kumoai.trainer.ModelArchitecturePlan

kumoai.trainer.ModelArchitecturePlan#

class kumoai.trainer.ModelArchitecturePlan[source]#

Configuration parameters that define how the Kumo graph neural network is architected. Please see the Kumo documentation for more information.

Variables:

  • model – (list[GNNModelPlan | GraphTransformerModelPlan]) A list of models to use. (default: [GNNModelPlan(...)]). Supported Task Types: All

  • channels – (list[int]) Deprecated in favor of channels in GNNModelPlan and GraphTransformerModelPlan. A list of potential dimension of layers in the graph neural network model (default: [64, 128, 256]). Supported Task Types: All

  • num_pre_message_passing_layers – (list[int]) A list of potential number of multi-layer perceptron layers before message passing layers in the graph neural network model (default: [0, 1, 2]). Supported Task Types: All

  • num_post_message_passing_layers – (list[int]) A list of potential number of multi-layer perceptron layers after message passing layers in the graph neural network model (default: [1, 2]). Supported Task Types: All

  • aggregation – (list[list["sum" | "mean" | "min" | "max" | "std"]]) Deprecated in favor of aggregation in GNNModelPlan. A nested list of aggregation operators in the graph neural network aggregation process (default: [ ["sum", "mean", "max"], ["sum", "mean", "min", "max", "std"]]). Supported Task Types: All

  • activation – (list["relu" | "leaky_relu" | "elu" | "gelu"]) A list of activation functions to use during AutoML (default: ["relu", "leaky_relu", "elu", "gelu"]). Supported Task Types: All

  • normalization – (list[None | "layer_norm" | "batch_norm"]) The normalization layer to apply (default: ["layer_norm"]). Supported Task Types: All

  • module – ("ranking" | "embedding") The link prediction module to use (default: ["ranking"]). Supported Task Types: Link Prediction

  • handle_new_target_entities – (bool) Whether to make link prediction models be able to handle predictions on new target entities at batch prediction time (default: False). Supported Task Types: Link Prediction

  • target_embedding_mode – (["lookup" | "feature" | "shallow_feature" | fusion]) Specifies how target node embeddings are embedded (default: ["lookup"]). Supported Task Types: Link Prediction

  • output_embedding_dim – ([int]) The output embedding dimension for link prediction models (default: [32]). Supported Task Types: Link Prediction

  • ranking_embedding_loss_coeff – ([float]) The coefficient of the embedding loss applied to train ranking-based link prediction models link prediction models (default: [0.0]). Supported Task Types: Temporal Link Prediction

  • distance_measure – (["dot_product" | "cosine"]) Specifies the distance measure between node embeddings to use in the final link prediction calculation (default: ["dot_product"]). Supported Task Types: Link Prediction

  • use_seq_id – ([bool]) Specifies whether to use postional encodings of the sequence order of facts as an additional model feature (default: [False]). Supported Task Types: All

  • prediction_time_encodings – ([bool]) Specifies whether to encode the absolute prediction time as an additional model feature (default: [False]). Supported Task Types: Temporal Node Prediction

  • past_encoder – (["decomposed" | "normalized" | "mlp" | "transformer"]) Specifies how to encode auto-regressive labels if present (default: ["decomposed"]). Supported Task Types: Temporal Regression

  • handle_new_entities – (bool) Whether to make forecasting models transductive by learning entity-specific heads. This can improve performance in case entities stay static over time, but will decrease performance on new entities arising during batch prediction time (default: True). Supported Task Types: Forecasting

Contents