tusgan-v5

About Model

TUS-GAN v5 is a conditional Wasserstein GAN with Gradient Penalty (WGAN-GP) designed to generate synthetic time-use diary sequences, developed as part of research on time-use pattern modeling under Dr. Agnivesh Pani at SCULPT-Lab, IIT (BHU) Varanasi. Architecture: The model consists of a Generator (1.4M parameters) and a Critic (902K parameters). Diary sequences are represented as tensors of shape (B, 3, 48, 1) — 3 channels encoding activity code, location state, and day-type, across 48 half-hour slots covering a 24-hour period anchored at 04:00. Generation is conditioned on an 82-dimensional vector encompassing demographic and geographic attributes, including embeddings for 71 districts and 36 states. Both 2019 and 2024 India Time Use Survey (ITUS) editions use a consistent 7-category location code scheme. Training: The model was trained on 445,268 diary records for 250 epochs on an NVIDIA A100 GPU (JarvisLabs). Training stability was achieved through WGAN-GP's gradient penalty formulation, addressing mode collapse and critic-generator imbalance issues observed in earlier iterations. Evaluation: Model quality was assessed using t-SNE projections to visualize the distribution of generated vs. real diary sequences, alongside quantitative diversity metrics across checkpoints (epochs 140, 200, 250). Use Case: TUS-GAN enables researchers and policymakers to augment incomplete or sparse time-use survey data, supporting downstream analysis of activity patterns, gender time-use gaps, and regional labor behavior — particularly useful where survey non-response or missing entries limit direct analysis.