ltntorch 0.1.0

checksums.yaml

@@ -0,0 +1,7 @@
+---
+SHA256:
+  metadata.gz: f9989fe438d5906e6c64b19ad2fd67c5b5582a4747ceb9381d2feba4bee35e3c
+  data.tar.gz: 9a0740b3c81d2767e2fd3f6ea22c025613830aa4041b9c69af398c2cbe8bd0f9
+SHA512:
+  metadata.gz: a1cc22e1a1937a4d9a959a4b6e949cf0b4770e4b484c67d2acd283466870891cdec4798c96497570af62b6aeecdb1bfbc0004e32af3f64fa2cb519b594f047d3
+  data.tar.gz: 840e1596485a959e7d08b5628b04d0bd0b78eb98126b9b8fd84a22079f4dca0af402831b34e90e6a136560fbb1ceedb956e2e56bc1c4a6c9c7f783c504b6a709

data/AGENTS.md

@@ -0,0 +1,34 @@
+# Tech stack
+
+The present project does not use mainstream Ruby gems. Instead it uses:
+
+- the [sus](https://github.com/socketry/sus) gem for testing (instead of, e.g., Rspec)
+- the [bake](https://github.com/ioquatix/bake) gem for task management (instead of, e.g., rake)
+- the [standard](https://github.com/standardrb/standard) gem for linting/formatting
+- the [covered](https://github.com/socketry/covered) gem for coverage (instead of, e.g., SimpleCov) -- wired into `sus` via `config/sus.rb`; run with `COVERAGE=BriefSummary bundle exec sus` locally, or `bundle exec bake covered:validate --minimum <ratio>` to check the persisted `.covered.db` against a threshold (what CI does)
+- a gems.rb file instead of a Gemfile
+
+# Porting notes
+
+This library is a port of [LTNtorch](https://github.com/tommasocarraro/LTNtorch) to [torch.rb](https://github.com/ankane/torch.rb). Unless a difference is documented, behavior (including numerical results) should match LTNtorch's: when in doubt, check the Python source.
+
+## torch.rb pitfalls
+
+torch.rb 0.27.0 differs from PyTorch in ways that silently change results (see `TODO_torch-rb.md`):
+
+- `<`, `<=`, `>`, `>=` and `==` on tensors are not element-wise: use `lt`, `le`, `gt`, `ge`, `eq`.
+- Write float/tensor arithmetic tensor-first (`x * c`, `Torch.rsub(x, 1.0)` for `1.0 - x`): `Float op Tensor` rounds the float to float32.
+- Use `LTN::FuzzyOps.where` instead of `Torch.where`: with 0-dim tensors that do not require grad, `Torch.where` returns float32.
+- `Torch.tensor(data, requires_grad: true)` returns a non-leaf tensor that never receives gradients: use `Torch.tensor(data).requires_grad!(true)`.
+- `Torch.norm(x, dim: 1)` fails: pass `p` explicitly, `Torch.norm(x, 2, dim: 1)`.
+- `Torch.randint` returns float32 (PyTorch: int64): pass `dtype: :int64` for indices or labels. There is no `one_hot`: use `Torch.eye(n)[labels]`.
+
+## LTN pitfalls
+
+- `LTN.diag` changes the labels of its variables, so call it before building the formulas that use them, as in `forall.call(LTN.diag(x, y), p.call(x, y))` (where Ruby evaluates the arguments left to right); a formula built earlier keeps the variables' own labels, and the quantifier leaves them free.
+
+## Examples
+
+`examples/` and `tutorials/` port LTNtorch's. Each example is a class whose `#run(epochs:)` returns its final metrics (`test/examples.rb` runs them briefly, `benchmark/examples.rb` fully); `LTNExamples.main` runs it with LTNtorch's settings when the file is executed. Coverage (`config/covered.rb`) only measures `lib/`.
+
+Reference values come from LTNtorch itself: `test/fixtures/*.py` generate `test/fixtures/*.json` (see their headers), and new behavior should get a fixture case there, float64 and 0-dim variants included.

