This document covers how to install Detectron, its dependencies (including Caffe2), and the COCO dataset.
- For general information about Detectron, please see
- NVIDIA GPU, Linux, Python2
- Caffe2, various standard Python packages, and the COCO API; Instructions for installing these dependencies are found below
- Detectron operators currently do not have CPU implementation; a GPU system is required.
- Detectron has been tested extensively with CUDA 8.0 and cuDNN 6.0.21.
To install Caffe2 with CUDA support, follow the installation instructions from the Caffe2 website. If you already have Caffe2 installed, make sure to update your Caffe2 to a version that includes the Detectron module.
Please ensure that your Caffe2 installation was successful before proceeding by running the following commands and checking their output as directed in the comments.
# To check if Caffe2 build was successful python -c 'from caffe2.python import core' 2>/dev/null && echo "Success" || echo "Failure" # To check if Caffe2 GPU build was successful # This must print a number > 0 in order to use Detectron python -c 'from caffe2.python import workspace; print(workspace.NumCudaDevices())'
caffe2 Python package is not found, you likely need to adjust your
PYTHONPATH environment variable to include its location (
build is the Caffe2 CMake build directory).
Install the COCO API:
# COCOAPI=/path/to/clone/cocoapi git clone https://github.com/cocodataset/cocoapi.git $COCOAPI cd $COCOAPI/PythonAPI # Install into global site-packages make install # Alternatively, if you do not have permissions or prefer # not to install the COCO API into global site-packages python setup.py install --user
Note that instructions like
# COCOAPI=/path/to/install/cocoapi indicate that you should pick a path where you'd like to have the software cloned and then set an environment variable (
COCOAPI in this case) accordingly.
Clone the Detectron repository:
# DETECTRON=/path/to/clone/detectron git clone https://github.com/facebookresearch/detectron $DETECTRON
Install Python dependencies:
pip install -r $DETECTRON/requirements.txt
Set up Python modules:
cd $DETECTRON && make
Check that Detectron tests pass (e.g. for
That's All You Need for Inference
At this point, you can run inference using pretrained Detectron models. Take a look at our inference tutorial for an example. If you want to train models on the COCO dataset, then please continue with the installation instructions.
Detectron finds datasets via symlinks from
detectron/datasets/data to the actual locations where the dataset images and annotations are stored. For instructions on how to create symlinks for COCO and other datasets, please see
After symlinks have been created, that's all you need to start training models.
Advanced Topic: Custom Operators for New Research Projects
Please read the custom operators section of the
For convenience, we provide CMake support for building custom operators. All custom operators are built into a single library that can be loaded dynamically from Python.
Place your custom operator implementation under
detectron/ops/ and see
detectron/tests/test_zero_even_op.py for an example of how to load custom operators from Python.
Build the custom operators library:
cd $DETECTRON && make ops
Check that the custom operator tests pass:
We provide a
Dockerfile that you can use to build a Detectron image on top of a Caffe2 image that satisfies the requirements outlined at the top. If you would like to use a Caffe2 image different from the one we use by default, please make sure that it includes the Detectron module.
Build the image:
cd $DETECTRON/docker docker build -t detectron:c2-cuda9-cudnn7 .
Run the image (e.g. for
nvidia-docker run --rm -it detectron:c2-cuda9-cudnn7 python detectron/tests/test_batch_permutation_op.py
In case of Caffe2 installation problems, please read the troubleshooting section of the relevant Caffe2 installation instructions first. In the following, we provide additional troubleshooting tips for Caffe2 and Detectron.
Caffe2 Operator Profiling
Caffe2 comes with performance
support which you may find useful for benchmarking or debugging your operators
BatchPermutationOp test for example usage).
Profiling support is not built by default and you can enable it by setting
-DUSE_PROF=ON flag when running Caffe2 CMake.
CMake Cannot Find CUDA and cuDNN
Sometimes CMake has trouble with finding CUDA and cuDNN dirs on your machine.
When building Caffe2, you can point CMake to CUDA and cuDNN dirs by running:
cmake .. \ # insert your Caffe2 CMake flags here -DCUDA_TOOLKIT_ROOT_DIR=/path/to/cuda/toolkit/dir \ -DCUDNN_ROOT_DIR=/path/to/cudnn/root/dir
Similarly, when building custom Detectron operators you can use:
cd $DETECTRON mkdir -p build && cd build cmake .. \ -DCUDA_TOOLKIT_ROOT_DIR=/path/to/cuda/toolkit/dir \ -DCUDNN_ROOT_DIR=/path/to/cudnn/root/dir make
Note that you can use the same commands to get CMake to use specific versions of CUDA and cuDNN out of possibly multiple versions installed on your machine.
Caffe2 uses protobuf as its serialization format and requires version
3.2.0 or newer.
If your protobuf version is older, you can build protobuf from Caffe2 protobuf submodule and use that version instead.
To build Caffe2 protobuf submodule:
# CAFFE2=/path/to/caffe2 cd $CAFFE2/third_party/protobuf/cmake mkdir -p build && cd build cmake .. \ -DCMAKE_INSTALL_PREFIX=$HOME/c2_tp_protobuf \ -Dprotobuf_BUILD_TESTS=OFF \ -DCMAKE_CXX_FLAGS="-fPIC" make install
To point Caffe2 CMake to the newly built protobuf:
cmake .. \ # insert your Caffe2 CMake flags here -DPROTOBUF_PROTOC_EXECUTABLE=$HOME/c2_tp_protobuf/bin/protoc \ -DPROTOBUF_INCLUDE_DIR=$HOME/c2_tp_protobuf/include \ -DPROTOBUF_LIBRARY=$HOME/c2_tp_protobuf/lib64/libprotobuf.a
You may also experience problems with protobuf if you have both system and anaconda packages installed. This could lead to problems as the versions could be mixed at compile time or at runtime. This issue can also be overcome by following the commands from above.
Caffe2 Python Binaries
In case you experience issues with CMake being unable to find the required Python paths when building Caffe2 Python binaries (e.g. in virtualenv), you can try pointing Caffe2 CMake to python library and include dir by using:
cmake .. \ # insert your Caffe2 CMake flags here -DPYTHON_LIBRARY=$(python -c "from distutils import sysconfig; print(sysconfig.get_python_lib())") \ -DPYTHON_INCLUDE_DIR=$(python -c "from distutils import sysconfig; print(sysconfig.get_python_inc())")
Caffe2 with NNPACK Build
Detectron does not require Caffe2 built with NNPACK support. If you face NNPACK related issues during Caffe2 installation, you can safely disable NNPACK by setting the
-DUSE_NNPACK=OFF CMake flag.
Caffe2 with OpenCV Build
Analogously to the NNPACK case above, you can disable OpenCV by setting the
-DUSE_OPENCV=OFF CMake flag.
COCO API Undefined Symbol Error
If you encounter a COCO API import error due to an undefined symbol, as reported here, make sure that your python versions are not getting mixed. For instance, this issue may arise if you have both system and conda numpy installed.
CMake Cannot Find Caffe2
In case you experience issues with CMake being unable to find the Caffe2 package when building custom operators,
make sure you have run
make install as part of your Caffe2 installation process.