kMKLDNN;
library = framework::LibraryType::kMKLDNN;
layout = framework::DataLayout::kMKLDNN;
}
#endif
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
layout_, library_);
layout, library);
}
};
......
paddle/fluid/operators/conv_op.cc
浏览文件 @ fa002cd6
......@@ -75,9 +75,8 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvOp::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
framework::LibraryType library{framework::LibraryType::kPlain};
std::string data_format = ctx.Attr<std::string>("data_format");
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
std::string data_format = ctx.Attr<std::string>("data_format");
framework::DataLayout layout = framework::StringToDataLayout(data_format);
#ifdef PADDLE_WITH_CUDA
......
paddle/fluid/operators/detail/grpc_client.cc
浏览文件 @ fa002cd6
......@@ -34,6 +34,12 @@ void GRPCClient::InitEventLoop() {
client_thread_.reset(new std::thread(std::bind(&GRPCClient::Proceed, this)));
}
void GRPCClient::SendComplete() {
for (auto& it : channels_) {
this->AsyncSendComplete(it.first);
}
}
GRPCClient::~GRPCClient() {
Wait();
cq_.Shutdown();
......@@ -210,6 +216,19 @@ void GRPCClient::AsyncSendFetchBarrier(const std::string& ep,
req_count_++;
}
void GRPCClient::AsyncSendComplete(const std::string& ep, int64_t time_out) {
const auto ch = GetChannel(ep);
BatchBarrierProcessor* s = new BatchBarrierProcessor(ch);
s->Prepare(time_out);
sendrecv::VariableMessage req;
req.set_varname(COMPLETE_MESSAGE);
auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
req_count_++;
}
void GRPCClient::Wait() {
std::unique_lock<std::mutex> lk(sync_mutex_);
sync_cond_.wait(lk, [this] { return req_count_ == 0; });
......
paddle/fluid/operators/detail/grpc_client.h
浏览文件 @ fa002cd6
......@@ -195,6 +195,8 @@ class GRPCClient : public RPCClient {
void Wait() override;
void SendComplete() override;
protected:
void InitImpl() override;
......@@ -204,6 +206,9 @@ class GRPCClient : public RPCClient {
void Proceed();
void AsyncSendComplete(const std::string& ep,
int64_t time_out = RPCClient::rpc_time_out);
std::shared_ptr<grpc::Channel> GetChannel(const std::string& ep);
private:
......
paddle/fluid/operators/detail/grpc_server.cc
浏览文件 @ fa002cd6
......@@ -162,16 +162,19 @@ class RequestPrefetch final : public RequestBase {
void Process() override {
// prefetch process...
std::string varname = request_->OutVarname();
VLOG(3) << "RequestPrefetch " << varname;
std::string in_var_name = request_->Varname();
std::string out_var_name = request_->OutVarname();
VLOG(3) << "RequestPrefetch, in_var_name: " << in_var_name
<< " out_var_name: " << out_var_name;
auto scope = request_->GetMutableLocalScope();
auto invar = scope->FindVar(varname);
framework::Variable* outvar = nullptr;
auto invar = scope->FindVar(in_var_name);
// out var must be created in local scope!
framework::Variable* outvar = scope->Var(out_var_name);
request_handler_->Handle(varname, scope, invar, &outvar);
request_handler_->Handle(in_var_name, scope, invar, &outvar, out_var_name);
SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(),
SerializeToByteBuffer(out_var_name, outvar, *request_handler_->dev_ctx(),
&reply_);
Finish(reply_, &responder_);
}
......@@ -287,7 +290,7 @@ void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name,
} else if (rpc_name == kRequestPrefetch) {
b = new RequestPrefetch(&service_, cq.get(), handler, req_id);
} else {
PADDLE_ENFORCE(false, "not surpported rpc");
PADDLE_ENFORCE(false, "not supported rpc");
}
reqs[req_id] = b;
......
paddle/fluid/operators/detail/request_handler.h
浏览文件 @ fa002cd6
......@@ -40,6 +40,7 @@ constexpr char kRequestPrefetch[] = "RequestPrefetch";
#define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV"
#define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV"
#define FETCH_BARRIER_MESSAGE "FETCH_BARRIER@RECV"
#define COMPLETE_MESSAGE "COMPLETE@RECV"
class RPCServer;
......@@ -60,9 +61,12 @@ class RequestHandler {
void SetDevCtx(const platform::DeviceContext* dev_ctx) { dev_ctx_ = dev_ctx; }
void SetProgram(framework::ProgramDesc* program) { program_ = program; }
void SetExecutor(framework::Executor* executor) { executor_ = executor; }
// Used for dist lookup table prefetch
void SetPrefetchPreparedCtx(
std::unique_ptr<framework::ExecutorPrepareContext> prepared) {
prefetch_ctx_.reset(prepared.release());
std::unordered_map<
std::string, std::shared_ptr<framework::ExecutorPrepareContext>>* g) {
prefetch_var_name_to_prepared_ctx_ = g;
}
// Used for async.
......@@ -78,9 +82,6 @@ class RequestHandler {
bool sync_mode() { return sync_mode_; }
framework::Scope* scope() { return scope_; }
const platform::DeviceContext* dev_ctx() { return dev_ctx_; }
framework::ExecutorPrepareContext* prefetch_ctx() {
return prefetch_ctx_.get();
}
framework::ProgramDesc* program() { return program_; }
framework::Executor* executor() { return executor_; }
......@@ -99,8 +100,8 @@ class RequestHandler {
// *request_handler_->dev_ctx(), &reply_);
// }
virtual bool Handle(const std::string& varname, framework::Scope* scope,
framework::Variable* var,
framework::Variable** outvar) = 0;
framework::Variable* var, framework::Variable** outvar,
const std::string& out_var_name = "") = 0;
protected:
const bool sync_mode_;
......@@ -109,12 +110,17 @@ class RequestHandler {
framework::Executor* executor_;
framework::Scope* scope_;
framework::ProgramDesc* program_;
std::unique_ptr<framework::ExecutorPrepareContext> prefetch_ctx_;
// used for distribute lookup table prefetch
std::unordered_map<std::string,
std::shared_ptr<framework::ExecutorPrepareContext>>*
prefetch_var_name_to_prepared_ctx_;
// Used for async.
std::unordered_map<std::string,
std::shared_ptr<framework::ExecutorPrepareContext>>*
grad_to_prepared_ctx_;
RPCServer* rpc_server_;
};
......
paddle/fluid/operators/detail/request_handler_impl.cc
浏览文件 @ fa002cd6
......@@ -30,7 +30,8 @@ namespace detail {
bool RequestSendHandler::Handle(const std::string& varname,
framework::Scope* scope,
framework::Variable* invar,
framework::Variable** outvar) {
framework::Variable** outvar,
const std::string& out_var_name) {
VLOG(4) << "RequestSendHandler:" << varname;
// Async
......@@ -49,6 +50,9 @@ bool RequestSendHandler::Handle(const std::string& varname,
if (varname == BATCH_BARRIER_MESSAGE) {
VLOG(3) << "sync: recv batch barrier message";
rpc_server_->IncreaseBatchBarrier(kRequestSend);
} else if (varname == COMPLETE_MESSAGE) {
VLOG(3) << "sync: recv complete message";
rpc_server_->DecreaseClientNum();
} else {
VLOG(3) << "sync: received var_name: " << varname;
if (sync_mode_) {
......@@ -79,7 +83,8 @@ void RequestSendHandler::ResetSparseVarRecorder() {
bool RequestGetHandler::Handle(const std::string& varname,
framework::Scope* scope,
framework::Variable* invar,
framework::Variable** outvar) {
framework::Variable** outvar,
const std::string& out_var_name) {
VLOG(4) << "RequestGetHandler:" << varname;
if (varname != FETCH_BARRIER_MESSAGE) {
......@@ -102,13 +107,14 @@ bool RequestGetHandler::Handle(const std::string& varname,
bool RequestPrefetchHandler::Handle(const std::string& varname,
framework::Scope* scope,
framework::Variable* invar,
framework::Variable** outvar) {
framework::Variable** outvar,
const std::string& out_var_name) {
VLOG(4) << "RequestPrefetchHandler " << varname;
auto var_desc = program_->Block(0).FindVar(varname);
*outvar = scope->FindVar(varname);
auto var_desc = program_->Block(0).FindVar(out_var_name);
InitializeVariable(*outvar, var_desc->GetType());
executor_->RunPreparedContext(prefetch_ctx_.get(), scope);
executor_->RunPreparedContext(
(*prefetch_var_name_to_prepared_ctx_)[varname].get(), scope);
return true;
}
......