data/CONTRIBUTING.md

@@ -0,0 +1,66 @@
+# Contributing to LTNtorch-rb
+
+This document covers the practical mechanics of working on this codebase: setup, running tests, linting, and coverage. For the pull request process, the Developer Certificate of Origin, and the code of conduct, see the "Contributing" section of `readme.md` and `code_of_conduct.md`.
+
+## Tech stack
+
+This project uses the same, deliberately less-mainstream, tech stack as [reinforce](https://github.com/DSG-UniFE/reinforce) (see `AGENTS.md` for the short version):
+
+- [`sus`](https://github.com/socketry/sus), not RSpec, for testing.
+- [`bake`](https://github.com/ioquatix/bake), not Rake, for task running.
+- [`standard`](https://github.com/standardrb/standard) for linting/formatting.
+- [`covered`](https://github.com/socketry/covered), not SimpleCov, for coverage.
+- `gems.rb`, not `Gemfile`. `gems.locked` is intentionally **not** committed: this is a gem, not an application.
+
+## Setup
+
+This project depends on [torch.rb](https://github.com/ankane/torch.rb), which needs a matching LibTorch install to compile its native extension. `.github/workflows/ci.yml` is the source of truth for the exact version (`LIBTORCH_VERSION`); as of this writing, LibTorch 2.15.0 for torch.rb 0.27.x:
+
+```bash
+# macOS (Apple Silicon), either via Homebrew...
+brew install pytorch
+# ...or with the official archive
+curl -L https://download.pytorch.org/libtorch/cpu/libtorch-macos-arm64-2.15.0.zip -o libtorch.zip
+unzip -q libtorch.zip
+bundle config set build.torch-rb --with-torch-dir="$(pwd)/libtorch"
+
+# Linux (CPU)
+curl -L https://download.pytorch.org/libtorch/cpu/libtorch-shared-with-deps-2.15.0%2Bcpu.zip -o libtorch.zip
+unzip -q libtorch.zip
+bundle config set build.torch-rb --with-torch-dir="$(pwd)/libtorch"
+```
+
+Then install dependencies:
+
+```bash
+bundle install
+```
+
+## Running tests
+
+```bash
+bundle exec sus
+```
+
+## Linting
+
+```bash
+bundle exec standardrb          # check
+bundle exec standardrb --fix    # auto-fix what's safe to auto-fix
+```
+
+CI runs this as a blocking check with zero tolerated offenses.
+
+## Coverage
+
+```bash
+COVERAGE=BriefSummary bundle exec sus              # overall percentage + least-covered files
+COVERAGE=PartialSummary bundle exec sus            # only the uncovered code snippets
+bundle exec bake covered:validate --minimum 0.90   # what CI actually checks
+```
+
+## Code conventions
+
+- `lib/<path>.rb` and `test/<path>.rb` mirror each other 1:1 (e.g. `lib/ltn/fuzzy_ops.rb` corresponds to `test/ltn/fuzzy_ops.rb`).
+- This is a port: keep LTNtorch's semantics, including numerical results, unless a deviation is deliberate. Document every deliberate deviation in the code and in `changelog.md`.
+- Ruby idioms replace Python ones where that doesn't change behavior: keyword arguments instead of optional positional ones, `#call` (so `op.(x, y)` works) instead of `__call__`, blocks or lambdas instead of Python `lambda`s, and symbols instead of string flags.

