config/callbacks/model_checkpoint.yaml

# https://lightning.ai/docs/pytorch/stable/api/lightning.pytorch.callbacks.ModelCheckpoint.html

model_checkpoint:
  # _target_: utils.modelcheckpoint_timestamp.DynamicTimestampModelCheckpoint
  _target_: lightning.pytorch.callbacks.ModelCheckpoint
  dirpath: artifacts/checkpoints # We need to modify this base on distributed or not
  filename: "model-best-step{step:08d}-loss{loss/mse_epoch:.6f}" # checkpoint filename
  monitor: "trainer/global_step" # name of the logged metric which determines when model is improving
  verbose: True # verbosity mode
  save_last: True # additionally always save an exact copy of the last checkpoint to a file last.ckpt
  save_top_k: 50 # save k best models (determined by above metric)
  mode: "max" # "max" means higher metric value is better, can be also "min"
  auto_insert_metric_name: False # when True, the checkpoints filenames will contain the metric name
  save_weights_only: False # if True, then only the model’s weights will be saved
  save_on_train_epoch_end: False # whether to run checkpointing at the end of the training epoch or the end of validation
  every_n_train_steps: ${trainer.val_check_interval}