purine / purine2 Goto Github PK

I used the demo nin_cifar_10, the iteration is 50000, decay the learning by div 10.0 every 20000 iterator, and just get 87% accuracy rate.

Specialize Op template for mpi recv and send operations, test communication between two machines.

It seem that test_layers.cpp use the header file in caffe. But after I paste that files, still can't compile the test. There are many redefinition errors, some things are already defined in the common folder. So, which file do I need to paste?

make mpi.cpp with Error like this:
/home/yaocx/workspace/purine2/operations/src/mpi.cpp:14:44: error: invalid conversion from ‘const void_’ to ‘void_’ [-fpermissive]

I just add (void_) in line (14) of mpi.cpp file to make
MPI_CHECK(MPI_Isend((void_)inputs_[0]->cpu_data(), inputs_[0]->size().count(),
MPI_FLOAT, dest, tag, MPI_COMM_WORLD, &request));

or maybe you can change the makefile with flag -fpermissive

Any good advice?

create all the operations and layers.

I post my question here, hope more people will look at it.
I am using OpenMPI 1.6.5

(pyenv)[zxx@ga85 examples]$ make googlenet
[ 2%] Built target googlenet.o
[ 10%] Built target purine_cu
[ 15%] Built target proto
[100%] Built target purine
Linking CXX executable ../test/googlenet
CMakeFiles/googlenet.o.dir/googlenet.cpp.o: In function MPI::Intracomm::Intracomm()': googlenet.cpp:(.text._ZN3MPI9IntracommC2Ev[_ZN3MPI9IntracommC5Ev]+0x14): undefined reference toMPI::Comm::Comm()'
CMakeFiles/googlenet.o.dir/googlenet.cpp.o: In function MPI::Intracomm::Intracomm(ompi_communicator_t*)': googlenet.cpp:(.text._ZN3MPI9IntracommC2EP19ompi_communicator_t[_ZN3MPI9IntracommC5EP19ompi_communicator_t]+0x19): undefined reference toMPI::Comm::Comm()'
CMakeFiles/googlenet.o.dir/googlenet.cpp.o: In function MPI::Op::Init(void (*)(void const*, void*, int, MPI::Datatype const&), bool)': googlenet.cpp:(.text._ZN3MPI2Op4InitEPFvPKvPviRKNS_8DatatypeEEb[_ZN3MPI2Op4InitEPFvPKvPviRKNS_8DatatypeEEb]+0x24): undefined reference toompi_mpi_cxx_op_intercept'
CMakeFiles/googlenet.o.dir/googlenet.cpp.o:(.rodata._ZTVN3MPI3WinE[_ZTVN3MPI3WinE]+0x48): undefined reference to MPI::Win::Free()' CMakeFiles/googlenet.o.dir/googlenet.cpp.o:(.rodata._ZTVN3MPI8DatatypeE[_ZTVN3MPI8DatatypeE]+0x78): undefined reference toMPI::Datatype::Free()'
collect2: error: ld returned 1 exit status
make[3]: *** [test/googlenet] Error 1
make[2]: *** [examples/CMakeFiles/googlenet.dir/all] Error 2
make[1]: *** [examples/CMakeFiles/googlenet.dir/rule] Error 2
make: *** [examples/CMakeFiles/googlenet.dir/rule] Error 2

Hi,

when I try to compile the purine code with "make" command, it show there is no definition of "cudnnGetErrorString", please check

Hi,
After struggling through all the compile problems and a quick lesson about MPI, I can finally use the nin_cifar10 example, but I find that the training process may be hung and the GPU utilization falls to 0% at any time during training(sometimes after hundreds of iterations sometimes even more). Can that be the problem about async()/ sync()?
Thanks

We test the lrn_layer and find this layer may cause crash on "Titan X" while work well on "k40". Is this layer debugging?

2.411124 0.132812
2.542028 0.132812
2.302772 0.125000
2.312779 0.109375
2.310245 0.117188
2.302449 0.101562
2.306294 0.085938
2.301918 0.148438
2.302899 0.031250
2.302605 0.132812
2.302575 0.132812
2.302580 0.109375
2.302587 0.093750
2.302583 0.117188
2.302583 0.054688
2.302583 0.093750
2.302583 0.070312
2.302583 0.109375
2.302583 0.078125
2.302583 0.125000
2.302583 0.109375
2.302583 0.062500
2.302583 0.109375
2.302583 0.093750
2.302583 0.125000
2.302583 0.109375
2.302583 0.070312
2.302583 0.093750

I print the loss and accuracy and find loss becomes const after several iterator and net does not convergence. I think it is a bug.

By the way, example goolent is work.

Some parameters (like "${Atlas_INCLUDE_DIR}" and "${CMAKE_BINARY_DIR}/test") are passed to CMake commands in your build scripts without enclosing them by quotation marks. I see that these places will result in build difficulties if the contents of the used variables will contain special characters like semicolons.

I would recommend to apply advices from a wiki article.

As the picture shows in paper, which tool was used to draw this figure?

hi,

I am try the googlenet example. I found it is running, but there is no log output. I don't know why. I also add some test line in the main funcition

LOG(INFO)<< "test";
MPI_LOG(<<"test";);

There is no output either.

Any one knows how to compile the examples? After “make all” at ${purine}, it built purine.a at ${purine}. But when I type “make nin_cifar10”, it report error. Am I doing it correctly? Please help me..

when I move the batch_normalize form branch of caffe to purine, I found convlayer create a shared tensor temp data:
conv_layer.hpp
93 Blob* tmp_data = create("before_act", top_[0]->shared_tensor());
94 Blob* tmp_diff = create("before_act_diff", top_[1]->shared_tensor());
and bind activation:
103 B{ tmp_data, tmp_diff } >> *act >> top_;
because of shared tensor, memory of tmp_data is top[0]'s memory, after forward, activation 'bottom[0] will be cover, so will lead some error of in activation_down....

I finish my batch_noramlization in purine, when I shared data just like convLayer, my train result(googlenet with batch_normalization) is bad,
and use temp data independent is ok,
so Is the conv layer right?

In file Inception.hpp

bottom_ >> *one_;                                                                                                                                                                                             
bottom_ >> *three_reduce_ >> *three_;
bottom_ >> *five_reduce_ >> *five_;
bottom_ >> *max_pool_ >> *pool_proj_;

when compute down, data bottom, I can not see any merge the 4 branch data, just recover, Why? I compare other framwork （cxxnet example）to build googlenet, they has the split layer after bottom

is inc_in in node.cpp thread safe?

thxs

inter and intra machine data copying graph.

Have you have published the relevant papers?