visual_dynamics

Algorithms used in the paper Learning Visual Servoing with Deep Features and Fitted Q-Iteration.

The goal is for a quadcopter/drone to follow a car around the city using only image observations. The following images are first-person views of the quadcopter successfully following the car. These are test executions of our policy based on VGG conv4_3 feature dynamics. The executions on the left use the cars seen during training and the ones on the right use novel cars.

The following executions are longer executions for each of the 5 novel cars.

Our policy was trained with the fitted Q-iteration algorithm that we propose using only 20 trajectories for reinforcement learning. To see executions of other methods, check out the paper's website.

Installation instructions

Install bleeding-edge version of Theano and apply patch

git clone git://github.com/Theano/Theano.git
cd Theano
git apply patches/theano_matrix_inverse.patch
python setup.py develop --prefix=~/.local

Install bleeding-edge version of Lasagne and apply patch

git clone https://github.com/Lasagne/Lasagne.git
cd Lasagne
git apply patches/lasagne_dilation.patch
pip install -r requirements.txt
pip install --editable . --user

Install OpenCV

sudo apt-get install python-opencv

Install CitySim3D and its dependencies

Follow the instructions from the CitySim3D site.

Install visual_dynamics and its dependencies

git clone [email protected]:alexlee-gk/visual_dynamics.git
cd visual_dynamics
pip install -r requirements.txt

Advanced installation instructions: Use pyenv and install dependencies from source

Set up a new python environment using pyenv

Install desired version of python 3 (e.g. 3.5.2). Make sure to use the --enable-shared flag to generate python shared libraries, which will later be linked to.

env PYTHON_CONFIGURE_OPTS="--enable-shared" pyenv install 3.5.2

Install Theano

git clone git://github.com/Theano/Theano.git
cd Theano
pyenv local 3.5.2
python setup.py develop

Install Lasagne

git clone https://github.com/Lasagne/Lasagne.git
cd Lasagne
pyenv local 3.5.2
pip install -r requirements.txt
pip install --editable .

Install OpenCV

Make sure python-dev is installed for the python version being used, e.g.

sudo apt-get install python3.5-dev

git clone [email protected]:opencv/opencv.git
mkdir opencv_build
cd opencv_build
pyenv local 3.5.2
cmake \
-DWITH_CUDA=OFF \
-DCMAKE_BUILD_TYPE=RELEASE \
-DPYTHON3_EXECUTABLE=~/.pyenv/versions/3.5.2/bin/python3.5 \
-DPYTHON3_INCLUDE_DIR=~/.pyenv/versions/3.5.2/include/python3.5m \
-DPYTHON3_INCLUDE_DIR2=~/.pyenv/versions/3.5.2/include/python3.5m \
-DPYTHON_INCLUDE_DIRS=~/.pyenv/versions/3.5.2/include/ \
-DPYTHON3_LIBRARY=~/.pyenv/versions/3.5.2/lib/libpython3.so \
-DPYTHON3_NUMPY_INCLUDE_DIRS=~/.pyenv/versions/3.5.2/lib/python3.5/site-packages/numpy/core/include \
-DPYTHON3_PACKAGES_PATH=~/.pyenv/versions/3.5.2/lib/python3.5/site-packages \
-DINSTALL_PYTHON_EXAMPLES=ON \
-DINSTALL_C_EXAMPLES=OFF \
-DBUILD_EXAMPLES=ON \
-DBUILD_opencv_python3=ON \
../opencv
make -j4
sudo make install
ln -s /usr/local/lib/python3.5/site-packages/cv2.cpython-35m-x86_64-linux-gnu.so ~/.pyenv/versions/3.5.2/lib/python3.5/site-packages/cv2.so

