Awesome

SPIRAL

Overview

This repository contains agents and environments described in the ICML'18 paper "Synthesizing Programs for Images using Reinforced Adversarial Learning". For the time being, we are providing two simulators: one based on libmypaint and one based on Fluid Paint (NOTE: our implementation is written in C++ whereas the original is in javascript). Additionally, we supply a Sonnet module for the unconditional agent as well as pre-trained model snapshots (9 agents from a single population for libmypaint and 1 agent for Fluid Paint) available from TF-Hub.

If you feel an immediate urge to dive into the code the most relevant files are:

Reference

If this repository is helpful for your research please cite the following publication:

@inproceedings{ganin2018synthesizing,
  title={Synthesizing Programs for Images using Reinforced Adversarial Learning},
  author={Ganin, Yaroslav and Kulkarni, Tejas and Babuschkin, Igor and Eslami, SM Ali and Vinyals, Oriol},
  booktitle={ICML},
  year={2018}
}

Quickstart with Docker

The easiest way to get SPIRAL up and running on your machine is to use a pre-built Docker image.

Installation

This section describes how to build and install the package on Ubuntu (16.04 or newer). The following instructions (with slight modifications) might also work for other Linux distributions.

Clone this repository and fetch the external submodules:

git clone https://github.com/deepmind/spiral.git
cd spiral
git submodule update --init --recursive

Install required packages:

apt-get install cmake pkg-config protobuf-compiler libjson-c-dev intltool libpython3-dev python3-pip
pip3 install six setuptools numpy scipy tensorflow==1.14 tensorflow-hub dm-sonnet==1.35

WARNING: Make sure that you have cmake 3.14 or later since we rely on its capability to find numpy libraries. If your package manager doesn't provide it follow the installation instructions from here. You can check the version by running cmake --version .

Finally, run the following command to install the SPIRAL package itself:

python3 setup.py develop --user

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

Optionally, in order to be able to try out the package in the provided jupyter notebook, you’ll need to install the following packages:

pip3 install matplotlib jupyter

Usage

For a basic example of how to use the package please follow this notebook.

Sampling from a pre-trained model

We provide pre-trained models for unconditional 19-step generation of CelebA-HQ images. Here is an example of how you can sample from an agent interacting with the libmypaint environment:

import matplotlib.pyplot as plt

import spiral.agents.default as default_agent
import spiral.agents.utils as agent_utils
import spiral.environments.libmypaint as libmypaint


# The path to a TF-Hub module.
MODULE_PATH = "https://tfhub.dev/deepmind/spiral/default-wgangp-celebahq64-gen-19steps/agent4/1"
# The folder containing `libmypaint` brushes.
BRUSHES_PATH = "the/path/to/libmypaint-brushes"

# Here, we create an environment.
env = libmypaint.LibMyPaint(episode_length=20,
                            canvas_width=64,
                            grid_width=32,
                            brush_type="classic/dry_brush",
                            brush_sizes=[1, 2, 4, 8, 12, 24],
                            use_color=True,
                            use_pressure=True,
                            use_alpha=False,
                            background="white",
                            brushes_basedir=BRUSHES_PATH)


# Now we load the agent from a snapshot.
initial_state, step = agent_utils.get_module_wrappers(MODULE_PATH)

# Everything is ready for sampling.
state = initial_state()
noise_sample = np.random.normal(size=(10,)).astype(np.float32)

time_step = env.reset()
for t in range(19):
    time_step.observation["noise_sample"] = noise_sample
    action, state = step(time_step.step_type, time_step.observation, state)
    time_step = env.step(action)

# Show the sample.
plt.close("all")
plt.imshow(time_step.observation["canvas"], interpolation="nearest")

Converting a trained agent into a TF-Hub module

import spiral.agents.default as default_agent
import spiral.agents.utils as agent_utils
import spiral.environments.libmypaint as libmypaint


# This where we're going to put our TF-Hub module.
TARGET_PATH = ...
# A path to a checkpoint of the trained model.
CHECKPOINT_PATH = ...

# We will need to create an environment in order to obtain the specifications
# for the agent's action and the observation.
env = libmypaint.LibMyPaint(...)

# Here, we wrap a Sonnet module constructor for our agent in a function.
# This is to avoid contaminating the default tensorflow graph.
def agent_ctor():
  return default_agent.Agent(action_spec=env.action_spec(),
                             input_shape=(64, 64),
                             grid_shape=(32, 32),
                             action_order="libmypaint")

# Finally, export a TF-Hub module. We need to specify which checkpoint to use
# to extract the weights for the agent. Since the variable names in the
# checkpoint may differ from the names in the Sonnet module produced by
# `agent_ctor`, we may also want to provide an appropriate name mapping
# function.
agent_utils.export_hub_module(agent_ctor=agent_ctor,
                              observation_spec=env.observation_spec(),
                              noise_dim=10,
                              module_path=TARGET_PATH,
                              checkpoint_path=CHECKPOINT_PATH,
                              name_transform_fn=lambda name: ...)

Disclaimer

This is not an official Google product.