data/changelog.md

@@ -0,0 +1,68 @@
+# Changelog
+
+All notable changes to this project will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [Unreleased]
+
+### Added
+
+- `LTN::FuzzyOps`, a port of LTNtorch's `ltn.fuzzy_ops`: the projections
+  `pi_0`/`pi_1`; the connective operators (`NotStandard`, `NotGodel`,
+  `AndMin`, `AndProd`, `AndLuk`, `OrMax`, `OrProbSum`, `OrLuk`,
+  `ImpliesKleeneDienes`, `ImpliesGodel`, `ImpliesReichenbach`,
+  `ImpliesGoguen`, `ImpliesLuk`, `Equiv`); the aggregators (`AggregMin`,
+  `AggregMean`, `AggregPMean`, `AggregPMeanError`), with masks for guarded
+  quantification; and `SatAgg`. Optional arguments are keywords
+  (`AndProd.new(stable: false)`, `agg.call(xs, dim: [0], p: 3)`), and
+  Python's ValueError becomes ArgumentError.
+- `LTN::LTNObject`, a grounding together with the labels of its free
+  variables.
+- The core of LTN, ported from LTNtorch's `ltn.core`: `LTN::Constant`
+  (optionally trainable), `LTN::Variable`, `LTN.process_ltn_objects` (LTN
+  broadcasting), `LTN::Predicate` and `LTN::Function` (grounded by a
+  `Torch::NN::Module`, a function given as `func:`, or a block),
+  `LTN::LambdaModel`, `LTN.diag`/`LTN.undiag`, `LTN::Connective` and
+  `LTN::Quantifier` (with guarded quantification, the guard given as
+  `cond_fn:` or a block). Quantifiers take `:forall`/`:exists` (LTNtorch's
+  `"f"`/`"e"` work too).
+- Ruby ports of LTNtorch's seven examples (`examples/`, with the Iris,
+  crabs and real estate datasets; MNIST is downloaded on first use) and four
+  tutorials (`tutorials/`), smoke-tested by `test/examples.rb`, plus a
+  seeded learning benchmark of the examples (`benchmark/examples.rb`), run
+  by CI as a separate job.
+- Example 4 takes `--default-init` (`init_weights: false`) to keep PyTorch's
+  default weight initialization: about 98% test accuracy on MNIST addition,
+  as in the LTN paper, instead of the about 48% of LTNtorch's notebook
+  initialization, which stays the default.
+- `test/ltn/training_parity.rb`: training a small knowledge base from the
+  same data and weights follows LTNtorch's loss at every epoch, from
+  `test/fixtures/training.py`.
+- Reference results for the core (broadcasting, predicates, functions,
+  connectives, quantifiers, guarded and diagonal quantification, and the
+  gradients of a small knowledge base), generated from LTNtorch by
+  `test/fixtures/core.py`; float64 and scalar cases for the fuzzy operators.
+- Reference values for every operator and aggregator, gradients included,
+  generated from LTNtorch by `test/fixtures/fuzzy_ops.py`.
+- Project scaffolding: `ltntorch` gem with the `LTN` module, `LTN.device`
+  configuration, sus/covered test setup, standard linting, and CI on
+  Ubuntu (every push) and macOS (scheduled).
+
+### Changed (compared to LTNtorch)
+
+- `LTN.process_ltn_objects`, `Quantifier.compute_mask` and
+  `Quantifier.transpose_vars` never modify their arguments. LTNtorch rebinds
+  the value and free variables of intermediate results it processes, so
+  reusing a formula after, e.g., a connective sees the broadcast grounding;
+  the computed values are the same.
+- `LTN::Connective#call` passes keyword arguments on to its operator
+  (e.g. `stable: false`); LTNtorch accepts and ignores them.
+- `LTN::Variable` converts any float64 individuals to float32; LTNtorch
+  only converts CPU ones.
+
+### Fixed (compared to LTNtorch)
+
+- `SatAgg` reports the shapes of formulas that are not closed; LTNtorch
+  raises an unrelated error while building that message.