For python 2, the cmake command is the following:

cmake \
-DWITH_CUDA=OFF \
-DCMAKE_BUILD_TYPE=RELEASE \
-DPYTHON2_EXECUTABLE=~/.pyenv/versions/2.7.12/bin/python2.7 \
-DPYTHON2_INCLUDE_DIR=~/.pyenv/versions/2.7.12/include/python2.7 \
-DPYTHON2_INCLUDE_DIR2=~/.pyenv/versions/2.7.12/include/python2.7 \
-DPYTHON_INCLUDE_DIRS=~/.pyenv/versions/2.7.12/include/ \
-DPYTHON2_LIBRARY=~/.pyenv/versions/2.7.12/lib/libpython2.7.so \
-DPYTHON2_NUMPY_INCLUDE_DIRS=~/.pyenv/versions/2.7.12/lib/python2.7/site-packages/numpy/core/include \
-DPYTHON2_PACKAGES_PATH=~/.pyenv/versions/2.7.12/lib/python2.7/site-packages \
-DINSTALL_PYTHON_EXAMPLES=ON \
-DINSTALL_C_EXAMPLES=OFF \
-DBUILD_EXAMPLES=ON \
-DBUILD_opencv_python2=ON \
../opencv

The library can be installed only for the local user by specifying a local install prefix, e.g. -DCMAKE_INSTALL_PREFIX=~/.local, in which case make install should be run without root priviledges and the last symbolic linking step might not needed.

Common installation problems

• After running cmake, the python2 OpenCV module appears next to 'Unavailable' instead of 'To be built'. Omit the flags that define PYTHON2_EXECUTABLE and PYTHON2_LIBRARY in the cmake command and then fix them with ccmake afterwards.
• The file Python.h is not found even though it is in the specified PYTHON3_INCLUDE_DIR, fatal error: Python.h: No such file or directory. Explicitly expanding the home directory ~ to ${HOME} might solve this.
• Installation for python 2 causes the compilation error error: invalid conversion from ‘const char*’ to ‘Py_ssize_t {aka long int}’. In this case, disable python 2 support with the option -DBUILD_opencv_python2=OFF.
• Importing cv2 gives the error ImportError: dynamic module does not define module export function (PyInit_cv2) because it is using the wrong cv2 library. Make sure the path for the newly built cv2 package appears first in the PYTHONPATH, export PYTHONPATH=~/.pyenv/versions/3.5.2/lib/python3.5/site-packages:$PYTHONPATH.

In Mac OS X, replace the cmake command with this one:

cmake \
-DWITH_CUDA=OFF \
-DCMAKE_BUILD_TYPE=RELEASE \
-DPYTHON3_EXECUTABLE=~/.pyenv/versions/3.5.2/bin/python3.5 \
-DPYTHON3_INCLUDE_DIR=~/.pyenv/versions/3.5.2/include/python3.5m \
-DPYTHON3_INCLUDE_DIR2=~/.pyenv/versions/3.5.2/include/python3.5m \
-DPYTHON3_LIBRARY=~/.pyenv/versions/3.5.2/lib/libpython3m.dylib \
-DPYTHON3_LIBRARY_DEBUG=~/.pyenv/versions/3.5.2/lib/libpython3m.dylib \
-DPYTHON3_NUMPY_INCLUDE_DIRS=~/.pyenv/versions/3.5.2/lib/python3.5/site-packages/numpy/core/include \
-DPYTHON3_PACKAGES_PATH=~/.pyenv/versions/3.5.2/lib/python3.5/site-packages \
../opencv

The option WITH_CUDA=OFF might be necessary if Caffe is used. See this issue for more information.

Links

1. https://gist.github.com/pohmelie/cf4eda5df24303325b16
2. http://stackoverflow.com/questions/33250375/compiling-opencv3-with-pyenv-using-python-3-5-0-on-osx

Example usage

Generate training and validation data

mkdir -p data
python scripts/generate_data.py config/environment/simplequad.yaml config/policy/random_quad_back.yaml -n100 -t100 -o data/simplequad_train_data
python scripts/generate_data.py config/environment/simplequad.yaml config/policy/random_quad_back.yaml -n10 -t100 -o data/simplequad_val_data

Train multiscale bilinear dynamics for a particular feature representation

python scripts/train.py config/predictor/multiscale_dilated_vgg_local_level1_scales012.yaml config/transformer/transformer_128.yaml config/solver/adam_gamma0.9_level1scales012.yaml config/data/simplequad.yaml

Learn a weighting of the servoing features using fitted Q-iteration reinforcement learning

python scripts/learn_visual_servoing.py models/theano/multiscale_dilated_vgg_local_level1_scales012/transformer_128/adam_gamma0.9_level1scales012/simplequad/_iter_10000_model.yaml config/algorithm/fqi_nooptfitbias.yaml