[Update on 03/20/2021]: The code for training Bongard LOGO baselines was released here.
Bongard-LOGO
is a Python code repository with the purpose of generating synthetic Bongard problems on a large scale with little human intervention. It mainly consists of a Bongard-LOGO
dataset that has 12,000 synthetic Bongard problems, and a bongard
Python library that is used for the synthesis of the Bongard problems. The user could use the Bongard-LOGO
dataset for concept learning and symbolic learning, or use the bongard
library for custom Bongard problem synthesis.
The Bongard-LOGO
dataset contains 12,000 Bongard problems solely based on the shape concept
, by considering as nuisances shape positions, orientations and sizes, etc. It is divided into three types of problems: 3,600 Freeform Shape Problem
s, 4,000 Basic Shape Problem
s and 4,400 Abstract Shape Problem
s.
Furthermore, each image is paired with an action program, with which we can recover the shape(s) in the image, by using the provided rendering method.
Below are examples of each type of Bongard-LOGO
problems.
Freeform Shape Problem | Basic Shape Problem | Abstract Shape Problem |
---|---|---|
The Bongard-LOGO
dataset could be directly downloaded from Google Drive.
For more details of how to generate the Bongard-LOGO
dataset, please refer to the Bongard-LOGO example.
It is strongly recommended to use Docker container. The Docker container has the bongard
library installed and the user does not have to install it manually.
To build the Docker image, please run the following command in the terminal.
docker build -f docker/bongard.Dockerfile --tag=bongard:0.0.1 .
To run the Docker container, please run the following command in the terminal.
docker run -it --rm bongard:0.0.1
To install the bongard
library from the latest source code, please run the following command in the terminal.
git clone https://github.com/NVlabs/Bongard-LOGO.git
cd Bongard-LOGO
pip install .
To install the bongard
library from the published packages on PyPI, please run the following command in the terminal.
pip install bongard
The Bongard library consists of four typical classes, BongardProblem
, BongardImage
, OneStrokeShape
, and BasicAction
. There exists a hierarchy for the four classes.
Each BongardProblem
consists of several positive BongardImage
s that represents a certain concept and several negative BongardImage
s that violates the concept in the positive images.
Each positive or negative BongardImage
consists of several OneStrokeShape
s. Each OneStrokeShape
consists of several consecutive BasicAction
s which are performed by a Python turtle.
In short, the data structure hierarchy is BongardProblem
-> BongardImage
-> OneStrokeShape
-> BasicAction
.
Note that BongardImage
is not really an image, but a schema for the painter to create image. This also applies to the above other classes.
The BasicAction
is performed by a Python turtle.
There are two types of BasicAction
s that have been implemented, LineAction
and ArcAction
. LineAction
is moving following a line with optionally some features in the trajectory, such as zigzag and stamps, followed by an optional change in the moving directions.
Similarly, ArcAction
is moving following an arc with optionally some features in the trajectory, such as zigzag and stamps, followed by an optional change in the moving directions.
from bongard import LineAction, ArcAction
# Create an instance of LineAction
# line_length: a float number between 0 and 1.0.
# line_type: a string indicating the type of lines. Currently, the BongardPainter supported "normal", "zigzag", "square", "arrow", "circle", "triangle", and "classic".
# turn_direction: a string indicating turning left or right after moving. It has to be "L" or "R". "L" for left and "R" for right.
# turn_angle: a float number between 0 and 180.0.
line_action = LineAction(line_length=0.5, line_type="normal", turn_direction="R", turn_angle=90)
# Create an instance of ArcAction
# arc_angle: a float number between -360 and 360.
# arc_radius: a float number between 0 and 1.0.
# arc_type: a string indicating the type of lines. Currently, the BongardPainter have supported "normal", "zigzag", "square", "arrow", "circle", "triangle", and "classic".
# turn_direction: a string indicating turning left or right after moving. It has to be "L" or "R". "L" for left and "R" for right.
# turn_angle: a float number between 0 and 180.0.
arc_action = ArcAction(arc_angle=90, arc_type=n"zigzag", turn_direction="R", turn_angle=0, arc_radius=0.5)
The OneStrokeShape
could be created using a sequence of BasicAction
s.
from bongard import LineAction, OneStrokeShape
# Create a rectangle
def create_rectangle(w, h):
action_0 = LineAction(line_length=w, line_type="normal", turn_direction="R", turn_angle=90)
action_1 = LineAction(line_length=h, line_type="normal", turn_direction="R", turn_angle=90)
action_2 = LineAction(line_length=w, line_type="normal", turn_direction="R", turn_angle=90)
action_3 = LineAction(line_length=h, line_type="normal", turn_direction="R", turn_angle=90)
rectangle_actions = [action_0, action_1, action_2, action_3]
# basic_actions: a list of consecutive `BasicAction`.
# start_coordinates: (float, float). The start coordinates on the canvas for the turtle. This would be set later when we start to generate images.