data/code_of_conduct.md

@@ -0,0 +1,84 @@
+# Contributor Covenant Code of Conduct
+
+## Our Pledge
+
+We as members, contributors, and leaders pledge to make participation in our community a harassment-free experience for everyone, regardless of age, body size, visible or invisible disability, ethnicity, sex characteristics, gender identity and expression, level of experience, education, socio-economic status, nationality, personal appearance, race, religion, or sexual identity and orientation.
+
+We pledge to act and interact in ways that contribute to an open, welcoming, diverse, inclusive, and healthy community.
+
+## Our Standards
+
+Examples of behavior that contributes to a positive environment for our community include:
+
+* Demonstrating empathy and kindness toward other people
+* Being respectful of differing opinions, viewpoints, and experiences
+* Giving and gracefully accepting constructive feedback
+* Accepting responsibility and apologizing to those affected by our mistakes, and learning from the experience
+* Focusing on what is best not just for us as individuals, but for the overall community
+
+Examples of unacceptable behavior include:
+
+* The use of sexualized language or imagery, and sexual attention or
+  advances of any kind
+* Trolling, insulting or derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or email
+  address, without their explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Enforcement Responsibilities
+
+Community leaders are responsible for clarifying and enforcing our standards of acceptable behavior and will take appropriate and fair corrective action in response to any behavior that they deem inappropriate, threatening, offensive, or harmful.
+
+Community leaders have the right and responsibility to remove, edit, or reject comments, commits, code, wiki edits, issues, and other contributions that are not aligned to this Code of Conduct, and will communicate reasons for moderation decisions when appropriate.
+
+## Scope
+
+This Code of Conduct applies within all community spaces, and also applies when an individual is officially representing the community in public spaces. Examples of representing our community include using an official e-mail address, posting via an official social media account, or acting as an appointed representative at an online or offline event.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be reported to the community leaders responsible for enforcement at mauro.tortonesi@unife.it. All complaints will be reviewed and investigated promptly and fairly.
+
+All community leaders are obligated to respect the privacy and security of the reporter of any incident.
+
+## Enforcement Guidelines
+
+Community leaders will follow these Community Impact Guidelines in determining the consequences for any action they deem in violation of this Code of Conduct:
+
+### 1. Correction
+
+**Community Impact**: Use of inappropriate language or other behavior deemed unprofessional or unwelcome in the community.
+
+**Consequence**: A private, written warning from community leaders, providing clarity around the nature of the violation and an explanation of why the behavior was inappropriate. A public apology may be requested.
+
+### 2. Warning
+
+**Community Impact**: A violation through a single incident or series of actions.
+
+**Consequence**: A warning with consequences for continued behavior. No interaction with the people involved, including unsolicited interaction with those enforcing the Code of Conduct, for a specified period of time. This includes avoiding interactions in community spaces as well as external channels like social media. Violating these terms may lead to a temporary or permanent ban.
+
+### 3. Temporary Ban
+
+**Community Impact**: A serious violation of community standards, including sustained inappropriate behavior.
+
+**Consequence**: A temporary ban from any sort of interaction or public communication with the community for a specified period of time. No public or private interaction with the people involved, including unsolicited interaction with those enforcing the Code of Conduct, is allowed during this period. Violating these terms may lead to a permanent ban.
+
+### 4. Permanent Ban
+
+**Community Impact**: Demonstrating a pattern of violation of community standards, including sustained inappropriate behavior,  harassment of an individual, or aggression toward or disparagement of classes of individuals.
+
+**Consequence**: A permanent ban from any sort of public interaction within the community.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 2.0,
+available at https://www.contributor-covenant.org/version/2/0/code_of_conduct.html.
+
+Community Impact Guidelines were inspired by [Mozilla's code of conduct enforcement ladder](https://github.com/mozilla/diversity).
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see the FAQ at
+https://www.contributor-covenant.org/faq. Translations are available at https://www.contributor-covenant.org/translations.

