LaMer. Implementation of "Non-Parallel Text Style Transfer with Self-Parallel Supervision" (ICLR 2022) — mines roughly parallel sentences from non-parallel data via scene graphs, trains BART with MLE, and refines with contrastive imitation learning.