train.conf.yaml

# This file provides the default values for all parameters of the training procedure.
# You can OVERRIDE these defaults by specifying values in respective model config files (such as gcn.conf.yaml).

# ---
# Training procedure parameters.

# Maximum number of epochs to train for. Should be large enough for early stopping to take effect.
num_epochs: 100000

# The early stopping criterion to use. Must be one of
#     - gcn        Early stopping criterion used in [1]. Early stopping if validation loss is larger than the mean of
#                  the validation losses of the last <patience> epochs.
#     - gat        Early stopping criterion used in [2]. Early stopping if neither validation loss nor validation
#                  accuracy improve, reset of variables after stopping.
#     - gnnbench   Early stopping criterion used in our paper. Early stopping only by improvement of validation loss,
#                  reset of variables after stopping.
# Set to null to disable early stopping.
#
# [1] Kipf & Welling (2016). Semi-supervised classification with graph convolutional networks.
# [2] Velickovic et al. (2017). Graph attention networks.
early_stopping_criterion: "gnnbench"

# The tolerance (or patience) of the early stopping criterion. To correspond to the criterion in [1] this must be set to
# 10, to correspond to the criterion in [2] this must be set to 100.
early_stopping_tolerance: 50

# Always set to 0.0. Only needed for LabelProp where it is updated in corresponding config files.
# This "dummy parameter" is needed here in order for the parameter to be defined in Sacred before the model is
# instantiated. Only needed for model-specific training parameters, not for model parameters.
improvement_tolerance: 0.0

# Initial learning rate.
learning_rate: 0.01

# Strength of the L2 regularization.
weight_decay: !!float 5e-4

# Only needed for MoNet. Number of consecutive steps to either optimize filter weights or kernel weights.
alternating_optimization_interval: 12

# Only needed for MoNet. Multiplicative factor to decay learning rate after each entry of lr_decay_steps. Set to 0.0 to
# disable learning rate decay.
lr_decay_factor: 0.0

# Only needed for MoNet. List of steps after which to decay the learning rate. Set to empty list or null to disable
# learning rate decay.
lr_decay_steps: null


# ---
# Dataset parameters

# Number of samples (i.e. nodes) in the train/val/test sets.
# For each of the [train, val, test] sets use one of
#     - <set_name>_examples_per_class    Randomly sample this many examples per class. Total number of <set_name> samples
#                                        is then num_classes * <set_name>_examples_per_class.
#     - <set_name>_size                  Randomly samples this many examples in total.
# If test_size and test_examples_per_class are both null, the remainder of the samples is used as test set.
split:
    train_examples_per_class: 20
    val_examples_per_class: 30
    test_size: null
# For example, in the Planetoid paper the following configuration is used.
#split:
#    train_examples_per_class: 20
#    val_size: 500
#    test_size: 1000

# Should the input graph be standardized?
# Standardizing includes:
#  - Making the graph undirected
#  - Making the graph unweighted
#  - Selecting the largest connected component (LCC)
standardize_graph: true

# Should the attribute matrix (X) be row-normalized s.t. each row has L2 norm equal to 1?
normalize_features: false


# ---
# Logging parameters
# The number of epochs after which to print a log message.
report_interval: 10