train.yaml

# @package _global_

# specify here default configuration
# order of defaults determines the order in which configs override each other
defaults:
  - _self_
  - dataset: conll2003.yaml
  - datamodule: default.yaml
  - taskmodule: transformer_token_classification.yaml
  - model: transformer_token_classification.yaml
  - callbacks: default.yaml
  - logger: null # set logger here or use command line (e.g. `python train.py logger=tensorboard`)
  - trainer: default.yaml
  - paths: default.yaml
  - extras: default.yaml
  - hydra: default.yaml

  # experiment configs allow for version control of specific hyperparameters
  # e.g. best hyperparameters for given model and taskmodule
  - experiment: null

  # config for hyperparameter optimization
  - hparams_search: null

  # optional local config for machine/user specific settings
  # it's optional since it doesn't need to exist and is excluded from version control
  - optional local: default.yaml

  # debugging config (enable through command line, e.g. `python train.py debug=default)
  - debug: null

# task name, determines output directory path
pipeline_type: "training"

# default name for the experiment, determines output directory path
# (you can overwrite this in experiment configs)
name: "default"

# tags to help you identify your experiments
# you can overwrite this in experiment configs
# overwrite from command line with `python train.py tags="[first_tag, second_tag]"`
# appending lists from command line is currently not supported :(
# https://github.com/facebookresearch/hydra/issues/1547
tags: ["dev"]

# set False to skip model training
train: True

# evaluate on test set, using best model weights achieved during training
# lightning chooses best weights based on the metric specified in checkpoint callback
test: False

# metric with mode (minimize or maximize) to monitor for checkpointing and early stopping callbacks
monitor_metric: "val/f1"
monitor_mode: "max"

# seed for random number generators in pytorch, numpy and python.random
seed: null

# simply provide checkpoint path to resume training
ckpt_path: null

# push the model and taskmodule to the huggingface model hub when training has finished
push_to_hub: False

# where to save the trained model and taskmodule
model_save_dir: ${paths.save_dir}/models/${name}/${now:%Y-%m-%d_%H-%M-%S}