LayCoder-Data

This repository contains the processed dataset subsets DS1 and DS2 used in the paper:

LayCoder: UI Layout Completion with an Encoder-Only Transformer and Layout Tokenizer

---

📘 Overview

The LayCoder project focuses on UI layout completion using an encoder-only transformer architecture.
This repository provides the processed layout data used for model training, validation, and testing.

• The data is derived from the LayoutTransformer++ dataset (Microsoft Research),
  which itself is based on the Rico dataset (https://interactionmining.org/rico.html).
• We further preprocessed the original layouts to create two subsets, DS1 and DS2, designed for different experimental configurations.
• Each subset contains train, validation, and test partitions serialized as .pkl files.

---

📂 Dataset Structure

Folder	Description
DS1/	Subset 1 – used for baseline training and primary evaluation
DS2/	Subset 2 – alternative processed version for comparison and ablation experiments

Each .pkl file contains a Python list of layout objects, formatted for direct use with the LayCoder data loaders.

---

⚙️ Usage Example

import pickle

# Load training data from DS1
with open('DS1/train_DS1.pkl', 'rb') as f:
    train_data = pickle.load(f)

print(f"Loaded DS1 training set with {len(train_data)} samples.")