paddle/fluid/operators/detail/request_handler_impl.h
浏览文件 @ fa002cd6
......@@ -39,7 +39,8 @@ class RequestSendHandler final : public RequestHandler {
explicit RequestSendHandler(bool sync_mode) : RequestHandler(sync_mode) {}
virtual ~RequestSendHandler() {}
bool Handle(const std::string& varname, framework::Scope* scope,
framework::Variable* var, framework::Variable** outvar) override;
framework::Variable* var, framework::Variable** outvar,
const std::string& out_var_name = "") override;
void ResetSparseVarRecorder();
private:
......@@ -52,7 +53,8 @@ class RequestGetHandler final : public RequestHandler {
explicit RequestGetHandler(bool sync_mode) : RequestHandler(sync_mode) {}
virtual ~RequestGetHandler() {}
bool Handle(const std::string& varname, framework::Scope* scope,
framework::Variable* var, framework::Variable** outvar) override;
framework::Variable* var, framework::Variable** outvar,
const std::string& out_var_name = "") override;
};
class RequestPrefetchHandler final : public RequestHandler {
......@@ -60,7 +62,8 @@ class RequestPrefetchHandler final : public RequestHandler {
explicit RequestPrefetchHandler(bool sync_mode) : RequestHandler(sync_mode) {}
virtual ~RequestPrefetchHandler() {}
bool Handle(const std::string& varname, framework::Scope* scope,
framework::Variable* var, framework::Variable** outvar) override;
framework::Variable* var, framework::Variable** outvar,
const std::string& out_var_name = "") override;
};
} // namespace detail
......
paddle/fluid/operators/detail/rpc_client.h
浏览文件 @ fa002cd6
......@@ -53,6 +53,11 @@ class RPCClient {
virtual void AsyncSendFetchBarrier(const std::string& ep,
int64_t time_out = rpc_time_out) = 0;
// SendComplete tells all the server that current trainer have no more data
// to train, so that the pserver can reduce it's barrier count, and continue
// to train with other trainers.
virtual void SendComplete() = 0;
virtual void Wait() = 0;
static constexpr int64_t rpc_time_out = 120 * 1000;
......
paddle/fluid/operators/detail/rpc_server.cc
浏览文件 @ fa002cd6
......@@ -43,7 +43,7 @@ void RPCServer::SavePort() const {
void RPCServer::WaitBarrier(const std::string& rpc_name) {
std::unique_lock<std::mutex> lock(this->mutex_);
barrier_cond_.wait(lock, [=] {
barrier_cond_.wait(lock, [this, &rpc_name] {
return (barrier_counter_[rpc_name] >= client_num_ || exit_flag_.load());
});
......@@ -53,19 +53,23 @@ void RPCServer::WaitBarrier(const std::string& rpc_name) {
void RPCServer::IncreaseBatchBarrier(const std::string rpc_name) {
VLOG(3) << "RPCServer begin IncreaseBatchBarrier " << rpc_name;
int b = 0;
{
std::unique_lock<std::mutex> lock(mutex_);
b = ++barrier_counter_[rpc_name];
}
VLOG(3) << "RPCServer IncreaseBatchBarrier " << rpc_name
<< ", barrier_count:" << b << ", fan_in" << client_num_;
std::unique_lock<std::mutex> lock(mutex_);
b = ++barrier_counter_[rpc_name];
if (b >= client_num_) {
lock.unlock();
barrier_cond_.notify_all();
lock.lock();
}
}
void RPCServer::DecreaseClientNum() {
{
std::unique_lock<std::mutex> lock(mutex_);
client_num_--;
}
barrier_cond_.notify_all();
}
void RPCServer::ResetBarrierCounter() {
VLOG(3) << "RPCServer ResetBarrierCounter ";
std::unique_lock<std::mutex> lock(mutex_);
......
paddle/fluid/operators/detail/rpc_server.h
浏览文件 @ fa002cd6
......@@ -60,7 +60,7 @@ class RPCServer {
void SetCond(const std::string& rpc_name);
void WaitCond(const std::string& rpc_name);
void IncreaseBatchBarrier(const std::string rpc_name);
void DecreaseClientNum();
void ResetBarrierCounter();
protected:
......@@ -79,8 +79,7 @@ class RPCServer {
std::string bind_address_;
std::atomic<int> exit_flag_;
int selected_port_;
const int client_num_;
int client_num_;
std::unordered_map<std::string, RequestHandler*> rpc_call_map_;
std::unordered_map<std::string, int> rpc_thread_num_;
......
paddle/fluid/operators/detail/rpc_server_test.cc
浏览文件 @ fa002cd6
......@@ -98,11 +98,17 @@ void StartServer() {
framework::Executor exe(place);
platform::CPUDeviceContext ctx(place);
auto* block = AppendPrefetchBlcok(&program);
auto prepared = exe.Prepare(program, block->ID());
std::string in_var_name("ids");
std::vector<int> prefetch_block_ids{block->ID()};
auto prepared = exe.Prepare(program, prefetch_block_ids);
InitTensorsOnServer(&scope, &place, 10);
std::unordered_map<std::string,
std::shared_ptr<framework::ExecutorPrepareContext>>
prefetch_var_name_to_prepared;
prefetch_var_name_to_prepared[in_var_name] = prepared[0];
g_req_handler->SetProgram(&program);
g_req_handler->SetPrefetchPreparedCtx(std::move(prepared));
g_req_handler->SetPrefetchPreparedCtx(&prefetch_var_name_to_prepared);
g_req_handler->SetDevCtx(&ctx);
g_req_handler->SetScope(&scope);
g_req_handler->SetExecutor(&exe);
......
paddle/fluid/operators/elementwise_op.h
浏览文件 @ fa002cd6
......@@ -66,40 +66,41 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker {
.SetDefault(-1)
.EqualGreaterThan(-1);
AddComment(string::Sprintf(R"DOC(
Limited Elementwise %s Operator.
Limited Elementwise %s Operator
The equation is:
$$%s$$
$X$ is a tensor of any dimension and the dimensions of tensor $Y$ must be
smaller than or equal to the dimensions of $X$.
- $X$: a tensor of any dimension.
- $Y$: a tensor whose dimensions must be less than or equal to the dimensions of $X$.
There are two cases for this operator:
1. The shape of $Y$ is same with $X$;
2. The shape of $Y$ is a congiguous subsequencet of $X$. The trailing dimensions
of size 1 for $Y$ will be ignored for the consideration of subsequence.
1. The shape of $Y$ is the same with $X$.
2. The shape of $Y$ is a continuous subsequence of $X$.
For case 2:
$Y$ will be broadcasted to match the shape of $X$ and axis should be
set to index of the start dimension to broadcast $Y$ onto $X$.
1. Broadcast $Y$ to match the shape of $X$, where $axis$ is the start dimension index
for broadcasting $Y$ onto $X$.
2. If $axis$ is -1 (default), $axis = rank(X) - rank(Y)$.
3. The trailing dimensions of size 1 for $Y$ will be ignored for the consideration of
subsequence, such as shape(Y) = (2, 1) => (2).
If axis is -1, it is treated as axis=rank(X)-rank(Y).
For example:
For example
.. code-block:: python
shape(X) = (2, 3, 4, 5), shape(Y) = (,)
shape(X) = (2, 3, 4, 5), shape(Y) = (5,)
shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)
shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5), with axis=-1(default) or axis=2
shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
shape(X) = (2, 3, 4, 5), shape(Y) = (2, 1), with axis=0
Either of the inputs $X$ and $Y$ or none can carry the LoD (Level of Details)
information. However, the output only shares the LoD information with input $X$.
The inputs $X$ and $Y$ can carry the different LoD information.
But the output only shares the LoD information with the input $X$.
)DOC",
GetName(), GetEquation()));
......
paddle/fluid/operators/gen_nccl_id_op.cc
浏览文件 @ fa002cd6
......@@ -67,6 +67,10 @@ class GenNCCLIdOp : public framework::OperatorBase {
client->AsyncSendVar(ep, dev_ctx, *scope, NCCL_ID_VARNAME);
}
client->Wait();
for (auto& ep : endpoint_list) {
client->AsyncSendBatchBarrier(ep);
}
client->Wait();
VLOG(3) << "sending completed...";
}
......
paddle/fluid/operators/listen_and_serv_op.cc
浏览文件 @ fa002cd6
......@@ -96,19 +96,22 @@ static int64_t GetTimestamp() {
return tp.tv_sec * 1000 + tp.tv_usec / 1000;
}
void ListenAndServOp::RunSyncLoop(framework::Executor *executor,
framework::ProgramDesc *program,
framework::Scope *recv_scope,
framework::BlockDesc *prefetch_block) const {
void ListenAndServOp::RunSyncLoop(
framework::Executor *executor, framework::ProgramDesc *program,
framework::Scope *recv_scope,
const std::vector<int> &prefetch_block_id_list) const {
size_t num_blocks = program->Size();
PADDLE_ENFORCE_GE(num_blocks, 2,
"server program should have at least 2 blocks");
std::vector<int> block_list;
for (size_t blkid = 1; blkid < num_blocks; ++blkid) {
block_list.push_back(blkid);
std::vector<int> optimize_block_id_list;
for (int blkid = 1; blkid < num_blocks; ++blkid) {
if (std::find(prefetch_block_id_list.begin(), prefetch_block_id_list.end(),
blkid) == prefetch_block_id_list.end()) {
optimize_block_id_list.push_back(blkid);
}
}
auto optimize_prepared = executor->Prepare(*program, block_list);
auto optimize_prepared = executor->Prepare(*program, optimize_block_id_list);
// Insert placeholder for block0 which holds current op itself.
