convert-caffemodel

The tutorial shows that how to convert caffemodel into tensorflow. It runs successfully in the environment:

• ubuntu 16.04 LTS
• cuda8.0 & cudnn6.0
• gcc-5.4

This conversion is required Caffe and MMdnn. Please check you have installed them, or follow the instructions to help you:

How to install Caffe

1. (optional) In order to install happily, it's recommanded to use Anaconda environment when you are going to install Caffe. Follow here to install Anaconda, and create a environment pycaffe and activate it.

   # for python3 user
   conda create -n pycaffe python=3 anaconda
   # for python2 user
   conda create -n pycaffe python=2 anaconda
   
   source activate pycaffe

2. Get the lastest Caffe, and change directory:

   git clone https://github.com/BVLC/caffe ; cd caffe

3. Build it. In this step, you can select Makefile or cmake to do this task:

   • Option-1: Makefile

     a. Edit the Makefile.config. Uncomment lines you want to use.

     cp Makefile.config.example Makefile.config
     vim Makefile.config
     # uncomment lines you want

     Or you can copy Makefile.config in this repo to the caffe folder:

     cp ../Makefile.config .

     b. Install required python packages. Make sure you are under the conda-env pycaffe:

     source activate pycaffe
     cd python
     for req in $(cat requirements.txt); do pip install $req; done
     cd ..

     c. Install python version, and assign the pycaffe path to PYTHONPATH.

     make pycaffe -j8
     export PYTHONPATH=/home/bass/data_server_storage/caffe/python:$PYTHONPATH

     d. Test your pycaffe.

     make pytest
     (...)
     python -c 'import caffe; print(caffe.__version__)

   • Option-2: cmake (WARNING: may cause errors)

     a. Deactivate the anaconda environment (we need to use system libraries).

     source deactivate

     b. Create the build folder, and make it.

     mkdir build; cd build;
     cmake .. -DBLAS=open -Dpython_version=3 -Wno-dev
     source activate pycaffe
     make all install -j8

     If you are a python2 user, change to set -Dpython_version=2. If you do not have GPU, add the argument -DCPU_ONLY=ON.

     c. Install required python packages:

     cd ../python
     for req in $(cat requirements.txt); do pip install $req; done

     d. Install python version, and assign the pycaffe path to PYTHONPATH.

     make pycaffe -j8
     export PYTHONPATH=/home/bass/data_server_storage/caffe/python:$PYTHONPATH

     e. Test your pycaffe.

     make pytest
     (...)
     python -c 'import caffe; print(caffe.__version__)

   🔥 Errors & solutions:

   1. No module named 'google':

      ======================================================================
      ERROR: test_coord_map (unittest.loader._FailedTest)
      ----------------------------------------------------------------------
      ImportError: Failed to import test module: test_coord_map
      Traceback (most recent call last):
        File "/home/bass/anaconda3/lib/python3.6/unittest/loader.py", line 428, in _find_test_path
          module = self._get_module_from_name(name)
        File "/home/bass/anaconda3/lib/python3.6/unittest/loader.py", line 369, in _get_module_from_name
          __import__(name)
        File "/home/bass/data_server_storage/caffe/python/caffe/test/test_coord_map.py", line 6, in <module>
          import caffe
        File "/home/bass/data_server_storage/caffe/python/caffe/__init__.py", line 4, in <module>
          from .proto.caffe_pb2 import TRAIN, TEST
        File "/home/bass/data_server_storage/caffe/python/caffe/proto/caffe_pb2.py", line 6, in <module>
          from google.protobuf.internal import enum_type_wrapper
      ModuleNotFoundError: No module named 'google'
      

      👉 solution: source activate pycaffe

   2. libboost-python problem

      👉 solution: edit your Makefile.config:

      # (for python3 user) in Makefile.config
      --- PYTHON_LIBRARIES := boost_python python3.5m
      +++ PYTHON_LIBRARIES := boost_python-py35 python3.5m     

      Or change linkpaths libboost_python.a and libboost_python.so to direct to libboos_python-py35. If you use Anaconda, you can link them to your own anaconda library folder. For example:

      cd $HOME/anaconda3/lib/
      ln -s /usr/lib/x86_64-linux-gnu/libboost_python-py35.a libboost_python3.a
      ln -s /usr/lib/x86_64-linux-gnu/libboost_python-py35.so libboost_python3.so
      export LD_LIBRARY_PATH=$HOME/anaconda3/lib/:$LD_LIBRARY_PATH

      Make sure your anaconda library path is in $LD_LIBRARY_PATH.