data/examples/1_binary_classification.rb

@@ -0,0 +1,108 @@
+#!/usr/bin/env ruby
+# frozen_string_literal: true
+
+# Released under the MIT License.
+# Copyright, 2026, by Mauro Tortonesi.
+# Ported from LTNtorch's examples/1-binary_classification.ipynb, copyright 2021-2024 by Tommaso Carraro.
+
+# Binary classification: learn a predicate A(x) that holds for the points of
+# [0, 1]^2 within distance 0.3 of (0.5, 0.5), from the knowledge base
+#
+#   forall x_A: A(x_A)
+#   forall x_not_A: not A(x_not_A)
+#
+# where x_A and x_not_A are the positive and negative training examples.
+#
+#   bundle exec ruby examples/1_binary_classification.rb
+
+require_relative "support/example"
+
+module LTNExamples
+  class BinaryClassification
+    # The grounding of A: a multi-layer perceptron ending in a sigmoid.
+    class ModelA < Torch::NN::Module
+      def initialize
+        super
+        @layer1 = Torch::NN::Linear.new(2, 16)
+        @layer2 = Torch::NN::Linear.new(16, 16)
+        @layer3 = Torch::NN::Linear.new(16, 1)
+        @elu = Torch::NN::ELU.new
+        @sigmoid = Torch::NN::Sigmoid.new
+      end
+
+      def forward(x)
+        x = @elu.call(@layer1.call(x))
+        x = @elu.call(@layer2.call(x))
+        @sigmoid.call(@layer3.call(x))
+      end
+    end
+
+    attr_reader :a, :dataset, :labels
+
+    def initialize(seed: 2020, output: $stdout)
+      @output = output
+      LTNExamples.seed(seed)
+
+      @dataset = Torch.rand(100, 2)
+      @labels = Torch.sum(Torch.square(@dataset - 0.5), dim: 1).lt(0.09)
+
+      @a = LTN::Predicate.new(ModelA.new)
+      @not = LTN::Connective.new(LTN::FuzzyOps::NotStandard.new)
+      @forall = LTN::Quantifier.new(LTN::FuzzyOps::AggregPMeanError.new(p: 2), :forall)
+      @sat_agg = LTN::FuzzyOps::SatAgg.new
+    end
+
+    # @returns [Hash] the final satisfaction levels and accuracies.
+    def run(epochs: 1000, log_every: 20)
+      train, test = [@dataset[0...50], @labels[0...50]], [@dataset[50..], @labels[50..]]
+      optimizer = Torch::Optim::Adam.new(@a.parameters, lr: 0.001)
+
+      epochs.times do |epoch|
+        train_loss = 0.0
+        batches = LTNExamples.batches(train, batch_size: 64, shuffle: true)
+        batches.each do |data, labels|
+          optimizer.zero_grad
+          loss = 1.0 - satisfaction(data, labels)
+          loss.backward
+          optimizer.step
+          train_loss += loss.item
+        end
+        next unless (epoch % log_every).zero?
+
+        @output.puts format(" epoch %d | loss %.4f | Train Sat %.3f | Test Sat %.3f | Train Acc %.3f | Test Acc %.3f",
+          epoch, train_loss / batches.size, sat_level(*train), sat_level(*test), accuracy(*train), accuracy(*test))
+      end
+
+      {train_sat: sat_level(*train), test_sat: sat_level(*test),
+       train_accuracy: accuracy(*train), test_accuracy: accuracy(*test)}
+    end
+
+    # The satisfaction level of the knowledge base on the given examples.
+    def satisfaction(data, labels)
+      x_a = LTN::Variable.new("x_A", data[labels])
+      x_not_a = LTN::Variable.new("x_not_A", data[labels.logical_not])
+      @sat_agg.call(
+        @forall.call(x_a, @a.call(x_a)),
+        @forall.call(x_not_a, @not.call(@a.call(x_not_a)))
+      )
+    end
+
+    def sat_level(data, labels)
+      Torch.no_grad { satisfaction(data, labels).item }
+    end
+
+    def accuracy(data, labels)
+      predictions = Torch.no_grad { @a.model.call(data).squeeze(1).gt(0.5) }
+      LTNExamples.accuracy(predictions, labels)
+    end
+  end
+end
+
+LTNExamples.main(LTNExamples::BinaryClassification.new, __FILE__) do |example|
+  points, labels = example.dataset[50..], example.labels[50..]
+  predicted = Torch.no_grad { example.a.model.call(points).squeeze(1).gt(0.5) }
+  [["Ground truth (test data)", labels], ["A(x) > 0.5 (test data)", predicted]].each do |title, mask|
+    puts
+    puts LTNExamples.scatter({"A" => ["#", points[mask]], "not A" => [".", points[mask.logical_not]]}, title: title)
+  end
+end