optimize_prepared.insert(
optimize_prepared.begin(),
......@@ -135,16 +138,17 @@ void ListenAndServOp::RunSyncLoop(framework::Executor *executor,
std::vector<size_t> parallel_blkids;
parallel_blkids.push_back(1);
double ts = GetTimestamp();
for (size_t blkid = 2; blkid < num_blocks; ++blkid) {
if (blkid != static_cast<size_t>(prefetch_block->ID())) {
if (program->Block(blkid).Parent() != last_parent_blkid) {
ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared,
program, recv_scope);
parallel_blkids.clear();
last_parent_blkid = program->Block(blkid).Parent();
}
parallel_blkids.push_back(blkid);
for (size_t i = 1; i < optimize_block_id_list.size(); ++i) {
// skip the first optimize block because it is already in the
// parallel_blkids.
int blkid = optimize_block_id_list[i];
if (program->Block(blkid).Parent() != last_parent_blkid) {
ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared,
program, recv_scope);
parallel_blkids.clear();
last_parent_blkid = program->Block(blkid).Parent();
}
parallel_blkids.push_back(blkid);
}
ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared, program,
recv_scope);
......@@ -210,18 +214,19 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor,
} // while(true)
}
static void FillRequestCtx(detail::RequestHandler *h, framework::Scope *scope,
platform::DeviceContext *dev_ctx,
framework::Executor *executor,
framework::ProgramDesc *program,
framework::ExecutorPrepareContext *prefetch_ctx,
detail::RPCServer *rpc_server) {
static void FillRequestCtx(
detail::RequestHandler *h, framework::Scope *scope,
platform::DeviceContext *dev_ctx, framework::Executor *executor,
framework::ProgramDesc *program,
std::unordered_map<std::string,
std::shared_ptr<framework::ExecutorPrepareContext>>
*prefetch_ctx,
detail::RPCServer *rpc_server) {
h->SetScope(scope);
h->SetDevCtx(dev_ctx);
h->SetExecutor(executor);
h->SetProgram(program);
h->SetPrefetchPreparedCtx(
std::unique_ptr<framework::ExecutorPrepareContext>(prefetch_ctx));
h->SetPrefetchPreparedCtx(prefetch_ctx);
h->SetRPCServer(rpc_server);
}
......@@ -255,17 +260,42 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
request_prefetch_handler_.get());
auto *optimize_block = Attr<framework::BlockDesc *>(kOptimizeBlock);
auto *prefetch_block = Attr<framework::BlockDesc *>(kPrefetchBlock);
auto *program = optimize_block->Program();
framework::Executor executor(dev_place);
// prepare for prefetch
VLOG(3) << "prefetch block id is " << prefetch_block->ID();
auto prefetch_prepared = executor.Prepare(*program, prefetch_block->ID());
std::vector<int> prefetch_block_id_list;
std::unordered_map<int, std::string> block_id_to_prefetch_var_name;
auto prefetch_var_name_to_block_id_str =
Attr<std::vector<std::string>>(kPrefetchVarNameToBlockId);
for (const auto &prefetch_var_name_and_id :
prefetch_var_name_to_block_id_str) {
std::vector<std::string> pieces;
split(prefetch_var_name_and_id, ':', &pieces);
VLOG(3) << "after split, prefetch_var = " << pieces[0]
<< ", id=" << pieces[1];
PADDLE_ENFORCE_EQ(pieces.size(), 2);
int block_id = std::stoi(pieces[1]);
prefetch_block_id_list.push_back(block_id);
block_id_to_prefetch_var_name[block_id] = pieces[0];
}
auto prefetch_prepared = executor.Prepare(*program, prefetch_block_id_list);
std::unordered_map<std::string,
std::shared_ptr<framework::ExecutorPrepareContext>>
prefetch_var_name_to_prepared_ctx;
for (size_t i = 0; i < prefetch_block_id_list.size(); ++i) {
auto block_id = prefetch_block_id_list[i];
auto prefetch_var_name = block_id_to_prefetch_var_name[block_id];
prefetch_var_name_to_prepared_ctx[prefetch_var_name] = prefetch_prepared[i];
}
auto f = std::bind(FillRequestCtx, std::placeholders::_1, &recv_scope,
&dev_ctx, &executor, program, prefetch_prepared.release(),
rpc_service_.get());
&dev_ctx, &executor, program,
&prefetch_var_name_to_prepared_ctx, rpc_service_.get());
f(request_send_handler_.get());
f(request_get_handler_.get());
......@@ -283,7 +313,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
// Write to a file of server selected port for python use.
SavePort();
if (sync_mode) {
RunSyncLoop(&executor, program, &recv_scope, prefetch_block);
RunSyncLoop(&executor, program, &recv_scope, prefetch_block_id_list);
} else {
RunAsyncLoop(&executor, program);
}
......@@ -309,8 +339,9 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker {
AddAttr<bool>("sync_mode", "if works at sync_mode or not").SetDefault(true);
AddAttr<framework::BlockDesc *>(kOptimizeBlock,
"BlockID to run on server side.");
AddAttr<framework::BlockDesc *>(kPrefetchBlock,
"prefetch block to run on server side.");
AddAttr<std::vector<std::string>>(kPrefetchVarNameToBlockId,
"prefetch blocks to run on server side.")
.SetDefault({});
AddAttr<int>("Fanin", "How many clients send to this server.")
.SetDefault(1);
}
......
paddle/fluid/operators/listen_and_serv_op.h
浏览文件 @ fa002cd6
......@@ -18,6 +18,7 @@ limitations under the License. */
#include <atomic>
#include <set>
#include <string>
#include <vector>
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/lod_tensor.h"
......@@ -30,7 +31,7 @@ namespace paddle {
namespace operators {
constexpr char kOptimizeBlock[] = "OptimizeBlock";
constexpr char kPrefetchBlock[] = "PrefetchBlock";
constexpr char kPrefetchVarNameToBlockId[] = "prefetch_var_name_to_block_id";
void RunServer(std::shared_ptr<detail::RPCServer> service);
......@@ -46,7 +47,7 @@ class ListenAndServOp : public framework::OperatorBase {
void RunSyncLoop(framework::Executor* executor,
framework::ProgramDesc* program,
framework::Scope* recv_scope,
framework::BlockDesc* prefetch_block) const;
const std::vector<int>& prefetch_block_id_list) const;
void RunAsyncLoop(framework::Executor* executor,
framework::ProgramDesc* program) const;
......
paddle/fluid/operators/math/blas.h
浏览文件 @ fa002cd6
......@@ -20,13 +20,16 @@
#ifdef PADDLE_WITH_MKLML
#include <mkl_cblas.h>
#include <mkl_lapacke.h>
#include <mkl_service.h>
#include <mkl_vml_functions.h>
#endif
#ifdef PADDLE_USE_OPENBLAS
#include <cblas.h>
#ifdef LAPACK_FOUND
#include <lapacke.h>
#endif
#endif
#ifndef LAPACK_FOUND
extern "C" {
......@@ -46,6 +49,18 @@ namespace paddle {
namespace operators {
namespace math {
static void SetNumThreads(int num_threads) {
#ifdef PADDLE_USE_OPENBLAS
int real_num_threads = num_threads > 1 ? num_threads : 1;
openblas_set_num_threads(real_num_threads);
#elif defined(PADDLE_WITH_MKLML)
int real_num_threads = num_threads > 1 ? num_threads : 1;
mkl_set_num_threads(real_num_threads);
#else
PADDLE_ENFORCE(false, "To be implemented.");
#endif
}
/**
* Matrix Descriptor of a memory buffer.
