How genaibiomed/genaibiomedassignment1 works
This repository is an assignment-stage protein sequence language-modeling project, rather than the MkDocs preview workflow described in the README. Its intended user workflow is to load protein sequences, train a pretrained ProGen causal language model, then generate sequence designs. Training and inference remain incomplete: both contain NotImplementedError placeholders, and inference also leaves checkpoint loading and prediction unfinished. The architecture reflects only relationships visible in the supplied source excerpts and README; data files are inputs, not separately implemented subsystems.
Main entry points: inference.py Inference workflow · dataset.py Dataset and batches · model.py Causal LM · train.py Training workflow
Components 10
Sequence data
- Protein splits
data - Dataset and batches
dataset.py - Protein vocabulary
progen_vocab.py
Model runtime
- Causal LM
model.py
Training and design
- Training workflow
train.py - Inference workflow
inference.py - Best checkpoint
train.py - Design sequences
inference.py
Other components
- Assignment user
- Pretrained ProGen
Connections 12
- Assignment user → Training workflow: runs training
train.py - Assignment user → Inference workflow: runs design
inference.py - Protein splits → Dataset and batches: loads splits
dataset.py - Dataset and batches → Protein vocabulary: encodes sequences
dataset.py - Training workflow → Dataset and batches: requests batches
train.py - Training workflow → Causal LM: loads and trains
train.py - Causal LM → Best checkpoint: saves weights
train.py - Inference workflow → Dataset and batches: requests test batches
inference.py - Inference workflow → Causal LM: loads model
inference.py - Pretrained ProGen → Causal LM: provides weights
- Inference workflow → Best checkpoint: loads weights
- Inference workflow → Design sequences: writes designs
inference.py