data/examples/2_multi_class_single_label_classification.rb

@@ -0,0 +1,130 @@
+#!/usr/bin/env ruby
+# frozen_string_literal: true
+
+# Released under the MIT License.
+# Copyright, 2026, by Mauro Tortonesi.
+# Ported from LTNtorch's examples/2-multi_class_single_label_classification.ipynb, copyright 2021-2024 by
+# Tommaso Carraro.
+
+# Multi-class single-label classification of the Iris dataset: learn a
+# predicate P(x, l), "example x has label l", from the knowledge base
+#
+#   forall x_A: P(x_A, l_A)
+#   forall x_B: P(x_B, l_B)
+#   forall x_C: P(x_C, l_C)
+#
+# where l_A, l_B, l_C are one-hot constants for the three species. P is a
+# softmax classifier, so each example has exactly one label.
+#
+#   bundle exec ruby examples/2_multi_class_single_label_classification.rb
+
+require_relative "support/example"
+
+module LTNExamples
+  class MultiClassSingleLabelClassification
+    # Returns the logits of the classes of x (the softmax is applied by
+    # LogitsToPredicate, so that the logits can be used for accuracy).
+    class MLP < Torch::NN::Module
+      def initialize(layer_sizes = [4, 16, 16, 8, 3])
+        super()
+        @elu = Torch::NN::ELU.new
+        @dropout = Torch::NN::Dropout.new(p: 0.2)
+        @linear_layers = Torch::NN::ModuleList.new(
+          layer_sizes.each_cons(2).map { |input, output| Torch::NN::Linear.new(input, output) }
+        )
+      end
+
+      def forward(x, training: false)
+        @linear_layers.to_a[0...-1].each do |layer|
+          x = @elu.call(layer.call(x))
+          x = @dropout.call(x) if training
+        end
+        @linear_layers.to_a.last.call(x)
+      end
+    end
+
+    # The truth value of P(x, l): the softmax probability of the class l
+    # (one-hot) of example x.
+    class LogitsToPredicate < Torch::NN::Module
+      def initialize(logits_model)
+        super()
+        @logits_model = logits_model
+        @softmax = Torch::NN::Softmax.new(dim: 1)
+      end
+
+      def forward(x, l, training: false)
+        probabilities = @softmax.call(@logits_model.call(x, training: training))
+        Torch.sum(probabilities * l, dim: 1)
+      end
+    end
+
+    def initialize(seed: 2020, output: $stdout)
+      @output = output
+      LTNExamples.seed(seed)
+
+      @train = load("iris_training.csv")
+      @test = load("iris_test.csv")
+
+      @l_a = LTN::Constant.new(Torch.tensor([1, 0, 0]))
+      @l_b = LTN::Constant.new(Torch.tensor([0, 1, 0]))
+      @l_c = LTN::Constant.new(Torch.tensor([0, 0, 1]))
+
+      @mlp = MLP.new
+      @p = LTN::Predicate.new(LogitsToPredicate.new(@mlp))
+      @forall = LTN::Quantifier.new(LTN::FuzzyOps::AggregPMeanError.new(p: 2), :forall)
+      @sat_agg = LTN::FuzzyOps::SatAgg.new
+    end
+
+    def run(epochs: 500, log_every: 20)
+      optimizer = Torch::Optim::Adam.new(@p.parameters, lr: 0.001)
+
+      epochs.times do |epoch|
+        train_loss = 0.0
+        batches = LTNExamples.batches(@train, batch_size: 64, shuffle: true)
+        batches.each do |data, labels|
+          optimizer.zero_grad
+          loss = 1.0 - satisfaction(data, labels, training: true)
+          loss.backward
+          optimizer.step
+          train_loss += loss.item
+        end
+        next unless (epoch % log_every).zero?
+
+        @output.puts format(" epoch %d | loss %.4f | Train Sat %.3f | Test Sat %.3f | Train Acc %.3f | Test Acc %.3f",
+          epoch, train_loss / batches.size, sat_level(*@train), sat_level(*@test), accuracy(*@train), accuracy(*@test))
+      end
+
+      {train_sat: sat_level(*@train), test_sat: sat_level(*@test),
+       train_accuracy: accuracy(*@train), test_accuracy: accuracy(*@test)}
+    end
+
+    def satisfaction(data, labels, training: false)
+      x_a = LTN::Variable.new("x_A", data[labels.eq(0)])
+      x_b = LTN::Variable.new("x_B", data[labels.eq(1)])
+      x_c = LTN::Variable.new("x_C", data[labels.eq(2)])
+      @sat_agg.call(
+        @forall.call(x_a, @p.call(x_a, @l_a, training: training)),
+        @forall.call(x_b, @p.call(x_b, @l_b, training: training)),
+        @forall.call(x_c, @p.call(x_c, @l_c, training: training))
+      )
+    end
+
+    def sat_level(data, labels)
+      Torch.no_grad { satisfaction(data, labels).item }
+    end
+
+    def accuracy(data, labels)
+      predictions = Torch.no_grad { Torch.argmax(@mlp.call(data), dim: 1) }
+      LTNExamples.accuracy(predictions, labels)
+    end
+
+    private
+
+    def load(name)
+      _, rows = LTNExamples.read_table(name)
+      [Torch.tensor(rows.map { |row| row[0...4].map(&:to_f) }), Torch.tensor(rows.map { |row| Integer(row[4]) })]
+    end
+  end
+end
+
+LTNExamples.main(LTNExamples::MultiClassSingleLabelClassification.new, __FILE__)