*
......
paddle/fluid/operators/math/math_function.h
浏览文件 @ fa002cd6
......@@ -21,8 +21,10 @@ limitations under the License. */
#ifdef PADDLE_USE_OPENBLAS
#include <cblas.h>
#ifdef LAPACK_FOUND
#include <lapacke.h>
#endif
#endif
#ifndef LAPACK_FOUND
extern "C" {
......
paddle/fluid/operators/mean_iou_op.cc 0 → 100644
浏览文件 @ fa002cd6
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/mean_iou_op.h"
namespace paddle {
namespace operators {
class MeanIoUOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Predictions"),
"Input (Predictions) of MeanIoU op should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Labels"),
"Input (labels) of MeanIoU op should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("OutMeanIou"),
"Output (OutMeanIou) of MeanIoU op should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("OutWrong"),
"Output (OutWrong) of MeanIoU op should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("OutCorrect"),
"Output (OutWrong) of MeanIoU op should not be null.");
int64_t num_classes =
static_cast<int64_t>(ctx->Attrs().Get<int>("num_classes"));
ctx->SetOutputDim("OutMeanIou", {1});
ctx->SetOutputDim("OutWrong", {num_classes});
ctx->SetOutputDim("OutCorrect", {num_classes});
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Predictions")->type()),
ctx.GetPlace());
}
};
class MeanIoUOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("Predictions",
"(Tensor), A Tensor of prediction results for semantic labels"
" with type int32 or int64. The rank should be greater than 1.");
AddInput(
"Labels",
"(Tensor), A Tensor of ground truth labels with type int32 or int64."
"Its shape should be the same as Input(Predictions).");
AddInput("InWrongs",
"(vector<Tensor>), A list of Tensor with shape "
"[num_classes]. They are used to collect wrong number among "
"batches. Empty list is also valid here.")
.AsDuplicable()
.AsDispensable();
AddInput(
"InCorrects",
"(vector<Tensor>), A list of Tensor with shape "
"[num_classes]. They are used to collect correct number among batches. "
"Empty list is also valid here.")
.AsDuplicable()
.AsDispensable();
AddInput("InMeanIou",
"(vector<Tensor>), A list of Tensor that Output(mean_iou) should "
"be added to. Empty list is also valid here.")
.AsDuplicable()
.AsDispensable();
AddOutput("OutMeanIou",
"(vector<Tensor>), A Tensor representing the"
" mean intersection-over-union with shape [1].");
AddOutput("OutWrong", "(Tensor), A Tensor with shape [num_classes]. ");
AddOutput("OutCorrect", "(Tensor), A Tensor with shape [num_classes]. ");
AddAttr<int>("num_classes", "(int), The possible number of labels.");
AddComment(R"DOC(
mean-IOU Operator.
Mean Intersection-Over-Union is a common evaluation metric for
semantic image segmentation, which first computes the IOU for each
semantic class and then computes the average over classes.
IOU is defined as follows:
IOU = true_positive / (true_positive + false_positive + false_negative).
It is based on pixel level area while "IOU Similarity Operator"
is based on area of rectangle.
)DOC");
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(mean_iou, ops::MeanIoUOp, ops::MeanIoUOpMaker,
paddle::framework::EmptyGradOpMaker);
REGISTER_OP_CPU_KERNEL(mean_iou, ops::MeanIoUKernel<int>,
ops::MeanIoUKernel<int32_t>,
ops::MeanIoUKernel<int64_t>);
paddle/fluid/operators/mean_iou_op.cu 0 → 100644
浏览文件 @ fa002cd6
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/mean_iou_op.h"
#include "paddle/fluid/platform/cuda_primitives.h"
#include "paddle/fluid/platform/gpu_info.h"
namespace paddle {
namespace operators {
using platform::PADDLE_CUDA_NUM_THREADS;
#define CUDA_1D_KERNEL_LOOP(i, n) \
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \
i += blockDim.x * gridDim.x)
template <typename T>
__global__ void CountCUDAKernel(const int num_classes, const int count,
const T* predictions, const T* labels,
int* wrong, int* correct) {
extern __shared__ int blcok_cache[];
int* wrong_c = blcok_cache;
int* correct_c = blcok_cache + num_classes;
// init cache
for (int i = threadIdx.x; i < num_classes * 2; i += blockDim.x) {
blcok_cache[i] = 0;
}
__syncthreads();
T pred;
T label;
CUDA_1D_KERNEL_LOOP(i, count) {
pred = predictions[i];
label = labels[i];
if (pred == label) {
atomicAdd(correct_c + pred, 1);
} else {
atomicAdd(wrong_c + pred, 1);
atomicAdd(wrong_c + label, 1);
}
}
__syncthreads();
for (int i = threadIdx.x; i < num_classes; i += blockDim.x) {
atomicAdd(wrong + i, wrong_c[i]);
atomicAdd(correct + i, correct_c[i]);
}
}
__global__ void ComputeIoUCUDAKernel(const int num_classes, int* wrong,
int* correct, float* ious, float* iou) {
__shared__ int valid_count_c;
if (threadIdx.x == 0) {
valid_count_c = 0;
}
__syncthreads();
CUDA_1D_KERNEL_LOOP(i, num_classes) {
int wrong_n = wrong[i];
int correct_n = correct[i];
int denominator = wrong_n + correct_n;
if (denominator > 0) {
atomicAdd(&valid_count_c, 1);
ious[i] = static_cast<float>(correct_n) / denominator;
} else {
ious[i] = 0;
}
}
__syncthreads();
if (threadIdx.x == 0) {
float iou_sum = 0;
for (int i = 0; i < num_classes; ++i) {
iou_sum += ious[i];
}
iou[0] += iou_sum / valid_count_c;
}
}
template <typename T>
class MeanIoUCUDAOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto& place = *ctx.template device_context<platform::CUDADeviceContext>()
.eigen_device();
// get input and output tensor
auto* predictions = ctx.Input<Tensor>("Predictions");
auto* labels = ctx.Input<Tensor>("Labels");
auto* out_mean_iou = ctx.Output<Tensor>("OutMeanIou");
auto* out_wrong = ctx.Output<Tensor>("OutWrong");
auto* out_correct = ctx.Output<Tensor>("OutCorrect");
int num_classes = static_cast<int>(ctx.Attr<int>("num_classes"));
// Get data ptr
const T* predictions_data = predictions->data<T>();
const T* labels_data = labels->data<T>();
int* out_wrong_data = out_wrong->mutable_data<int>(ctx.GetPlace());
int* out_correct_data = out_correct->mutable_data<int>(ctx.GetPlace());
float* out_mean_iou_data =
out_mean_iou->mutable_data<float>(ctx.GetPlace());
// Get Eigen tensor
auto out_mean_iou_t = EigenTensor<float, 1>::From(*out_mean_iou);
auto out_wrong_t = EigenTensor<int, 1>::From(*out_wrong);
auto out_correct_t = EigenTensor<int, 1>::From(*out_correct);
// Temporary tensor
Tensor ious;
float* ious_data = ious.mutable_data<float>(
{static_cast<int64_t>(num_classes)}, ctx.GetPlace());
auto ious_t = EigenTensor<float, 1>::From(ious);
// Init out_wrong, out_correct and out_mean_iou
out_wrong_t.device(place) = out_wrong_t.constant(0);
out_correct_t.device(place) = out_correct_t.constant(0);
out_mean_iou_t.device(place) = out_mean_iou_t.constant(0.0f);
// collect pre wrong, correct and mean_iou
auto in_mean_ious = ctx.MultiInput<Tensor>("InMeanIou");
for (int i = 0; i < in_mean_ious.size(); ++i) {
out_mean_iou_t.device(place) +=
EigenTensor<float, 1>::From(*in_mean_ious[i]);
}
auto in_wrongs = ctx.MultiInput<Tensor>("InWrongs");
for (int i = 0; i < in_wrongs.size(); ++i) {
out_wrong_t.device(place) += EigenTensor<int, 1>::From(*in_wrongs[i]);
}
auto in_corrects = ctx.MultiInput<Tensor>("InCorrects");
for (int i = 0; i < in_corrects.size(); ++i) {
out_correct_t.device(place) += EigenTensor<int, 1>::From(*in_corrects[i]);
}
// compute
auto stream = ctx.cuda_device_context().stream();
int block = PADDLE_CUDA_NUM_THREADS;
int grid = (predictions->numel() + block - 1) / block;
int cache_size = (num_classes * 2 + 1) * sizeof(int);
CountCUDAKernel<T><<<grid, block, cache_size, stream>>>(
num_classes, predictions->numel(), predictions_data, labels_data,
out_wrong_data, out_correct_data);
ctx.device_context().Wait();
ComputeIoUCUDAKernel<<<1, block, 0, stream>>>(num_classes, out_wrong_data,
out_correct_data, ious_data,
out_mean_iou_data);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(mean_iou, ops::MeanIoUCUDAOpKernel<int>,
ops::MeanIoUCUDAOpKernel<int64_t>,
