veas_pilot_forecast_cv_rnn.yaml

# @package _global_

# example hyperparameter optimization of some experiment with Optuna:
# python train.py -m hparams_search=mnist_optuna experiment=example

defaults:
  - veas_pilot_cv_base
  - override /experiment: veas_pilot_forecast
  - override /model: veas_pilot_rnn_forecast

use_dtw: null
model:
  loss_fn:
    _target_: ${eval:'"src.metrics.soft_dtw.SoftDTWMetric" if ${use_dtw} else "torch.nn.MSELoss"'}

log_hyperparameters_custom:
  - "use_dtw"

#trainer:
#  accelerator: ${eval:'"cpu" if ${use_dtw} else "auto"'}

# here we define Optuna hyperparameter search
# it optimizes for value returned from function with @hydra.main decorator
# docs: https://hydra.cc/docs/next/plugins/optuna_sweeper
hydra:
  sweeper:
    study_name: veas_pilot-forecast-cv-rnn
    # define hyperparameter search space
    params:
      ++model.hidden_dim: range(1, 200)
      ++model.dropout: range(0,0.6,0.1)
      ++model.n_rnn_layers: range(1,5)
      ++model.input_chunk_length: range(1, 288)
      use_dtw: choice(True, False)
      ++model.optimizer_kwargs.lr: interval(1e-5,1e-2)
      use_inputs.oxygen: choice(True, False)
      use_inputs.ammonium: choice(True, False)
      use_inputs.filterpressure_1: choice(True, False)
      use_inputs.filterpressure_8: choice(True, False)
      use_inputs.turb: choice(True, False)
      use_inputs.temp: choice(True, False)
      use_inputs.methanol: choice(True, False)
      use_inputs.orto-p: choice(True, False)
      use_inputs.tunnelwater: choice(True, False)