data/examples/3_multi_class_multi_label_classification.rb

@@ -0,0 +1,158 @@
+#!/usr/bin/env ruby
+# frozen_string_literal: true
+
+# Released under the MIT License.
+# Copyright, 2026, by Mauro Tortonesi.
+# Ported from LTNtorch's examples/3-multi_class_multi_label_classification.ipynb, copyright 2021-2024 by
+# Tommaso Carraro.
+
+# Multi-class multi-label classification of the Leptograpsus crabs dataset:
+# each crab has a color (blue or orange) and a sex (male or female). Learn a
+# predicate P(x, l), "crab x has label l", from the knowledge base
+#
+#   forall x_blue: P(x_blue, l_blue)        (and likewise for orange, male, female)
+#   forall x: not (P(x, l_blue) and P(x, l_orange))
+#   forall x: not (P(x, l_male) and P(x, l_female))
+#
+# P is a sigmoid classifier, so the labels are learned independently, and the
+# last two axioms state that they are mutually exclusive. After training, the
+# script also queries three formulas the knowledge base does not state:
+#
+#   phi1: forall x: P(x, l_blue) -> not P(x, l_orange)   (implied by the axioms)
+#   phi2: forall x: P(x, l_blue) -> P(x, l_orange)       (contradicted by them)
+#   phi3: forall x: P(x, l_blue) -> P(x, l_male)         (independent of them)
+#
+#   bundle exec ruby examples/3_multi_class_multi_label_classification.rb
+
+require_relative "support/example"
+
+module LTNExamples
+  class MultiClassMultiLabelClassification
+    # Returns the logits of the four labels of x.
+    class MLP < Torch::NN::Module
+      def initialize(layer_sizes = [5, 16, 16, 8, 4])
+        super()
+        @elu = Torch::NN::ELU.new
+        @dropout = Torch::NN::Dropout.new(p: 0.2)
+        @linear_layers = Torch::NN::ModuleList.new(
+          layer_sizes.each_cons(2).map { |input, output| Torch::NN::Linear.new(input, output) }
+        )
+      end
+
+      def forward(x, training: false)
+        @linear_layers.to_a[0...-1].each do |layer|
+          x = @elu.call(layer.call(x))
+          x = @dropout.call(x) if training
+        end
+        @linear_layers.to_a.last.call(x)
+      end
+    end
+
+    # The truth value of P(x, l): the sigmoid probability of the label l
+    # (one-hot) of example x.
+    class LogitsToPredicate < Torch::NN::Module
+      def initialize(logits_model)
+        super()
+        @logits_model = logits_model
+        @sigmoid = Torch::NN::Sigmoid.new
+      end
+
+      def forward(x, l, training: false)
+        Torch.sum(@sigmoid.call(@logits_model.call(x, training: training)) * l, dim: 1)
+      end
+    end
+
+    # Label codes, as in LTNtorch: colors 0 (blue) and 1 (orange), sexes 2 (male) and 3 (female).
+    CODES = {"B" => 0, "O" => 1, "M" => 2, "F" => 3}.freeze
+
+    def initialize(seed: 2020, output: $stdout)
+      @output = output
+      LTNExamples.seed(seed)
+
+      header, rows = LTNExamples.read_table("crabs.dat", separator: /\s+/)
+      rows = rows.values_at(*Torch.randperm(rows.size).to_a) # shuffle, as LTNtorch does
+      features = Torch.tensor(rows.map { |row| %w[FL RW CL CW BD].map { |column| Float(row[header.index(column)]) } })
+      sex = Torch.tensor(rows.map { |row| CODES.fetch(row[header.index("sex")]) })
+      color = Torch.tensor(rows.map { |row| CODES.fetch(row[header.index("sp")]) })
+      @train = [features[0...160], sex[0...160], color[0...160]]
+      @test = [features[160..], sex[160..], color[160..]]
+
+      @l_blue, @l_orange, @l_male, @l_female = Torch.eye(4).to_a.map { |onehot| LTN::Constant.new(Torch.tensor(onehot)) }