ops::MeanIoUCUDAOpKernel<int32_t>);
paddle/fluid/operators/mean_iou_op.h 0 → 100644
浏览文件 @ fa002cd6
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <algorithm>
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T, int D, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
template <typename T>
class MeanIoUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto& place = *ctx.template device_context<platform::CPUDeviceContext>()
.eigen_device();
// get input and output tensor
auto* predictions = ctx.Input<Tensor>("Predictions");
auto* labels = ctx.Input<Tensor>("Labels");
auto* out_mean_iou = ctx.Output<Tensor>("OutMeanIou");
auto* out_wrong = ctx.Output<Tensor>("OutWrong");
auto* out_correct = ctx.Output<Tensor>("OutCorrect");
int num_classes = static_cast<int>(ctx.Attr<int>("num_classes"));
// get data ptr
const T* predictions_data = predictions->data<T>();
const T* labels_data = labels->data<T>();
float* out_mean_iou_data =
out_mean_iou->mutable_data<float>(ctx.GetPlace());
int* out_wrong_data = out_wrong->mutable_data<int>(ctx.GetPlace());
int* out_correct_data = out_correct->mutable_data<int>(ctx.GetPlace());
// get eigen tensor
auto out_mean_iou_t = EigenTensor<float, 1>::From(*out_mean_iou);
auto out_wrong_t = EigenTensor<int, 1>::From(*out_wrong);
auto out_correct_t = EigenTensor<int, 1>::From(*out_correct);
// Tmp tensor
Tensor denominator;
Tensor valid_count;
Tensor iou_sum;
// get data ptr of tmp tensor
int* denominator_data = denominator.mutable_data<int>(
{static_cast<int64_t>(num_classes)}, ctx.GetPlace());
int* valid_count_data = valid_count.mutable_data<int>({1}, ctx.GetPlace());
float* iou_sum_data = iou_sum.mutable_data<float>({1}, ctx.GetPlace());
// get eigen tensor of tmp tensor
auto denominator_t = EigenTensor<int, 1>::From(denominator);
auto valid_count_t = EigenTensor<int, 1>::From(valid_count);
auto iou_sum_t = EigenTensor<float, 1>::From(iou_sum);
// init out_wrong, out_correct and out_mean_iou
out_wrong_t = out_wrong_t.constant(0);
out_correct_t = out_correct_t.constant(0);
out_mean_iou_t = out_mean_iou_t.constant(0);
// collect pre wrong, correct and mean_iou
auto in_mean_ious = ctx.MultiInput<Tensor>("InMeanIou");
for (size_t i = 0; i < in_mean_ious.size(); ++i) {
out_mean_iou_t.device(place) +=
EigenTensor<float, 1>::From(*in_mean_ious[i]);
}
auto in_wrongs = ctx.MultiInput<Tensor>("InWrongs");
for (size_t i = 0; i < in_wrongs.size(); ++i) {
out_wrong_t.device(place) += EigenTensor<int, 1>::From(*in_wrongs[i]);
}
auto in_corrects = ctx.MultiInput<Tensor>("InCorrects");
for (size_t i = 0; i < in_corrects.size(); ++i) {
out_correct_t.device(place) += EigenTensor<int, 1>::From(*in_corrects[i]);
}
// compute
for (int64_t i = 0; i < predictions->numel(); ++i) {
if (predictions_data[i] == labels_data[i]) {
out_correct_data[predictions_data[i]] += 1;
} else {
out_wrong_data[labels_data[i]] += 1;
out_wrong_data[predictions_data[i]] += 1;
}
}
denominator_t = out_wrong_t + out_correct_t;
valid_count_t =
(denominator_t > denominator_t.constant(0.0f)).cast<int>().sum();
for (int i = 0; i < num_classes; ++i) {
if (denominator_data[i] == 0) {
denominator_data[i] = 1;
}
}
iou_sum_t =
(out_correct_t.cast<float>() / denominator_t.cast<float>()).sum();
out_mean_iou_data[0] += (iou_sum_data[0] / valid_count_data[0]);
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/merge_ids_op.cc 0 → 100644
浏览文件 @ fa002cd6
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/merge_ids_op.h"
namespace paddle {
namespace operators {
class MergeIdsOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}");
AddInput(
"X",
"(LoDTensors) multi input tensor with shape{batch_num, N}, N is the "
"size of embedding table")
.AsDuplicable();
AddOutput("Out", "(LoDTensor) The merged outputs of the input tensors.");
AddComment(R"DOC(
Merge multi LoDTensor's into one according to Ids's shard num.
split_ids_op -> prefetch_op -> merge_ids_op
merge_ids_op should be used after split_ids_op and prefetch_op, split_ids_op
will split input Ids into multiple tensors according to Id's shard number.
prefetch_op will send them to parameter server to prefetch embedding value
back. During split, the order of ids is disordered. In merge_ids_op we use
the original Ids to restore the order of the fetched embedding value and
also pass the lod information to the merged output.
Example:
Ids = [1,2,3,4,5,6] # 3 shared
split_ids_op ->
Id0 = [3, 6] # id % 3 == 0
Id1 = [1, 4] # id % 3 == 1
Id2 = [2, 5] # id % 3 == 2
prefetch_op ->
X0 = [[0.3 0.3] # 3
[0.6 0.6]] # 6
X1 = [[0.1 0.1] # 1
[0.4 0.4]] # 4
X2 = [[0.2 0.2] # 2
[0.5 0.5]] # 5
merge_ids_op ->
Out = [[0.1 0.1] # 1
[0.2 0.2] # 2
[0.3 0.3] # 3
[0.4 0.4] # 4
[0.5 0.5] # 5
[0.6 0.6]] # 6
)DOC");
}
};
class MergeIdsOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Ids"), "MergeIdsOp must has input Ids.");
PADDLE_ENFORCE(ctx->HasInputs("X"), "MergeIdsOp must has input X.");
PADDLE_ENFORCE(ctx->HasOutput("Out"), "MergeIdsOp must has output Out.");
auto ids_var_type = ctx->GetInputsVarType("Ids").front();
auto ids_dims = ctx->GetInputDim("Ids");
if (ids_var_type == framework::proto::VarType::LOD_TENSOR) {
PADDLE_ENFORCE_EQ(ids_dims.size(), 2);
PADDLE_ENFORCE_EQ(ids_dims[1], 1);
}
auto x_var_type = ctx->GetInputsVarType("X");
for (auto &var_type : x_var_type) {
PADDLE_ENFORCE_EQ(var_type, framework::proto::VarType::LOD_TENSOR,
"input X only support lod tensors");
}
ctx->ShareLoD("Ids", "Out");
}
private:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext &ctx) const override {
return framework::OpKernelType(
framework::ToDataType(
ctx.MultiInput<framework::Tensor>("X").front()->type()),
ctx.GetPlace());
}
};
class MergeIdsOpInferVarType : public framework::VarTypeInference {
public:
void operator()(const framework::OpDesc &op_desc,
framework::BlockDesc *block) const override {
auto *input_var = block->Var(op_desc.Input("Ids")[0]);
for (auto &out_var : op_desc.Output("Out")) {
block->Var(out_var)->SetType(input_var->GetType());
}
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(merge_ids, ops::MergeIdsOp, ops::MergeIdsOpMaker,
ops::MergeIdsOpInferVarType);
REGISTER_OP_CPU_KERNEL(
merge_ids, ops::MergeIdsOpKernel<paddle::platform::CPUPlace, float>);
paddle/fluid/operators/merge_ids_op.h 0 → 100644
浏览文件 @ fa002cd6
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
namespace paddle {
namespace operators {
template <typename DeviceContext, typename T>
class MergeIdsOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
auto place = ctx.GetPlace();
if (!platform::is_cpu_place(place)) {
PADDLE_THROW("MergeIds do not support GPU kernel");
}
VLOG(3) << "run in MergeIdsOpKernel";
const auto *ids_var = ctx.InputVar("Ids");
PADDLE_ENFORCE(ids_var->IsType<framework::LoDTensor>(),
"only support to merge Ids of LoDTensor");
const auto &ids_tensor = ids_var->Get<framework::LoDTensor>();
const auto &ids_dims = ids_tensor.dims();
const int64_t *ids = ids_tensor.data<int64_t>();
auto x_tensors = ctx.MultiInput<framework::LoDTensor>("X");
auto *out = ctx.Output<framework::LoDTensor>("Out");
int batch_size = 0;
int embedding_size = 0;
for (auto &input : x_tensors) {
if (framework::product(input->dims()) != 0) {
if (embedding_size == 0) {
embedding_size = input->dims()[1];
}
PADDLE_ENFORCE_EQ(embedding_size, input->dims()[1],
"embedding size of all input should be the same");
batch_size += input->dims()[0];
}
}
PADDLE_ENFORCE_EQ(
batch_size, ids_dims[0],
"the batch size of ids and merged embedding value should be the same");
const size_t shard_num = x_tensors.size();
if (shard_num == 1) {
VLOG(3) << "only one shard, we can copy the data directly";
TensorCopy(*x_tensors[0], place, out);
} else {
std::vector<int> in_indexs(shard_num, 0);
auto *out_data = out->mutable_data<T>(
framework::make_ddim({batch_size, embedding_size}), place);
// copy data from ins[shard_num] to out.
for (int i = 0; i < ids_dims[0]; ++i) {
int64_t id = ids[i];
size_t shard_id = static_cast<size_t>(id) % shard_num;
int index = in_indexs[shard_id];
memcpy(out_data + embedding_size * i,
x_tensors[shard_id]->data<T>() + index * embedding_size,
sizeof(T) * embedding_size);
in_indexs[shard_id] += 1;
}
for (size_t i = 0; i < shard_num; ++i) {
PADDLE_ENFORCE_EQ(in_indexs[i], x_tensors[i]->dims()[0],
"after merge, all data in x_tensor should be used");
}
}
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/reader/create_batch_reader_op.cc
浏览文件 @ fa002cd6
......@@ -20,7 +20,7 @@ namespace reader {
class BatchReader : public framework::DecoratedReader {
public:
BatchReader(ReaderBase* reader, int batch_size)
BatchReader(const std::shared_ptr<ReaderBase>& reader, int batch_size)
: DecoratedReader(reader), batch_size_(batch_size) {
buffer_.reserve(batch_size_);
}
......
paddle/fluid/operators/reader/create_custom_reader_op.cc
浏览文件 @ fa002cd6
......@@ -22,7 +22,8 @@ namespace reader {
class CustomReader : public framework::DecoratedReader {
public:
CustomReader(ReaderBase* reader, const framework::BlockDesc& sub_block,
CustomReader(const std::shared_ptr<ReaderBase>& reader,
const framework::BlockDesc& sub_block,
const std::vector<std::string>& source_var_names,
const std::vector<std::string>& sink_var_names)
: DecoratedReader(reader),
......
paddle/fluid/operators/reader/create_double_buffer_reader_op.cc
浏览文件 @ fa002cd6
......@@ -34,7 +34,8 @@ static constexpr size_t kChannelSize = 1; // kCacheSize - 2
class DoubleBufferReader : public framework::DecoratedReader {
public:
explicit DoubleBufferReader(
ReaderBase* reader, platform::Place target_place = platform::CPUPlace())
const std::shared_ptr<ReaderBase>& reader,
platform::Place target_place = platform::CPUPlace())
: DecoratedReader(reader), place_(target_place) {
cpu_tensor_cache_.resize(kCacheSize);
gpu_tensor_cache_.resize(kCacheSize);
......
paddle/fluid/operators/reader/create_multi_pass_reader_op.cc
浏览文件 @ fa002cd6
......@@ -21,7 +21,7 @@ namespace reader {
class MultiPassReader : public framework::DecoratedReader {
public:
MultiPassReader(ReaderBase* reader, int pass_num)
MultiPassReader(const std::shared_ptr<ReaderBase>& reader, int pass_num)
: DecoratedReader(reader), pass_num_(pass_num), pass_count_(0) {}
void ReadNext(std::vector<framework::LoDTensor>* out) override {
......
paddle/fluid/operators/reader/create_shuffle_reader_op.cc
浏览文件 @ fa002cd6
......@@ -23,7 +23,8 @@ namespace reader {
class ShuffleReader : public framework::DecoratedReader {
public:
ShuffleReader(ReaderBase* reader, size_t buffer_size, size_t seed = 0)
ShuffleReader(const std::shared_ptr<ReaderBase>& reader, size_t buffer_size,
size_t seed = 0)
: DecoratedReader(reader), buffer_size_(buffer_size), seed_(seed) {
VLOG(10) << "Create shuffle reader of " << reader_;
if (seed_ == 0) {
......
paddle/fluid/operators/reader/create_threaded_reader_op.cc
浏览文件 @ fa002cd6
......@@ -21,7 +21,8 @@ namespace reader {
class ThreadedReader : public framework::DecoratedReader {
public:
explicit ThreadedReader(ReaderBase* reader) : DecoratedReader(reader) {}
explicit ThreadedReader(const std::shared_ptr<ReaderBase>& reader)
: DecoratedReader(reader) {}
void ReadNext(std::vector<framework::LoDTensor>* out) override {
std::lock_guard<std::mutex> lock(mutex_);
......
paddle/fluid/platform/cpu_info.cc
浏览文件 @ fa002cd6
......@@ -21,12 +21,17 @@ limitations under the License. */
#include <unistd.h>
#endif
#include <algorithm>
#include "gflags/gflags.h"
DEFINE_double(fraction_of_cpu_memory_to_use, 1,
"Default use 100% of CPU memory for PaddlePaddle,"
"reserve the rest for page tables, etc");
DEFINE_uint64(
initial_cpu_memory_in_mb, 500,
"Default initial 500MB of CPU memory for PaddlePaddle, in MD unit.");
DEFINE_double(
fraction_of_cuda_pinned_memory_to_use, 0.5,
"Default use 50% of CPU memory as the pinned_memory for PaddlePaddle,"
......@@ -54,7 +59,10 @@ inline size_t CpuTotalPhysicalMemory() {
size_t CpuMaxAllocSize() {
// For distributed systems, it requires configuring and limiting
// the fraction of memory to use.
return FLAGS_fraction_of_cpu_memory_to_use * CpuTotalPhysicalMemory();
return std::min(
static_cast<size_t>(FLAGS_fraction_of_cpu_memory_to_use *
CpuTotalPhysicalMemory()),
static_cast<size_t>(FLAGS_initial_cpu_memory_in_mb * 1 << 20));
}
size_t CpuMinChunkSize() {
......
paddle/fluid/platform/device_tracer.cc
浏览文件 @ fa002cd6
......@@ -322,7 +322,6 @@ class DeviceTracerImpl : public DeviceTracer {
DisableActivity();
dynload::cuptiUnsubscribe(subscriber_);
CUPTI_CALL(dynload::cuptiGetTimestamp(&end_ns_));
PADDLE_ENFORCE(dynload::cuptiFinalize());
enabled_ = false;
}
......
paddle/fluid/platform/dynload/cupti.h
浏览文件 @ fa002cd6
......@@ -72,7 +72,6 @@ extern void *cupti_dso_handle;
__macro(cuptiGetResultString); \
__macro(cuptiActivityGetNumDroppedRecords); \
__macro(cuptiActivityFlushAll); \
__macro(cuptiFinalize); \
__macro(cuptiSubscribe); \
__macro(cuptiUnsubscribe); \
__macro(cuptiEnableCallback); \
......
paddle/fluid/platform/nccl_helper.h
浏览文件 @ fa002cd6
......@@ -41,6 +41,11 @@ inline ncclDataType_t ToNCCLDataType(std::type_index type) {
}
}
// NOTE(minqiyang): according to the ncclGroupEnd documentations:
// https://docs.nvidia.com/deeplearning/sdk/nccl-api/ncclapidoc.html,
// ncclGroupEnd will wait for all communicators to be initialized, which will
// cause blocking problem when a runtime_error was thrown, so try only guard
// NCCL actions when use it.
class NCCLGroupGuard {
public:
static std::mutex &NCCLMutex() {
......
paddle/fluid/pybind/pybind.cc
浏览文件 @ fa002cd6
......@@ -413,6 +413,9 @@ All parameter, weight, gradient are variables in Paddle.
py::class_<framework::Executor>(m, "Executor")
.def(py::init<const platform::Place &>())
#ifdef PADDLE_WITH_DISTRIBUTE
.def("complete", &Executor::Complete)
#endif
.def("run",
(void (Executor::*)(const ProgramDesc &, Scope *, int, bool, bool)) &
Executor::Run);
......
paddle/math/MathFunctions.h
浏览文件 @ fa002cd6
......@@ -12,8 +12,7 @@ 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. */
#ifndef MATHFUNCTIONS_H_
#define MATHFUNCTIONS_H_
#pragma once
#ifdef PADDLE_WITH_MKLML
#include <mkl_cblas.h>
......@@ -21,7 +20,7 @@ limitations under the License. */
#include <mkl_vml_functions.h>
#endif
#if defined(PADDLE_USE_VECLIB)
#ifdef PADDLE_USE_VECLIB
extern "C" {
#include <cblas.h>
#include <clapack.h>
......@@ -30,8 +29,10 @@ extern "C" {
#ifdef PADDLE_USE_OPENBLAS
#include <cblas.h>
#ifdef LAPACK_FOUND
#include <lapacke.h>
#endif
#endif
#ifndef LAPACK_FOUND
extern "C" {
......@@ -126,5 +127,3 @@ template <class T>
void vTanh(const int n, const T* a, T* r);
} // namespace paddle
#endif // MATHFUNCTIONS_H_
paddle/scripts/paddle_build.sh
浏览文件 @ fa002cd6
......@@ -132,7 +132,8 @@ EOF
-DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake \
-DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \
-DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
-DWITH_CONTRIB=${WITH_CONTRIB:-ON}
-DWITH_CONTRIB=${WITH_CONTRIB:-ON} \
-DWITH_ANAKIN=ON
}
function abort(){
......
python/paddle/batch.py
浏览文件 @ fa002cd6
......@@ -15,7 +15,7 @@
__all__ = ['batch']
def batch(reader, batch_size, drop_last=False):
def batch(reader, batch_size, drop_last=True):
"""
Create a batched reader.