   3. hdf5 problem

      👉 solution: refer to here

   4. complier problems:

      In file included from ./include/caffe/common.hpp:6:0,
               from ./include/caffe/blob.hpp:8,
               from ./include/caffe/caffe.hpp:7,
               from python/caffe/_caffe.cpp:17:
      ./include/caffe/net.hpp: In member function ‘const std::vector<caffe::Blob<Dtype>*>& caffe::Net<Dtype>::ForwardPrefilled(Dtype*)’:
      /usr/local/include/glog/logging.h:917:30: warning: typedef ‘INVALID_REQUESTED_LOG_SEVERITY’ locally defined but not used [-Wunused-local-typedefs]
                                    INVALID_REQUESTED_LOG_SEVERITY);           \
                                    ^
      /usr/local/include/glog/logging.h:912:73: note: in definition of macro ‘GOOGLE_GLOG_COMPILE_ASSERT’
         typedef google::glog_internal_namespace_::CompileAssert<(bool(expr))> msg[bool(expr) ? 1 : -1]
                                                                               ^
      ./include/caffe/net.hpp:41:5: note: in expansion of macro ‘LOG_EVERY_N’
           LOG_EVERY_N(WARNING, 1000) << "DEPRECATED: ForwardPrefilled() "
           ^

      👉 solution: make sure your gcc version is 5.4. source deactivate

   5. dynamic module does not define init function:

      ======================================================================
      ERROR: test_solver (unittest.loader.ModuleImportFailure)
      ----------------------------------------------------------------------
      ImportError: Failed to import test module: test_solver
      Traceback (most recent call last):
        File "/usr/lib/python2.7/unittest/loader.py", line 254, in _find_tests
          module = self._get_module_from_name(name)
        File "/usr/lib/python2.7/unittest/loader.py", line 232, in _get_module_from_name
          __import__(name)
        File "/home/bass/data_server_storage/caffe/python/caffe/test/test_solver.py", line 7, in <module>
          import caffe
        File "caffe/__init__.py", line 1, in <module>
          from .pycaffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, RMSPropSolver, AdaDeltaSolver, AdamSolver, NCCL, Timer
        File "caffe/pycaffe.py", line 13, in <module>
          from ._caffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, \
      ImportError: dynamic module does not define init function (init_caffe)

      👉 solution: using /usr/bin/gcc. Another way that removes the original conda-env, and recreates it.

      conda-env remove -n pycaffe ; conda-env create -n pycaffe python=3 anaconda

   6. python-dateutil version problem:

      pandas 0.22.0 has requirement python-dateutil>=2, but you'll have python-dateutil 1.5 which is incompatible.
      matplotlib 2.2.2 has requirement python-dateutil>=2.1, but you'll have python-dateutil 1.5 which is incompatible.

      👉 solution: pip install --upgrade python-dateutil

   7. No module named 'pydotplus':

      ======================================================================
      ERROR: test_draw (unittest.loader._FailedTest)
      ----------------------------------------------------------------------
      ImportError: Failed to import test module: test_draw
      Traceback (most recent call last):
        File "/home/bass/data_server_storage/caffe/python/caffe/draw.py", line 20, in <module>
          import pydotplus as pydot
      ModuleNotFoundError: No module named 'pydotplus'
      
      During handling of the above exception, another exception occurred:
      
      Traceback (most recent call last):
        File "/home/bass/anaconda3/envs/pycaffe/lib/python3.6/unittest/loader.py", line 428, in _find_test_path
          module = self._get_module_from_name(name)
        File "/home/bass/anaconda3/envs/pycaffe/lib/python3.6/unittest/loader.py", line 369, in _get_module_from_name
          __import__(name)
        File "/home/bass/data_server_storage/caffe/python/caffe/test/test_draw.py", line 6, in <module>
          import caffe.draw
        File "/home/bass/data_server_storage/caffe/python/caffe/draw.py", line 22, in <module>
          import pydot
      ModuleNotFoundError: No module named 'pydot'

      👉 solution: pip install pydotplus

   8. glog/logging.h:

      ...include/glog/logging.h(362): warning: using-declaration ignored -- it refers to the current namespace

      👉 solution: refer to PaddlePaddle/Paddle#3277

   9. protobuf problems:

      lib/libcaffe.so: undefined reference to `google::collect2: error: ld returned 1 exit status
      ...
      .build_release/lib/libcaffe.so: undefined reference to `google::protobuf::Message::InitializationErrorString() const'

      👉 solution: Remove the original conda-env, and recreate it.

      conda-env remove -n pycaffe ; conda-env create -n pycaffe python=3 anaconda

If you failed to install Caffe with above steps, try to make clean and then install the corresponding packages with conda before make. Or follow official installation instructions to install from beginning.

How to install MMdnn

Follow here to install MMdnn.

---

Convert caffemodel

1. Prepare your environment.

export PYTHONPATH=/home/bass/data_server_storage/caffe/python:$PYTHONPATH
# for anaconda user
source activate tensorflow 

2. Download your own caffemodel and prototxt. For example, FCN8s:

wget https://raw.githubusercontent.com/simonguist/testing-fcn-for-cityscapes/master/final_model_url.txt
# download the caffemodel from above url

wget https://raw.githubusercontent.com/simonguist/testing-fcn-for-cityscapes/master/train/deploy_8s.prototxt

3. Using MMdnn to convert, for example, into tensorflow.

mmconvert -sf caffe -in deploy_8s.prototxt -iw cityscapes-fcn8s-2x.caffemodel -df tensorflow -om cityscapes-fcn8s-2x

4. Finally, you get the checkpoint named cityscapes-fcn8s-2x, and some random-named files including .pd, .py, .json, .npy. You can rename and save them if you need.