Release 2.1.0

2.1.0 (2025-01-07)

Feat

• prepare: outputs additional column signature_morgans
• learning: add transformer code
• dataset: add code to compute model tokens
• dataset: add code for download and prepare datasets
• transformer/train: additional arg for setting source / target max length
• transformer/train: implement gradient accumulation
• transformer/train: define num of data loader workers from args
• transformer/train: make modele compilation by Torch optional
• transformer/train: generalize mixed precision scaler usage
• transformer/model: refine state_dict Module's method
• transformer/train: check for NaNs in loss
• transformer/train: model dir output as arg
• transformer/train: experimentation with mixed precision floats
• transformer/train: make use of pin_memory=true in dataloaders expected to increase GPU perf
• transformer/train: first working version
• transformer: in dev code
• new code to download and make use of the signature code (#10)

Fix

• prepare: remove deprecated import
• get_smiles: remove superflous Hs
• prepare: sanitize molecule after stereo-isomer enumeration
• prepare: add missing header
• update changelog on version bump
• attempt to trigger GA
• main instead of master branch name
• dataset: remove unused code
• transformer/train: load_checkpoint
• transformer/train: effective batch indexes
• transformer/train: duplicated loss normalization
• transformer/train: wrong arg name
• transformer/train: take into account remaining remaining batches for the sceduler counts
• transformer/train: propagate gradient for last batches of epoch
• transformer/train: remove multiple calls to unscale_
• transformer/train: use save_checkpoint
• transformer/train: refine save and load methods
• transformer/train: correct seq length arg
• transformer/train: stop sending to preset device
• dataset/utils.py: forward pass logger in recursive calls
• tokenizer: allow additional depictions

Refactor

• remove old code
• .env: ignore local env file
• erase old code
• transformer: sweep code
• dataset: clean deprecated code
• transformer: remove deprecated code
• transformer/train: refine gradient accumulation
• transformer/config: reduce learning rate to prevent NaN / Inf values
• transformer/train: make GPU pinned memory an option
• transformer/train: add few debug messages
• transformer/config: update
• transformer/config: update
• transformer/train: get the number of epochs from config
• transformer/train: better log Nan / Inf value issues
• transformer/config: increase learning rate
• transformer/config: increase learning rate
• transformer/config: reduce learning rate
• transformer/train: update default log level
• transformer/train: better handle device arg
• transformer/config.yaml: update training values
• model: remove unecessary code
• dataset/utils.py: don't sort config keys
• download: update paths

Perf

• transformer/train: AdamW optimizer instead of Adam, OneCycleLR scheduler