......
python/paddle/fluid/framework.py
浏览文件 @ fa002cd6
......@@ -382,7 +382,7 @@ class Operator(object):
'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', 'recv',
'listen_and_serv', 'parallel_do', 'save_combine', 'load_combine',
'ncclInit', 'channel_create', 'channel_close', 'channel_send',
'channel_recv', 'select'
'channel_recv', 'select', 'gen_nccl_id'
}
def __init__(self,
......
python/paddle/fluid/layers/nn.py
浏览文件 @ fa002cd6
此差异已折叠。
python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py
浏览文件 @ fa002cd6
......@@ -96,10 +96,11 @@ def train(use_cuda, train_program, params_dirname):
train_reader = paddle.batch(
paddle.reader.shuffle(
cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10),
batch_size=BATCH_SIZE)
batch_size=BATCH_SIZE,
drop_last=False)
test_reader = paddle.batch(
paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE)
paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE, drop_last=False)
def event_handler(event):
if isinstance(event, fluid.EndStepEvent):
......
python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py
浏览文件 @ fa002cd6
......@@ -73,10 +73,11 @@ def train(use_cuda, train_program, params_dirname):
train_reader = paddle.batch(
paddle.reader.shuffle(
cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10),
batch_size=BATCH_SIZE)
batch_size=BATCH_SIZE,
drop_last=False)
test_reader = paddle.batch(
paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE)
paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE, drop_last=False)
def event_handler(event):
if isinstance(event, fluid.EndStepEvent):
......
python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py
浏览文件 @ fa002cd6
......@@ -87,7 +87,9 @@ def train(use_cuda, train_program, params_dirname):
def event_handler(event):
if isinstance(event, fluid.EndEpochEvent):
test_reader = paddle.batch(
paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE)
paddle.dataset.imdb.test(word_dict),
batch_size=BATCH_SIZE,
drop_last=False)
avg_cost, acc = trainer.test(
reader=test_reader, feed_order=['words', 'label'])
......@@ -113,7 +115,8 @@ def train(use_cuda, train_program, params_dirname):
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.imdb.train(word_dict), buf_size=25000),
batch_size=BATCH_SIZE)
batch_size=BATCH_SIZE,
drop_last=False)
trainer.train(
num_epochs=1,
......
python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py
浏览文件 @ fa002cd6
......@@ -56,7 +56,7 @@ BATCH_SIZE = 200
# fix the order of training data
train_reader = paddle.batch(
paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE)
paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE, drop_last=False)
# train_reader = paddle.batch(
# paddle.reader.shuffle(
......
python/paddle/fluid/tests/unittests/test_mean_iou.py 0 → 100644
浏览文件 @ fa002cd6
# 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 division
import unittest
import numpy as np
from op_test import OpTest
def compute_mean_iou(predictions, labels, num_classes, in_wrongs, in_corrects,
in_mean_ious):
assert predictions.shape == labels.shape
predictions = predictions.flatten()
labels = labels.flatten()
out_wrong = np.zeros([num_classes]).astype("int32")
for _, wrong in in_wrongs:
out_wrong += wrong
out_correct = np.zeros([num_classes]).astype("int32")
for _, correct in in_corrects:
out_correct += correct
for pred, label in zip(predictions, labels):
if pred == label:
out_correct[pred] += 1
else:
out_wrong[pred] += 1
out_wrong[label] += 1
denominator = out_wrong + out_correct
valid_count = (denominator != 0).sum()
denominator = np.where(denominator > 0, denominator,
np.ones(denominator.shape))
mean_iou = (out_correct / denominator).sum() / valid_count
for _, in_mean_iou in in_mean_ious:
mean_iou += in_mean_iou
return mean_iou, out_wrong, out_correct
class TestMeanIOUOp(OpTest):
def setUp(self):
self.config()
self.op_type = "mean_iou"
predictions = np.random.randint(0, self.num_classes,
self.image_size).astype("int32")
labels = np.random.randint(0, self.num_classes,
self.image_size).astype("int32")
in_wrongs = []
for i in range(self.in_wrong_num):
in_wrongs.append(("in_wrong_%d" % i, np.random.randint(
0, 10, [self.num_classes]).astype("int32")))
in_corrects = []
for i in range(self.in_correct_num):
in_corrects.append(("in_correct_%d" % i, np.random.randint(
0, 10, [self.num_classes]).astype("int32")))
in_mean_ious = []
for i in range(self.in_mean_iou_num):
in_mean_ious.append(("in_mean_iou_%d" % i, np.random.uniform(
0, 1, [1]).astype("float32")))
self.inputs = {
'Predictions': predictions,
'Labels': labels,
'InWrongs': in_wrongs,
'InCorrects': in_corrects,
'InMeanIou': in_mean_ious
}
self.attrs = {'num_classes': long(self.num_classes)}
mean_iou, out_wrong, out_correct = compute_mean_iou(
predictions, labels, self.num_classes, in_wrongs, in_corrects,
in_mean_ious)
self.outputs = {
'OutMeanIou': mean_iou,
'OutWrong': out_wrong,
'OutCorrect': out_correct
}
def config(self):
self.num_classes = 10
self.image_size = [128, 128]
self.in_wrong_num = 0
self.in_correct_num = 0
self.in_mean_iou_num = 0
def test_check_output(self):
self.check_output()
class TestCase1(TestMeanIOUOp):
def config(self):
self.num_classes = 5
self.image_size = [100, 128]
self.in_wrong_num = 2
self.in_correct_num = 2
self.in_mean_iou_num = 2
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/tests/unittests/test_merge_ids_op.py 0 → 100644
浏览文件 @ fa002cd6
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
import numpy as np
from op_test import OpTest
class TestMergeIdsOp(OpTest):
def setUp(self):
self.op_type = "merge_ids"
ids = np.array([[0], [2], [2], [3], [5], [5], [6]]).astype('int64')
x0 = np.array([[0.1, 0.2], [0.2, 0.3], [0.3, 0.4]]).astype('float32')
x1 = np.array([]).astype('float32')
x2 = np.array([[0.4, 0.5], [0.4, 0.5], [0.5, 0.6],
[0.5, 0.6]]).astype('float32')
out = np.array([[0.1, 0.2], [0.4, 0.5], [0.4, 0.5], [0.2, 0.3],
[0.5, 0.6], [0.5, 0.6], [0.3, 0.4]]).astype('float32')
self.inputs = {'Ids': ids, "X": [('x0', x0), ('x1', x1), ('x2', x2)]}
self.outputs = {'Out': out}
def test_check_output(self):
self.check_output()
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/transpiler/distribute_transpiler.py
浏览文件 @ fa002cd6
......@@ -515,35 +515,38 @@ class DistributeTranspiler:
grad_to_block_id, None)
# process distributed lookup_table
prefetch_block = None
prefetch_var_name_to_block_id = []
if self.has_distributed_lookup_table:
pserver_index = self.pserver_endpoints.index(endpoint)
table_opt_block = self._create_table_optimize_block(
pserver_index, pserver_program, pre_block_idx, grad_to_block_id)
prefetch_block = self._create_prefetch_block(
prefetch_var_name_to_block_id = self._create_prefetch_block(
pserver_index, pserver_program, table_opt_block)
# NOTE: if has_distributed_lookup_table is False, then prefetch_block will
# not be executed, so it's safe to use optimize_block to hold the place
if self.has_distributed_lookup_table:
assert prefetch_block is not None
assert len(prefetch_var_name_to_block_id) > 0
else:
assert prefetch_block is None
prefetch_block = pserver_program.global_block()
assert len(prefetch_var_name_to_block_id) == 0
attrs = {
"OptimizeBlock": pserver_program.block(1),
"endpoint": endpoint,
"Fanin": self.trainer_num,
"sync_mode": self.sync_mode,
"grad_to_block_id": grad_to_block_id
}
if len(prefetch_var_name_to_block_id) > 0:
attrs['prefetch_var_name_to_block_id'] \
= prefetch_var_name_to_block_id
# step5 append the listen_and_serv op
pserver_program.global_block().append_op(
type="listen_and_serv",
inputs={'X': recv_inputs},
outputs={},
attrs={
"OptimizeBlock": pserver_program.block(1),
"endpoint": endpoint,
"Fanin": self.trainer_num,
"PrefetchBlock": prefetch_block,
"sync_mode": self.sync_mode,
"grad_to_block_id": grad_to_block_id
})
attrs=attrs)
pserver_program.sync_with_cpp()
return pserver_program
......@@ -608,8 +611,15 @@ class DistributeTranspiler:
def _replace_lookup_table_op_with_prefetch(self, program,
pserver_endpoints):
# 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op
self.prefetch_input_vars = None
self.prefetch_output_vars = None
