Trainer

AIDO.ModelGenerator uses the LightningCLI for configuring runs with the PyTorch Lightning Trainer. The entrypoint for the CLI is mgen, which can be used with the fit, test, validate, and predict commands and the --model, --data, and --trainer arguments and their sub-arguments.

mgen fit --model ConditionalDiffusion --model.backbone aido_dna_300m \
  --data ConditionalDiffusionDataModule --data.path "genbio-ai/100m-random-promoters" \
  --trainer.max_epochs 1 --trainer.accelerator auto --trainer.devices auto

For detailed information about the LightningCLI, see the LightningCLI documentation.

# Example Trainer Configuration
trainer:
  accelerator: auto
  strategy: lightning.pytorch.strategies.DDPStrategy
  devices: auto
  num_nodes: 1
  precision: bf16-mixed
  logger: null
  callbacks:
  - class_path: lightning.pytorch.callbacks.ModelCheckpoint
    init_args:
      filename: best_val:{step}-{val_loss:.3f}-{train_loss:.3f}
      monitor: val_loss
      save_top_k: 1
  fast_dev_run: false
  max_epochs: 100
  limit_val_batches: null
  val_check_interval: null
  check_val_every_n_epoch: 1
  log_every_n_steps: 50
  accumulate_grad_batches: 1
  gradient_clip_val: 1
  gradient_clip_algorithm: null
  detect_anomaly: false
  default_root_dir: logs
model:
  ...
data:
  ...