stldec_random

Materials for the paper "Bridging Logic and Learning: Decoding Temporal Logic Embeddings via Transformers" (Candussio et al.) @ ECML-PKDD 2025

TL;DR:

• (trained) models are available at: https://huggingface.co/collections/saracandu/stldec-ecml-pkdd-2025-686fe174a16915bc32aa53eb
• code, results, and other details can be found in this repo.

The goal of STLdecoder is to take a NeSy embedding of a Signal Temporal Logic (STL) formula and recover a semantically equivalent formula.

The encoder.py file allows you to obtain the NeSy embeddings of (a list of) formulae with respect to a predefined anchor set, which you can find in the anchor_sets/ folder. More details on this procedure can be found at https://ebooks.iospress.nl/doi/10.3233/FAIA240638
This class also relies on the following files: phis_generator.py, traj_measure.py, kernel.py, stl.py, anchor_set_generation.py, custom_typing.py, trajectories.py.

The decoder.py component aims at translating a vector (i.e., the encoding of a formula, as done by encoder.py) into a string (i.e., an STL formula consisting of a hybrid syntax made of numbers, parentheses, and words, whose vocabulary can be found in the tokenizer_files/ folder).
This is practically implemented in the modeling_stldec.py file, as we perform the aforementioned procedure using a decoder-only Transformer architecture. This process requires autoregressively generating the tokens of the STL formula and embedding them in order to merge this information with the initial semantic vector through the cross-attention block. The configuration.py file serves as a crystallized structure guiding the transformers classes.

In order to train this architecture, we can use the training.py file, leveraging the different training settings available in the training_config/ folder.