default.yaml

# See available models keys at: downstream_tasks/test_modules/__init__.py
MODEL: "secoeco"

# See available datasets keys at: downstream_tasks/datasets/__init__.py
DATASET: "biomes"

# Evaluation mode, either 'linear' or 'knn'
PROBE: "linear"

USE_PRECOMPUTED_EMBEDDINGS: False
USE_PRECOMPUTED_TEST_PREDICTIONS: False

SAVE_NAME_linear: "${oc.env:SECOECO_ROOT}/experiment_files/${MODEL}_${DATASET}_linear"
SAVE_NAME_knn: "${oc.env:SECOECO_ROOT}/experiment_files/${MODEL}_${DATASET}_${K_PROBE}-nn"
SAVE_NAME: "${oc.select:SAVE_NAME_${PROBE}, ${oc.env:SECOECO_ROOT}/experiment_files/${MODEL}_${DATASET}_default}"

EPOCHS: 1000
EARLY_STOPPING_PATIENCE: 10
EARLY_STOPPING_SENSITIVITY: 0.0
BATCH_SIZE: 256
LEARNING_RATE: 1e-3
K_FOLDS: 10  # implicitly rules the size of the test split
NUM_USED_FOLDS: 10  # may be lower than K_FOLDS if we do not want to run all folds
NUM_WORKERS: 16
TRAIN_VAL_SPLIT: [0.9, 0.1]
SEED: 42
DEVICE: "cuda"
K_PROBE: -1  # number of neighbors used when PROBE='knn'. Setting <1 will use the optimal value for the dataset
TEMP_PROBE: 0.07  # number of neighbors used when PROBE='knn'

SAVE_TSNE: False # True False

hydra:
  # When running a single config, then the output path is defined here
  run:
    dir: "${SAVE_NAME}/output/${now:%Y-%m-%d-%H-%M-%S}"
  # When doing a multirun, the output path is defined here
  # subdir can be adapted to keep the sweeped arguments apart (e.g. putting LR here instead of model-dataset)
  sweep:
    dir: "${oc.env:SECOECO_ROOT}/multirun/${now:%Y-%m-%d-%H-%M-%S}"
    subdir: ${MODEL}_${DATASET}_${PROBE}