A PyTorch implementation of Unsupervised Doodling and Painting with Improved SPIRAL by Mellor, Park, Ganin et al.

For further details, see https://learning-to-paint.github.io for paper with generation videos.

The easiest way to build and install all of PolyBeast's dependencies and run it is to use Docker:

$ docker build -t spiralpp .
$ docker run --name spiralpp -it -p 8888:8888 spiralpp /bin/bash

or

$ docker run -it -p 8888:8888 urw7rs/spiralpp:latest /bin/bash

To run PolyBeast directly on Linux, follow this guide.

Create a new Conda environment, and install PolyBeast's requirements:

$ conda create -n spiralpp python=3.7
$ conda activate spiralpp
$ pip install -r requirements.txt

Install spiral-envs

Install required packages:

$ apt-get install cmake pkg-config protobuf-compiler libjson-c-dev intltool
$ pip install six setuptools numpy scipy gym

WARNING: Make sure that you have cmake 3.14 or later since we rely on its capability to find numpy libraries.

Install cmake by running:

$ conda install cmake

Finally, run the following command to install the spiral-gym package itself:

$ git submodule update --init --recursive
$ pip install -e spiral-envs/

You will also need to obtain the brush files for the libmypaint environment to work properly. These can be found here. For example, you can place them in third_party folder like this:

$ wget -c https://github.com/mypaint/mypaint-brushes/archive/v1.3.0.tar.gz -O - | tar -xz -C third_party

Finally, the Fluid Paint environment depends on the shaders from the original javascript implementation. You can obtain them by running the following commands:

$ git clone https://github.com/dli/paint third_party/paint
$ patch third_party/paint/shaders/setbristles.frag third_party/paint-setbristles.patch

PolyBeast requires installing PyTorch from source.

PolyBeast also requires gRPC, which can be installed by running:

$ conda install -c anaconda protobuf
$ ./scripts/install_grpc.sh

Compile the C++ parts of PolyBeast:

$ pip install nest/
$ export LD_LIBRARY_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}/lib:${LD_LIBRARY_PATH}
$ python setup.py build develop

To start both the environment servers and the learner process, run

$ python -m torchbeast.polybeast \
     --dataset celeba-hq \
     --env_type libmypaint \
     --canvas_width 64 \
     --use_pressure \
     --use_tca \
     --num_actors 64 \
     --total_steps 30000000 \
     --policy_learning_rate 0.0004 \
     --entropy_cost 0.01 \
     --batch_size 64 \
     --episode_length 40 \
     --xpid example

Results are logged to ~/logs/torchbeast/latest and a checkpoint file is written to ~/logs/torchbeast/latest/model.tar.

The environment servers can also be started separately:

$ python -m torchbeast.polybeast_env --num_actors 10

Start another terminal and run:

$ python -m torchbeast.polybeast --no_start_servers

The provided jupyter notebook will load checkpoints at a specified path to draw a single sample.

If you're using docker run

$ jupyter notebook --ip 0.0.0.0 --no-browser --allow-root

make sure you opened a port for the container.

libtorchbeast: C++ library that allows efficient learner-actor communication via queueing and batching mechanisms. Some functions are exported to Python using pybind11. For PolyBeast only.

nest: C++ library that allows to manipulate complex nested structures. Some functions are exported to Python using pybind11.

third_party: Collection of third-party dependencies as Git submodules. Includes gRPC.

torchbeast: Contains monobeast.py, and polybeast.py and polybeast_env.py. (monobeast.py is currently unavailable)

spiral-envs: spiral-envs is a libmypaint and fluidpaint based environments. ported to openai gym from spiral.

spiralpp is released under the Apache 2.0 license.