terminator.models.layers.condense.EdgeFeatures

class terminator.models.layers.condense.EdgeFeatures(hparams, in_dim, hidden_dim, feature_mode='shared_learned', compress='project')

Bases: Module

Module which computes edge features for TERMs

Variables:

• embedding (nn.Embedding or equivalent, conditionally present) – Layer to embed TERM match residue identities

• lin (nn.Linear, conditionally present) – Input embedding layer

• cnn (Conv2DResNet, conditionally present) – CNN that generates 2D features by convolution over matches

• W (nn.Linear or nn.Sequential(nn.Linear, nn.ReLU, nn.Linear)) – Output layer

__init__(hparams, in_dim, hidden_dim, feature_mode='shared_learned', compress='project')

Parameters:

• hparams (dict) – Dictionary of model hparams (see ~/scripts/models/train/default_hparams.json for more info)

• in_dim (int) – Dimensionality of input feature vectors

• hidden_dim (int) – Hidden dimension

• feature_mode (string from ['shared_learned', 'all_raw', 'aa_learned', 'aa_count']) –

  Generate initial covariation matrix by computing covariation on

  • 'shared_learned': inputted match features without preprocessing

  • 'all_raw': raw counts as well as inputted match features

  • 'aa_learned': features in learned embedding for residue identity

  • 'aa_count': raw residue identity counts

• compress (string from [‘project’, ‘ffn’, ‘ablate’]) –

  Method to compress covariance matrix to vector. Flatten, then

  • 'project': project to proper dimensionality with a linear layer

  • 'ffn': use a 2 layer FFN with proper output dimensionality

  • 'ablate': return a zero vector of proper dimensionality

Methods

__init__(hparams, in_dim, hidden_dim[, ...])	Parameters: hparams (dict) -- Dictionary of model hparams (see ~/scripts/models/train/default_hparams.json for more info)
forward(matches, term_lens, rmsds, mask[, ...])	Generate embeddings for weighted covariation features between TERM residues

Attributes

T_destination	alias of TypeVar('T_destination', bound=Mapping[str, Tensor])
dump_patches	This allows better BC support for load_state_dict().

forward(matches, term_lens, rmsds, mask, features=None)

Generate embeddings for weighted covariation features between TERM residues

Parameters:

• matches (torch.Tensor) – Matches, either as ints representing TERM match residue identities, or featurized matches Shape: n_batches x sum_term_len (x n_in if already featurized)

• term_lens (list of (list of int)) – Length of TERMs per protein

• rmsds (torch.Tensor) – RMSD associated with each match Shape: n_batches x sum_term_len

• mask (torch.ByteTensor) – Mask for TERM residues Shape: n_batches x sum_term_len

Returns:

cov_features – Embeddings for covariance matrices between TERM residues Shape: n_batches x n_terms x max_term_len x max_term_len x n_hidden

Return type:

torch.Tensor