jova_hyperparams.py

def default_hparams_bopt(flags):
    """
    protein model types:
    --------------------------------------------------------------------------------------------
    short name  | full name
    ------------|-------------------------------------------------------------------------------
    psc         | Protein Sequence Composition
    ------------|-------------------------------------------------------------------------------
    p2v         | Protein to Vector / Embeddings using n-gram amino acid 'words'.
    ------------|-------------------------------------------------------------------------------
    rnn         | Uses embeddings and an RNN variant (e.g. LSTM) to learn protein features.
    ------------|-------------------------------------------------------------------------------
    pcnn        | Protein CNN: https://academic.oup.com/bioinformatics/article/35/2/309/5050020
                | The final output is a 1D vector for each protein in a batch.
    ------------|-------------------------------------------------------------------------------
    pcnn2d      | A variant of PCNN that returns a 2D tensor for each protein in a batch.
    ------------|-------------------------------------------------------------------------------
    NOTE: All protein models, except 'psc' and 'p2v', use embeddings from the :class:Prot2Vec module.
    """
    return {
        "explain_mode": flags.explain,
        "attn_heads": 2,
        "attn_layers": 1,
        "lin_dims": [1033, 1481, 1800],
        "output_dim": len(flags.tasks),
        "latent_dim": 512,

        # weight initialization
        "kaiming_constant": 5,

        # dropout
        "dprob": 0.2,

        "tr_batch_size": 256,
        "val_batch_size": 128,
        "test_batch_size": 128,

        # optimizer params
        "optimizer": "adamax",
        "optimizer__global__weight_decay": 0.0007,
        "optimizer__global__lr": 0.0004,
        "prot": {
            "model_types": flags["prot_model_types"],
            "vocab_size": flags["prot_vocab_size"],
            "window": 11,
            "pcnn_num_layers": 2,
            "embedding_dim": 33,
            "psc_dim": 8421,
            "rnn_hidden_state_dim": 27
        },
        "weave": {
            "dim": 50,
            "update_pairs": False,
        },
        "gconv": {
            "dim": 512,
        },
        "ecfp8": {
            "dim": 1024,
        },
        "gnn": {
            "fingerprint_size": len(flags["gnn_fingerprint"]) if flags["gnn_fingerprint"] else 0,
            "num_layers": 3,
            "dim": 100,
        }
    }