terminator.utils.model.default_hparams

Default hyperparameter set for TERMinator

Parameters:

  • matches (str, default=’transformer’) – How to processes singleton statistics.

    Options
    ‘resnet’

    process using a convolutional ResNet

    ‘transformer’

    process using MatchAttention

    ‘ablate’

    perform no processing

  • term_hidden_dim (int, default=32) – Hidden dimensionality for TERM Information Condenser (e.g. net1, self.bot, or CondenseMSA variant)

  • enrgies_hidden_dim (int, default=32) – Hidden dimensionality for GNN Potts Model Encoder (e.g. net2, self.top, or PairEnergies variant)

  • gradient_checkpointing (bool, default=True) – Enable gradient checkpointing at most memory-intensive steps (currently, MatchAttention)

  • cov_features (str, default=’all_raw’) – What features to include for covariance matrix computation. See terminator.models.layers.condense.EdgeFeatures for more information.

    Options
    ‘shared_learned’:

    use those produced by the ResidueFeatures module.

    ‘all_raw’:

    concatenate a one-hot encoding of residue identity and the fed-in additional features.

    ‘all_learned’:

    ‘all_raw’, but fed through a dense layer.

    ‘aa_learned’:

    use an embedding matrix for residue identity

    ‘aa_counts’:

    use a one-hot encoding of residue identity

    ‘cnn’:

    use a 2D convolutional neural net on fed-in matches (WARNING: not tested due to extreme memory consumption)

  • cov_compress (str, default=’ffn’) – The method the covariance matrix is compressed into a vector.

    Options
    ‘ffn’

    Use a 2-layer dense network

    ‘project’

    Use a linear transformation

    ‘ablate’

    Use a 0 vector

  • num_pair_stats (int, default=28) – [DEPRECIATED] Number of precomputed pairwise match statistics fed into TERMinator

  • num_sing_stats (int, default=0) – [DEPRECIATED] Number of precomputed singleton match statistics fed into TERMinator

  • resnet_blocks (int, default=4) – Number of ResNet blocks to use if matches='resnet'

  • term_layers (int, default=4) – Number of TERM MPNN layers to use.

  • term_heads (int, default=4) – Number of heads to use in TERMAttention if term_use_mpnn=False

  • conv_filter (int, default=3) – Length of convolutional filter if code:matches=’resnet’

  • matches_layers (int, default=4) – Number of Transformer layers to use in MatchesCondensor if matches='transformer'

  • matches_num_heads (int, default=4) – Number of heads to use in MatchAttention if matches='transformer'

  • k_neighbors (int, default=30) – What k is for kNN computation

  • k_cutoff (int, default=None) – When outputting a kNN potts model, take the top k_cutoff edges and output the truncated etab

  • contact_idx (bool, default=True) – Whether or not to include contact indices in computation

  • cie_dropout (float, default=0.1) – Dropout rate for sinusoidal encoding of contact index

  • cie_scaling (int, default=500) – Multiplicative factor by which to scale contact indices

  • cie_offset (int, default=0) – Additive factor by which to offset contact indices

  • transformer_dropout (float, default=0.1) – Dropout rate for Transformers used in the TERM Information Condensor

  • term_use_mpnn (bool, default=True) – If set to True, use a feedforward network to compute TERM graph messages. Otherwise, update TERM graph representations using an Attention-based mechanism.

  • energies_protein_features (str, default=’full’) – Feature set for coordinates fed into the GNN Potts Model Encoder

  • energies_augment_eps (float, default=0) – Scaling factor for Gaussian noise added to coordinates before featurization

  • energies_encoder_layers (int, default=6) – Number of {node_update, edge_update} layers to include in the GNN Potts Model Encoder

  • energies_dropout (float, default=0.1) – Dropout rate in the GNN Potts Model Encoder

  • energies_use_mpnn (bool, default=False) – If set to True, use a feedforward network to compute kNN graph messages. Otherwise, update kNN graph representations using an Attention-based mechanism.

  • energies_output_dim (int, default=400) – Output dimension of GNN Potts Model Encoder

  • energies_geometric (bool, default=False) – Use Torch Geometric version of GNN Potts Model Encoder instead

  • energies_gvp (bool, default=False) – Use GVP version of GNN Potts Model Encoder instead

  • energies_full_graph (bool, default=True) – [DEPRECIATED] Update both node and edge representations in the GNN Potts Model Encoder. GNN Potts Model Encoder always updates node and edge representations now, making this option do nothing.

  • res_embed_linear (bool, default=False) – Replace the singleton matches residue embedding layer with a linear layer.

  • matches_linear (bool, default=False) – Remove the Matches Condensor

  • term_mpnn_linear (bool, default=False) – Remove the TERM MPNN

  • struct2seq_linear (bool, default=False) – Linearize the GNN Potts Model Encoder

  • use_terms (bool, default=True) – Whether or not to use the TERM Information Condensor / net1

  • term_matches_cutoff (int or None, default=None) – Use the top term_matches_cutoff TERM matches for featurization. If None, apply no cutoff.

  • test_term_matches_cutoff (int, optional) – Apply a different term_matches_cutoff for validation/evaluation

  • use_coords (bool, default=True) – Whether or not to use coordinate-based features in the GNN Potts Model Encoder

  • train_batch_size (int or None, default=16) – Batch size for training

  • shuffle (bool, default=True) – Whether to do a complete shuffle of the data

  • sort_data (bool, default=False) – Create deterministic batches by sorting the data according to the specified length metric and creating batches from the sorted data. Incompatible with shuffle=True and semi_shuffle=True.

  • shuffle (bool, default=True) – Shuffle the data completely before creating batches. Incompatible with sort_data=True and semi_shuffle=True.

  • semi_shuffle (bool, default=False) – Sort the data according to the specified length metric, then partition the data into semi_shuffle_cluster_size-sized partitions. Within each partition perform a complete shuffle. The upside is that batching with similar lengths reduces padding making for more efficient computation, but the downside is that it does a less complete shuffle.

  • regularization (float, default=0) – Amount of L2 regularization to apply to the internal Adam optimizer

  • max_term_res (int or None, default=55000) – When train_batch_size=None, max_term_res>0, max_seq_tokens=None, batch by fitting as many datapoints as possible with the total number of TERM residues included below max_term_res. Calibrated using nn.DataParallel on two V100 GPUs.

  • max_seq_tokens (int or None, default=None) – When train_batch_size=None, max_term_res=None, max_seq_tokens>0, batch by fitting as many datapoints as possible with the total number of sequence residues included below max_seq_tokens.

  • term_dropout (str or None, default=None) – Let t be the number of TERM matches in the given datapoint. Select a random int n from 1 to t, and take a random subset n of the given TERM matches to keep. If term_dropout='keep_first', keep the first match and choose n-1 from the rest. If term_dropout='all', choose n matches from all matches.

  • num_features (int, default=9) – The number of non-sequence TERM-based features included per TERM residue.

  • loss_config (dict of str (loss component) -> float (scaling factor)) – Dictionary that describes how to construct a loss function. An example dictionary follows: .. code-block :

    {

    ‘nlcpl’: 1, ‘etab_norm_penalty’: 0.01

    }

  • finetune (bool) – Whether or not to train the model in finetuning mode (i.e. freezing all weights but the output layer)