Models Reference

Import path

from phylognn.models import (
    BaseGATNet,
    BasePhyloGNN,
    GATBiLSTMNet,
    TemporalBiLSTMEncoder,
)

Base classes

class BasePhyloGNN

Abstract base class for phylogenetic graph models.

validate_data(data, require_batch=False): Validate x, edge_index, and optionally batch.

get_num_parameters(trainable_only=True): Count model parameters.

freeze_encoder(): Freeze modules returned by get_encoder_modules().

unfreeze_all(): Make all parameters trainable.

class BaseGATNet(input_dim, preprocess_dim, gat_hidden_dim=64, gat_heads=4, num_gat_layers=3, dropout_prob=0.2, use_preprocessing=True, encoder_type='res_gat'): Reusable GAT encoder base with optional feature preprocessing.

End-user model

class GATBiLSTMNet(input_dim, output_dim, preprocess_dim=32, gat_hidden_dim=64, gat_heads=4, num_gat_layers=3, dropout_prob=0.2, use_preprocessing=True, encoder_type='res_gat', temporal_mode='lstm', num_time_bins=None, temporal_hidden_dim=128, temporal_fc_hidden_dims=None, num_lstm_layers=2, temporal_aggregation='mean', graph_pool='sum', head_hidden_dim=64, output_positive=False)

End-user model combining GAT node encoding with graph pooling, FC temporal encoding, or BiLSTM temporal encoding.

Inputs follow the graph field contract documented in Graph Conversion. The output tensor has shape [batch_size, output_dim].

Temporal encoder

class TemporalBiLSTMEncoder(input_dim, hidden_dim, num_layers=1, dropout_prob=0.0, aggregation='mean'): Bidirectional LSTM temporal encoder exported as a stable public component.

Excluded internals

Low-level layers such as GATBlock, ResidualGATStack, PositionalEncoding, and MLPHead are not package-level user APIs.