[docs]
def get_default_config():
"""
Returns the hyperparameters configuration dictionary used to initialize and train the COSIE model.
Returns
-------
config : dict
A dictionary containing the following sections:
- GraphAutoencoder:
* hidden_dim (list of int): Hidden layer dimensions for the graph autoencoder.
* activations (str): Activation function used (default: 'relu').
- Prediction:
* hidden_dim (list of int): List of hidden layer dimensions used in the dual-prediction module.
- training:
* seed (int): Random seed for reproducibility.
* start_dual_prediction (int): Epoch to start dual-prediction loss.
* start_cross_section_integration (int): Epoch to start cross-section integration.
* epoch (int): Total number of training epochs.
* lr (float): Learning rate for optimizer.
* gamma (float): Weight for entropy regularization in contrastive loss.
* lambda1 (float): Weight for contrastive loss.
* lambda2 (float): Weight for prediction loss.
* lambda3 (float): Weight for triplet loss.
* knn_neighbors_spatial (int): Number of neighbors in spatial graph construction.
* knn_neighbors_feature (int): Number of neighbors in feature graph construction.
* print_num (int): Interval (in epochs) to print training progress.
"""
return dict(
Prediction=dict(
hidden_dim=[512, 512]
),
GraphAutoencoder=dict(
hidden_dim=[256, 128],
activations='relu',
),
training=dict(
seed=8,
start_dual_prediction=100,
start_cross_section_integration=200,
epoch=600,
lr=1.0e-4,
gamma=5,
lambda1=0.1,
lambda2=0.2,
lambda3=1.,
knn_neighbors_spatial=5,
knn_neighbors_feature=30,
print_num=50,
),
)