A PyTorch-based VAE-Transformer model for predicting reactants from products in chemical reactions. Includes SMILES tokenization, latent-space encoding, and autoregressive decoding. Supports sampling, checkpointing, and experiment tracking with WandB.
- VAE-Transformer architecture with continuous latent space.
- Autoregressive sampling with argmax decoding for retrosynthesis.
- SMILES tokenization with halogen simplification ([Cl] → L, [Br] → R) and restoration.
- WandB integration for training logs, learning rate, and model checkpoints.
- Checkpoint loading for fine-tuning or inference.
- Supports long SMILES sequences (up to 1300 tokens).
The dataset has been compressed for easier upload.
- Dataset:
uspto.zip(containsuspto.csv)