# start_orientation: a float number between 0 and 360.0. The start orientation on the canvas for the turtle. This would be set later when we start to generate images.
rectangle = OneStrokeShape(basic_actions=rectangle_actions, start_coordinates=None, start_orientation=None)
return rectangle
rectangle = create_rectangle(w=0.5, h=0.3)
The BongardImage
could be created using a list of OneStrokeShape
s.
from bongard import OneStrokeShape, BongardImage
# Create an image that consists of two rectangles
def create_random_rectangle_image():
shapes = []
for _ in range(2):
w = np.random.uniform(low=0.3, high=1.0)
h = np.random.uniform(low=0.3, high=1.0)
shapes.append(create_rectangle(w=w, h=h))
# one_stroke_shapes: a list of `OneStrokeShape`s
bongard_image = BongardImage(one_stroke_shapes=shapes)
return bongard_image
rectangle_image = create_random_rectangle_image()
The BongardImage
could be created using a list of positive BongardImage
s and a list of negative BongardImage
s.
# Create an Bongard problem that has rectangle images as positive images and circle images are negative images.
from bongard import BongardImage, BongardProblem
bongard_problem_positive_images = [create_random_rectangle_image() for _ in range(7)]
bongard_problem_negative_images = [create_random_circle_image() for _ in range(7)]
# positive_bongard_images: a list of `BongardImage`s
# negative_bongard_images: a list of `BongardImage`s
# problem_name: string.
# problem_description: string.
# positive_rules: string.
# negative_rules: string.
bongard_problem = BongardProblem(positive_bongard_images=bongard_problem_positive_images
negative_bongard_images=bongard_problem_negative_images,
problem_name=None, problem_description=None,
positive_rules=None, negative_rules=None)
We could also extract normalized action program from BongardProblem
or create BongardProblem
from normalized action program.
# Extract action program
action_program = bongard_problem.get_action_string_list()
# Create `BongardProblem` from action program
bongard_problem_recovered = BongardProblem.import_from_action_string_list(action_string_list=action_program)
A typical action program for BongardProblem
is like this. The actions from the positive images comes first followed by the actions from the negative images.
[[[['line_triangle_0.715-0.250', 'line_triangle_0.891-0.250', 'line_triangle_0.715-0.250', 'line_zigzag_0.891-0.250']], [['line_circle_0.752-0.250', 'line_normal_0.606-0.250', 'line_normal_0.752-0.250', 'line_normal_0.606-0.250'], ['line_triangle_0.568-0.250', 'line_circle_0.854-0.250', 'line_normal_0.568-0.250', 'line_zigzag_0.854-0.250']], [['line_normal_0.885-0.250', 'line_triangle_0.536-0.250', 'line_normal_0.885-0.250', 'line_triangle_0.536-0.250'], ['line_normal_0.970-0.250', 'line_circle_0.398-0.250', 'line_triangle_0.970-0.250', 'line_normal_0.398-0.250']], [['line_circle_0.623-0.250', 'line_triangle_0.846-0.250', 'line_normal_0.623-0.250', 'line_zigzag_0.846-0.250'], ['line_triangle_0.400-0.250', 'line_circle_0.961-0.250', 'line_normal_0.400-0.250', 'line_triangle_0.961-0.250']], [['line_normal_0.430-0.250', 'line_circle_0.816-0.250', 'line_normal_0.430-0.250', 'line_normal_0.816-0.250'], ['line_normal_0.527-0.250', 'line_normal_0.405-0.250', 'line_zigzag_0.527-0.250', 'line_triangle_0.405-0.250']], [['line_normal_0.777-0.250', 'line_zigzag_0.552-0.250', 'line_zigzag_0.777-0.250', 'line_triangle_0.552-0.250']], [['line_triangle_0.979-0.250', 'line_circle_0.757-0.250', 'line_circle_0.979-0.250', 'line_triangle_0.757-0.250'], ['line_normal_0.825-0.250', 'line_zigzag_0.725-0.250', 'line_circle_0.825-0.250', 'line_normal_0.725-0.250']]], [[['arc_normal_0.497_0.625-0.500', 'arc_triangle_0.497_0.625-0.500', 'arc_triangle_0.497_0.625-0.500', 'arc_normal_0.497_0.625-0.500'], ['arc_normal_0.248_0.625-0.500', 'arc_normal_0.248_0.625-0.500', 'arc_normal_0.248_0.625-0.500', 'arc_circle_0.248_0.625-0.500']], [['arc_circle_0.342_0.625-0.500', 'arc_triangle_0.342_0.625-0.500', 