error when running the demo about tffrcnn

[nsivaramakrishnan]

Hi,
I am getting the same error while trying to run demo.py:
tensorflow.python.framework.errors_impl.NotFoundError: /home/fmc/rcnn/Faster-RCNN_TF/tools/../lib/roi_pooling_layer/roi_pooling.so: undefined symbol: _Z22ROIPoolBackwardLaucherPKffiiiiiiiS0_PfPKiRKN5Eigen9GpuDeviceE
I added "-D_GLIBCXX_USE_CXX11_ABI=0" in make,sh. I use g++ version 5.4.0 and TF V0.12. Btw, am trying to run this on CPU. Any help is highly appreciated.
-Siva

[jwnsu]

for gcc5, simply add this to compile flag "-D_GLIBCXX_USE_CXX11_ABI=0", then it will work fine. It's documented issue by tensor flow: https://www.tensorflow.org/versions/r0.10/how_tos/adding_an_op/index.html, search for "gcc5".

[CharlesShang]

If it does not work, try to install a different version of g++, via sudo apt-get install g++-4.x or sudo apt-get install g++-5.x
Then run
sudo rm /usr/bin/g++
sudo ln -s /usr/bin/g++-4.x /usr/bin/g++

[CharlesShang]

Hi, @xinghedyc , I think you first should go to ./lib/ dir and run make

[CharlesShang]

Maybe, you should edit the g++ config according to your platform to ensure the roi_pooing.so is correct.
See ./lib/make.sh

[xinghedyc]

@CharlesShang thanks for your reply, but the make command seems right to generate roi_pooing.so.

[xinghedyc]

@CharlesShang thanks for your reply, but the make command seems right to generate roi_pooing.so.

[CharlesShang]

It happened to me too, the roi_pooing.so seems all right, but it just could not work.
Check this issue: Failed when running tf.load_op_library(roi_pooling_so)
It's very embarrassing to say that I've not figured out the real solution. Alternatively, you can download a already-compiled .so file, or try re-start your machine.

[xinghedyc]

@CharlesShang thanks a lot, but your link to download the roi_pooling_so seems unreachable?
https://github.com/CharlesShang/TFFRCNN/tree/master/lib/roi_pooling_layer/roi_pooling.so
it shows 404 not found?

[CharlesShang]

Try this link so version

[xinghedyc]

@CharlesShang it's embarrassingto say that the same error occurs when using your .so file.
It's so strange.
Thanks you anyway!

[xinghedyc]

python faster_rcnn/demo.py VGGnet_fast_rcnn_iter_70000.ckpt
I tensorflow/stream_executor/dso_loader.cc:125] successfully opened CUDA library libcublas.so locally
I tensorflow/stream_executor/dso_loader.cc:125] successfully opened CUDA library libcudnn.so locally
I tensorflow/stream_executor/dso_loader.cc:125] successfully opened CUDA library libcufft.so locally
I tensorflow/stream_executor/dso_loader.cc:125] successfully opened CUDA library libcuda.so.1 locally
I tensorflow/stream_executor/dso_loader.cc:125] successfully opened CUDA library libcurand.so locally
faster_rcnn
Traceback (most recent call last):
File "faster_rcnn/demo.py", line 12, in
from lib.networks.factory import get_network
File "/home/dyc/workspace/TFFRCNN/lib/networks/init.py", line 8, in
from .VGGnet_train import VGGnet_train
File "/home/dyc/workspace/TFFRCNN/lib/networks/VGGnet_train.py", line 2, in
from network import Network
File "/home/dyc/workspace/TFFRCNN/lib/networks/network.py", line 4, in
from ..roi_pooling_layer import roi_pooling_op as roi_pool_op
File "/home/dyc/workspace/TFFRCNN/lib/roi_pooling_layer/init.py", line 7, in
import roi_pooling_op
File "/home/dyc/workspace/TFFRCNN/lib/roi_pooling_layer/roi_pooling_op.py", line 5, in
_roi_pooling_module = tf.load_op_library(filename)
File "/home/dyc/anaconda2/lib/python2.7/site-packages/tensorflow/python/framework/load_library.py", line 63, in load_op_library
raise errors._make_specific_exception(None, None, error_msg, error_code)
tensorflow.python.framework.errors.NotFoundError: /home/dyc/workspace/TFFRCNN/lib/roi_pooling_layer/roi_pooling.so: undefined symbol: _ZN10tensorflow7strings6StrCatB5cxx11ERKNS0_8AlphaNumE

[CharlesShang]

I googled it, found: GPflow/GPflow#173
According to james, the problem is tf and roi_pooling are built with different versions of g++,
Try to add -D_GLIBCXX_USE_CXX11_ABI=0 in your make.sh will solve this issue

I've already push the fixed make.sh to master branch, you can go to TFFRCNN root dir and run git pull, then complie the roi_pooling.so

[xinghedyc]

@CharlesShang thanks very much for your help!
it works now, really thanks you!

[CharlesShang]

Glad to hear that. Maybe we've found out the real solution to this kind of issue

[CharlesShang]

@jwnsu Thanks for your helpful discussion. Learned a lot .

[civilman628]

update as:

g++ -std=c++11 -shared -o roi_pooling.so roi_pooling_op.cc  -D_GLIBCXX_USE_CXX11_ABI=0  \
	roi_pooling_op.cu.o -I $TF_INC  -D GOOGLE_CUDA=1 -fPIC $CXXFLAGS \
	-lcudart -L $CUDA_PATH/lib64 

[chaochaow]

@nsivaramakrishnan I got the same problem like yours when running on CPU. Do you have any luck to solve this issue?

[louisquinn]

@nsivaramakrishnan @chaochaow
I'm getting a similar issue when attempting to build for CPU:
Faster-RCNN_TF/tools/../lib/roi_pooling_layer/roi_pooling.so: undefined symbol: _ZN10tensorflow8internal21CheckOpMessageBuilder9NewStringEv

This looks like a missing tensorflow symbol.
I'm using g++ v5.4.0 as well. Not sure if downgrading to v4.x will help.

[chaochaow]

GPU version does not have this problem

[kalanityL]

I had the issue with GPU
solution given by @Griovwe here worked:
you have to add "-D_GLIBCXX_USE_CXX11_ABI=0" to both the g++ lines for roi_pooling layer and psroi_pooling layer if you encountered this problem.