# self.all_prefetch_input_vars =
# [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1]
# [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]]
self.all_prefetch_input_vars = []
# self.all_prefetch_input_vars =
# [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1]
# [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]]
self.all_prefetch_output_vars = []
continue_search_lookup_table_op = True
while continue_search_lookup_table_op:
......@@ -619,26 +629,27 @@ class DistributeTranspiler:
if op.type == LOOKUP_TABLE_TYPE:
continue_search_lookup_table_op = True
op_index = list(all_ops).index(op)
lookup_table_op_index = list(all_ops).index(op)
ids_name = op.input("Ids")
out_name = op.output("Out")
if self.prefetch_input_vars is None:
ids_var = program.global_block().vars[ids_name[0]]
self.prefetch_input_vars = self.create_splited_vars(
source_var=ids_var,
block=program.global_block(),
tag="_prefetch_in_")
if self.prefetch_output_vars is None:
out_var = program.global_block().vars[out_name[0]]
self.prefetch_output_vars = self.create_splited_vars(
source_var=out_var,
block=program.global_block(),
tag="_prefetch_out_")
ids_var = program.global_block().vars[ids_name[0]]
prefetch_input_vars = self.create_splited_vars(
source_var=ids_var,
block=program.global_block(),
tag="_prefetch_in_")
self.all_prefetch_input_vars.append(prefetch_input_vars)
out_var = program.global_block().vars[out_name[0]]
prefetch_output_vars = self.create_splited_vars(
source_var=out_var,
block=program.global_block(),
tag="_prefetch_out_")
self.all_prefetch_output_vars.append(prefetch_output_vars)
# insert split_ids_op
program.global_block().insert_op(
index=op_index,
index=lookup_table_op_index,
type="split_ids",
inputs={
'Ids': [
......@@ -646,14 +657,14 @@ class DistributeTranspiler:
for varname in ids_name
]
},
outputs={"Out": self.prefetch_input_vars})
outputs={"Out": prefetch_input_vars})
# insert prefetch_op
program.global_block().insert_op(
index=op_index + 1,
index=lookup_table_op_index + 1,
type="prefetch",
inputs={'X': self.prefetch_input_vars},
outputs={"Out": self.prefetch_output_vars},
inputs={'X': prefetch_input_vars},
outputs={"Out": prefetch_output_vars},
attrs={
"epmap": pserver_endpoints,
RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
......@@ -661,16 +672,21 @@ class DistributeTranspiler:
# insert concat_op
program.global_block().insert_op(
index=op_index + 2,
type="concat",
inputs={'X': self.prefetch_output_vars},
index=lookup_table_op_index + 2,
type="merge_ids",
inputs={
'Ids': [
program.global_block().vars[varname]
for varname in ids_name
],
'X': prefetch_output_vars
},
outputs={
"Out": [
program.global_block().vars[varname]
for varname in out_name
]
},
attrs={"axis": 0})
})
# delete lookup_table_op
delete_ops(program.global_block(), [op])
......@@ -709,30 +725,34 @@ class DistributeTranspiler:
optimize_block):
# STEP: create prefetch block
table_var = pserver_program.global_block().vars[self.table_name]
prefetch_block = pserver_program.create_block(optimize_block.idx)
trainer_ids = self.prefetch_input_vars[pserver_index]
pserver_ids = pserver_program.global_block().create_var(
name=trainer_ids.name,
type=trainer_ids.type,
shape=trainer_ids.shape,
dtype=trainer_ids.dtype)
trainer_out = self.prefetch_output_vars[pserver_index]
pserver_out = pserver_program.global_block().create_var(
name=trainer_out.name,
type=trainer_out.type,
shape=trainer_out.shape,
dtype=trainer_out.dtype)
prefetch_block.append_op(
type="lookup_sparse_table",
inputs={'Ids': pserver_ids,
"W": table_var},
outputs={"Out": pserver_out},
attrs={
"is_sparse": True, # has no effect on lookup_table op
"is_distributed": True,
"padding_idx": -1
})
return prefetch_block
prefetch_var_name_to_block_id = []
for index in range(len(self.all_prefetch_input_vars)):
prefetch_block = pserver_program.create_block(optimize_block.idx)
trainer_ids = self.all_prefetch_input_vars[index][pserver_index]
pserver_ids = pserver_program.global_block().create_var(
name=trainer_ids.name,
type=trainer_ids.type,
shape=trainer_ids.shape,
dtype=trainer_ids.dtype)
trainer_out = self.all_prefetch_output_vars[index][pserver_index]
pserver_out = pserver_program.global_block().create_var(
name=trainer_out.name,
type=trainer_out.type,
shape=trainer_out.shape,
dtype=trainer_out.dtype)
prefetch_block.append_op(
type="lookup_sparse_table",
inputs={'Ids': pserver_ids,
"W": table_var},
outputs={"Out": pserver_out},
attrs={
"is_sparse": True, # has no effect on lookup_table op
"is_distributed": True,
"padding_idx": -1
})
prefetch_var_name_to_block_id.append(trainer_ids.name + ":" + str(
prefetch_block.idx))
return prefetch_var_name_to_block_id
def _create_table_optimize_block(self, pserver_index, pserver_program,
pre_block_idx, grad_to_block_id):
......
python/paddle/trainer_config_helpers/attrs.py
浏览文件 @ fa002cd6
......@@ -240,14 +240,15 @@ class ExtraLayerAttribute(object):
:type error_clipping_threshold: float
:param drop_rate: Dropout rate. Dropout will create a mask on layer output.
The dropout rate is the zero rate of this mask. The
details of what dropout is please refer to `here
<https://www.cs.toronto.edu/~hinton/absps/
JMLRdropout.pdf>`_.
details of what dropout is please refer to `JMLRdropout
<https://www.cs.toronto.edu/~hinton/absps/JMLRdropout.pdf
>`_.
:type drop_rate: float
:param device: device ID of layer. device=-1, use CPU. device>=0, use GPU.
The details allocation in parallel_nn please refer to `here
<http://www.paddlepaddle.org/doc/ui/cmd_argument/
use_case.html#case-2-specify-layers-in-different-devices>`_.
The details allocation in parallel_nn please refer to `use_case
<https://github.com/PaddlePaddle/Paddle/blob/develop/doc/v2
/howto/cmd_parameter/use_case_en.md#case-2-specify-layers-in
-different-devices>`_.
:type device: int
"""
......
python/paddle/trainer_config_helpers/layers.py
浏览文件 @ fa002cd6
......@@ -2556,7 +2556,7 @@ def img_conv_layer(input,
the output will be obtained by concatenating the two results.
The details of grouped convolution, please refer to:
`ImageNet Classification with Deep Convolutional Neural Networks
`ImageNet Classification With Deep Convolutional Neural Networks
<http://www.cs.toronto.edu/~kriz/imagenet_classification_with_deep_convolutional.pdf>`_
The example usage is:
......@@ -5678,8 +5678,8 @@ def warp_ctc_layer(input,
<https://github.com/baidu-research/warp-ctc>`_ library, which is used in
`Deep Speech 2: End-toEnd Speech Recognition in English and Mandarin
<https://arxiv.org/pdf/1512.02595v1.pdf>`_, to compute Connectionist Temporal
Classification (CTC) loss. Besides, another `warp-ctc
<https://github.com/gangliao/warp-ctc>`_ repository, which is forked from
Classification (CTC) loss. Besides, another `warp-ctc repository
<https://github.com/gangliao/warp-ctc>`_ , which is forked from
the official one, is maintained to enable more compiling options. During the
building process, PaddlePaddle will clone the source codes, build and
install it to :code:`third_party/install/warpctc` directory.
......
python/paddle/v2/minibatch.py
浏览文件 @ fa002cd6
......@@ -15,7 +15,7 @@
__all__ = ['batch']
def batch(reader, batch_size, drop_last=False):
def batch(reader, batch_size, drop_last=True):
"""
Create a batched reader.
......
tools/codestyle/pylint_pre_commit.hook
浏览文件 @ fa002cd6
......@@ -7,13 +7,13 @@ DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
export PYTHONPATH=$DIR:$PYTHONPATH
# The trick to remove deleted files: https://stackoverflow.com/a/2413151
for file in $(git diff --cached --name-status | awk '$1 != "D" {print $2}'); do
for file in $(git diff --name-status | awk '$1 != "D" {print $2}'); do
pylint --disable=all --load-plugins=docstring_checker \
--enable=doc-string-one-line,doc-string-end-with,doc-string-with-all-args,doc-string-triple-quotes,doc-string-missing,doc-string-indent-error,doc-string-with-returns,doc-string-with-raises $file;
TOTAL_ERRORS=$(expr $TOTAL_ERRORS + $?);
done
#exit $TOTAL_ERRORS
exit $TOTAL_ERRORS
#For now, just warning:
exit 0
#exit 0
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