+
+      @mlp = MLP.new
+      @p = LTN::Predicate.new(LogitsToPredicate.new(@mlp))
+      @not = LTN::Connective.new(LTN::FuzzyOps::NotStandard.new)
+      @and = LTN::Connective.new(LTN::FuzzyOps::AndProd.new)
+      @implies = LTN::Connective.new(LTN::FuzzyOps::ImpliesReichenbach.new)
+      @forall = LTN::Quantifier.new(LTN::FuzzyOps::AggregPMeanError.new(p: 2), :forall)
+      @sat_agg = LTN::FuzzyOps::SatAgg.new
+    end
+
+    def run(epochs: 500, log_every: 20)
+      optimizer = Torch::Optim::Adam.new(@p.parameters, lr: 0.001)
+
+      epochs.times do |epoch|
+        train_loss = 0.0
+        batches = LTNExamples.batches(@train, batch_size: 64, shuffle: true)
+        batches.each do |data, sex, color|
+          optimizer.zero_grad
+          loss = 1.0 - satisfaction(data, sex, color)
+          loss.backward
+          optimizer.step
+          train_loss += loss.item
+        end
+        next unless (epoch % log_every).zero?
+
+        @output.puts format(" epoch %d | loss %.4f | Train Sat %.3f | Test Sat %.3f | Train Acc %.3f | " \
+          "Test Acc %.3f | Test Sat Phi 1 %.3f | Test Sat Phi 2 %.3f | Test Sat Phi 3 %.3f",
+          epoch, train_loss / batches.size, sat_level(*@train), sat_level(*@test), accuracy(*@train),
+          accuracy(*@test), *(1..3).map { |i| phi(i, @test[0]) })
+      end
+
+      {train_sat: sat_level(*@train), test_sat: sat_level(*@test),
+       train_accuracy: accuracy(*@train), test_accuracy: accuracy(*@test),
+       phi1: phi(1, @test[0]), phi2: phi(2, @test[0]), phi3: phi(3, @test[0])}
+    end
+
+    def satisfaction(data, sex, color)
+      x = LTN::Variable.new("x", data)
+      x_blue = LTN::Variable.new("x_blue", data[color.eq(0)])
+      x_orange = LTN::Variable.new("x_orange", data[color.eq(1)])
+      x_male = LTN::Variable.new("x_male", data[sex.eq(2)])
+      x_female = LTN::Variable.new("x_female", data[sex.eq(3)])
+      @sat_agg.call(
+        @forall.call(x_blue, @p.call(x_blue, @l_blue)),
+        @forall.call(x_orange, @p.call(x_orange, @l_orange)),
+        @forall.call(x_male, @p.call(x_male, @l_male)),
+        @forall.call(x_female, @p.call(x_female, @l_female)),
+        @forall.call(x, @not.call(@and.call(@p.call(x, @l_blue), @p.call(x, @l_orange)))),
+        @forall.call(x, @not.call(@and.call(@p.call(x, @l_male), @p.call(x, @l_female))))
+      )
+    end
+
+    def sat_level(data, sex, color)
+      Torch.no_grad { satisfaction(data, sex, color).item }
+    end
+
+    # The truth value of one of the queries phi1, phi2, phi3 on the given examples.
+    def phi(index, data)
+      Torch.no_grad do
+        x = LTN::Variable.new("x", data)
+        blue = @p.call(x, @l_blue)
+        consequent = {1 => -> { @not.call(@p.call(x, @l_orange)) }, 2 => -> { @p.call(x, @l_orange) },
+                      3 => -> { @p.call(x, @l_male) }}.fetch(index).call
+        @forall.call(x, @implies.call(blue, consequent), p: 5).value.item
+      end
+    end
+
+    # One minus the Hamming loss of the four labels. Like LTNtorch, it
+    # thresholds the logits (not the probabilities) at 0.5.
+    def accuracy(data, sex, color)
+      predictions = Torch.no_grad { @mlp.call(data).gt(0.5) }
+      onehot = Torch.stack([color.eq(0), color.eq(1), sex.eq(2), sex.eq(3)], dim: -1)
+      predictions.eq(onehot).float.mean.item
+    end
+  end
+end
+
+LTNExamples.main(LTNExamples::MultiClassMultiLabelClassification.new, __FILE__)