'arc_triangle_0.342_0.625-0.500', 'arc_zigzag_0.342_0.625-0.500'], ['arc_normal_0.402_0.625-0.500', 'arc_zigzag_0.402_0.625-0.500', 'arc_zigzag_0.402_0.625-0.500', 'arc_zigzag_0.402_0.625-0.500']], [['arc_zigzag_0.311_0.625-0.500', 'arc_circle_0.311_0.625-0.500', 'arc_normal_0.311_0.625-0.500', 'arc_zigzag_0.311_0.625-0.500'], ['arc_circle_0.451_0.625-0.500', 'arc_normal_0.451_0.625-0.500', 'arc_zigzag_0.451_0.625-0.500', 'arc_triangle_0.451_0.625-0.500']], [['arc_normal_0.353_0.625-0.500', 'arc_triangle_0.353_0.625-0.500', 'arc_zigzag_0.353_0.625-0.500', 'arc_zigzag_0.353_0.625-0.500']], [['arc_circle_0.212_0.625-0.500', 'arc_triangle_0.212_0.625-0.500', 'arc_circle_0.212_0.625-0.500', 'arc_triangle_0.212_0.625-0.500'], ['arc_circle_0.289_0.625-0.500', 'arc_zigzag_0.289_0.625-0.500', 'arc_circle_0.289_0.625-0.500', 'arc_circle_0.289_0.625-0.500']], [['arc_triangle_0.475_0.625-0.500', 'arc_circle_0.475_0.625-0.500', 'arc_circle_0.475_0.625-0.500', 'arc_triangle_0.475_0.625-0.500'], ['arc_circle_0.460_0.625-0.500', 'arc_triangle_0.460_0.625-0.500', 'arc_circle_0.460_0.625-0.500', 'arc_zigzag_0.460_0.625-0.500']], [['arc_circle_0.373_0.625-0.500', 'arc_circle_0.373_0.625-0.500', 'arc_triangle_0.373_0.625-0.500', 'arc_normal_0.373_0.625-0.500']]]]
The BongardProblemPainter
could be used for plotting Bongard problems with variations from shape sizes, positions and orientations. Given a BongardProblem
, assuming each BongardImage
has no more than two OneStrokeShape
s, it could generate random size, position and orientation for each OneStrokeShape
with minimal chance of OneStrokeShape
collisions.
from bongard import BongardProblem, BongardProblemPainter
# Set random seed for `BongardProblemPainter` to make the painting process reproducible.
bongard_problem_painter = BongardProblemPainter(random_seed=random_seed)
# `BongardProblemPainter` will generate `ps` and `png` format images for the `BongardProblem` provided.
bongard_problem_painter.create_bongard_problem(bongard_problem=bongard_problem, bongard_problem_ps_dir=bongard_problem_ps_dir, bongard_problem_png_dir=bongard_problem_png_dir)
More than six hundreds of shapes have been designed for Bongard problem synthesis. The action program annotations and shape attributed were saved in data/human_designed_shapes.tsv
and data/human_designed_shapes_attributes.tsv
. The users are encourage to create new shapes and attributes, and contribute to this project.
The preview of human designed shapes could be found here.
This is the most basic example for showing how to use the Bongard library. Every user is strongly recommended to go through the short code.
This is a slightly more advanced example for showing how to take advantage of the human designed shapes we provided for creating shapes.
These are the scripts for generating the Bongard-LOGO dataset. Advanced Bongard problem samplers were implemented.
This is a simple example for showing how to manually assign locations and sizes of the shapes in Bongard images.
The Bongard library has dependencies on Python Turtle is a X client. To use the Bongard library without display, we recommend to use Virtual X Server xvfb
and ghostscript
. To install xvfb
and ghostscript
on Ubuntu, please run the following commands.
[sudo] apt update
[sudo] apt install -y xvfb ghostscript
To use the Bongard library without display, please add xvfb-run
at the beginning of the command. For example,
xvfb-run python [script_name.py]
If the user is not using the Docker container we provided, it is likely that there might be some missing dependencies. For this particular dependency, please install it via apt
on Ubuntu.
[sudo] sudo apt install python3-tk
To cite this work, please use
@INPROCEEDINGS{Nie2020Bongard,
author = {Nie, Weili and Yu, Zhiding and Mao, Lei and Patel, Ankit B and Zhu, Yuke and Anandkumar, Animashree},
title = {Bongard-LOGO: A New Benchmark for Human-Level Concept Learning and Reasoning},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2020}
}