configs.single_train_test.toml

############################################ 
# ------------------------------------------
#       configs_single_training.toml
# ------------------------------------------
#
# Configuration file for single training
#
############################################


# Run name
run_name = 'single_gegen_m01'

# Run description (optional)
run_description = """
single_training
no early stopping
with validation set (10%)
"""

# Output folder to save results
output_folder = 'results'

# Dataset path 
data_folder = 'data'

dataset = 'sea_surface_temperature'
# dataset = 'pm25_concentration'
# dataset = 'intel_dataset'

# Cuda setting
no_cuda = false

# gpu to use (if available and if no_cuda is false)
gpu = 0


[preprocessing]

# sampling density 
sampling_density = 0.1

# data normalization
# 'none' : no normalization
# 'minmax' : calls sklearn.preprocessing.MinMaxScaler
# 'standard' : calls sklearn.preprocessing.StandardScaler
# 'robust' : calls sklearn.preprocessing.RobustScaler
normalization = 'none'

# neighbor normalization from `hdla-gcn`
# deep Neural Message Passing With Hierarchical Layer Aggregation and Neighbor Normalization
layer_normalization = "neighbornorm"


[reproducibility]

# Development seed
dev_seed = 0

# Validation seeds
val_seeds = [1, 2]

# Testing seeds
test_seeds = [3, 4, 5, 6, 7]


[training]
# validation size 
# a value between (0-1) representing a percentage
# if 0, no validation set will be used
validation_size = 0.3

# save checkpoint
save_checkpoint = true

# Epochs to save checkpoint
epochs_to_save_checkpoint = 100

# Early stopping
# only if there is validation set
early_stopping = false

# Patience for early stopping
patience = 100

[training.model]
# Model
# `gegen`, `gegenbauer`, `gegenconv`
# `cheby`, `chebyshev`, `chebyconv`
# `gcn`, `gcnnet`
# `gat`, `gatnet`
# `transformer`, `transformernet`
# `hdla`, `gcnhdla`, `gcn-hdla`
# `ffk`, `ffkgcnii`, 'ffk-gcnii'
# `mr`, `mrgnn`, `mr-gnn`
model = 'gegen'

# model kwargs
model_kwargs = {}#{concat=false}

[training.logging]
# verbose
verbose = {train=true, val=false, test=true}


[hparams]

# Maximum number of epochs to train
# A value will be randomly chosen from the list
epochs = 5000

# Range for initial learning rate
# A value will be randomly sampled from the interval [lr_min, lr_max]
lr = 0.026

# perturbation for Laplacian 
epsilon = 0.05

# Range for filter order
filter_order = 4

# Range for alpha
alpha = 1

# Range for heads
heads = 2

# Range for dropout rate (1 - keep probability)
# A value will be randomly sampled from the interval [dropout_min, dropout_max]
dropout = 0

# max_k parameter from:
# - Multiresolution Reservoir Graph Neural Network 'mr-gnn'
# or k parameter for normalization trick in `ffk-gcn`
# - Graph Neural Networks With High-Order Polynomial Spectral Filters `ffk-gcn`
k = 1

# gamma parameter for normalization trick in `ffk-gcn`
# - Graph Neural Networks With High-Order Polynomial Spectral Filters `ffk-gcn`
gamma_norm = 1.0

# output parameter from:
# - Multiresolution Reservoir Graph Neural Network 'mr-gnn' 
output= 'funnel'


[hparams.reg]
# Range for the regularization parameter for the Laplacian
# A value will be randomly sampled from the interval [lambda_min, lambda_max]
lambda_param = 6.10822003e-5


[hparams.layers]
# Range for number of layers
# A value will be randomly sampled from the interval [n_layers_min, n_layers_max]
n_conv_layers = 1
# Range for number of hidden units
# A value will be randomly sampled from the interval [n_hidden_min, n_hidden_max]
n_conv_hidden = 4

# Range for number of layers
# A value will be randomly sampled from the interval [n_layers_min, n_layers_max]
n_linear_layers = 0
# Range for number of hidden units
# A value will be randomly sampled from the interval [n_hidden_min, n_hidden_max]
n_linear_hidden = 12

[hparams.optim]
# Optimizer
# Set optimizer and **kwargs
# Check documentation for specific **kwargs
# https://pytorch.org/docs/stable/optim.html 
# 'adam'
# 'adamw'
# 'sgd'
# 'adamax'
optim = 'adam'
# Range for weight decay (L2 loss on parameters)
# A value will be randomly sampled from the interval [weight_decay_min, weight_decay_max]
weight_decay = 0.00016795
optim_kwargs = {}

[hparams.lr_scheduler]
# Learning rate type and **kwargs
# Check documentation for specific **kwargs
# https://pytorch.org/docs/stable/optim.html 
# 'none' : No scheduler
# 'cosine' : `CosineAnnealingLR`
# 'exponential' : `ExponentialLR`
# 'step' : `StepLR`
# 'plateau' : `ReduceLROnPlateau`
# 'none'
lr_scheduler = 'plateau'
lr_scheduler_kwargs = {mode = 'min', factor = 0.5, patience = 500, min_lr = 1e-3}

[hparams.criterion]
# Select the metric and **kwargs
# 'crossentropy' : `nn.CrossEntropyLoss`
# 'mse' : `nn.MSELoss`
# 'l1' : `nn.L1Loss`
criterion = 'mse'
criterion_kwargs = {reduction = 'mean'}