Cross product of vectors 'v1' and 'v2'.

Dot product of vectors 'v1' and 'v2'.

Computes norm of vectors 'v'. Args: v: vectors to be normalized. epsilon: small regularizer added to squared norm before taking square root

Executes Module. This returns a set of rigid with the same shape as activations, projecting the channel dimension, rigid_shape controls the t

Compute geometric aware attention. Given a set of query residues (defined by affines and associated scalar features), this function computes

Predict sidechains using multi-rigid representations. Args: rigid: The Rigid's for each residue (translations in angstoms) representa

Run the AlphaFold model.

Builds TemplatePairStack module. Arguments: pair_act: Pair activations for single template, shape [N_res, N_res, c_t]. pair_mask: Pai

Builds Transition module. Arguments: act: A tensor of queries of size [batch_size, N_res, N_channel]. mask: A tensor denoting the mas

Builds Attention module. Arguments: q_data: A tensor of queries, shape [batch_size, N_queries, q_channels]. m_data: A tensor of memor

Builds GlobalAttention module. Arguments: q_data: A tensor of queries with size [batch_size, N_queries, q_channels] m_data: A

Builds MSARowAttentionWithPairBias module. Arguments: msa_act: [N_seq, N_res, c_m] MSA representation. msa_mask: [N_seq, N_res] mask

Builds MSAColumnAttention module. Arguments: msa_act: [N_seq, N_res, c_m] MSA representation. msa_mask: [N_seq, N_res] mask of non-pa

Builds MSAColumnGlobalAttention module. Arguments: msa_act: [N_seq, N_res, c_m] MSA representation. msa_mask: [N_seq, N_res] mask of

Builds TriangleAttention module. Arguments: pair_act: [N_res, N_res, c_z] pair activations tensor pair_mask: [N_res, N_res] mask of n

Builds MaskedMsaHead module. Arguments: representations: Dictionary of representations, must contain: * 'msa': MSA representation,

Builds ExperimentallyResolvedHead module. Arguments: representations: Dictionary of representations, must contain: * 'structure_mod

Builds PredictedAlignedErrorHead module. Arguments: representations: Dictionary of representations, must contain: * 'pair': pair re

Builds ExperimentallyResolvedHead module. Arguments: representations: Dictionary of representations, must contain: * 'single': Sing

Builds TriangleMultiplication module. Arguments: left_act: Pair activations, shape [N_res, N_res, c_z] left_mask: Pair mask, shape [N_

Builds DistogramHead module. Arguments: representations: Dictionary of representations, must contain: * 'pair': pair representation

Builds OuterProductMean module. Arguments: act: MSA representation, shape [N_seq, N_res, c_m]. mask: MSA mask, shape [N_seq, N_res].

Builds EvoformerIteration module. Arguments: activations: Dictionary containing activations: * 'msa': MSA activations, shape [N_seq

Build the single template embedding. Arguments: query_embedding: Query pair representation, shape [N_res, N_res, c_z]. batch: A batch

Build TemplateEmbedding module. Arguments: query_embedding: Query pair representation, shape [N_res, N_res, c_z]. template_batch: A ba

Compute geometry-aware attention. Given a set of query residues (defined by affines and associated scalar features), this function computes g

Predict side chains using multi-rigid representations. Args: affine: The affines for each residue (translations in angstroms). repres

Generate an embedding for a set of templates. Args: query_embedding: [num_res, num_res, num_channel] a query tensor that will be us

Build the single template embedding graph. Args: query_embedding: (num_res, num_res, num_channels) - embedding of the query sequenc

Build a single iteration of the template embedder. Args: act: [num_res, num_res, num_channel] Input pairwise activations. pair_mask:

Parallel computation of full transformer layer

Parallel computation of full transformer layer

Parallel computation of full transformer layer

Parallel computation of full transformer layer

Graph-conditioned sequence model

LayerNorm is applied either before or after the self-attention/ffn modules similar to the original Transformer implementation.

Args: model_name: Name of the pretrained IgLM model (e.g. "IgLM" or "IgLM-S"). seq_length: The length of the variable

Configure temperature and device; set scFv split attributes externally.

Constructs Linear Module. Args: num_points: number of points to project. Can be tuple when outputting multiple dimensions g

Initialize. Args: config: iterative Fold Head Config global_config: Global Config of Model. dist_epsilon: Small value to avoid