gustiks / neuralogic Goto Github PK

View Code? Open in Web Editor NEW

93.0 5.0 14.0 3.46 MB

Deep relational learning through differentiable logic programming.

License: MIT License

Java 97.55% ANTLR 0.15% Python 2.30%

relational-learning deep-learning logic-programming machine-learning differentiable-programming

neuralogic's People

Contributors

Stargazers

Watchers

Forkers

mmmmeow9 kunc lukaszahradnik neuralogic kairanithin vishalbelsare kangalice hussien harshakokel krausjakub keshava dtrycode noticeable neumannjan

neuralogic's Issues

Inconsistent Value formatting

Current behavior:

ScalarValue::toString() formats the value to 10 decimal places (Settings.detailedNumberFormat).
VectorValue::toSring() formats the value to 2 decimal places (Settings.shortNumberFormat).
MatrixValue::toString() does not use number formatter - formats numbers to 16 decimal places.
One::toString() formats the value to 0 decimal places.
Zero::toString() formats the value to 0 decimal places.

This behavior results in inconsistent numbers rendering and larger rendered graphs (because matrices strings are too wide).
The drawing/rendering looks like it is supposed to be customizable via Drawer.numberFormat property, but this property is not used anywhere.

Expected behavior:

At least non-constant values (Scalar, Vector, and Matrix) use the same (parametrized) number formatter.

SoftEntropy overflow resulting in NaN

The bug description

The SoftEntropy error function produces NaN values for higher input values.
I identified problems in three different places:

Softmax computation https://github.com/GustikS/NeuraLogic/blob/master/Algebra/src/main/java/cz/cvut/fel/ida/algebra/functions/specific/SoftEntropy.java#L75
Math.exp goes to infinity - should be fixable with exp(input[i] - max(input))
Scalar softentropy - https://github.com/GustikS/NeuraLogic/blob/master/Algebra/src/main/java/cz/cvut/fel/ida/algebra/functions/specific/SoftEntropy.java#L33
Math.exp goes to infinity
Vector softentropy -
https://github.com/GustikS/NeuraLogic/blob/master/Algebra/src/main/java/cz/cvut/fel/ida/algebra/functions/specific/SoftEntropy.java#L47
Math.log goes to -Infinity for zero and multiplication with zero targets results into NaN

Steps to reproduce

        VectorValue in = new VectorValue(
                new double[] {
                        19056.858719905307, 4960.272544490949, 11971.046875371912,
                        5381.27145922252, 12381.873594408318, 4820.585859773736
                }
        );

        VectorValue target = new VectorValue(new double[] {0, 0, 0, 0, 0, 1});

        System.out.println(new SoftEntropy().evaluate(in, target));
        System.out.println(new SoftEntropy().differentiate(in, target));

Outputs:

NaN
[NaN,NaN,NaN,NaN,NaN,NaN]

Maven repository?

Is this nice library available as via some Maven Repository?

If not, may I suggest to consider using one othe following?

Maven Central - https://central.sonatype.org/publish/publish-guide/
GitHub Packages - https://github.com/features/packages
JitPack - https://jitpack.io/

This would greatly impact the re-usability of you library, IMHO.

Comparison method violates its general contract

The Value::compareTo violates the transitivity contract. For example:

VectorValue a = new VectorValue(new double[]{0.047629895714159626, 0.047604189443657435, 0.04854122567285466, 0.04761906321919965, 0.04738330389768908, 0.048210586527278, 0.047356453205930266, 0.04779385977269353, 0.04749880890690395, 0.049574893497338526, 0.0473965862144424, 0.04752080406323121, 0.05822692203068567, 0.04901807601925808, 0.04765455362607377, 0.05002604412125539, 0.05739356645498594, 0.047385976587184166, 0.0688271909926611, 0.047338000032517556});
VectorValue b = new VectorValue(new double[]{0.05230714877347989, 0.04768642416075893, 0.04828379450589333, 0.04748746042330981, 0.047413899100790274, 0.0480960111883369, 0.04760666402944034, 0.04780428311901668, 0.04741611488536501, 0.048049785025318044, 0.06407256569995881, 0.04745754447958131, 0.05822264359013865, 0.04743569428412833, 0.05136507865966237, 0.050132734147586, 0.04761751121968647, 0.05404798031492382, 0.0481002774579048, 0.04939638493472022});
VectorValue c = new VectorValue(new double[]{0.047687861504175824, 0.04853932979708919, 0.05375328444286901, 0.06789273268891113, 0.04752737473067581, 0.04765000054474707, 0.04752381588854653, 0.05315426060101445, 0.04825011418905211, 0.048603366978242836, 0.0482348444133695, 0.04745493855841848, 0.05201354770686542, 0.048210370136723586, 0.04875627392006865, 0.048451973613676286, 0.04850025765899207, 0.048359046978719435, 0.05203654183998997, 0.047400063807852756});

System.out.println("A  B " + a.compareTo(b)); // Outputs 0 - A == B
System.out.println("B  C " + b.compareTo(c)); // Outputs 0 - B == C
System.out.println("A  C " + a.compareTo(c)); // Outputs -1 - A < C

This becomes an issue when I want to assign VectorValue as a label for a query (for example, in the case of clutrr). There is sorting of a collection of result (Vector) values happening here, which requires transitivity property and crashes NeuraLogic (does not happen in the Python frontend, as this part is not used).

gustiks / neuralogic Goto Github PK

neuralogic's People

Contributors

Stargazers

Watchers

Forkers

neuralogic's Issues

Inconsistent Value formatting

Current behavior:

Expected behavior:

SoftEntropy overflow resulting in NaN

The bug description

Steps to reproduce

Maven repository?

Comparison method violates its general contract

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent