parameter_descriptions.md

Specifying keywoards and valid parameters that can be used to create datasets and configure ML-Methods and training:

During data set creation, set up of the ML methods and the training, choices can be made. The following table gives a list of possible parameters and valid configurations.

Datasets (dataset_description):

Name	Purpose
DSET_NB	A number assigne to the dataset. Used to create multple datasets with the same configuration
GRID_TYPE	"Flat" or "Ico"
CALENDAR	Stores the calendar types of the used climate model (filled automatically)
T_UNITS	Stores the time units used by the climate model (filled automatically)
REFERENCE_DATE	Stores the reference date used by the climate model, i.e. in reference to what date timesteps are stored (filled automatically)
DO_SHUFFLE	Whether or not the data should be shuffled before splitting into test and training set (our default: don't shuffle)
TEST_FRACTION	What fraction of the dataset is split of as training set
SPLIT_YEAR	Alternatively to giving a TEST_FRACTION, one can also provide a year, based on which the data will be split into test and training set
CLIMATE_MODEL	Specify which climate model is used: Valid arguments: "iHadCM3", "GISS", "ECHAM5", "isoGSM", "iCESM"
LATITUDES_SLICE	If using a flat grid, exclude this number of pixels at bottom and top of latitude, e.g. [1,-1] excludes the first and last latitude value. (flat grid only)
LATITUDES	Array containing the latitudes of the used flat grid (flat grids only), is filled automatically.
LONGITUDES	Array containing the longitudes of the used flat grid (flat grids only), is filled automatically.
START_YEAR	Year with which the valid part of the last millennium run starts
END_YEAR	Year with which the valid part of the last millennium run ends
DATASETS_NO_GAPS	Datasets containing variables of which we require variables to be present at each time step.
DATASETS_USED	Used datasets. Names need to match with file names, e.g. "tsurf" to use "tsurf.nc" or "tsurf_yearly.nc"
PREDICTOR_VARIABLES	Which of the variables are used as predictors, dict containing DATASETS_USED as keys and variable name als values.
TARGET_VARIABLES	Which of the variables are used as targets, dict containing DATASETS_USED as keys and variable name als values.
TIMESCALE	What timescale we want to work on: "yearly" or "monthly"
MONTHS_USED	If on monthly timescale: one can decide to only use certain months in the dataset, e.g [0,1] would indicate a dataset only containing January and February timesteps.
MONTHS_USED_IN_PREDICTION	For each timestep, it is possible to include timesteps from future or past in the emulation: To use the current and the previous months in the prediction, e.g. np.sort([0,-1]).tolist() or [0] to only use the current months
RESOLUTION	When using icosahedral data: Refinement level of the grid (interpolated data on the given resolution required)
INTERPOLATE_CORNERS	Ico: Whether or not we interpolate the corners pixels of the icosahedron
INTERPOLATION	The type of interpolation used by cdo, usually cons1 (first order conservative), NN (nearest neighbor) possible to.
INDICES_TEST	Indices of the test set in the dataset (automatically generated)
INDICES_TRAIN	Indices of the training set in the dataset (automatically generated)
TIMESTEPS_TEST	Time steps of the test set in the dataset (automatically generated)

| PRECIP_WEIGHTING | Whether or not to weight individual months by precipitation amount when creating yearly datasets (only for yearly time scale) | | RESULTS_INTERPOLATED | Whether or not the emulation results have been interpolated from one grid to another after the emulation (set automatically during interpolation, GRID_TYPE is changed as well) | | RESULTS_RESCALED | Whether or not the predictions were rescaled during the interpolation (set automatically during interpolation) |

ML models and training (model_training_description):

Name	Purpose
MODELTYPE	Type of model we want to use: Valid choices: UNet_Flat, UNet_Ico, LinReg_Pixelwise, RandomForest_Pixelwise, PCA_Flat, PCA_Ico
CREATE_VALIDATIONSET	Whether or not we want to create a validationset. The validation set is split of the training set.
SHUFFLE_VALIDATIONSET	Whether to shuffle the training set before creating the validation set or split it off chronologically
DATASET_FOLDER	Folder that was given as "base folder" when creating the corresponding dataset, i.e. the folder in which the data set folder is stored
RUN_NR	Number of run with the given configuration of data set, model and training (to run the same configuration multiple times)
S_MODE_PREDICTORS	Standardization mode for prediction variables: Can be set individually for each variable. Valid choices: "None", "Pixelwise", "Global_mean_pixelwise_std", "Pixelwise_mean_global_std", "Global"
S_MODE_TARGETS	Standardization mode for target variables. Valid choices: "None", "Pixelwise", "Global_mean_pixelwise_std", "Pixelwise_mean_global_std", "Global"
N_PC_PREDICTORS	PCA-regression: Number of principle components for predictor variables
BATCHSIZE	UNet Batchsize of UNet Models (UNets only)
LEARNING_RATE	UNet learning rate (UNets only)
IN_CHANNELS	Number of input channels of the UNet (UNets only)
CHANNELS_FIRST_CONV	Number of output channels of the first convolutional layer of the UNet (UNets only)
OUT_CHANNELS	Number of output channels of the UNet (UNets only)
FMAPS	Number of filter maps at every depth in the UNet-arch, tuple (UNets only)
ACTIVATION	Type of activation function used in the UNet (UNets only)
NORMALIZATION	Normalization type used in the UNet, e.g. torch.nn.BatchNorm2d (UNets only)
OPTIMIZER	Numerical optimizer used to train the UNet, e.g. "Adam" (UNets only)
DEVICE	Device the training took place on (UNets only)
DEPTH	Depth of UNets Models (UNets only)
NUM_EPOCHS	Number of epochs for training UNets, can be set to "early_stopping" or an integer val
PATIENCE	If using early stopping, this parameter determines for how many epochs we train without an improvement of the global minimum loss before aborting the training
USE_CYLINDRICAL_PADDING	Whether or not to use cylindrical padding (flat UNet only)
USE_COORD_CONV	Whether or not to use coordconv (flat UNet only)
LOSS	Loss function to use. Implemented are two choices for flat UNet: A masked MSE loss and an area weighted loss.
N_PC_TARGETS	PCA-regression: Number of principle components for target variables
REGTYPE	PCA-regression: Type of regression model used in the reduced space. Valid choices are "linreg" and "lasso"